Thanks for your comments here, I think they perfectly complement the impression I had from other conversations with Chomskian linguists. It's typically toxic, in bad faith, overclaiming their (mostly theoretical) achievements, and it lacks any sort of self-irony and applicability. Also they constantly reference to their supreme leader more than anyone else, who is the greatest example of this behavior. I mean, Darwin didn't refer to his predeccessors like "mere collectors", neither did Linney, Mendeleev or other global theorists. They typically expressed some gratitude and admiration to those who 'collected stuff' for them. I think now that the success of this approach was less 'grammarian generative' and more 'dissertation generative', because it opens an infinite field of theorizing. I mean, you'd probably expect that good theory would somehow predict how we learn languages, how we model them, and what languages can and cannot exist, or how they evolve. In all of this Ch. didn't provide much. I don't devaluate their work fully, that'd be a very low bar given probably majority of linguists still following this paradigm. But now I think it did more harm than good, suppressing less dogmatic research and thinking.
Thanks for that link. I never quite understood the distinctions (or non-distinctions) between self-embedding and MERGE, but the intro explains it rather well.
Yes, what is going on here? And not just AI research but Deep Learning in particular (Bengio, Hinton). This area has now captured a significant majority of researchers for all time?
In some of these cases it’s hard to say what “the research community” is. I would have expected Marx to be higher, since the research community of all Soviet scientists working on any subject was larger than any particular modern research community.
Maybe conducting AI research is less expensive because it's primarily digital. In STEM or social sciences researchers have to interact with meat space. In humanities here are no resource constraints so maybe artificial limits on volume of published work is necessary to maintain prestige.
But humanities researchers publish *much* less than science and engineering researchers (especially when measured by counting discrete publications, as citation counts do, so that being cited by 10 articles of 6 pages each shows up more than being cited on every page of a 200 page book), and there are far fewer humanities researchers than engineering researchers.
In part, the publishing conventions in AI (and adjacent fields) lean towards producing a large number of small papers. That is, the papers per researcher is very high, and the number of researchers is very high.
Because the number of researchers has exploded in the past 10 years, driven by massive investment from industry, rather than academia. Big AI conferences now get more than 10,000 submissions a year. Other areas of computer science are primarily academic, and are at least two orders of magnitude smaller.
Hinton's been around for a long time, but if you look at citations prior to 2014 or so, when neural nets were still mostly an academic curiosity, the number will look much more reasonable.
I did a little experiment. I was browsing some of the papers that cite Chomsky's most cited book, so I picked up one paragraph from the abstract of one of the top results that Google Scholar produced, fed it to an AI chatbot (Grok) and asked for reference suggestions:
"Based on the provided abstract and considering the context of attachment theory, I can suggest several references that would be relevant to your academic paper titled 'Patterns of Attachment: A Psychological Study of the Strange Situation':"
It gave me 8 suggestions. Call me cynical, but I think AI researchers are actively using AI to write their papers. A lot.
Did it suggest any papers actually cited by the allegedly human author? Were they even real papers (hallucinated citations are a real issue with LLMs)?
Because it is industry backed there is constant publishing. See "Two Minute Papers" in YouTube and you will see. There are always new techniques published specifically by companies such as Nvidia or universities that are backed by such companies in this specific area.
There is a big difference between the small amount of language a baby hears when learning to speak and the large amount of language artificial intelligence needs. It is not clear that artificial intelligence disproves Chomsky's theories.
A counter-argument to this “poverty of stimulus” line of thinking is that a baby doesn’t get merely exposed to some words. It also gets a deluge of visual, auditory, sensory etc. information, all perfectly synced, and _then_ it also gets to actively explore and _then_ it also gets multi-faceted types of feedback from the environment and from fellow humans. ML theory suggests that each of these differences can matter quite a bit.
This, indeed, doesn’t disprove Chomsky- but it means the poverty of stimulus argument is far from decisive, as the stimulus is not all that poor.
I am guessing that the baby has most of the deep layers already hard-coded (by evolution), and only requires some equivalent of a LoRA to get up to speed in his native language. The LLM has to be taught from scratch.
Also keep in mind that human language itself evolved to be easy to pick up by humans.
LLMs don't have that luxury: they need to learn a language that evolved not for ease of understanding by them, but by some aliens with completely different quirks in their underlying systems.
LLMs don't really have minds, though: they're just language-learning machines. That's what makes them ultimately unsuitable as the sole basis for AGIs.
Well, there's been a lot of handwaving and jumping up and down by behavioral geneticists about the FOXP2 gene and language acquisition in humans. And FOXP2 seems to play a role in song-bird song acquisition. Although FOXP2 may help with the ability to distinguish patterns in sound, it doesn't explain the emergency of language in the brain.
The human genome doesn't really have enough room to provide all that much hard coding. The entire thing is a couple of gigabytes, and most of it doesn't have anything to do with how particular synapses in the brain are wired. You've got basic reflexes, the learning system for training up the bulk of the mind mostly from scratch, and I'd guess not much else.
I suspect that human babies
1) Get a lot more training signal than just language. You've got to count all the visual information, audio information, etc. as well. This data is probably lower quality, but it still helps a lot.
2) Just have a more efficient architecture than transformers.
Since everyone's working on better architectures and better training procedures, I'd guess the AIs will catch up and then surpass the data efficiency of human brains not too long from now.
Information isn't infinitely compressible. If there are x different ways a brain can be (x different possible phenotypes), and if genes were the sole determinator of which of the x different brains you'd get, there'd need to be x different possible genome settings. Otherwise not all x phenotypes could be realised.
And the point here is that x<<2^11, or ca. 3 gigabytes. And I think that's likely not enough to encode all the possible different pieces of deep language heuristics.
Now in reality, the genes don't actually need to encode the entire blueprint for the language faculties. They just need to encode some initial setup that will predictably become a language faculty once the neural network starts training. This requires much less description length. You just need to encode the right optimiser and the right sort of architecture.
This suffices to explain why humans can learn language and other animals mostly can't. It's not that the whole language faculty is hard coded, the setup to be able to produce that language faculty is. Same as how you're not going to get a good language model out of a 2015 era MLP trained with the methods of the time, while a modern transformer or mamba model does great at it.
It can also explain why language faculties usually seem to be encoded in similar brain regions in different humans, and work in roughly the same way. You don't actually need hard coding of the entire language processing apparatus in the genome to explain something like this. I can see how people might have thought that twenty years ago, but a big lesson mechanistic interpretabillity of modern neural networks has taught us is that there's a lot of convergence in what is learned. If you train a bunch of different networks on the same sort of task, even if you do it with different initalisations and different data and even somewhat different architectures, they will often still converge to implement almost identical solutions.
'Not exactly, there is some variability introduced by the RNA transcription process, from what I can gather. Therefore, each combination of DNA could have more than one phenotype'
This variability can't be used to encode the information needed to specify a language faculty. Otherwise, organisms with the same DNA (humans) would sometimes come out having a language faculty, and sometimes come out not having a language faculty. For each bit that the transcription process contributes, the number of humans that possess the hard coded language abilities should be cut in half.
"The big difference however is that DNA is quaternary storage not binary like our computers."
This is already taken into account in the 'couple of gigabytes number I gave. Humans have ca. 3 billion base pairs, each of which takes 2 bits to encode. That's ca. 0.75 gigabytes total, I misremembered it to be a bit more.And that's not even taking any further compression into account, which would cut the actually available storage down even further. And then most of it is used to specify the rest of our bodies, rather than our brain initialization.
Presumably the human genome still contains some kind of a language learning bootstrapper, since human babies can easily learn at least one language, whereas babies of other ape species cannot.
This is precisely the question, no? Whether humans have an innate facility for language, or whether we cobble one together out of other cognitive adaptations.
Definitely. No problem fitting that into a couple of gigabytes.
But I expect that to be coding for a particular kind of brain architecture/training signal/more neurons of a particular type. Because those things probably don't require many bytes of description.
Taking the example of LLMs, encoding the LLM architecture, loss function and optimiser takes very little space. Whereas pretrained weights for hard-coding a bunch of deep language rules would take a lot of space, if our LLMs are any judge.
I don't really understand why Chomsky is so opposed to it or why it would disprove him. Statistical learning could just statistically learn grammar as well as or better than an expicit encoding. I suspect he didn't understand machine learning techniques and made some ill-considered comments, and his pride won't let him back off of it.
At some point, this becomes a matter of definition, but for most intents and purposes, we kinda do. (My go-to book for this is Zorina & Smirnova's "What speaking animals taught us", but it is in Russian.)
It just seems astounding unlikely that we are so unique ... but if you have a book with good arguments otherwise in a language I can read (just English, sorry!) then I'll gladly read it
I mean, there's "Why Only Us?", obviously, but it is a bit too condescending and abstract to my taste. Rumbaugh and Savage-Rumbaugh (two different people!) have a number of works, but probably on the more technical side. And Hockett (1966) is a classic.
We already know we're unique. We've built a giant civilization and all sorts of crazy technology. Our uniqueness including our language isn't that surprising.
That said, we also know we're not unique in terms of sheer brain power. I don't know what pilot whales and the like need all that brain power for, but it's at least plausible that they could use language.
Dolphins can at least understand basic syntax, in that they've been taught a language where changing word order can change the meaning.
I’ve long suspected marine mammals are quite intelligent, and it’s a small leap of imagination to think they have a language. Water is an excellent sound conductor; natural underwater waveguides allow sound propagation for literally thousands of miles. Perfect environment to develop a language.
This is very far from a proof that whale communication is language. I think it’s absolutely right to remain open-minded about whether whale communication is or isn’t like human language in whatever sense you mean. There might be a few bird species whose communication is possibly like human language.
But “using AI to identify a ‘phonetic inventory’ of codas” is very far from showing that this is language.
Did you read the whole piece? Where, for example, the researchers describe “[…]the sperm whale "phonemes" could be used in a combinatorial fashion, allowing the whales to construct a vast repertoire of distinct vocalisations. The existence of a combinatorial coding system, write the report authors, is a prerequisite for "duality of patterning" – a linguistic phenomenon thought to be unique to human language – in which meaningless elements combine to form meaningful words.”
They are pretty far along the road toward “yes whales have a language”. The help of machine learning pattern recognition is useful, but far from the only thing pointing toward it.
I hadn’t read the whole article. I refuse on principle to read BBC science articles, because they’re about as bad as they come with sensationalizing things. But the actual Nature article is interesting.
It’s interesting for sure, and seems like a breakthrough in interpreting this communication. But this is very far from saying that this is a language. It’s not possible to show anything significant about a claimed language if your database includes a total of 8000 phonemes (the equivalent of about 1500 words) across several dozen communicative exchanges.
No, there are 21 distinct coda-types - they have recordings of 8000 coda-tokens. It's like they have recordings of a few dozen people each saying a few words to each other in a language like Japanese, that has about 20 distinct phonemes. Definitely not enough to prove it's a language, but maybe enough to get its phoneme inventory if they're right.
Yeah I’m with you on science coverage, and not just BBC of course. This one seems reasonable enough.
One thing to note: I remember reading somewhere that the vocabulary of an English peasant in the Shakespearean comprised only about 600 words, so the whales have at least the same order of magnitude vocabulary. Sure we don’t know much about their language - yet - but that’s just a matter of time, IMHO.
Looking forward to someone dropping a hydrophone into the ocean and having the first dialog with a whale!
Note - it's not that they have 600 distinct "words" - it's that they have recordings of about "600 words" worth of conversation total. In any human language, about a hundred of those words would likely be repeats.
I could imagine a peasant from a few hundred years ago might not know the following words from the first 500 on that list: school, state, student, country, American, company, program, government, million, national, business, issue, provide, political, include, community, president, real, information, office, party, research, education, policy, process, nation, college, experience, former, development, economic, military, relationship, federal.
But a peasant would surely know several dozen names for animals and plants that aren't on that list, as well as tools and actions to do with them, even ignoring whatever religious and family vocabulary they might have.
I feel that is incredibly dependent on the definition of language. Chomsky's definition seems incredibly ridged, as is common in foundational research. "A structured system of communication that consists of grammar and vocabulary" is true of various bird songs. The objection then becomes that grammar is more than just structure but a structure of a certain complexity defined to separate human language from other animals, which seems very post-hoc to me.
>The objection then becomes that grammar is more than just structure but a structure of a certain complexity defined to separate human language from other animals, which seems very post-hoc to me.
You have a good point, but there are fairly obvious measures of the complexity of human language which I'd be surprised to see any other animal's language reach.
For instance, human sentences are sufficiently varied that, barring quotes, repeated sentences (except exceedingly short ones, "Damn!" is common) are rare. Does any other animal match this? This doesn't require any specific properties of the language's syntax.
The simplest and most distinctive features of language:
A very large fraction of utterances (e.g. sentences) are novel combinations of existing elements
The patterns by which these elements are combined are recursive (e.g., any two sentences can be combined with “and”, noun phrases can combine with a preposition to form a modifier within a bigger noun phrase).
Does anyone know what the Chomsky followers would say about the recent findings that non-human animals (a large number of them) clearly have methods of communication?
I'm not sure what important recent findings you're referring to; the fact of non-human communication is ancient knowledge. What Chomskyans would say is that human language is evolutionarily discontinuous (it is not an outgrowth of the same phenomenon in other species) because there are categorical differences. Nonhuman communication systems are syntactically basic (certainly non-recursive), goal-orientated, lack reference, conceptual labelling, symbol arbitrariness and so on. Claims to the contrary based on e.g. training apes to sign have always anthropomorphised the data and there is not wide acceptance that any species has demonstrated human-like capabilities.
Confirmed by a Chomskyan (me!). Also, note that language is not a tool for communication, that function is secondary and forced (see the quote in the beginning of the article), it is originally a tool for structuring thought.
"language is not a tool for communication, that function is secondary and forced (see the quote in the beginning of the article), it is originally a tool for structuring thought"
Can you summarize, for a complete lay person in this area, what the evidence is to support this claim?
Multiple homonymies, synonymies, and other inconsistencies between the likely syntax (and semantics, which is read off syntax in a not-quite-trivial way) of sentences and their observed morphophonologies.
Yeah, got too jargony there for a second (although that itself is a good illustration!). Like, take the "I saw a girl with a telescope" example above. It has a structural ambiguity (do you first join "girl" with "with a telescope" or "a girl" with "saw") _and_ a lexical ambiguity between instrumental and comitative (in Russian, for instance, these meanings are expressed differently). You can also show each of the homonymies separately. We think hierarchically and only then try to push that hierarchy into a linear sequence (a sentence); and, to add insult to injury, our lexical system is also ripe with various meaning switches (metonymies, metaphors, and so on): "I saw a mouse" is ambiguous between a computer gadget and a small animal, and it persists despite us usually not noticing a similarity between the two (although it is, of course, originally, a metonymy on visual similarity).
There's an often-found claim that most homonymies are not a problem in real speech (the one @jumpingjacksplash begins with below), but there remain quite enough.
Wait, the argument is that when I'd never get confused saying "witch/which" whether I mean "sorceress" or "that," but someone listening to me might? That's faintly interesting, but it's a feature of toy systems which aren't remotely for the purpose of thought.
For example, imagine two ships communicating with flag signals, but with a system where letters and concepts use the same flags and you have to guess what the other person means. The crew of ship one try to write "halt," but the flags for H=Denmark, A=Collision, L=Rock and T=West. The crew of ship two thinks a Danish ship has hit a rock west of them and sails off in that direction.
The real evolutionary pattern of language must be something like: 1. Simple sounds to indicate things; 2. more complex sounds to convey relations that can be thought non-linguistically (eg. this rock is on top of that rock; 3. the possibility of having thoughts that couldn't be expressed non-linguistically due to thinking in language (eg. "language is ontologically prior to thought"). 4. The utility of non-linguistic thoughts by allowing complex social co-ordination (e.g. "Obey the king of Uruk or face the wrath of Marduk"). This then just comes back to Wittgenstein's point that sentences of type 3 are basically the equivalent of "colourless green ideas sleep furiously" so far as external grounding is concerned, although I'd modify that to having powerful impacts on social co-ordination. Hence they stuck around and developed as opposed to being sanded off by evolution.
That's the evidence that's easiest to explain in non-technical terms (not that the comment you replied initially succeeded in it). But yeah, the main thrust of the argument is along the lines of "we see a bunch of people hammering nails with microscopes, but the microscopes' internal structure suggests it wasn't made for hammering".
(I originally came across it in Scott's review of "Origin of Consciousness in the Breakdown of the Bicameral Mind")
The person who wrote that claims that they used language to communicate but not to think. I realize it's one, self-reported Reddit post, so not exactly high-quality data, but I wonder if there are other claims like this, and if they have been investigated by linguists?
There are other claims like this, and they are true at some level and false at another. We often indeed don't think in language in the sense of "build a sentence, up to and including how it sounds/what the articulatory commands are" (the acoustic/articulatory debate is its own bag of worms), but the idea is that to formulate _complex_ thoughts specifically, we use a method to chain thoughts which is so similar to syntax it probably _is_ syntax, even if it never gets sent to the systems of articulation ("is never sent to spell-out", in jargon-y terms). Note how using language was tremendously helpful for the guy, because it helped organize thought.
I think I read that post differently than you. He seems to be claiming that he didn't have complex thoughts before he started thinking articulatory in language, e.g.
"I can only describe my past life as ...."Mindless"..."empty"....."soul-less".... As weird as this sounds, I'm not even sure what I was, If i was even human, because I was barely even conscious. I felt like I was just reacting to the immediate environment and wasn't able to think anything outside of it."
So we seem to have a case where language was used exclusively for communication, and not at all for thought. Doesn't this contradict the claim that language is primarily used for thought, with communication being secondary?
Imagine you have a microscope and use it as a hammer. Doesn't this contradict the claim that the primary use of microscope, as evidenced by its structure, is for something else than hammering in nails?
<i>I can only describe my past life as ...."Mindless"..."empty"....."soul-less".... As weird as this sounds, I'm not even sure what I was, If i was even human, because I was barely even conscious.</i>
Huh, maybe p-zombies aren't such an outlandish idea after all.
Also, I wonder if this kind of testimony might be useful in studying how animals experience the world. Obviously there are many more differences between humans and other kinds of animal than just thinking with language, but still, this might be the closest we can get to answering the question, "What's it like to be a bat?"
An interesting question, but, to be fair, other explanations are available. (Is she able to imagine to raise her hand without raising her hand, I wonder?)
>We often indeed don't think in language in the sense of "build a sentence, up to and including how it sounds/what the articulatory commands are" (the acoustic/articulatory debate is its own bag of worms), but the idea is that to formulate complex thoughts specifically, we use a method to chain thoughts which is so similar to syntax it probably is syntax, even if it never gets sent to the systems of articulation ("is never sent to spell-out", in jargon-y terms).
At the very least, some thinking can also be geometrical rather than syntactic. Visualizing a part dangling from a flexible support and envisioning where its center of gravity will be and how it will be oriented is a complex operation, but it certainly doesn't feel linguistic to me.
This reminds me of the other day I was rolling out a pie crust. The instructions said to roll out the size of a sheet of notepaper. (In fairness to the recipe writer, I had misread and this step was not the final step; rather she wanted us to roll this shape prior to doing several folds for flakiness.)
But the point is, I kept trying to do this shape, which was also hard and impossible for me because I was making only a small pie in an 8 inch pan so didn’t have the full amount to make a notebook-paper sized rectangle. Poor reasoning on my part.
Throughout this process my brain vexingly kept flashing in my head a *picture* of the round tin, as if to say, you’re on the wrong track, old girl. Can’t emphasize enough this was a picture only, that I had to keep suppressing: “She said notebook paper! She said rectangle!”
It's _complicated_, but I am not sure that when laypeople (not physicists who learned advanced ways of modeling such things through, well, language) do it, it is _complex_ in the sense of explicitly manipulating multiple parts of the system.
Inner mental experience is far richer than that - this is a massively under-researched and under-discussed topic in our society. The best detailed research into it is by Hurlburt using a technique called Descriptive Experience Sampling.
One result he found is that most people often think in rich unsymbolized concepts - I do this myself, and I often don't *have* words for the concepts. I have to invent language, or spend a long time crafting it to turn the thoughts into words. This to me makes it pretty clear my underlying mind isn't primarily linguistic
(I use inner speech a lot too, and it is very useful for certain kinds of thought. It is but one modality of thought.)
Similarly, lots of people think mainly with images, or mainly with emotions. Others are mainly paying detailed attention to sensations, or have no focussed inner mental experience at all.
There's a phenomena Hurlburt found which the complex chained thoughts you describe reminds me of, where people can have "wordless speech". This has the cadence and rhythym of a sentence, but with concepts in the awreness instead of linguistic words.
Lots going on to investigate here, and I think linguistics would hugely benefit from studying Hurlburt's research.
> This has the cadence and rhythym of a sentence, but with concepts in the awreness instead of linguistic words.
is indeed prima facie evidence for, not against, "to formulate _complex_ thoughts specifically, we use a method to chain thoughts which is so similar to syntax it probably _is_ syntax, even if it never gets sent to the systems of articulation". It is just a specific case where you don't send it to the systems of articulation because the specific terminal nodes correspond to concepts which don't have a phonological word to them. (Compare a situation when you're bilingual and know a word for the concept you need in one language but not in the other.)
Second,
> Similarly, lots of people think mainly with images, or mainly with emotions. Others are mainly paying detailed attention to sensations, or have no focussed inner mental experience at all.
Indeed, there are many non-linguistic ways of thought; the question is, though, whether any of them are _complex_, chaining ways. Like, images seem to just be, well, non-compositional: an image of a house, cognitively, is not images of windows imposed over image of a wall combined with an image of a roof (as evidenced by the default image of window probably differing from the parts of the default image of a house that correspond to the house's windows).
What is the Chomskyan response to Evelina Fedorenko's work that suggests that even when the language systems in adult brains are damaged, their other cognitive functions remain intact? For example, people with global aphasia can still engage in complex causal reasoning tasks such as solving arithmetic problems and reasoning about others' intentions.
From my own experience, I mostly think without using language (i.e. talking to myself). It's only when I have to break down complex tasks into logical subunits that I begin to talk to myself about what I'm doing (but that's mostly a placeholder narrative — "e.g. this is the screw that I need to insert here, after I get this thingy lined up with the thingamajig." Also when I need to write, I'll start thinking in language. So, I'll grant you, language is necessary for structuring thought that leads to communication with others, but I don't think it's integral to reasoning. I'm pretty sure I could get the screw into the thingy and thingamajig without overtly describing the process to myself.
Aphasias are a prime proof of linguistic module's separateness, so this is a very important question indeed. The problem is, there are at least six different kinds of aphasias (Luria's classification), and none of them looks like "break the syntactic module". Broca's aphasia (in Luria's terminology, efferent motor aphasia) is pretty clearly "broken PF" (syntax-phonology interface), Vernike's aphasia (don't remember the Luria name for it) is pretty clearly "broken LF", some others are way more specific (there is, for instance, an aphasia that specifically targets multivalent concepts like "under").
Name-like calls are not sentences with grammatical structure. They don’t imply recursion or an infinitude of possible utterances with a finite vocabulary.
They may communicate, but they don't seem to do language well. All the ape-language studies from the 70s and 80s failed miserably (the most famous of these research projects, Coco the gorilla, made a good mascot but terrible science, particularly as the researcher refused to share or release unedited video of Coco communicating). It seems like even smart animals, like chimps, can't grok language the way humans can.
Thank you, this is very informative. I'm thinking about results from whales and elephants. Do you have thoughts on whether this counts as "language" from a Chomskian view?
I don't know Chomsky well enough to comment. My only insight is that laymen (such as myself) often think that language is simpler and easier than it is, so much so that we are impressed by chimps that seem like they can do sign language, or elephants that seem like they have names for each other. There really does seem to be a huge gulf in capability between humans and all other animals when it comes to language.
>There really does seem to be a huge gulf in capability between humans and all other animals when it comes to language.
Agreed. From the point of view of resolving how language actually evolved, it is a pity that all of our fellow Hominini except the chimps and the bonobos are extinct. At least the hyoid bones can fossilize, consonants and vowels, not so much... I'm skeptical that we will ever really know.
Having a name for each elephant that they use to get each other’s attention is more sophisticated than a finite list of alarm calls, but it does not mean that there is productive recombination of utterances to express novel meanings. One of the major features of language is that you can (largely) understand novel expressions that you have never heard before, and the limitations on your understanding aren’t due to the novelty, but due to inherent uncertainties of context and intention.
They can't grok human language the way humans can. I don't understand why this is sufficient to decide that they can't grok language? Surely one would need to study how they communicate with each other to decide that?
We need to study the ways they communicate to decide that. There are many animals whose communication systems have been studied sufficiently to be sure that they don’t have the complex structure of human language. This is true for dogs, cats, primates, many birds, and many other animals. I think there are some animals whose communication system is complex enough that we know we haven’t yet understood it, so that we can’t be sure it’s not like human language. This is true for some whales, elephants, some birds, and some cephalopods. None of these has been shown to be as complex as human language, but they haven’t been shown not to be either.
One interesting note: apparently Ken Hale, late well respected field linguist who did a lot of work on the Australian language Walpiri among many other things, uncontroversially described Walpiri and some other Australian indigenous languages as lacking clausal embedding in the 70s. This Language Log blog post (http://itre.cis.upenn.edu/~myl/languagelog/archives/004592.html) muses on various possible reasons this basically went unnoticed whereas Everett immediately got in an academic flamewar, but the main reason seems to have basically been that recursion hadn't been proposed as a linguistic universal until the introduction of Chomsky's Minimalist Program in 1995, earlier Chomskyan models of universal grammar did not include it.
So the idea that this may not even be unique to Pirahã seems like an important counterpoint to some claims I've seen that Pirahã in fact probably does have syntactic recursion and the real problem is that its sole expert isn't describing the language correctly. I've seen Everett comment somewhere that a problem may be that field linguists are essentially trained to fit grammar into a certain framework and so may be trying to square peg a round theoretical hole (as he himself says he did in his original PhD thesis description of Pirahã), thus failing to properly document the true linguistic diversity that really exists in the world.
Disclaimer that I don't have any expert understanding of the technical issue here, just a longstanding hobbyist fascination with linguistics. I acknowledge that my amateur understanding is causing me to miss something important here. But overall my reactions to the Chomskyan team response to this, where Everett is probably wrong but even if he's right it actually doesn't matter for his theory, strikes me similarly to you: it feels like dragging a major hypothesis into an unfalsifiable territory where it gets hard to understand what definitive claims it's actually making about the nature of human language at all.
The claim that all human languages have sentences containing more than one word isn’t vacuously true, it’s profoundly true. Just compare this fact with the norm in the communications of other species. No equivalent of “the [color] bird” is known in how any other species communicates. That, is Merge.
Communicating in something composed of parts with their own semantic effect (like, "dog" in "A dog came" has a meaning but "g" doesn't; these parts are normally called morphemes but layman explanations often default to "words" because it's simpler) is _not_ trivial, most animal communication systems don't have morphemes. Of those that do, seemingly none except humans and apes trained by humans are in the primate tree.
Interestingly, no! We know the answer from ontogenesis (children development). There are three prolonged stages in humans: one-word, two-word (the involved words are short and probably analyzed as monomorphemic by the child, so ignore the word/morpheme angle), and then the so-called "grammatical explosion" where the child rapidly builds more and more complex structures. So, to answer the question in a pithy way, two doesn't make recursion (you can combine, but you can't use the combined thing as the input to the next combination, which is where recursion kicks in), but three does.
And see the dolphinspeak comment above on the linearly recursive vs. hierarchically recursive difference.
There’s a difference between obligatorily one-word utterances and a possibility of two-word utterances. What I think is actually the interesting difference though is the difference between utterances with a fixed upper bound on length and utterances that can be in principle unbounded (which even Everett says is true of Pirahã).
LLMs' way of picking language is interesting but certainly different from humans', both because humans need less material and because they show unexpected biases (for instance, non-conservative quantifiers are unlearnable for human children of the critical age but not for LLMs; you can trivially teach an LLM extraction from islands or binding violations; and so on and so forth).
"Mary wrote a letter to her" cannot mean "Mary wrote a letter to herself".
"Mary saw a snake in front of her" _can_ mean "Mary saw a snake in front of herself".
And, finally, "Mary knows that Penny saw herself" cannot mean "Mary knows that Penny saw her" (and vice versa, but the vice versa is uninteresting because it's the same as the first).
There are very pecuilar and technical laws surrounding both when a reflexive _can_ appear and when it _must_ appear.
>"Mary wrote a letter to her" cannot mean "Mary wrote a letter to herself".
Is that strictly true? What about this: "Mary reflected on the woman that she hoped she would become. Mary wrote a letter to her."
Is there some technical linguistic sense in which she's not writing a letter to herself? Is metaphorical meaning considered out-of-scope for linguistics? If so, how is that rigorously defined?
There is a purely technical linguistic sense in which she isn't writing a letter to herself, yes. The antecedent, to use a technical term, of the pronoun "her" is "the woman that she hoped she would become", which is a different phrase than "Mary" and, again, in a technical sense, we know it not to be co-indexed with Mary because the first sentence is possible (cf. #"Mary reflected on her"). The first sentence establishes the two women as different objects for the purposes of binidng theory.
To take a different example, note how "Joe Biden reflected on the president of the United States" is an infelicitous sentence if it happens now but would be felicitous if it happened four years ago.
(Note that nobody explicitly teaches children the laws, and, judging by CHILDES corpora, they don't get enough data to arrive at the very generalizations they tend to arrive at.)
Sorry, should've been clearer! I've written Hindi words in English, and ChatGPT is responding in Hindi written in English after "recognizing" that that's what I'm doing. Its ability to mimic that form of Hindi is what I was trying to highlight.
This is kind of a side rant but - I don't think this should be so surprising. As far as I can tell, there's sort of a "script is a transliteration with an army and navy" situation going on with romanized Hindi. It's pretty standardized and extremely commonly used, and it's mostly political and historical reasons that keep it from being recognized as a "real" script for the language. ChatGPT had plenty of examples to learn from online and no real reason not to learn it.
Also note though that you can converse in any number of encoding mechanisms, such as rot13, or, pig-latin, etc ... The patterns that have been learned aren't just at the level of specific examples but also at various levels of representation between the representational, syntactic and semantic.
Yeah it really shouldn't be surprising. What surprised me is that I followed up and chatgpt apparently has a hard time recognizing that the writing system is not the usual one and says it's writing Devanagari
ChatGPT(-4o) can do much more impressive stuff than that with transliterated language. Try asking it novel questions in romanized Hindi, or in English transliterated into devanagari. (e.g. "Who was Prime Minister of Australia on the day Björn Þórðarson became Prime Minister of Iceland?") I can't say for sure whether it will be able to handle these sorts of questions using romanized Hindi or devanagari English, but it seems to do pretty well when I ask questions in katakana English or romanized Japanese.
Somewhat weirdly, it answers romanized Japanese questions in romanized Japanese, but it answers katakana English in straight Japanese. (This behaviour is slightly different from what happens with GPT-3.5, which seems to have trouble recognizing katakana English as English, and responds to romanized Japanese in English rather than romanized Japanese.) I don't know exactly what is going on here.
My tentative conjecture right now is that it reads katakana English as borrowings into Japanese that it hasn’t seen yet (because the rate of novel English borrowing into Japanese is very high in recent texts) while it reads romanized Japanese as Japanese, because there is more use of romanized Japanese by actual Japanese speakers than there is novel borrowing of Japanese into English.
I don't think there is enough romanized Japanese produced by native Japanese speakers to train an LLM effectively. I think it's more likely is that the GPT researchers generated the data artificially by converting their entire Japanese-language corpus into romaji and feeding that in as additional training data.
Your explanation for the Japanese responses to katakana English seems plausible, but I'm still surprised at how well it handles these questions. It gives convincing answers to:
"If John is standing to the right of Jane, who is standing to the left of Jane?"
"Which game is easier to play: checkers or chess?"
"Do Australian Rules football players swim better than Canadian hockey players?"
"Who was Taoiseach of Ireland during the time when John Major was Prime Minister?"
After numerous attempts, I finally managed to confuse it by asking:
"During the prime ministership that followed the prime ministership of Tony Blair, which Taoiseach of Ireland finished his term of office?"
Even here, ChatGPT gets the answer right at least some of the time. But the answers seem inconsistent across separate chat sessions. It has no problem naming Gordon Brown as the next PM after Blair, but gets confused trying to make the connection to Ireland. Sometimes it correctly answers "Bernie Ahern"; sometimes it names Brian Cowen (i.e. Ahern's successor, who finished his term after Brown left office); sometimes it balks at the weird question and suggests that if I want to know about the Taoiseach(s) in that period, I should ask a "more concrete" question instead.
(Of course, being able to answer this question in any language is pretty impressive — let alone when it is posed in the pseudo-language of katakana English!)
I think this is a shallow take on a complex field. I appreciated that it's well-researched but the sources have not been well understood, though admittedly the review is from a non-linguist, and admittedly the discourse in the research literature is often unprofessional and opaque.
The objections raised against Chomsky's research program are mostly intuitive and therefore trivial - most were dealt with in the 1950s rejection of structuralism and behaviorism (which is foundational to almost all of cognitive science, which is why Chomsky is hugely cited and not as controversial as this makes out), while others betray a misunderstanding of what generative linguistics studies (this is a consequence of 'language' itself acceptably meaning many things to many people, though the review didn't adequately tease these meanings apart).
If anyone is interested to learn about the Chomskyan approach in an informed way, I would highly recommend exploring Norbert Hornstein's Faculty of Language blog. It's now no longer active but it is a treasure trove of explanation that helped me immensely when I was a graduate student trying to make sense of all this.
If the Chomskyan paradigm is so uncontroversial and secure, why the profound insecurity and un-science-like nature of the Chomskyan camp response to Everett? One or two or ten thin-skinned scientists would be expected and not a big deal, but why does Everett's claim ruffle feathers so universally?
The field is large and most working Chomskyan linguists have zero interest in Everett, which exactly matches the expectation that a secure field would not take note of him. Popular understanding of his impact is overestimated because media (even science media) loves drama.
The reason why those who do engage with Everett do so in unprofessional ways (and I happily criticise them for not controlling themselves) is because people like Everett rarely take on Chomsky's ideas as scientific hypotheses; they also want to be iconoclasts and paradigm-shifters, leading the charge against the big bad man who has hold of the field because of politics or mysticism or some other conspiracy. It's unscientific because it's personal, not because Everett's analyses can't be trivially dismissed.
This review claims that several ideas of Chomskyan linguistics seem obviously wrong. It would be very interesting to hear rebuttals to those points from a Chomskyan linguist!
I have already commented on "language isn't for communication" angle and recursion angle in the other comments. As for I-languages vs. E-languages, the "language is a dialect with an army" problem of distinguishing whether two I-languages belong to the same E-language or different ones as well as most aspects of language (especially non-lexical ones) being essentially in the brain and, importantly, out of conscious control, should be sufficient.
Deciding which are dialects and which are separate languages can be done the same way biologists decide the boundaries of genus, species, and subspecies: messily and with much arguing. Or perhaps biologists should just take the simpler route and declare every organism a unique species, save a lot of effort that way.
Yeah, the species debates are a very good analogy, and species is much less well-defined than population, making it a not too useful notion. (The analogue of population would be a dialect of a tight-knit community; of course, linguistic communities are generally more permeable than populations.) There is one benefit though for biologists: they are not generally under siege from nationalists to declare certain populations parts of the same or of different species. Try telling Arabs that Palestinian Arabic and Moroccan Arabic are fourteen centuries of development apart, making them more far from each other than Icelandic and Danish. Or, conversely, that the differences between Danish and Bokmål or between the four Yugoslavian literary varieties (Bosnian, Croatian, Serbian, Montenegran; Slovenian and Macedonian are genuinely separate, but Macedonian has its own problem of continuum with western Bulgarian dialects) are laughably small and don't count as languages by any linguistics-internal criterion.
Yes, I worded it imprecisely. Everett does do that, though he doesn't stop where that would be sufficient. He embeds it in the rest that I described, which is what baits bad behaviour.
Baits bad behavior? It almost sounds like you're blaming Everett for his critics being nasty to him. I also see a lot of dubious psychologizing, suggesting that he's motivated by a desire to be an iconoclast. You do that here when you say:
"they also want to be iconoclasts and paradigm-shifters, leading the charge against the big bad man who has hold of the field because of politics or mysticism or some other conspiracy."
...this kind of psychologizing is part of the bad behavior and unjustified nastiness directed at Everett, criticizing someone based on speculative assertions about their motives. If someone genuinely comes to hold iconoclastic views, they have every business sharing those views with the broader community. Maybe you've got the causality reversed here, and are supposing a desire to be iconoclastic motivates Everett's views, rather than Everett coming to those views automatically making him an iconoclast regardless of whether that's what he wanted? Why psychologize in this way?
>I happily criticise them for not controlling themselves
For me a rancorous style is possible evidence of an idea which is unparsimonious. It's not just about being rude in the abstract. Maintaining belief in an unstable hypothesis is exhausting and people resent having to explain themselves.
I agree. For me, this is why it's important that these people are a small minority. In the wider context of quality research being done in the field (imo), it makes more sense to conclude that these people are a type who wilfully indulge in undignified behaviour - I think there are people in lots of fields who have this trait even when they are broadly right on the technical detail.
Thanks for the reply. It’s the human condition for sure. I don't have much stake in this, except we have a problem with how 'elites' are viewed, and although a lot of that discourse is toxic, I have some sympathy with this having done a humanities degree and seeing what "cultural marxism", for want of a better word, can do to the study of beautiful things. Can it do the same thing to the true as to the beautiful? Well here you have a Marxist who apparently totally dominates an emerging science and his defenders will go beyond the confines of typical academic discussion to defend him and that raises my eyebrow, I suppose. A person can be completely on top of the technical detail and still be an ideologue. In fact that's mostly how the left has come to dominate academia. But I'm getting into unparsimonious ideas of my own!
You elucidate something subtle and important here, imo. I've always been puzzled by the fact that there's a positive correlation between Marxism and IQ despite the many obvious gaping flaws with Marx. My current best guess is that the answer is that many academics crave power and Marxism gives that to them. Marxism elevates the central planner over the common person and thus creates an implicit power structure in which high-IQ academic types are singularly adapted to thrive. "No no, the collective wisdom of the hoi polloi inevitably leads to revolution so the best answer is to let me order society for you." It's sort of a sad irony that the post-structuralists never turned their tools on themselves because that would have clearly revealed how the structure of the ideology perfectly reflects the megalomania of deflated narcissism that haunts marginalized academics. "Sure you fools make your money and ignore your intellectual superiors, but I'll show you! I'll turn the masses against you!"
Here is an accessible article by a couple of researchers. The dismissal is trivial because it turns on a definitional rather than an empirical issue. https://inference-review.com/article/fiat-recursio
To me, that article seems confused on the following point: It insists that Chomsky wasn't talking about "recursion" in the sense of "self-embedding", but was instead talking about recursion in the following sense:
> In 1963, Chomsky wrote that his grammatical theory expresses “the fact that … our [linguistic] process [is a] recursive specification”9 and, in 1968, that a “generative grammar recursively enumerates … sentences.”10 In 2000, he noted that our linguistic knowledge is “a procedure that enumerates an infinite class of expressions”11 and, in 2014, that “we can think of recursion as [an] enumeration of a set of discrete objects by a computable finitary procedure.”12
Presumably implying that _this_ sense isn't affected by Everett's study of Piraha.
But I thought that Everett's claims, if true, suggested that Piraha might have a longest sentence, and thereby only a finite set of sentences. And this would seem to contradict this other sense of "recursion" that the article's authors want the discussion to be about. (One of Chomsky's quotes refers directly to infinity.)
-----------
Seperately, there's another route of defense available, and the one that Chomsky seem to take in the interview referred to in the wikipedia article on Everett.
> Chomsky called Everett a "charlatan", and said that even if Pirahã had all the properties described by Everett, there would be no implications for universal grammar. "Everett hopes that readers will not understand the difference between UG in the technical sense (the theory of the genetic component of human language) and in the informal sense, which concerns properties common to all languages".
But on a quick skim of Everett's 2005 article, Everett seems perfectly aware of this distinction. His full critique of Universal Grammar is:
> For advocates of universal grammar the arguments here present a challenge defending an autonomous linguistic module that can be affected in many of its core components by the culture in which it "grows." If the form or absence of things such as recursion, sound structure, word structure, quantification, numerals, number, and so on is tightly constrained by a specific culture, as I have argued, then the case for an autonomous, biologically determined module of language is seriously weakened.
This paragraph seems to roughly assert something like: the lack of certain properties in Piraha is _evidence_ against those properties being genetically built-in. Which seems like a totally fair claim to me, and certainly not one that should be countered by accusations of being a "pure charlatan" by the leader of the field.
(Though Chomsky could be referring to some later claims by Everett. I haven't previously read anything by Everett.)
I think you've somewhat missed why the recursive operation is said to be innate. The prediction of UG is not that every existing natural language will exhibit surface properties that could only be described by recursive enumeration, the prediction is that every existing human (without disability) is capable at birth of learning any of the class of possible natural languages (impossible languages being conceivable languages that no human could learn to speak naturally), and that class includes languages that can only be described by recursive enumeration. That property in turn is not a cultural innovation, it's a cognitive computation. The role of culture is to use or not use recursion; not to invent it.
So long as we consider all humans at birth to be equally capable of learning any natural language, theorists must come to understand linguistic ability by looking at the class of possible languages, and so particular data from Piraha is not some kind of disproof, as the requirements for UG are already established by other languages. Everett would rather need to demonstrate that children of Piraha descent cannot acquire e.g. European languages, which is where the (I think crass) accusations of culturally relative racism come from.
It's a slightly coarse analogy but Everett's argument is in effect like saying that because the Piraha do not ride bicycles, they therefore would not only need to invent bicycles to ride them, they would also need to culturally develop the cognitive pre-requisites that are necessary to learn how to ride a bike. This is a category error. The cognitive pre-requisites must exist latently in every human because they cannot be culturally invented (you cannot learn how to make your cerebellum memorise complex motor co-ordination); the abilities are just awaiting the right cultural inventions to be put to use.
Did you actually read the linked reddit post? That was a Chomsky skeptic who still thought Everett was unscientific. There isn't the necessary connection between the things that you think there is.
...Have you heard of a guy called Fomenko and his "New Chronology"? Most historians ignore him, but those who do very rarely pretend that this is a respectable opponent.
Or take biologists' debates with creationists (fitting given Everett's missionary history). To the extent they react to that, it is an inherently undignifying thing to have to react to something like this, with obvious consequences to behavior.
Nevins feels (correctly, in my view, although I wouldn't even waste time on Everett personally) that is the proper comparison, not your garden variety "we have a disagreement".
It is trivial to build a case against young earth creationists and people of that sort. If you want to compare them to Everett, it should therefore be pretty trivial to disprove his point about recursion and pirahas.
I must admit I am very suspicious of the response to an article stating that "Chomski is the caliph of linguistics" being "the caliph does not need to respond to every beggar in the street". Especially when the beggar in question seems to have done groundwork in a way the Caliph did not
Nevins et al did respond. Hell, the guy also, independently got a whole issue of a journal discussing whether the claims on Piraha have any merit. Chomsky is not the one and only caliph precisely because he doesn't need to personally fight every battle, Chomskyan linguistics is a whole field.
The disproval of Fomenko (done, incidentally, by a linguist, Andrey Zaliznyak) was somewhat technical and took some fifty-ish pages. But it was "in the predictable direction". For me, the points by Everett are in the same bizarre category, and yet a formal disproval does take a bit longer than for young-earth creationists. It is no less damning for its being longer and more technical.
Did Nevins or anyone else trained in linguistics also become as competent in speaking Pirahã? I'd be curious to see what we'd find if we simply had more linguists doing field research and learning more languages.
His former wife, Karen, did. She disagrees with his claims. But then, obviously, she isn't exactly an unbiased observer, either. And yeah, a famously closed-off people wary of outsiders living in an area of Brazil that's further restricted for outsiders by the Brazilian government, is... let's say, poorly accessible for researchers. It's a serious problem with some language-rich areas (Amazonia, Iran, China, New Guinea), it's not because people don't want to do research (there was a common meme that Chomskyans only do research on English, but it is largely outdated; for instance, a lot of Dr. Preminger's career is done on fieldwork in Mayan languages, Dr. Nevins is a co-author of a monograph on Basque, and I have a favorite quote somewhere listing a PhD recruiting data from fifty-ish disparate languages, although in that case, often second-hand).
>Or take biologists' debates with creationists (fitting given Everett's missionary history).
An odd little aside, which comes aside as just a drive-by attempt to associate with Everett with creationism even though there's no seeming connection.
I claim that Everett's claims are to linguistics what creationism is to biology. And I do have a hunch that he is more comfortable with such breakage of science because of his connection to another field that famously engages in such.
Should we go over that quotation set again? The equivalence is in level of iconoclasm and holism as well as in Russell's teapot quality of rejection of analyses more "traditional" ("Your data don't show what you claim!" — "They do if you analyze them this way!" — "But you don't have to analyze them this way!" — "Yes, you do, I know because I lived in the tribe for decades and you didn't, so obviously you don't understand!"). I certainly don't claim that there's specific, concept-to-concept mapping to creationism.
(Regarding holism specifically, he refuses to contemplate separation of language and culture because they interact; imagine telling a biologist that they can't study genes because they interact with environment!)
I'd argue anticolonialist politics was probably a big factor: Everett was an American that travelled to Brazil to study a remote tribe and made an ethnographic account, at a time it was becoming increasingly clear that this had happened before and the result was questionable methodology and disastrous consequences for the natives. (Napoleon Chagnon and the Yanomami)
Also, Everett the charming missionary certainly look the part of the gringo with an agenda.
And Kuhn scientific paradigms: if you take Universal Grammar too seriously, Everett's claims sound a lot like "Pirahã people are too stupid to parse tree-like language structures", and this counts as evidence of Everett being both racist and bad at his job *because* of racism.
Well, to give credit where credit's due, Everett specifically rejects the idea of their being mentally deficient. (Whether it is actually logically consistent with his ideas is a tougher question, but people are known to espouse inconsistent or not-proven-to-be-consistent ideas.)
The mental capacity for language? Seemingly so. And then, by the usual genetic drift (probably the version for weakly-beneficial mutations, because being able to do complex thoughts is beneficial, including for intra-group competition), spread to the rest of the population in one of the bottleneck events.
I do tentatively agree, and one piece of evidence relevant here is "anatomically modern humans" in Africa who hadn't done much not-present-in-other-hominids human behaviour in 200000 years, with the parsimonous explanation being their not having that mutation.
Isn't the parsimonious conclusion there that the key factor is IQ, not language? Or is the position that complex cognition and language are inextricably intertwined?
Partly but not fully. The problem is, their brain size _and_ their brain _zones_, to the extent they're recognizable from the cranium, are the same as in modern humans. So, we wouldn't expect huge difference in "horsepower" of the mind.
"one piece of evidence relevant here is "anatomically modern humans" in Africa who hadn't done much not-present-in-other-hominids human behaviour in 200000 years, with the parsimonous explanation being their not having that mutation."
Can you clarify? They haven't done much in what sense and in what timeline compared to what other oganisms? And what is the claim evidence for and why is it evidence for that claim?
And just to affirm: is the claim that whatever distinctive genetic capacities allow modern humans to engage in language as we do arose via mutation in a single mutation relatively recently? If not, what is the claim, and if so, how recently?
*On one hand, all Homo sapiens including Khoisan have language, which means that language must have appeared before the Khoisan split some 150000 years ago.
*On the other hand, there's the whole "cultural artifacts appearing" thing, that seems to happen either after that or around that time. (Scarcity of African archeological record is limiting here.)
*Crucially, anatomically modern humans strongly precede those first artifacts.
*The claim is that the crucial capacity for syntax and its interpretation appeared first, in what amounts to a single mutation, and then that system got hooked up to communicative system in what is plausibly a _different_, later mutation. So, two mutations: to think and to articulate the thoughts.
*Both happened and propagated via a bottleneck event relatively recently (but before Khoisan split — not 50000 years ago, as some places claim, that's a later and different bottleneck that wasn't decisive for language for the simple reason it only affected non-African population). A million years ago, probably neither mutation was present yet. So, we have a timeframe of ~1000000-~150000 for both.
*In the timeframe between a million years ago and 150000 years ago, all the three modern-ish Homo species developed: us, Neanderthals, and Denisovians. (There may have been more, lost to history, just like we hadn't known of Denisovians for a long time: again, scarcity of remains...) There were secondary interminglings, the speciation there was _not_ simple, but, to the best of our understanding of what parts of brain do what, it is likely that Neanderthals first separated before the two language mutations, and certainly at least before the second one. This pushes the date downward, closer to us.
Oh, and note that if someone proves Neanderthals did have language, it pushes the lower bound further behind (makes the mutations less recent) but doesn't change the substantive claims on how it happened — if anything, it ends up _narrowing_ the window for mutation. (Then it could align with the 900000-850000 bottleneck, which currently looks too far away.)
Pedantry alert: Genetic drift is when it's random (e.g. you kill off half the population, turns out this gene was more common in one half than the other, so now it's more common in descendant populations). Language reaching fixation would be directional selection since, as you note, it would be beneficial.
There was no major bottleneck even in the last 150,000+ years and the Khoisan split off before 150,000 years and they do language just fine. There has been minimal gene flow between them and other humans since then and the genes that have been selected for when they interbred with some Cushitic pastoralists 2000 years ago aren't related to language.
The Neanderthals also had small amounts of gene flow 200k-250k and another pulse from 120k-100k, after the most basal, language using, modern humans split off.
The Khoisan are far too genetically diverse to have gone through a bottle neck then. That theory predates our ability to cheaply sequence dna and the dna evidence we have has refuted it.
This is partially a matter of point of comparison: if (as is virtually certain) out-of-Africa population has gone through an additional bottleneck, our baseline of what is high variability in humans changes (and cross-species comparisons obviously have their own problems). Like, there is no serious theory on the market that wouldn't predict that relatively highest diversity would be in Africa, and within Africa — in Khoisan specifically.
(There are a lot of components in how the entire edifice of language arose: for instance, it required greater control of mouth and/or head movements, and the hominid excavations suggest that this arose earlier, which confuses scores of researchers into thinking that's when language arose, including even such brilliant minds as Svetlana Burlak - but it is quite obvious that better control of fine motorics would benefit tasks beyond language.)
Yes. I personally find this implausible and I think he comes to postulate it because he backs himself into a corner by over-simplifying other (non-structural) features of linguistic ability. I nonetheless think the recursive structure-building operation he proposes is essential to explaining some complex properties of semantics, so I end up in the Chomskyan camp even though I (and many others who would be characterised as sympathisers) think we need a strong refinement of the theory.
I think a lot of those features are relegated to interface between language and communication (Phonological Form, Distributed Morphology/Nanosyntax, all that jazz). There is a discussion to be had - between professionals - to what extent that relegation is plausible (for instance, whether Marantz-style case assignment happens in syntax or in the interface), and, of course, the less one can relegate, the less plausible the theory becomes (although, continuing to use the same example, one can fully imagine that the protolanguage, well, didn't have m-case just like many modern languages don't, and the machinery for that developed later). But that's a discussion on a whole other level than needed here.
I'm a big Nanosyntax fan - what it boils down to for me is that I think the T-model with the two interfaces is a conceptual mistake. I think if the foundational assumptions under that are corrected (which requires work at the boundary between linguistics and philosophy of language), Merge will be preserved but it won't sit within cognition as an autonomous syntax module. That then has ramifications for the makeup of the narrow/broad language faculties and their historical development, but it's indeed another conversation :)
Nanosyntax is very much a T-model model though. (Well, it's a Syntax - Lexicon - Interfaces model rather than Lexicon - Syntax - Interfaces model, and there's a bit more lookahead and lookbehind than usual, but it is a T-model nonetheless, after the insertion phonology and semantics operate on the result separately.)
Indeed. I don't believe there are many well-developed non-T-models in generative grammar (however much the specific interactions are toyed with) because that model has been taken for granted ever since Aspects. But I'm not committed to rejecting all T-models outright; only to claiming that they have explanatory potential while misdescribing certain phenomena, such that a future, more explanatory model will incorporate their strengths and reconceptualise their weaknesses (as is characteristic of scientific development).
The 1950s rejection of behaviorism was important and powerful, but it relied on a lot of arguments about what could and couldn’t in principle be done that have since been falsified by large language models and other work. Many of Chomsky’s arguments early on involve the idea that you can’t make sense of probability as anything but a frequency or some simple observable feature like pitch or volume, which is an appropriate refutation of Skinner, but not of any of the many constructions of Bayesian probability in the second half of the 20th century. He also argues that you need to study language separately from the world, but modern large language models seem to do a good job of creating a model of the language-world system, whether or not they adequately separate it into the two components.
Many of the criticisms of behaviourism had nothing to do with token sequence probabilities, or even linguistics, which is why it failed as a paradigm in all of psychological science. Its foundational problems have not been addressed in any way by machine learning, and the capability of LLMs to do better next token prediction than Chomsky expected doesn't mean much in an integrated theory of human cognition, language included.
Saying that behaviorism “failed as a paradigm” doesn’t mean we were right to dismiss every one of its claims. Every shift of paradigm usually involves rejecting the problematic claims of that old paradigm, but also rejecting some of the things it got right. Behaviorism had a lot of valuable ideas that we are properly learning to appreciate again after decades of rejecting them because of their association with its over-ambitious claims.
English is also full of this kind of thing if learned as a second language. E.g. write/right/rite/riot, not to mention such wonderful pairs as come/cum, sheet/shit, beach/bitch, can't/cunt, fox/fucks.
> These two points are not unrelated. Everett’s core idea is that language should properly be thought of as an invention rather than an innate human capability. You might ask: who invented it? Who shaped it? Lots of people, collaboratively, over a long time. In a word, culture. As Everett notes in the preface, “Language is the handmaiden of culture.”
I feel like the existence of Nicaraguan Sign Language - a natural language constructed by deaf children in a society without any preexisting sign language use - very strongly suggests some intrinsic biological quality to language. That is, rather than it being an invention that propagated the same way as writing or metallurgy, it is a product of human brains on some level which children can reinvent, including features like grammar.
I don't understand the relevance of the question of whether it's one or the other. Obviously on some level, there is [information encoding of perceived real world things] occurring in the human (or animal) brain, dependent on the [info encoding - "brain"] capabilities of that creature. Is that an internal language? I don't know, does the categorization matter? I'm a computer programmer and certainly it sounds like a language to me, for whatever that's worth.
This means that you could reverse the process and do something in the real world (sounds, calls, whistles) that mirrors that encoding in your brain - in a way that other brains could interpret in the same, real-world verifiable way, for evolutionarily beneficial reasons like coordination on hunting. This seems to of course be both invented/learned (i.e. I point at the tiger and make the same sound, and you understand that it is a reference to it) and innate (people all around the world seem to understand that the word "squirrel" is more likely to describe a small fluffy thing than "hippopotamos").
But I don't get why it matters. And by matters I mean "in a way that, if we prove it's one or the other, that's a thing that can be objective proven, rather than just vague category definitions, and then informs our understanding of the science such that we can apply it to productive new insights."
To me it's obviously sufficiently both, not because I'm a genius linguist (I am not) but because treating it as a thing that obviously has many influences from our innate biology AND as an invented thing is necessary to (me!) understanding it in a way that we (I?) couldn't if I had to declare it 100% one or the other.
> But I don't get why it matters. And by matters I mean "in a way that, if we prove it's one or the other, that's a thing that can be objective proven, rather than just vague category definitions, and then informs our understanding of the science such that we can apply it to productive new insights."
Understanding where language comes from will probably eventually be useful to genetically modifying other species to be capable of talking. That's the only end use that immediately comes to mind.
If it's just brain power + culture, that implies you can just keep sticking more neurons in a dog's brain, and at 86 billion it'll be able to use a full language. If there's some specialized structures, you'll either need a lot more than 86 billion neurons (so the resultant creature can use nonspecialized bits, like how we solve math problems) or to introduce analogous structures.
But yes, definitely I would agree that "understanding where language comes from" would be important in that context, but I'm not sure why the exact ratio matters.
"If it's just brain power + culture, that implies you can just keep sticking more neurons in a dog's brain, and at 86 billion it'll be able to use a full language"
So replace "dog" with "Chatgpt" and obviously we've got a thing that is pretty language competent, right? I feel like LLMs are the example/thing we're all implicitly referring to?
" If there's some specialized structures, you'll either need a lot more than 86 billion neurons (so the resultant creature can use nonspecialized bits, like how we solve math problems) or to introduce analogous structures."
But here's where the categorization breaks down for me. Obviously for human language there ARE specialized bio/neuro structures that are necessary for actual spoken human language: the parts of your brain that encode information, voice boxes, those parts of the brain the article above talks about, etc... for a LLM, though, do we categorize them as having those things or not? Does the screen and text output sub-module and hard drive structure count? Why does it matter, at least in an important way?
See the other comment of mine on LLMs :) LLMs are the pinnacle of "every time I fire a linguist, productivity goes up", because, well, constraints of natural language learning by a child are quite different from constraints on computers. Look up CHILDES corpus: it is small and unrepresentative (poverty of stimulus argument), if you try to teach LLMs on what children actually hear, rather than on what gazillions on balanced and generally-more-complicated data OpenAI feeds them, you won't get as good results as children routinely get, and you won't get biases toward learnability of existing structures over non-existing. This suggests that LLMs do just brute-force the problem.
Yep, my favorite feature of LLMs have been the discussions they inspire about the extent to which they are doing the same thing as humans.
"This suggests that LLMs do just brute-force the problem."
That gets right at what I'm curious about with recursion-as-a-physical-process. But I guess, is the analogy that informs the conclusion "brute-force" really "kids vs LLMs, two equivalent things" or "kids, who have billions of years of evolution that worked on tools for understanding things, vs LLMs, whose tools are computer submodules designed by hardware/software people for mostly different things." The fact that kids do way better with less training data at a young age, but at "adulthood", ChatGPT seems to, having been trained on petabytes of data in its "childhood", now has tools that let it do better at some things.
Did kids (and therefore humanity and animals over millions of years) just brute-force problems in the distant past, solving them with (now built in) informational/genetic structures (brains, memory, eyes, neurons or.... recursion neural tools?) that they then passed on to their descendents, so that kids today can learn important things with vastly less data? What if the comparison doesn't work because each side isn't at the same starting point, and can't possibly ever be? Is that better described as "they just think 2 different ways" or "they think the same base level way, but they just have millions of years of history and developed tools, so it looks totally different now"
My bet is on "2 different ways", but this is one thing I am _not_ ready to debate, and at some point it falls into semantics (like with pointers-to-pointers: in some sense, it is a "developed tool", right?).
>But I don't get why it matters. And by matters I mean "in a way that, if we prove it's one or the other, that's a thing that can be objective proven, rather than just vague category definitions, and then informs our understanding of the science such that we can apply it to productive new insights."
This has been made moot by LLMs' linguistic success, but, if that _hadn't_ happened, then the question of "Does language competence require some special structure, inborn in humans, or is it purely a cultural artifact, learnable by the same general purpose learning algorithms that we use for everything else?" would have mattered to AI design. AFAIK, we now get to ask the same sort of question, but probably about e.g. executive function, instead.
I don't see how humans couldn't have evolved language faculties once they had language, but developing a cludgy capacity to be able to communicate with one another, then gradually evolving dedicated capacities to massively upscale it to the point where it's helpful for thinking seems way more likely than developing full-on linguistic thought in one step then using it to communicate (presumably originally just among the kids of whatever super-hominid developed this ability).
Well, chimpanzees also have a kludgy capacity to communicate with one another, so obviously our ancestors did. But the most complicated language encoding we know of in the animal kingdom is, AFAIK, that of prairie dogs, where they can tell the others about the size, speed, color, heading, and species of an intruder (and even if a human has a gun or not!). This is obviously hugely different from the thing we're doing right now in a way that the thing we're doing right now isn't hugely different from chatting around the fireplace.
It's possible this is a slow-and-gentle evolution. But the fact that "behavioral modernity" pretty much appears all at once seems to suggest there was some big sudden shift.
There's some debate on if it's just an artifact of how the evidence is recorded, versus actual near-simultaneous arrival, but it reeeeally seems to be near-simultaneous arrival.
If the "I-Language" theory is true, as I understand it.... couldn't the "big sudden shift" be communication ?
If people were developing grammar/syntax/whatever initially for "having complex thoughts" then you'd expect all the hard work on recursiveness to get done on this without communication. It can continue gradually evolving, just as most other human capabilities gradually evolved, and would not neccessarily leave any physical trace. Getting really quite complex, and close to full modern lanuague complexity as its always adaptive to have beeter, more well structured, internal thoughts this can drive itself for hundreds of thousands of years.
Then .... one day.... one human or another discovers he can make sounds to share this "intuitive personnal i-langauge" with someone else, and the "someone else" groks it !!! Between the two of them they are now able to *share* the highly complex thought patterns that have been evolving for generations for the first time.
Because the huge infrastructure of grammar had been built up through standard gradual evolution there'd be no need of a "big bang" genetic evolution, which seems to me unlikely on basic genetic principles. Genetic I-Language could have developed as slowly as the mammal eye.
But then once they can *actually communicate ideas* it can move explosively fast as it is now cultural! You can go from 0 to "burying your dead and painting deers on caves" in a few generations as humans explosively turn their fully intrinsic personally developed I-Language into a commonly used/commonly codified/common "this sound equals this concept"/collaborative E-Language.
Why does "behavioural modernity" under Chomsky I/E laguange distinction have to be related to the genetic development of the basics of the I-Language ? And not the titanic shift that the theory intrinsicly implies is there (the titanic day when for the first time a human communicated by sound a complex i-language thought that another i-language human was able to interpret into his thoughts...inventing the first compound sentence in the first e-language).
It would also seem to be generally supported about what that "cultural modernity" was.... which was an explosion of "humans making symbols" (i.e. cave paintings) or "humans acting symbolicly" (i.e. covering buried dead in ochre/having burial ceremonies/carving statues).
Surely thats the kind of behaviour you'd expect to see increase massively exponentially at the point that internally developed thought language got (at least partially) repurposed into a communication method for spreading thoughts by interspersing symbols <I-Language grammatical recursive thought in 1st human> <-> Symbol shared with another human either verbally or visually <-> <Identical I-Language grammatical recursive thought in 2nd human>.
I/E language implies there would have been the day of the creation of the first syntacticly/grammatically complex E-language.... You'd expect (if the I-Languages were already well developed by then) for this to lead to an explosion of symbols as an already well developed I-Language got rapidly translated into shared common symbols for use and communications.... and what we see with behavioural modernity is an explosion of those symbols.
If that explosion wasn't behavioural modernity.... where is it ? Because at any point I expect the first humans to first use verbal symbols to translate complex recursive I-Language to create the first proper e-language, I'd also expect to see an explosion in other symbols at that point as well as "the new tool" they've discovered culturally explodes across the landscape.
Doesn't it make sense for the genetics of this all to remain the slow, gradual, incremental process we see with all other genetics.... and the big/quick/explosion thing that appears at a pace more usual of cultural evolution to be a *cultural* outgrowth spandrel of that (explaining how it can spread so quickly *once discovered* as its no longer confined to the descendant line but can spread horizontally as well as vertically across hundreds of thousands of people every generation).
Or consider developmental verbal dyspraxia. A mutation in a single gene (FOXP2) causes it. People with this condition seem to exhibit ordinary non-linguistic intelligence, but only speak language at a very basic level. More interestingly, this seems to be because they have to learn it without any innate grammar. For example, if you learn a new word, "phlogon", and then you hear someone use the word "phlogons", you infer that "phlogons" means more than one "phlogon". They need to learn both words separately. (As adults, naturally, people teach them these rules and they can remember them, but they need to learn each grammatical rule consciously, instead of intuiting it the way that everyone else seems to)
Yeah, NSL is sometimes taken to be evidence of a biological wiring for language, but it’s also in some ways evidence of the importance of social interaction, since it emerged from a bunch of kids being put together in a school, who all independently had their own “home sign” systems (home sign could also be seen as evidence of biological wiring for language at work). I don’t think it’s either/or. But either way, really fascinating case of a language emerging!
Yeah. I think realistically, if language-as-we-think-of-it emerges all at once, you can't say "language is for thinking" OR "language is for communicating." It would have been a single thing that altered how human beings both thought AND communicated, at the same time, with the thinking element probably requiring interaction with others with the same architecture in order to reach its full potential.
You don't think those kids' parents ever pointed at things or gestured to them in any way?
Eta: I agree that the existence of home-signers is remarkable, and probably points to some sort of innate language ability. But these still aren't, like, feral children raised by wolves.
It’s probably useful to think of it on analogy with cooking. Humans started cooking before we adapted to the presence of cooked food, and after those adaptations cooking is now essential for almost all foods that we eat. But there’s not one single adaptation that explains this.
Agreed. Just as, after the initial invention of cooking, our digestion and cuisine co-evolved to become more and more dependent on cooking, and to more thoroughly exploit the advantages of that capability, if a linguistic "MERGE" capability resulted from a single mutation, I would expect that it would have e.g. increased the advantage that a larger vocabulary conferred, and, more generally, increased the frequencies of genes that provided complementary linguistic capabilities. So, in the end, many genes may be involved in our current linguistic capabilities, above what were present pre-"MERGE".
>Regardless of the details, a generic prediction should be that there is no longest sentence in a language whose grammar is recursive
Why then:
>He does explicitly claim (in the aforementioned paper and elsewhere) that Pirahã probably has no longest sentence, which is about the most generic anti-recursion statement one can make
Wouldn't a non recursive language *have* a longest sentence?
2.
>I find this notion of recursion completely insane, because it would imply (i) any language with more than one word in its sentences has recursion
From the quotes, it sounds like the linguisists in question emphasize the role of recursion as an iterative - not isolated phenomenon. A sentence with two words might be a primitive instance of recursion, but lack the iterative flexibility that characterizes it.
3.
>But this is kind of a dumb rhetorical move. Either the theory makes predictions about real languages or it doesn’t
This objection doesn't seem fully compelling. All models are wrong, but some are useful. The mere existence of a supposed language doesn't prove that the model isn't generally useful. Did not the vast majority of langauges discovered since the advent of the theory fit the model?
4.
>Is “chicken” the subject of the sentence, the object of the sentence, or something else? Well, it depends on the context
This needs some more elaboration as to how it demonstrates a clear lack of noun/verb distinction. After all, similar ambiguous phrases could be constructed in other languages. Is the point that these are typical, rather than exceptional for the language?
5.
>He mostly seems to take issue with the idea that some region of our brain is specialized for language
Is that actually necessary for Chomsky's model? That is, does a sudden take off of language due to a novel mutation facilitating universal grammar necessitate a single specialized part of the brain?
6.
>ChatGPT...the brute empirical fact remains that it can handle language comprehension and generation pretty well. And this is despite the conception of language underlying it—language use as a statistical learning problem, with no sentence diagrams or grammatical transformations in sight—being somewhat antithetical to the Chomskyan worldview.
While Chomsky may personally have been skeptical of the success of LLMs, I'm not sure why they pose such a problem to his general model of language. After all, his theory, as presented, seems to be about the human faculty of language, and how evolves it. It seemingly doesn't preclude the possibility that a machine could mimic such its results using a completely different framework; after all, there are many tasks which are performed using quite differnet systems by humans and machines.
>It’s conceivable that a completely orthogonal system designed according to the principles of universal grammar could outperform LLMs built according to the current paradigm...
Again, why is this central to the Chomskyan model? That humans evolved the faculty of language as a particular mutation with particular characteristics doesn't preclude a different or even better system being developed.
>Chomsky’s claim that engineering success does not necessarily produce scientific insight is not uncommon, but a large literature speaks against it.
Again, how does "large literature" speak against it? Presumably the reviewer doesn't merely mean that many disagree with Chomsky, as that's trivial, but rather that the large literature about engineering sucesses is an existence proof against Chomsky. But is it? Following the review's presentation of Chomsky, would not Chomsky refer to such literature as the cataloging of butterflies bodies? Does it ultimately provide a parsimonious model of the concept of language, or insight into its evolution?
>Again, how does "large literature" speak against it?
The example that I've seen cited is that thermodynamics owed more to heat engines than the other way around. I would guess that this is probably true of mechanical engineering preceding the science of classical mechanics as well. I'd guess that electromagnetism is probably the other way around: There was a lot of scientific investigation before the first motor.
Metallurgy is probably the most extreme version of this. Smiths knew how to create metals with very specific properties with precision without understanding anything about the underlying chemistry or physics. The actual understanding of the chemistry of how to make steel and bronze and, why the specific heatings, coolings, hammerings and ratios of materials used were necessary, came far later.
This was a well written overview of a fascinating topic, but I don't think it's a very good book review. Like so many of these reviews, the author uses it primarily as an opportunity to review a topic they are interested in, rather than a book about that topic. Obviously, one needs to have some background knowledge of the topic, but when the the actual book under review only takes up one out of five sections, it's not really about the book or the ideas it discusses.
This is neatly illustrated by this sentence: "There’s a lot of interesting stuff in there I haven’t talked about..." But isn't talking about the interesting stuff the main point of a book review?
Exactly my thoughts, great article, not a great book review. Also think some editing on subject ordering would have been good (define recursion when it was mentioned, maybe move the chomsky stuff a little later)
I’m afraid Chomsky wins this one because of the work in computation theory that he did which the reviewer doesn’t reference at all. Chomsky’s biggest insight was what is now known as the Chomsky Hierarchy of languages or grammars:
Regular Languages, recognized by finite-state machines (animals have these)
Context-Free Languages, recognized by machines with a simple memory element such as a pushdown stack which allows objects to be merged or linked to each other (requires a brain modification; perhaps humpback whales and other animals can handle this)
Context-Sensitive Languages, recognized by linearly bounded automata that can mark or operate on their input data (allows language to interact with the external world so that we can build things following instructions)
Recursive (in the Godelian sense) languages, recognized by Turing machines and other automata which can use expanding amounts of rewritable memory to keep track of things, attained by humans with the invention of writing (pre-literate humans were not Turing Complete and the Church-Turing thesis is not actually about what humans can compute but about what technology-aided humans can compute).
Pirahã people (or at least the young children in it) have the capability to learn other human languages and the same brain improvements that the rest of us have. Their language is restricted in interesting ways but the only really important question about it, at the very top level, is whether they are capable of doing things that people in other societies can do, or whether their language makes this impossible.
My guess is that, when a phonetic alphabet is eventually developed for them, they will indeed become capable of reading and writing their own language, and doing the other things that writing enables.
A really good first test case of this: have they been taught to count yet, and imported or created number words that allow them to communicate the cardinality of a large collection of objects reliably so that they would know if something had been added or taken away?
Everett’s work may show that the language of the Pirahã is missing certain structures that other languages have; but I don’t think it falls outside Chomsky’s framework. If there is a concept which Pirahã cannot understand no matter how you attempt to translate it into their language, for example if they can’t be taught to count by giving them a few number words and showing how it works, that would be a much bigger deal.
Thank you so much for listing those 4 categories, very helpful.
Honest, a-political question for my own understanding, coming from a computer science background: how is that categorization not just a measure of "the amount of working memory available to the user of the language"? At least, in the context of what is actually going in the animal's brain. I 100% get how it's useful to us humans, seeking to categorize it by different features.
Regular languages (FSMs) words as you describe, don't necessarily refer to base primitives. The animal makes a call that says "there is a predator nearby". It would not be a language, or intelligble to the other animals, if they didn't understand the concept of predator or subconcepts like "predators are threats to our ability to pass on our genes." Sure, maybe it's not "memory" but "instincts" but (from a CS perspective it's still "info we already had")
"Context-Free Languages" add "memory" but it seems (naively, to me) like it's just more complex "memory."
For the "Context-Sensitive Languages" and "Recursive languages" - again, it just seems to me like the base difference is "even more memory, better configured or more complex perhaps." What is the actual physical brain difference between processing the reference to "predator" in the animal call "there's a predator nearby" and the highest level of language "the symbology of X in War and Peace is evocative of the concepts of Y in the bible," except in the complexity of the concepts being referenced? Does the brain actually do anything different when it references "the bible" than "predator", except in scale/complexity?
Working memory increase was certainly a facilitator (this happens to be Markov the biologist's favorite claim on the matter). However, "better configured" already betrays that quantity is not yet quality. You can have an enormous working memory but if you cannot use it to make, as you put it in the other thread, pointers to pointers, it's moot.
(Also, people whose Daneman-Carpenter test returns 1 or 2 for working memory can still use language pretty well - although they do have more problems with complex sentences, if you simplify the number of elements used enough, you can still get some recursive hierarchies.)
"However, "better configured" already betrays that quantity is not yet quality. "
How do we distinguish though? When I typed that, I recognized that it would imply a quantity/quality issue, but for...
"people whose Daneman-Carpenter test returns 1 or 2 for working memory can still use language pretty well"
... how do we know it's not just straight memory capacity? Wouldn't that be exactly what we see when you say:
"although they do have more problems with complex sentences, if you simplify the number of elements used enough, you can still get some recursive hierarchies"
Wouldn't that be consistent with "it's working memory" as the relevant concept? Let's say you've got 100 bits, where you can get "the sky is blue" because sky takes up 20 bits and blue takes up 40 bits, and 60 is less than 100. But for "I remember when last week, Steve said that Alice thought the sky was green" doesn't work for you, because sky is 20, green is 40, last week is 30, Steve and Alice each take 50, and the concept of Alice thinking herself (recursive) about the sky being green is another 110 bits (Alice + sky + green). So it's hard for you to understand?
I mean on some level, the brain kinda has to work that way right? I could see if the "quality" was an LLM-like capability of linking/associating bits together to differing degrees (green is 5 distance from blue, but Alice is 100 distance from Steve) but that.... is surely physically just the capability of determining (perhaps in the brain with true physical distance) the difference between different groups of bits. You will need to be able to handle more of them in order to do so. Is it really "recursive" as a fundamentally different concept, or just that recursion (mathematically, like multiplication) requires much more working memory, specifically to keep the relationships between the different levels of recursion accurately intact?
To bring back up to a higher level - is higher intelligence empirically correlated with better memory? Memory not only of specific facts but of the interplay/relationships between those facts, and the sub-facts that are implied by them?
Sorry for rambling, this is just fascinating to me, and thank you for taking the time to explain to a layman.
Your Steve-Alice example is exactly what we'd expect if memory were the only relevant issue. However, a person who can't reliably hold a sequence of "Steve, Alice, apple" in their memory in Daneman-Carpenter test is still able to hold "Steve thinks that Alice ate an apple", even though without recursive Merge, one would expect that the latter takes more memory, not less. (While in reality, it seems that it is composed as "[Steve [thinks [that [Alice [ate [an apple]]]]]]", with only two elements combined at each moment.) This suggests some privileged mechanism for exploiting "pointers to pointers" to ultimately use less memory.
The Chomsky hierarchy is an important mathematical result, but I don’t see how this tells us about human language. We don’t actually have as good evidence as we think that human minds are Turing complete or even at the level of context free grammar. We know for sure that a finite neural net like an LLM *cannot* reproduce a context-free grammar (that is one of Chomsky’s important results!) and yet it seems to be able to imitate one just as well as humans do. This is strongly suggestive to me that the human mind (which we know is also implemented by a finite collection of neurons, though we don’t know if the architecture is really a neural net) is actually a finite state machine that has somehow managed to implement within itself a rough approximation of a large initial fragment of context-free grammars and Turing machines. If so, then it’s quite possible that in whatever sense a human mind is able to implicitly define a full Turing machine, a neural net may be able to too, despite its in-principle limitations.
You didn’t read the part where I said that unaided human minds are NOT Turing-complete. They are only Turing-complete when combined with the TECHNOLOGY of writing.
Whether LLMs and neural nets can eventually do the equivalent is an interesting question—quantity compensates for quality so their limitations may not be easily apparent. But computers can obviously be Turing complete and I expect that the next generation of AIs will have more sophisticated architecture that can emulate human cognition more directly and require less brute force than now.
> We know for sure that a finite neural net like an LLM *cannot* reproduce a context-free grammar
I don't know much about linguistics, but modern LLMs are basically Turing complete. At least in the sense that a normal home computer, or a human with a finite amount of memory and paper to write on are basically Turing complete.
> To be clear, this usage of ‘recursion’ seems consistent with how many other Chomskyan linguists have used the term. And with all due respect to these researchers, I find this notion of recursion completely insane, because it would imply (i) any language with more than one word in its sentences has recursion, and that (ii) all sentences are necessarily constructed recursively.
No. The claim is actually somewhat more technical. Dolphins have the equivalent of sentences, but they are linear, not recursively hierarchical. An illustration: in English, "I saw a girl with a telescope" has two readings, "I [saw [a girl with a telescope]]" and "I [[saw a girl] with a telescope]" (i.e. using a telescope). Such structural ambiguity is not possible in dolphinspeak, where linear position _is_ hierarchical position but is possible in human languages.
More generally, Everett is exactly as much of a buffoon as Nevins claims he is, and "common sense" is a poor guide to linguistics.
This is an honest, a-political question, coming from a computer programming background: what is the point of the term "recursion" in the context of "lowest level description of language?" I know how *we* use it, and the difference in programming situations that are and are not recursive, and many of the ways it is used in this discussion seems similar, but at the lowest level of understanding language and how it relates to the brain, my understanding of the similarity breaks down: what is the point? Isn't everything expressed in language (and I mean this as "everything" like down to the atoms) recursive? When I say "my friend Bob" my brain (and the people I'm talking to) understands that I "actually" mean "that sack of water and carbon over there, made up of certain atoms and molecules, and electrical neural structures that contain the memories of our relationship" - and then all the various recursions down to "water is H20" and so on.
Even a single word: "telescope" must be "recursive" on a number of different mental concepts and the way they're encoded in the brain of the speaker/listener. Certainly a more complex sentence like above: "I thought you said you saw a telescope" encodes far more complex recursion across multiple things, but it's all just hierarchical: "this refers to this [ which refers to this, which refers to this...] and that [which refers to this, which refers to this...]". This is reflected (haha) in programming where recursion is "really" just "go to this part of the code and run it" where the code and the parameters and the variables can be self-referenced and iteratively changed - it is described as "recursive" for human understanding. The computer doesn't care.
So I guess I'm asking: in the context of understanding language at its lowest level, what does "recursive as a concept" get us? That would not be just as easily described as "a thing that refers to another thing, which naturally could refer to another thing (or things), and so on down to whatever the lowest level of info encoding."? Yes, I can totally see why "referencing a thing that itself references a ton of other things, especially things that reference ever more things, including itself" is an advanced concept that you might need more brain power (or GPU cycles) to understand, but I don't get why it's *different*.
That's an interesting perspective, but, using computer science jargon, I think you mix the ability for having pointers with the ability for a function (namely, Merge) to call itself. You can (and probably do) use the former for the latter, but it doesn't mean they are one and the same, and in practice, not every system with pointers is a system with recursion (although the fairly unique ability of humans for symbolic pointers was probably a prerequisite).
And more generally, "the lowest level" is neurons firing electrons and ions, which is no more useful than "silicon pieces firing electrons" for computer, we don't actually describe what computer does that way. (I am reading "Gödel, Escher, Bach" now, and there's a related discussion.)
Right, but in computer science, at a very low level, recursion IS just clever, self-referential (critically: self-referential as perceived by humans) pointers.
" not every system with pointers is a system with recursion"
Right, recursion (as a CS principle for humans) is pointers pointing at themselves... but it's still pointers (for computers). The abstraction of how recursion works on the code and variables it is affecting is important for humans to understand - not the computer. The computer still sees it as a one-way stream of 1s and 0s - though when you add recursion to your language, you often do need to add a new layer of debugging in order to accurately represent the concept of recursion so the human can understand it.
"And more generally, "the lowest level" is neurons firing electrons and ions, which is no more useful than "silicon pieces firing electrons" for computer, we don't actually describe what computer does that way."
Well, sure, when we are.... talking about the computer's behavior at a high level! High level = meaning dependent on embedded (recursive?) concepts. But I'm interested in understanding how that lowest level can be built up into that highest level, and which parts are "base", and which parts are just other higher order (potentially recursive) structures. That lowest level for computers is "is this bit a 1 or 0". For LLMs, a key additional lowest level is added - "what is the distance between the concept this 1 represents, to the concept that 0 represents (represented, of course, as additional 1s and 0s)" - which I think is probably something that the human brain relies on as well. That basic structure will let you create brains and language - and, I think, amoebas and crystals and the universe.
Sure, if you know how to build - aka if you have a blueprint - aka if you have a higher-level structure. I do not think the lowest level is informative on itself, for either human language, other aspects of human cognition, or computers. At the lowest lowest level, we're all just a bunch of interacting quarks. We build models of higher levels because they actually reflect useful generalizations. Your mileage of what's a useful generalization may vary, but just like I speak of a processor as an abstract machine that performs calculations not as of bunch of quarks organized into a bunch of atoms and electrons interacting in such a way that they happen to correspond to said calculations, I speak of a function calling itself not of a line of 1s and 0s that encodes a piece of memory that... you get the idea.
"A map is not the territory, but you cannot put the territory into your pants' pocket".
"Sure, if you know how to build - aka if you have a blueprint - aka if you have a higher-level structure. "
But at some point, we didn't. Isn't that an important distinction? The things that existed right before amoebas, or crystals or brains or languages did not have those blueprints. They created (evolved) those higher level structures emergently, based only on [whatever the lowest level is] and [whatever previous higher level structures they had built on top of it].
"We build models of higher levels because they actually reflect useful generalizations."
Totally agree, where "we" is humans who need to understand it.
"Your mileage of what's a useful generalization may vary, but just like I speak of a processor as an abstract machine that performs calculations not as of bunch of quarks organized into a bunch of atoms and electrons interacting in such a way that they happen to correspond to said calculations, I speak of a function calling itself not of a line of 1s and 0s that encodes a piece of memory that... you get the idea."
Again, totally agree in the context of "letting humans understand what's going on." But in the context of the debate around "recursion - do they [a language, the piraha, animals, etc...] have it?" it seems like it might matter a lot if recursion is just a concept that humans created to describe the salient (to us) aspects of "clever pointers in our brain", and the slow build of complexity around languages was "ever more clever pointers" rather than "the adoption of this importantly new break-from-the-past category of thinking."
Evolution throws things around almost at random and looks what "sticks" in terms of helping/not harming genetic fitness. But how gradual evolution is is very different for different things. Getting a pointer to a pointer is "importantly new break-from-the-past category of thinking", for the same reason RNA+DNA system is an importantly new break-from-the-past category of replicator (over RNA world).
Why is it that so many critics of Everett and defenders of Chomsky that I see don't simply critique Everett, but call him names and say nasty things? You do that here when you call him a "buffoon." Whatever the merits of either position, I find it really unappealing and off-putting and it makes me not like people who defend Chomsky. At least from the outside, it gives me the vibe that people who defend Chomsky are dogmatic and insecure, that raises a big red flag for me and probably for others that make us more sympathetic to Everett than we might otherwise be.
Another thread here goes into detail on this. In short, it's because Everett (unlike, say, Kenneth Hale with the non-configurational suggestion or Mary Dalrymple with Lexical-Functional Grammar) "started it first", in a sense, by playing the big, Fomenko-style iconoclast who has seen past Chomskyans' delusion. (All the more surprising given that in personal - well, Facebook - communication he doesn't seem an unpleasant person you'd expect from this characterization.) Think biologists that have to respond to creationists. Could one, in principle, be a bigger person and respond the way you respond to normal scientific disagreement? Sure. Is it the least bit surprising that this isn't what actually happens? No.
Oh, let's see. From the lasst pages of Everett 2005:
> 2. Linguistic fieldwork should be carried out in a cultural community of speakers, because only by studying the culture and the grammar together can the linguist (or ethnologist) understand either.
Read: "Every one of you who hasn't done what I did, living with the group for many years, is a charlatan who doesn't understand the language they study".
> 3. Studies that merely look for constructions to interact with a particular thesis by looking in an unsophisticated way at data from a variety of grammars are fundamentally untrustworthy because they are too far removed from the original situation.
Once again but with passion: "You, yes you who wrote your thesis on binding violations or person-case constraint or whatever, are fundamentally untrustworthy".
How did the review put it? "This is the closest you come to insulting someone in a published paper"?
A bit later:
> This beautiful language and culture, fundamentally different from anything the Western world has produced, have much to teach us about linguistic theory, about culture, about human nature, about living for each day and letting the future take care of itself, about personal fortitude, toughness, love, and many other values too numerous to mention here.
...What is this? This is not how scientific (or linguistic, or historical) articles are written. This is sometimes how anthropologists write, but this is a condemnation for them, too.
> The need is more urgent than ever for field researchers to document these languages and for more individuals and foundations to follow the lead of the Hans Rausing Endangered Languages Document Project and donate to support research on them.
Oh, and call to action. "Here's how cool my sponsors are, gimme money money money". This isn't a grant proposal.
(And there's also a strong inconsistency where Piraha speakers are described as monolingual in some parts but as bilingual in other.)
Thanks for providing these quotes. Previously, I had said this: “What in particular did he do that you see as "starting it"? You say "playing the big, Fomenko-style iconoclast." What does that involve?”
I do not believe any of the quotes you provide present significant evidence of what I asked for. Some of the remarks may be subject to the criticisms you present, but they don’t seem to me to make a case for the claim you initially made. I’ll go through each and comment on why that is.
Quote #1: “> 2. Linguistic fieldwork should be carried out in a cultural community of speakers, because only by studying the culture and the grammar together can the linguist (or ethnologist) understand either.”
You respond to this by interpreting it as “Read: "Every one of you who hasn't done what I did, living with the group for many years, is a charlatan who doesn't understand the language they study".
Where are you getting this stuff about being a “charlatan” from? If one’s view is that language is integrally tied to culture in inextricable ways, then it’s simply true that one cannot fully understand a language without understanding culture. Nothing about making such a claim requires you to believe that anyone who doesn’t do this is a “charlatan.” All it would entail is that those who don’t do so won’t have a complete understanding of the language. That trivially follows from Everett’s perspective on language. Suggesting that those who take a different perspective on language won’t be able to have a complete understanding of it doesn’t mean that those people are “charlatans.”
I asked for remarks from Everett that support iconoclasm. What I see is a remark that contains no insults or critiques of those who hold views contrary to Chomsky’s, followed by you adding commentary proposes I interpret Everett as implying this. If what Everett says here is correct, then it may very well be that those who hold contrary views about the importance of culture are mistaken. This doesn’t mean they’re charlatans. It just means they’re wrong. Incidentally, Chomsky has explicitly called Everett a charlatan.
Quote #2: > 3. Studies that merely look for constructions to interact with a particular thesis by looking in an unsophisticated way at data from a variety of grammars are fundamentally untrustworthy because they are too far removed from the original situation.
This looks like a claim alluding to the importance of understanding language as it is used in social contexts and raising doubts about approaches to language that don’t do this. It reminds me a bit of remarks I’ve seen about usage-based models of language. Nothing about this strikes me as evidence that the person who said it is objectionably striving for fame as an iconoclast.
You interpret it this way: “Once again but with passion: "You, yes you who wrote your thesis on binding violations or person-case constraint or whatever, are fundamentally untrustworthy".”
Much of my own research focuses on questioning the methods and validity of research in my field. I frequently argue that a study’s methods are invalid, and I may say things like “We should not trust the results of this study.” This in no way indicates that the authors of the study aren’t “trustworthy.” Calling people untrustworthy implies those people are deceptive or bad people. Saying the results of their studies are “untrustworthy” means something quite different: it means we should have doubts about whether the findings of those studies are true. It has absolutely nothing to do with the honesty or character of the authors of those studies.
What I see in this remark is a comment about the trustworthiness of studies, not the authors of those studies. Given this, not only does this not strike me as evidence that Everett is seeking to be an iconoclast or acting like a jerk, it strikes me as evidence of what I suspected: that you and other critics of Everett are projecting, imagining hostility and iconoclasm where it isn’t present. Everett’s remark strikes me as completely innocuous and in no way an indication that he’s seeking fame as a paradigm-shifter.
Quote #3: > This beautiful language and culture, fundamentally different from anything the Western world has produced, have much to teach us about linguistic theory, about culture, about human nature, about living for each day and letting the future take care of itself, about personal fortitude, toughness, love, and many other values too numerous to mention here.
You say of this, “...What is this? This is not how scientific (or linguistic, or historical) articles are written. This is sometimes how anthropologists write, but this is a condemnation for them, too.”
Even if a remark like this is inappropriate for a scientific article, well, okay. Then it’s inappropriate for such an article. But that’s not very good evidence of “starting it” or seeking iconoclasm. It also doesn’t strike me as a good reason to insult Everett or treat him with contempt or hostility.
Quote #4: > The need is more urgent than ever for field researchers to document these languages and for more individuals and foundations to follow the lead of the Hans Rausing Endangered Languages Document Project and donate to support research on them.
You interpret this as “Oh, and call to action. ‘Here's how cool my sponsors are, gimme money money money’. This isn't a grant proposal.”
Your interpretation again strikes me as dubious and more than a little cynical. It may be a bit weird to suggest that we should donate money for a cause in an article like this, but again, this does not strike me as evidence of someone “starting” anything or seeking fame via iconoclasm. This looks like someone who thinks endangered languages are important sources of insight and that believes we should donate money to preserving those languages and cultures. Incidentally, I completely agree with Everett about this.
From looking at what Everett says in this article, it looks like he writes in a way that doesn’t conform to conventional norms about scientific writing in certain fields and perhaps more closely resembles the commentary one might observe in others. He makes personal remarks about his feelings towards cultures, he makes practical remarks about how we should donate to causes, and so on. To me, it reads as someone who doesn’t care much about academic conventions and who is either intentionally defying them or isn’t aware of them. Nothing about that strikes me as someone striving for fame via iconoclasm. It looks to me like someone who has little patience for those norms. I see pros and cons in that. I think scientists should strive for objectivity and professional detachment in their work, but I also think if we lean too heavily in that direction we can lose sight of insights that are obtained via personal experience and the gestalt of one’s impression about whatever it is you’re studying (societies, languages, ecologies, etc.). What I see here is someone whose work is very human and passionate. I do not see a raging jerk starting fights with Chomskyeans or calling them charlatans.
You say this: “(And there's also a strong inconsistency where Piraha speakers are described as monolingual in some parts but as bilingual in other.)”
Inconsistencies are not evidence that a person is starting anything or being some kind of iconoclast.
What I’d have expected is explicit denunciations of a Chomskyan approach to language, calling people names, insults, fist-pumping revolutionary declarations that his approach to language is the only way and that previous approaches are stupid and terrible, or something along those lines. Instead, what you’ve given me are either (a) innocuous remarks that you’ve interpreted in a highly cynical way, but for which I see no reason to share that interpretation (b) remarks that may be subject to legitimate criticisms but aren’t relevant to what I was asking for, e.g. “this is unscientific” or “this is inconsistent” is not evidence that Everett “started it” or that he’s seeking fame through iconoclasm. If such an accusation were true, I’d have expected better evidence. That you haven’t provided what strikes me as a single piece of even slightly convincing evidence actually serves as evidence to the contrary: it suggests to me that your depiction of Everett is less likely to be true. If it were, I’d have expected better evidence than this.
> If one’s view is that language is integrally tied to culture in inextricable ways, then it’s simply true that one cannot fully understand a language without understanding culture. Nothing about making such a claim requires you to believe that anyone who doesn’t do this is a “charlatan.” All it would entail is that those who don’t do so won’t have a complete understanding of the language. If one’s view is that language is integrally tied to culture in inextricable ways, then it’s simply true that one cannot fully understand a language without understanding culture. Nothing about making such a claim requires you to believe that anyone who doesn’t do this is a “charlatan.” All it would entail is that those who don’t do so won’t have a complete understanding of the language.
This paragraph is ripe with inconsistencies. Yes, the iconoclasm is in some sense a trivial consequence of Everett's views. It doesn't make them any less iconoclastic. If your view is "the only way to understand X is to do Y", then you automatically exclude from sciencehood those who don't do Y, which, in this case, is the supermajority of the field (including supermajority of linguists who do fieldwork, because this particular way isn't how linguistic fieldwork is normally done). And you also try to sneak in "complete", which wasn't in the quote. He doesn't say "can ... completely understand either". He says that if you don't do this, you don't understand the phenomenon, period. YES THIS IS ICONOCLASM. Same iconoclasm that Chomsky himself applied to behaviorists half a century before Everett. This is a call for full reestablishing of a scientific paradigm, with (implied but inescapable) condemnation of those who won't jump to the new paradigm. Compare this to, say, an introduction to LFG (which is, in its own way, no less revolutionary inside):
> Towards the end of the twentieth century, new formal ideas began to achieve prominence in linguistic theory, making use of parallel rather than serial structures and computations, and comparative evaluation of multiple overt structures rather than serial derivation of individual overt structures. These ideas are embodied in a family of nonderivational, constraint-based linguistic theories and in approaches based on optimization (both statistical and discrete). These newer theories are compatible with different linguistic epistemologies drawing on structuralist and functional/typological ideas which have both predated and coexisted with generative grammar. One such theory is lexical-functional grammar (LFG) (Kaplan and Bresnan 1982), which is the topic of this book.
Also, again, Everett 2005 is a published paper, _explicit_ insults wouldn't have been allowed even if he were in the mood for them, which he probably wouldn't be because of his kind demeanor. But kind is not right, and good is not nice.
As an outsider looking in, 3 of those 4 quotes strike me as obvious potential methodological problems in how linguistics research is performed. The odd one out is the sappy quote about fortitude and beautiful cultures, which I think is just fluff (although I disagree with your evaluation that at least to me seems to leave open the normative interpretation that science papers should be dry and boring). Your examples don't seem to me (a random person) to be particularly felicitous.
I was not aware we had that level of knowledge or confidence regarding the structure of dolphin communication; could you please point me in the correct direction to learn more?
Hey, Ive read a lot of your comments on this article and wanted to collect my thoughts into a single response.
First, I think a lot of the clashes that you have here are due to different priorities. While I dont know too much about linguistics, Ive seen those grammar tree diagramms, and they seem like a reasonable thing for linguists to do, and like something that could produce real insight about "grammar". But Chomskyanism also includes a lot of general claims, and often the defense of these looks like a continuous retreat from falsifiability. For someone not interested in getting deeper into linguistics, these general claims with potential takeaways for elsewhere will be the main thing they care about, and they will give a lot more weight to arguments against them from other fields.
I think your examples of Everett "starting it" by being an iconoclast are pretty weak. The idea that there could be more relevant evidence than a papers worth of analysis examples, and that the chomskyian "reanalysis from a distance" only makes sense in terms of fitting them into their scheme and not in terms of what makes sense for studying that language - seems like the sort of thing that is generally accepted without outrage in other sciences, and I suspect it would have gone over much better in this case if Everett wasnt the sole authority on Piraha.
Why do you think homonymies and synonymies are evidence that language is for thinking more than communication? I agree that these are suboptimal for communication, but arent they even worse for thinking? If you have two concepts that are homonymous in thought, you just... wont be able to distinguish them, period. And if you have a synonymie, you can believe things about one but not the other, without realising that anything is wrong. I think both of those happen, and they are way worse than having to figure things out from context.
I dont agree that thinking in language is more potent than in images. You can understand interactions of gears with visual imagery that would take pages to explain in words, and noone could understand those without constructing their own images. Propositional thinking is applicable in cases where visual is not, but those are generally more removed from the evolutionary environment.
If modern human cognition developed in a single mutation, it can also disappear in a single mutation. Why is it so ridiculous as to dismiss out of hand that there is one isolated tribe somewhere in the world where this back-mutation reached fixation?
> I suspect it would have gone over much better in this case if Everett wasnt the sole authority on Piraha.
Very much so. (My bet is that we wouldn't be having this _discussion_ if he weren't - someone else would just come out and show their own work.)
> I agree that these are suboptimal for communication, but arent they even worse for thinking?
The point is that something which is pronounced homonymously is _not_ homonymous for the speaker's thought. You need to specifically _point out_ to people that, say, witch and which or "I [saw a girl] with a telescope" and "I saw [a girl with a telescope]" are homonymous, exactly because they're represented differently in the speaker's head and only get conflated when pronounced. Of course, true homonymy in thinking would suck.
> I dont agree that thinking in language is more potent than in images.
It has a specific capability of chaining which the image thinking does not (as the example of house image not being composed of wall image and windows images suggests). Of course there are other aspects in which imagery is more "potent", e.g., more expressive. One more case of a vague word :)
> Propositional thinking is applicable in cases where visual is not, but those are generally more removed from the evolutionary environment.
It is not as obvious as it seems at first glance. A very important part of any ape's evolutionary environment is social environment, dominance and prestige, all that jazz. And it would seem that thinking through coalitions and stuff like that is very much such a case.
> Why is it so ridiculous as to dismiss out of hand that there is one isolated tribe somewhere in the world where this back-mutation reached fixation?
Their children wouldn't have been able to acquire Portuguese if that happened, which they do.
>There's actually an interesting blogpost on the topic
The orbit of mercury is evidence against Newton and should reduce our confidence in him, increasingly so as we keep failing to find Vulcan, even absent a competing theory, and it doesnt count fully as evidence for Einstein until were sure there is no Vulcan.
Even if we took the argument at face value, I dont think this works as an excuse, because it doesnt explain why the issue would appear in linguistics more so than other sciences. In fact, the arguments apply to linguistics less than to physics, and theres not a general impression that physicist are slippery.
Generally, when you resort to philosophical arguments to defend a particular scientific theory, its because youre losing on the scientifc front. And even the philosophy here is pretty whack - the "theory first" post he linked implies apriori knowledge of medical facts, and he propably doesnt realise.
>Of course, true homonymy in thinking would suck.
Well, I think there is homonymy in thinking. I think the distinctions between e.g. precision and accuracy, or between different quanitifier orderings, are not ones that people can naturally use in their thinking. They need to be tought or discovered through great effort. Of course there are additional homonymies in speech, but that doesnt show thought is favoured - it just means that some popped up in every place they theoretically could. Obviously, the ones in thought are also reflected in speech.
>And it would seem that thinking through coalitions and stuff like that is very much such a case.
I think just playing through particular scenarios in your head can do a lot of lifting here, even without propositional thinking. Of course, once we start ordering society with rules that are themselves linguistic, and complex enough that you cant know the applications relevant to you by heart, thats when it becomes required. But the average isolated hunter-gather tribe we find today is not there. I dont think it kicks in until "memorising the rigveda" levels of cultural complexity.
>Their children wouldn't have been able to acquire Portuguese if that happened, which they do.
I hadnt heard. Do you know if they can count past seven now?
There absolutely is impression that physicists are slippery, especially when they work on string theory on somesuch.
As for the argument about things like precision and accuracy and similar — that is a reasonable point, but... There's a difference between homonymy (two concepts with the same reflection, in each specific case only one is meant) and vagueness (a reflection that fails to specify details that could distinguish two related concepts). Vagueness is inevitable, homonymy is not. It doesn't invalidate your point but mitigates it. (Example of vagueness in speech — Russian и fails to specify difference between BIT vowel and BEET vowel, so the actual pronunciations are across the whole continuum.)
String theory, maybe, but thats ~noone says thats settled science. Even *in his motivating example* people didnt think the physicists were slippery.
What detail distinguishes between precision and accuracy, such that they are otherwise the same? I agree that they are related, but... if two concepts can fool your mind into thinking theyre one, why wouldnt they still seem related after? Are they actually "more" related than the pan and the stove? Plus, the sitting-bank and the money-bank are also related, as shown by the etymology, but are usually an uncontroversial example of homonymy.
Most science isn't settled. Certainly most linguistics isn't settled.
Etymological links can be lost — or non-existent, if the merger is phonological, as in лук 'onion' and лук 'bow (weapon)' — without affecting.
To a layperson, neither precision nor accuracy refers to its statistical definition, both vaguely mean "how close the thing is to the truth/target/...".
As for Piraha children counting past seven once learning Portuguese, I don't technically know, but no one — certainly not Everett, who claims that the differences are cultural and teachable — suggested they can't.
I remember him saying that he tried a lot and failed to teach counting. Maybe he thinks young children are different? But that would be really weird to believe.
It wouldn't be weird to believe, if you think culture is the root cause (as Everett does), that young children haven't yet imbibed the culture and are indeed different.
In addition to being stubbornly wrong on linguistics, Chomsky is also known for being stubbornly wrong on politics, insisting that communism is the best way of organising human affairs. Infamously, Chomsky took a trip to Venezuela to congratulate Hugo Chávez on building a society for the future, a few years before its dramatic economic collapse.
He also never admitted he was wrong - a diehard “true communism has never been tried,” he doubled down on the Venezuela situation, saying “Capitalists were free to undermine the economy in all sorts of ways, like massive export of capital.”
(As per Chomsky, if you have enough resources to flee a failing state, the fact that it’s failing is now your fault.)
This is a clear ad hominem. You wouldn’t infer from Henry Ford’s adoration of fascism that he made crappy cars. The two arenas are completely unrelated.
Well if we’re using caveats then let me just add my own, that if you ***were*** to use Chomsky’s political views to disparage his academic record ***that would be*** a baseless ad hominem attack.
I agree, but it does reveal a tendency towards dogmatism. That doesn't mean Chomsky is wrong on linguistics, but it does raise the question of whether he'd ever admit it if he is?
Why is Left Libertarianism the only ideology labelled as dogmatic? Because of Venezuela? Has no other ideology made similar errors? Should the centre-right be disowned because Margaret Thatcher backed Pinochet?
You are reading *way* too much into my comment. I never even hinted that Left Libertarianism (not how I would describe Venezuela, but that's beside the point) is the only dogmatic ideology, or that other ideologies aren't dogmatic, or that we should disown anyone simply because they are dogmatic.
All I said is that we should keep in mind that a person who tends towards dogmatism in one area might tend towards it in others.
Chomsky is, from what was heard, a person prone to dogmatic and condescending responses, including to colleagues and including in detail (there was the whole "Semantic Wars" thing, for instance...). However, a world's most dogmatic person can tell you that two plus two equals four, and it won't make two plus two equal five.
Specifically, the work in linguistics based on his assumptions has mostly proved tremendously _productive_ (including when it shows that some specific _details_ of his proposal were wrong).
Not trying to argue at all, simple honest question: does Chomsky ever talk about how his understanding of linguistics underpins/justifies his political views? I am by no means an expert on the guy, but I do remember reading some of his stuff where he makes the connection. (Could be wrong!)
I never got round to reading Manufacturing Consent, if there is any crossover it will probably be there - how the powers that be use language to manipulate us etc.
Haven't read the entire corpus of his works, but I don't think so. There is an often-made joke that Chomsky the linguist and Chomsky the anarchist are two different people, and in Russian it is reinforced by the fact that his surname was traditionally translated as Хомский (with a [x] sound and declinable) for linguistic works but as Чомски (with the same-ish sound as in English and usually indeclinable) for political works.
And, personally, I'm a Chomskian (largely, there are important differences of detail, as there are bound to be when you do science) in linguistics but very much not Chomskian in politics.
In his debate with Foucault this comes out quite a bit. Universal grammar suggests a universal human need for things like creativity which feed into a more university moral and political stance. The debate is cool. You can watch it on YouTube or read the transcript online.
I didn't see any Ford quotes there. Instead there's a youtube video from "Turn Leftist & Chris Jeffries", which appears to be about various things not actually specific to fascism that Ford believed in.
Near the middle of that same page there's a whole section on how he earned it. He's mentioned so specifically in Mein Kampf, the wiki's editors felt it necessary to clarify that Hitler was already an anti-semite before encountering Ford's essays.
The analogy doesn't work because cars are a physical product that can be personally tested and used unlike his ideas and theories.
I used to be a firm believer in engaging arguments on their own merits. I've come to realize that it's too difficult to do that in fields I'm not personally very familiar with, and thus there is a lot of value in using a person's thoughts on other topics on which I'm more confident to judge how seriously I should take their thoughts on the topic in question.
It doesn't prove he's wrong about linguistics, but it provides a good reason to hate him.
Given that he's so hateable, though, the most reasonable thing to do seems to be to quit giving him so much credit for being (supposedly) the first person to formulate what is in retrospect a fairly obvious idea. Why do we always need to hear Chomsky's name every time innate grammar is mentioned? We don't hear Payne-Gaposchkin's name every time the hydrogen-helium composition of the sun is mentioned. (In fact I've never heard that name, I had to look it up as an example of a correct idea whose first proponent is not widely known or celebrated.
Let's discuss linguistics without ever mentioning Chomsky again.
The argument is that "stubborness" is a personality trait that Chomsky shares across multiple areas. One can easily think the Ford is "fascist" in politics, car design and dealing with workers without making the cars worse it just implies something about his personality.
Orwell thinks there is something inherently fascist about good matadors while fighting fascism (and liking bull fights).
I am wondering why the question of whether Piraha has a longest sentence got so much attention. English has a longest word, German doesnt. Piraha has a longest sentence (say), English doesnt. What should matter is that no language has a longest story. The focus on sentences seems arbitrary to me.
The claim is that sentence is the biggest object built by grammar, while it is not even clear what a word is, that's a hot debate: for instance, if you claim that "child seat" is a word that just happens to be written with a space (or, vice versa, that Kinderstuhl is two words that just happen to be written together), English suddenly becomes a lot like German.
The example of words was just meant for illustration, apologies. My point was:
One can easily extend any grammar for sentences to a grammar for stories. The latter grammar will always be recursive. So what is the point in claiming that the grammar for Pidaha sentences is not recursive when the more important grammar for Pidaha stories is recursive?
Well, again, the argument may be that stories aren't recursively hierarchical, they just chain sentences linearly. (I am not ready to discuss whether that argument would be true, mind.)
"To Everett, the most important feature of language is not grammar or any particular properties of grammar, but the fact that it involves communication using symbols. "
"Can’t lots of animals do all those things? Yes. Everett views the difference as one of degree, not necessarily of quality. "
Before today, I had never heard of all this tribal language beef. Nevertheless (being an inveterate prescriptivist, though untrained in formal linguistics), I arrived at identical versions of the above conclusions independently some years ago.
I'm still not sure whether this points to a sign of weakness in Everett's argument, but I'm calling it a small win.
I can't help feeling that being somewhat more open to Aristotelian 'folk' metaphysics, purely pragmatically, might take the edge off a lot of academic rancour. The human mind clearly has a great potential for language, which is nevertheless realised in act only in certain contexts. I'm sure you all have excellent jokes about bowling balls tied together with string but I'm looking here at an interminable debate about what "innate" means, and I do think at least having the vocabulary to talk about "non-being which in some way is" is a strength of Aristotelianism (and Thomism). It's not sufficient but it seems necessary, or just helpful, useful. Anyway, interesting review.
Can someone please explain why Chomsky continues to have a role in linguistics like Marx in communism. What he says has been proved to be wrong again and again yet he is still deified
The situation is not as black and white as there are Chomskyan linguists and non-Chomskyan ones, although people do organise themselves institutionally in these ways for contingent reasons. Few people agree with everything that Chomsky has claimed but not everything Chomsky has said is false. The useful stuff sticks around while people reapply it in new frameworks in an improved form but his personality casts a shadow because people like drama and it gives outsiders the wrong impression of where the field is intellectually.
I think Newton was a lot like this as well. And in paleontology at the turn of the 20th century, Cope and Marsh were two competing poles that similarly dominated a field by sheer force of will, developing useful theories that led to major progress despite being fundamentally flawed in ways that were known early on (in Newton’s case, by Leibniz).
This review got off on the wrong foot, for me, by initially pretending that all of linguistics has gone Chomskyan, which is the same order of mistake as pretending that all Americans are Republican. Chomskyans do like to pretend that they're the whole field (or at least the "theoretical core" of the field), just as politicians like to pretend that all Americans agree with them; but the rest of us don't have to believe it.
(In fairness, the review later implicitly backs off of this pretense, both by using phrases like "Chomskyan linguists" and by citing various non-Chomskyan linguists; but I don't think there was any reason for the pretense to begin with, and I think that an uninformed reader would still come away from this review with a *massive* overestimate of what proportion of the field is Chomskyan.)
In the 2020 election, 74.2 million Americans voted for Donald Trump, which is (by a large margin) more people than had ever voted for any candidate in any U.S. Presidential election. It's far more people than ever voted for Trudeau, Sunak, Macron, Merkel, Putin, Stalin, etc., etc., etc.
If you were given no other information besides that paragraph, you might conclude that Trump is overwhelmingly popular and everyone supported him.
But in fact, his main opponent, Joe Biden, won *even more* votes! Somehow that just didn't come up before. ;-)
Of course, in the case of Chomsky, there's no corresponding "anti-Chomsky" we can compare him to. (The opposite of a quasi-cult isn't a different quasi-cult.) And in fact, it wouldn't surprise me if many of Chomsky's cites are actually from scholars disagreeing with him.
But either way, what the citation counts show (at best) is the size of Chomskyan linguistics, not the proportion of linguistics that is Chomskyan.
It would be useful to have the proportions I agree, but it seems unlikely that Chomsky's citation share in Linguistics would be less than Marx's citation share in Economics, judging by the numbers in the review.
Sure -- and I'll gladly stipulate that there are more Chomskyan linguists than Marxist economists -- but are you really under the impression that most economists are Marxists??
No, I just chose Marx from the reviewer's list. My takeaway was: X(A) = expected % share of citations for an arbitrary 'famous' academic; X(C) = expected % share of citations for Chomsky. X(A) should be roughly equal to X(C) but they are way off. That's interesting even if Chomsky's actual share is, say, 30%.
Sorry, I'm not sure I understand what you're trying to say.
The review uses the citation numbers to help make its case that most linguists are Chomskyan; so when you asked me "What are your thoughts on the citation stat?", I thought that you were calling my attention to that supposed evidence, and asking for my rebuttal.
But from your latest comment, it seems that you don't actually agree with the review's conclusion — which is obviously fine — but then I'm not sure what you're trying to ask me?
Chomsky has had a huge impact on linguistics in the sense that his work has established central debates and research questions. So many linguists who are “non / anti Chomskyan” are in some ways working within the bounds of those questions (eg, about innateness vs empiricism, gradualism vs saltatory development of language, communication vs cognitive function of language, etc). Of course you’re right that the field is much broader than followers of Chomsky, but it really is helpful to understand his paradigmatic impact to contextualize Everett’s work, since some of what Everett argues feels like common sense to outsiders (eg, language is “for” communication).
Honestly, even that seems overstated to me — plenty of linguists are perfectly happy to do descriptive work with little or no reference to the questions you mention (just as plenty of Americans don't vote for either party) — but it's infinitely better than the review's claim that "linguists are those who believe Noam Chomsky is the rightful caliph."
I'm sure you're right, but I think the reviewer really does expect us to take claims like this one at face value:
> Since around 1957, Chomsky has dominated linguistics.
Especially given that it's repackaged later as a presupposition in claims like this one:
> [...] the dominance of Chomskyan linguistics has had downstream effects in adjacent fields like artificial intelligence (AI), evolutionary biology, and neuroscience.
toki pona is a minimalist constructed language, made by Sonja Lang about 20 years ago. It consists of about 130 words and a small grammar. Despite these limitations, it works surprisingly well for communication, and there are friendly, active online-communities.
toki pona seems similar to Pirahã in that it has very little recursion. Many things are broken down in parts. For example, a sentence like "Alice said that the cat died" would be translated as "jan Ali li toki e ni: soweli li moli" -- literally "person Ali says this: mammal die".
Interestingly, toki pona *is* an invention. One that has sparked a unique culture, complete with meetings, music videos, penpals, literature. The language has attracted many people who are neurodiverse or part of the LGBTQ+ culture, maybe because toki pona comes without some of the cultural baggage of older languages. In this case, one can truly say that language is the handmaiden of culture.
Yes, context is important in toki pona. It seems surprising at first that words have such broad semantic fields, like soweli which covers all mammals. Yet in practice, it is usually clear what is being talked about.
If anyone is interested in reading more linguistics beef, 'The Linguistics War' documents the acerbic debates on how linguists should study language in the 1970s, and Chomsky is, of course, the main character.
I'm not a .linguist, but having met linguists, I get the impression that Chomsky is not universally well-regarded, despite his massive citation count. Sp this review may be overestimating Chomsky's hold on the field.
It strokes me that there are weaker claims than full Choimskiansism that re still pretty interesting:
- pretty much all humans can learn all human languages, provided they start early and put in effort. (language learning easier in childhood). The specifics of particulars langauages seem noit to be genetically hard-wited at all.
- it sems that this group of people can learn languages like Portuguese with recursive structures even f (arguably) tgheir own language doesn't use them. So something like the potential to understand recursive langauges is bioliogically embedded in all humans
- on the other hand, many non-human animals cant learn our languages, So its clearly not trivial.
I agree that the FOXP2 mutation is not particularly related to language—I think it causes a general deficit in learned rapid precise motor control, and as it happens, speech is the area most saliently impacted by a general deficit in learned rapid precise motor control. See my blog post here for details including two textbooks that come to the same conclusion: https://www.lesswrong.com/posts/smgwRt5MbY9dwhecL/is-foxp2-speech-and-language-disorder-really-foxp2-forebrain
I got the impression that FOXP2 was one' of the steps on the way to language ... conveys the ability to make a wider ranger of squeaky noises (in mice) without grammar. Presumably, in humans, some other adaptation built upon it.
Chomsky consistently repeats the claim that his linguistics and his politics are utterly separate. I disagree. The through-line is in Chomsky’s implicit belief that the Good Life is one of mass democratic decision-making. Anarcho-syndicalism, a kind of radical egalitarian belief that all humans are created equal, would require that all people/workers/citizens be informed about and actively participate in the collective decision-making of their workplace and society. Basically, everyone needs to become middle-class, literate, and non-violent.
Theories of Universal Grammar also posit that all humans are created equal. But, by centering grammar as the core of language, they privilege the relationship that educated, articulate people have with language. Grammar just isn’t that important to the working class and the underclass, people who didn’t get their grammar corrected in school until they’ve learned to follow the rules.
Now, I know that Chomsky means something peculiar when he uses “grammar”: Deep Structure. But what evidence do we have for deep structure? Similar people probably have similar mental representations. But beyond that, can we really say with confidence that words and sentences encode meaning? No, meaning is pragmatic, and depends heavily on contextual information. The position that words have precise meanings, that we can literally evaluate sentences compositionally, is itself a social construct, and in particular, a hallmark of academic and quasi-academic discourse, such as this discussion.
In working-class and underclass cultures, it’s uncouth to be so literal. Also true at some upper class social events. Often, language is meant to signal communal bonds, not make truth claims.
To further unpack grammar, it’s worth noting that even in English, the boundaries between parts of speech are porous. Nouns are verbable. Verbs often become adjectives. Adjectives may be used as nouns or adverbs. Nonstandard words are informal, but confidently using informalities is power talk. Branding, marketing, sales, persuasion, propaganda, poetry, all benefit from novelty, which focuses the audience’s attention, and makes things memorable.
Frankly, I find it painful when people — or language models; ChatGPT is boringly perfectly grammatical; it almost never coins new words/concepts or goes on fun tangents or interrupts itself with higher-order ideas — are overly grammatical. Get to the point. Say what you mean. Rules are meant to be broken. Languages evolve, dialects form, communities divide, cultures emerge, and fun is had when we take ownership of the rules of communication, rather than follow the conventions set by older generations.
After all, language is about communication, which includes self-communication aka thought. But communication is hardly about words and sentences.
In reality, what you say matters less than how you say it (volume, rate, tone, prosody, pitch, etc.); how you speak matters less than your body language (do you project confidence or doubt? take up space or shrink? eye contact or not?); body language matters less than what you look like (age, gender, race, grooming, clothing, accessories; what’s your social status?); your identity matters less than that of your audience (who are you talking to?); and what you listen to matter more than what you say (before you produce information outputs, you must consume information inputs).
Also, as a technical point, strongly consider that natural language is not arbitrarily recursive! After 3 layers of nesting, the amount of semantic ambiguity gets too great to contextually disambiguate. This is not merely a problem of limited mental horsepower and working memory, it’s a combinatorial explosion of interactions between the pragmatic context, the first-order meaning of the composition of the words in the sentence, and the predicted intention of the speaker.
The book Decoding Chomsky may of interest as it argues the opposite; that Chomsky's linguistic theories serve the interests of the American establishment.
I contend that Chomsky’s leftism also serves the interest of the American establishment. It’s an intellectual stance that is of no threat to the ruling class; a particularly erudite form of champagne socialism/anti-imperialism, inaccessible to the masses required for a real revolutionary movement. Chomsky is of course more rationalistic than the critical theorists / continental philosophers he called obscurantist, eg Foucault and other pomo theorists that paved the way for 21st century identity politics and intersectionality — though they too are of course part of the Western elite discourse, and no threat to the world order; indeed, today, LGBTQ+ rights and racial justice are a justification and tool of Western power.
EDIT: To clarify, my point is that idealistic egalitarianism that imagines all humans as potential rational educated western middle-class knowledge-workers participating in democratic collective decision-making systems is actually not as liberatory as it seems. Advanced human society requires complex division of labor, including division of decision-making authority, and the hoarding and wielding of secrets. In other words, a social hierarchy; workers and bosses.
For all the criticisms of mass media as a means of elite control over democratic processes, we haven't yet found a better means of social coordination. The chaotic populism of social media — rampant misinformation and conspiracy theories, declining trust in authoritative institutions, a race to the bottom attention economy, incentivizing everyone to become an influencer — seems to be in many ways worse than old-school manufactured consent.
I think it is important to recognize that the way both Chomsky and Everett use the word grammar is quite different from the way laypeople use the word grammar. The kind of grammar that linguists are talking about is simply the organizational structure of speech. This means that Day Labourism professor is having exactly the same grammar, and it is equally important to all of them. Conscious knowledge of grammar that is learned in school is not talking about, it is certainly true that that kind of explicit knowledge is not terribly important to the average man in the street, but the average man in the street speaks coherent structure language, which is no different from that of the professor, perhaps for the number of fancy words being used.that is point on which Chomsky and Everett would agree
Yes, I understand that grammar has a technical meaning in linguistics. I just question whether this is actually carving reality at it's joints.
As I wrote above, from a perspective on language and communication that centers pragmatics and sociolinguistics, it doesn't make sense to imagine that sentences have a Deep Structure. Sentences don't mean anything outside of their context! Context — the who, where, when, why, and how of speech — matters more than compositional semantics — what was said.
The concept of grammar used by academic linguists is a kind of motte-and-bailey. Linguists retreat to the distinction you mention, between deep structure (an abstract formalism) and surface structure (the rules you learn in school) when challenged by the reality that there's scant evidence for their emphasis on grammar. But in practice, most linguists hate the distributional hypothesis, The Bitter Lesson, and probabilistic systems like LLMs; but they like Montague Semantics; and they like rule-based recursive parsers — converting surface structure to deep structure — which are a good model of computer languages but not natural language.
“Grammar just isn’t that important to the working class and the underclass”
That seems completely wrong to me. There is a formal grammar of written academic language that doesn’t matter to them. But if you look at the grammatical patterns of their natural language, you’ll see that they’re just as complex and just as important to them - someone who uses “ain’t” wrong, or fails to double-negate in the standard ways, or in any other way breaks the grammatical rules of their particular non-privileged varieties of English is going to sound silly, like an outsider who is interloping and pretending to speak their language.
That said, I agree with a lot of the other criticism, that people aren’t using language for the same kind of abstract thinking that Chomsky thinks it is essentially for.
I'm not saying that working- and underclass people don't have any linguistic regularity. That would be preposterous. Certainly, how you communicate signals your affinity with the group, your insider/outsider status, etc., but this is as much or more a matter of pragmatics and sociolinguistics as it is about syntax.
It's my experience that in less academic contexts, with working- and underclass folks, in both casual speech and text messages, there is indeed less emphasis on adhering to the grammatical norms of the non-prestige dialect. Certain conventions must be followed to — don't sound like an outsider — but beyond that there is more latitude for all sorts of disfluencies.
This is a somewhat controversial perspective, I acknowledge. Many good-hearted folks in linguistics want to maintain the view that non-prestige dialects are equivalently complex to standard dialects. Certainly, it's a fact that all languages can express all ideas, just not with equal ease. Now, one could hold that dialects may be morally equal without being equally complex in some senses. But that's a hard distinction to maintain, and not one that most academics are comfortable making, for reasons that mirror the complexities of the discourse around biological differences in ethnic groups.
When we zoom out from dialectical differences to differences between languages, it's then rather obvious that some languages are more grammar-heavy than others. Why would this be untrue of dialects?
For an example to make this more concrete, while African American English (AAE) does have some constructs that Standard English lacks, such as the habitual be and null copula, it lacks the subjunctive mood, and perhaps other grammatical constructs of Standard English. (Note that English's subjunctive is much simpler than French's.)
I contend that the habitual be and null copula are ways to communicate more ambiguous notions than are easily expressed in Standard English, while the subjunctive mood conveys more precise notions than are natural for AAE. And while the habitual be and null copula are now commonly used by younger non-black people, using "were" instead of "was" in an AAE context would signal outsiderness to "urban" black Americans, even to those who code-switch to Standard English professionally.
See, the reason why dialects are maintained by people who are also fluent in the prestige dialect is because they represent the culture and perspective of their community. And for a community to exist, it must be able to differentiate itself from the surrounding society.
AAE has constructs that enable expressing ambiguous notions because the African American community values the ability to make these statements. When black Americans want to get more precise, they usually code-switch to Standard English. And people of all races have adopted these ambiguous constructs because they too see the utility in them.
AAE constructs are expanding beyond black speakers because they're simplifications. Grammatical simplicity is often a good thing. It's part of why English is a great language compared to eg German or French.
I should add that I stuttered as a child, and had to go through speech therapy as a teenager. I "solved" my stutter by learning to circumlocute. Rather than attempt to say a word that would make me stutter, I substitute a synonymous word or phrase. Sometimes, this means that I choose the wrong part of speech, or get the prepositions wrong. When this happens, it's a minor faux pas in professional contexts, but just not an issue for folks with less book smarts.
"I should add that I stuttered as a child, and had to go through speech therapy as a teenager. I "solved" my stutter by learning to circumlocute. Rather than attempt to say a word that would make me stutter, I substitute a synonymous word or phrase."
I recall reading John Stossel's account of failing to do this with the word "dollars," as the initial hard consonant was a problem for him, and the only synonym he could think of, "bucks," had the same problem, in addition to being too informal. Did you do any better than him?
I thankfully didn’t have the experience of having certain words I would frequently stutter on. I can only imagine how stressful it would be to constantly avoid key words. For me, it’s more of an immediate feeling that the word(s) I’ve planned to say won’t come out the way I intended. The more I would try to force myself to say something, the more I would stutter.
The temperature parameter in large language models is a good analogy. Temp = 0 means the model always outputs the most likely next token. Higher temperatures mean it outputs less probable tokens with some probability. For me to speak fluently, I have to have a “higher temperature” — I often surprise myself with my choice of words.
When I did speech therapy at 15, I realized my stutter was minor compared to the handful of other teens in my month-long summer program. The other kids stutters were triggered in particular contexts like reading aloud or talking on the phone, and many of them had key words that they would predictably stutter on. For some it was their own names. And they seemed to be more self-conscious than me about it, more concerned with how others judge them. Maybe it was low self-esteem but I didn’t think other kids thought or cared about me much, nor was I that concerned with their judgments.
My stuttering then, as now, was only correlated with heightened emotional states, particularly anxiety.
Once I realized that I don’t stutter when I’m not anxious or otherwise emotionally stressed, I started to see my verbal fluency as a barometer of my mood. So when I do catch myself stuttering, I prioritize regulating my emotional state and solving the problem that is distressing me over trying to sound good.
Most mathematicians I've met claim to think mostly in geometric terms, definitely not the kind of grammar-structured language that Chomsky seems to have advocated through a long career. Maybe this way of thinking is sometimes compositional and therefore could be brought into a Merge framework, but I don't think compositionality is universal in such thought (or in any other styles of thought).
>A more parsimonious speculation is that they did not, and that the Great Leap was effectively instantaneous, in a single individual, who was instantly endowed with intellectual capacities far superior to those of others, transmitted to offspring and coming to predominate.
Isn't this basically what Catholics believe about Adam? That all modern humankind is descended from a single male (and a single female) ancestor, and that God granted Adam the gift of rationality, a mind made in his own image?
"He does explicitly claim (in the aforementioned paper and elsewhere) that Pirahã probably has no longest sentence, which is about the most generic anti-recursion statement one can make."
Was this meant to say that it probably DOES have a longest sentence? From what I read up to this point I'd have thought that all languages having no longest sentence is the standard Chomskyian view.
Ah I misread this! I thought it was saying he was conceding the one most generic *pro*-recursion point, that it doesn’t have a longest sentence, even though it doesn’t use any of the standard recursive methods for extending them.
taking this account at face value, it seems like it would be *extremely valuable* to the field of linguistics to have one more person actually interact with the Piraha and learn their language, which would presumably dissolve the questions about Everett's description of their language. I realize that getting in slow-burning asynchronous fights is more fun than that but it's kind of embarrassing to read about as an outsider
Your note about parse trees in the "mechanism" of LLM's is no hypothetical - such trees have actually been found, as early as 2019 (in BERT, a precursor to GPT)! This work finds that the trees are linearly embedded in a certain fixed subspace of the latent vectors. It's powerful evidence that such an algorithm emerged with no explicit programming, as simply a necessary mechanism to predict the statistics of language.
Excellent review! I'm gonna go out on a limb and say that it's Scott.
And call this a non-central objection if you must, but I'd like to respond to the joke in the link in 4th footnote. Friedman would have won the exchange if he'd simply used a binary search: ln(2^n) = n so he still would have been able to pin Yessenin-Volpin down in linear time.
I don’t think it’s Scott. He hasn’t expressed significant views on the Chomskyan debate before, and this seems like someone who has spent a long time thinking about it (perhaps someone like me who grew up Chomskyan but had a conversion some time in the past decade).
> I find this notion of recursion completely insane, because it would imply (i) any language with more than one word in its sentences has recursion, and that (ii) all sentences are necessarily constructed recursively.
I had to ctrl+f 'Goptjaam' and 'glossable' to see if anyone beat me to this, but it looks like no. Granted, Goptjaam is a *deliberately* cursed conlang (https://www.youtube.com/watch?v=ze5i_e_ryTk), but I think it's still an interesting example of how this statement needn't necessarily follow! Goptjaam isn't glossable - its words depend entirely on the context of the full sentence. The moment you put the 'same' word into a different context, it's a different word.
To be clear, this isn't meant as an argument; actual human languages aren't like this! I just find it fun to look at how one can break languages if one tries.
Isn't from Chomsky to Everett a Wittgensteinian turn?
I think Wittgenstein never thought language was inherent, but he tried to invent a logically consistent language to make truth statements about the world (an "I-Language")
He then realised he missed the point, language is something that is practised - and the practise improves it over time rather than one perfect design
Wasn't Wittgenstein answering a totally different question? He was interested in how language is used while linguists are interested in how language is constructed.
Maybe, but I don't think the distinction between how it's used and how it's constructed is the relevant difference
The "E-language" [Everett] part seems to me saying that "how it's used is how it is constructed" while Chomsky [I-Language] says "mental properties construct how it's used"
There's clearly some interaction but I think syntax and semantics are distinct. Building the car vs driving the car would be the rough analogy I'd use.
I think the important point of Wittgenstein (which descriptive linguists of any sort accept) is that languages are defined by how they are used, and whatever it is to say they are constructed has to be defined by that.
I know little of either Wittgenstein or linguistics, but whatever W was talking about didn't involve, for example, parse trees. That seems to me to be indicative of some sort of fundamental conceptual distinction.
I do sort of find echoes of W in things like word embeddings: the meaning of a word is defined by the words which surround it, which seems suspiciously like "meaning as use". Chomsky's view of word vectors and LLMs as antithetical to the more analytical approach of linguistics I feel lends credence to the notion that what he was doing and what W was doing were pretty distinct.
Great review. Linguistics is such an interesting topic area I appreciate the work that went into creating a layman's summary.
On a side note: can we create a corollary to Godwin's law stating that the longer an argument goes on the probability of calling someone a racist approaches 1? And that anyone that does so automatically loses the argument?
There are also conversations where mentioning Hitler or Nazi's is relevant - the vast majority of the time this isn't true and even in the rare case it is adds exactly nothing to the conversation anyway.
Calling someone a racist isn't an argument, it's a status claim. By doing that you assert that he's a Bad Person, and you're entitled to make that judgement. This maneuver succeeds or fails mainly based on your relative social standings.
Well, the original Godwin's law was about comparing something to Hitler/Nazis, not necessarily insulting the opponent directly, so it was a bit more subtle. A law for a more civilized age.
Some other people have pointed out parts of this, but it's bizarre how this book review leads in by pretending that non-Chomskyan linguistics doesn't exist. Firstly, his influence was a lot less in many subdisciplines of linguistics (phonetics/phonology, historical linguistics, etc.) than in his core syntax/morphology/child language acquisition specialties. Secondly, even in the disciplines where Chomsky has had more to say, there's always been a sizable number of linguists opposing him (see Randy Allen Harris's The Language Wars). So yeah, Everettian ideas are also building on an established tradition (as should be clear by the many references to pre-Everett field linguistics)
Yes, this seems right and important. But I also think it’s interesting to me to see an Everettian anti-Chomskyan take, because I’ve come to this from a different angle.
There are a bunch of cryptographers and linguists who are still trying to decipher the Voynch manuscript. A pattern observed in that community....
a) Someone will point out a statistical feature of the Voynich manuscript which they claim never occurs in any human language ever. This, they claim, is conclusive evidence that the Voynich manuscript is a meaningless forgery, not a simple cypher of some real language.(*)
b) Someone else will point out an example of an absolutely definitely real language that also exhibits the feature in question. Those guys have samples of lots of languages in machine-processable format.
(*) Historical context would suggest that if its a real language, its got to be either a real language known to Europeans around the time of Rudold II of Bavaria, or a conlang invented by them. This really cuts down the space of likely possibilities.
Had the Voynich manuscript been discovered much later, like in say 1960, after the widespread use of machine ciphers and with he existence of widely travelled linguists, our priors for "what kind of cipher of which natural language might this be" would be so, so much worse.
Exotic possibilities for the Voynich manuscript include: its a simple cipher padded with nulls, i.e. most of the ciphertext is random junk padding, but subset of it deciphers to something cohetrent.
> (*) Historical context would suggest that if its a real language, its got to be either a real language known to Europeans around the time of Rudold II of Bavaria, or a conlang invented by them. This really cuts down the space of likely possibilities.
How much does that really cut it down? It would rule out American and Oceanic languages sure, but isn't all of Afro-Eurasia still plausible in principle?
Just want to note that I really enjoyed Kingdom of Speech and found it highly rewarding to read, though I had a suspicion it started after Wolfe and Chomsky met each other at a NYC cocktail party and intensely despised each other.
Also to note on this:
"…well, okay, maybe I don’t entirely get Foucault’s number. Every humanities person must have an altar of him by their bedside or something. "
Basically yes, he dominates citations in the humanities/liberal arts world
It is worth pointing out that the standard transformer architecture used in modern LLMs is not recursive (or recurrent). That is to say, an LLM cannot construct a grammatically correct sequence of tokens that corresponds (parses) to an abstract syntax tree of unlimited depth. This is obvious on theoretical grounds from the basic architecture, but you can also demonstrate it quite trivially by training a transformer on toy languages described by context-free grammars. The maximum depth of recursion is limited by the number of layers in the transformer.
There are some variants (i.e. feedback transformer, universal transformer) that are fully recursive. We don't use those variants because in practice it doesn't matter. Humans can't process grammatical structures of unlimited depth either; once the complexity of a sentence starts to overflow working memory, the grammatical structure gets lost. So the difference between English and Pirahã is merely one of degree -- Pirahã has a maximum recursive depth of 1, while English has unlimited depth in theory, but in practice has a maximum recursive depth that is slightly higher, but still small (single digits).
Note that humans can still write code of unlimited length, even though the abstract syntax trees for human-written programs have finite (and typically small) depth. The trick is to do exactly what Everett did when translating the Bible to Pirahã -- break big expressions into smaller ones, and use variables (names/pronouns) instead of nested clauses. An extreme example is assembly language, where each expression (sentence) has exactly 3 "words" -- e.g. (operation register1 register2). Recursive grammar is unnecessary to represent complex ideas.
It’s also worth noting that half of the attempts to extend the context transformers may have run into the quadratic cost (complexity, memory) issue, and that probably the majority of proposed solutions to this problem are de-facto injecting recurrence, one way or another.
I have a vague memory that there's a field called psycholinguistics that studies things like...
although you might write down a grammar of English that sugges6s there are arbitrarily long valid sentences, there is in practise a limit to what actual human speakers will put up with, Experimentally ,measure how much stuff there can be on the parser stack before a human decides the sentence is unreasonable.
I strongly disliked this review. It is incredibly condescending and equally wrong (and often "not even wrong"). It honestly strikes me as odd that the reviewer, a self-admitted layman, believe they can so easily identify the supposed absurd and "insane" mistakes of thousands of brilliant researchers. The alternative hypothesis, that the reviewer might be misunderstanding something, should've a at least crossed their mind. This has all the red flags typical of a crank (possibly due to the source material, but still...).
Like the reviewer, I'm not a linguist, but even I could spot such misunderstandings in this review. These include conceptual misunderstandings (such as the misconception about the meaning and implications of "recursion"), contextual misunderstandings (e.g. regarding the long history of structuralism, from De Saussure onward), and disciplinary misunderstandings (there are many counter-Chomskyan linguists and even entire such linguistics departments; none subscribe to Everett's views to the best of my knowledge, for what seem like good reasons). Believing all the linguists are dense and mindless cult-followers lacking critical thinking is just silly.
I feel these interested amateur reviews are high risk high return. Sometimes you read something that shares a glimpse into another world and has an infectious feeling of excitement - and sometimes you get a smart person totally misunderstanding a really complex topic and leaving you feeling vaguely annoyed.
Do Chomsky's critics want to go back to the pre-Chomsky views within linguistics before 1958 or so they accept that Chomsky made progress in at least a negative sense by criticizing the conventional wisdom of the 1950s even if many of positive theories haven't panned out?
Any idea what percentage of the Chomsky citations are to his political polemics? Especially among grad students, his political writings are, because of their volume and the predictability of his opinions, an ‘easy cite’ for students looking to frame a debate.
I would expect that even his linguistic articles have _huge_ citation numbers, because half the articles include "Following the general premise of Minimalism (Chomsky, 2000)" or some such intro, which artificially inflates quotations even if the rest of the article doesn't directly engage with Chomsky's work at all.
I thought it was pretty clear that the opening paragraphs of the review, which purport to describe Chomsky's quasi-divine status, were tongue-in-cheek? - anyway, I very much enjoyed this review - I have only read Wolfe's book, but now I want to read Everett's own accounts.
I thought this was a great review, and I wanted to share another interesting thing about the Piraha. They are really happy. Relentlessly happy. In the 30 years Everett studied them, there were zero suicides. The idea was a completely foreign concept. Here’s Everett:
“I was still a very fervent Christian missionary, and I wanted to tell them how God had changed my life. So, I told them a story about my stepmother and how she had committed suicide because she was so depressed and so lost, for the word depressed I used the word sad. So she was very sad. She was crying. She felt lost. And she shot herself in the head, and she died. And this had a large spiritual impact on me, and I later became a missionary and came to the Piraha because of all of this experience triggered by her suicide. And I told this story as tenderly as I could, and tried to communicate that it had a huge impact on me. And when I was finished everyone burst out laughing.”
They burst out laughing! It just didn’t make any sense to them that someone would even consider doing that.
Does this have to do with their language? Well, language and culture are deeply intertwined, and we know that the Piraha only use language centred on the present. They have very limited vocabulary to describe the past and the future. So it suggests a culture that’s deeply focussed on the present. In Western culture we sometimes say that depression is living in the past, anxiety is living in the future, and getting someone out of their head and into the real world (the now) is the point of therapy and mindfulness training.
The Piraha may be very happy, but if their reaction to being told “this person was very sad and then she killed herself” is to burst out laughing, then I would describe them as somewhere between “zero emotional intelligence” and “stone cold sociopaths.” Not great role models!
I picture it like - imagine someone told you “this person was so sad that she ate a thousand bananas until her stomach exploded.” You might laugh, right? Just cause the concept is so ridiculous and unheard of, not because you don’t have empathy for people with exploding stomachs.
I love these ACX book reviews, and will read this one in its entirety later, but is it too much to ask them to be a little more brief? With the length of this one I feel it’s probably easier just to read the book itself.
The fact is that many (most?) winners/ front-runners have been very long, historically. So the revealed preference is for longer rather than shorter reviews.
Of course we’re all selected for liking Scott’s writing here- and it leans towards thoroughness at the expense of brevity.
The thing that I find striking about Chomsky is the claim that nothing other than grammatical analysis really matters. From a philosophy of language perspective, I am much more sympathetic to the Wittgensteinian view, as expanded by Kripke. Chomsky in the end seems quite reductionist, mistaking his understanding of how there can be formal systems that describe how languages function via grammar to mean that other aspects shouldn't be studied. His project reminds me of philosophical work trying to derive the basics of arithmetic from the rules of logic, and then declaring that nothing else you could do with that was worthwhile. Even without that, there were plenty of things to do in mathematics- and different logical foundations are possible for the empirical facts that can be observed. If the Chomsky project ends up just describing a very simple rule for recursion, it may not be all that useful as a foundation for study of more complex phenomena, as the complexity that must be built on top of that foundation in order to describe common uses of language makes those descriptions less useful.
Language is a complex interaction of many things. The claim is that the grammatical side (which is richer than just recursion because of interfaces - again, ChHF 2002 is a _very_ bad intro into UG, good linguistic articles are not published in Nature!) is basically the only (and existing) _proprietarily linguistic_ thing among them, while others are purview of psychology, sociology or whatever.
Define "pointless". It certainly leads to discoveries about how grammar works and indirectly bears on other cognitive science because language is often the window used to try and look into the brain.
" Perhaps the most interesting suffixes, however (though these are not unique to Pirahã), are what linguists call evidentials, elements that represent the speaker’s evaluation of his or her knowledge of what he or she is saying. There are three of these in Pirahã: hearsay, observation, and deduction."
This sort of thing is not so unusual - for instance, Quechua has them. (A page in a grammar book on my bookshelf lists the situations in which one particular evidential particle can be used - some examples from memory: I was there but I was asleep, I was there but I was very drunk, I was there but I was less than five years old...)
> How Chomsky attained this stranglehold on linguistics is an interesting sociological question, but not our main concern in the present work
This review is bizarre. Chomsky doesn't have a stranglehold on linguistics. It's like writing a long essay criticizing Aristotle's stranglehold on physics.
While it’s overstated, I think it’s right. There are chomskyan departments and anti-chomskyan departments, but everyone has opinions on Chomsky. There is no comparable figure in physics.
I totally agree!! After the first few paragraphs I was seriously confused about what this guy had actually read. It just seems the opposite of any recent “mass appeal” linguistics book.
> He does explicitly claim (in the aforementioned paper and elsewhere) that Pirahã probably has no longest sentence, which is about the most generic anti-recursion statement one can make.
Should that read 'probably has a longest sentence'?
Did Chomsky really declare that the first language evolved abruptly a mere 50,000 years ago? If that is a prediction of his theories then its epic stupidity largely discredits them on its own. Anatomically modern humans appeared over 300,000 years ago for a start, and it seems misguided, verging on cranky, to dispute that throughout that time they were fully capable of speech with a proper language, even if this may have lacked some modern refinements.
I mean one can teach chimps and gorillas to speak, at least simple sentences. They may grunt and huff and puff in trying to imitate human vocalizations, but when they hold up cards saying things like "Koco want banana" and suchlike, there is no doubt they are meaningfully communicating, and is cast iron evidence they are thinking in words up to a point. Even my cat clearly recognises a few words, such as her name and "din dins"!
If you assume that being anatomically modern means having a fully developed language ability, then you're making a circular argument (and ironically, you're kinda agreeing with Chomsky while doing it, though with respect to a slightly different of definition of "anatomically modern").
Also, writing appeared abruptly about 5,000 years ago, even though humans 300,000 years ago were anatomically capable of writing just like us. This shows that your reasoning is flawed, and it's reasonable to suggest a similar trajectory for speech. In fact, this makes more sense for speech: there is evidence (like types of aphasia) indicating that speech, unlike writing, isn't genetically stable and can break down isolatedly. While not definitive proof, this supports the hypothesis of an abrupt development of speech.
Finally, archaeological evidence (as far as I know) shows that about 50,000 years ago, there was a significant cultural leap in art and technology that spread quickly worldwide. The fact that there were no major anatomical changes during this time, as you mentioned, actually supports the idea that language suddenly emerged rather than contradicting it.
Not to pick on you but there is a lot of confident incorrectness about linguistic theory because we are prone to assuming that being able to speak language is a good qualification to start theorising about it but it isn't. Things like "word" and "sentence" have intuitive meanings to lay people but they are pre-theoretical terms. When you put in the work to analyse their role in a larger explanatory model (such as those starting with Eric Lenneberg and Charles Hockett, who made contributions to the biology of language long before Chomsky), you realise that these things can be counter-intuitive and you can reasonably come away with the conclusion that non-humans communicate with things we might not want to call "words".
In general, in this review and the comments, it ought to be suspicious how quickly people conclude that Chomsky is ideological or moronic in the face of counter-evidence, or that the field as a whole is just a cult of acceptance for everything he says, as if there haven't been hundreds of other linguists in the same tradition who have done fieldwork and animal studies and psychological experiments.
I mean, just open journals like Linguistic Inquiry, Syntax, Glossa or Natural Language and Linguistic Theory etc. - that should have been foundational to a piece which ultimately disparages the field alongside Chomsky but there is no evidence in those pages of what the review describes because the Everett drama is totally unrepresentative.
I think underlying all this, there's really a badly misunderstood model that people have for how academia works. Perhaps biased by reasonable critiques of the postmodernised humanities but that brush can't be used to tarnish everything, and people ought to think about other fields a little more carefully.
That's not what evidence from teaching apes language actually suggests, and anatomically modern humans that seemingly didn't engage in modern behaviour are an argument in the other direction, that is, that they were _not_ capable; I've already touched on both in other comments.
Must admit to skimming the last two sections. Three parts preface, one part actual review of the book in question, one part kludged-on red meat for rats (ratmeat?), with the former roughly 65% of the entire review, and cornucopious blockquotes throughout (so many blockquotes, mostly from...not-actually-the-book-being-reviewed)...it's just A Lot. Yet despite all the references and scene-setting, I still find myself with that nagging [Citation Needed] feeling of insufficiently-supported statements. This was not helped by c.f.'s to other works that the author admits to not having actually read/understood, nor ~all the quotes being from within the first few dozen pages of each work. I want to assume the full texts were indeed read, and their contexts explicated in prose, but...I dunno, it reminds me a lot of sundry mandatory Works Cited bibliographies in school stuffed with trite surface-level references. A palette of bon mots more than integral supporting research. One can certainly stitch together an argument this way, but it shambles rather than running, which is optimal for zombies but not rhetoric.
Ultimately I leave with no more knowledge than I started of what linguistics even is, or why it's important outside departmental funding disputes, or why anyone takes Chomsky seriously. "It's not my theory that's wrong, it's the world!" says the dismissively towering Ivyist who's never done any fieldwork. (Also, I still don't understand what an I-language is. The spherical cow of languages?) Despite reading it, it's Not For Me, sorry.
Without commenting on the rest, I-language is the mental module for language in each individual's brain, which gives rise to the individual's observed language, "idiolect".
This would have been helpful to know while reading, thank you. So language is theorized to be necessarily individual and arises from a particular <s>candidate gene</s> shared brain module, yet nevertheless follows certain universal patterns, despite the trivially observable vast differences between discrete..geocultural expressive groupings, or e-languages. (And maybe - maybe - some languages that don't follow all the axioms.) A meta-language, if you will, and an essential H. sapiens evolutionary stepping stone leading to enhanced cognition overall. That's at least the seed of a potentially interesting and falsifiable hypothesis. (I wonder what Chomsky thinks about Neuralink and similar proposed technologies...)
Should have amended that I...never actually learned proper grammar, and wasn't even introduced to the concept of certain fundamentals like "parts of speech" until late teens. Even today, the formalisms continue to baffle...some language simply seems intuitively correct, and other constructions not, without really being able to explicate why. Chomsky obviously means more than *just* the formal Rules Of Proper Language one learns in school when he refers to Grammarcy Hall, but that's definitely a part of it...so not having that, uh, brain module is perhaps a large part of why I bounce hard off linguistics. Not an impediment to actual speech-as-communication-or-internal-thinking (I assume perfect written/verbal SAT/WAIS measures *something* real), but probably a big stumbling block for understanding the theoretical underpinnings.
Not really? If the rules are part of the brain, you don't need to teach them. More generally, you can't explain how specialized modules of the brain work by the mere fact you possess them, any more than possessing a heart makes you a cardiologist.
Not sure I understand. If such knowledge is inherent to all brains, just as natural as no one needs to be taught how to make their heart beat or their legs walk, then...why do we teach grammar at all? Still-disappointing-but-gradually-increasing average literacy levels certainly seem to suggest there's some additional need being met (although such paucity remaining a common feature of the complacency of the learned is also not encouraging). I do notice that such deficiencies perhaps reflecting on the actual strength of one's internal thinking, rather than merely communicative barriers, is...kind of a dark road to go down as well. A conclusion to flinch away from. One wants to believe in the rich interior lives of other sentients...
Incidentally, I appreciate you bothering to patiently interlocute in good faith with the sundry ignorant commenters on this topic, self included obviously. Edification on impenetrable topics is a worthy endeavor even if it doesn't stick. "More of this, please," as the kids say these days.
Oof. What we teach in schools is an attempt to formalize and standardize what has already happened in heads, and the grammatical differences between how kids speak and what they are taught are differences between their grammar and "Standard English" grammar, not between presence and absence of grammar.
Language emerging suddenly is a somewhat separate claim from language being innate. If humans rapidly developed true language ~50 thousand years ago (as SF ideas go, I like the suggestion that we were taught language by sperm whales at that time, or if you're boring you can just go with aliens, but really I'm talking about an existing proto-language rapidly making a qualitative jump in some tribe), and other humans already had the capability for it because anyone smart enough can learn language, it could spread very quickly, as useful technologies do. And there really was a spurt of Homo sapiens expansion that started ~50 thousand years ago, for unclear reasons. But blaming it on a de novo mutation that rapidly spread to literally everyone seems dubious.
By contrast, you could also argue for a biologically innate language ability, hardcoded in, that nevertheless evolved gradually, through multiple levels of selection, from a proto-language towards today through millions of years of selection for better linguistic ability.
And, frankly, I feel that this would be a more logical alignment of the abrupt-gradual and hardcoded-general questions than the one the review describes linguists as going for. Hardcoded linguistic ability should emerge at the speed of evolution; language as a teachable invention, after a "hardware overhang", can spread at the speed of culture.
But not all evolution is gradual. If the intuition about the system is broadly correct, the recursive property is as abrupt as switching from RNA to RNA+DNA, and this lines up well with ontogenesis.
That's a strange analogy, given that we don't know how the RNA-DNA transition happened. But yes - evolutionary changes can happen quickly. However, quick by evolutionary standards is still pretty slow by cultural standards.
Sure, but, seeing as we speak about hominids, the proper comparison is paleontological standards not cultural, and evolution can be much quicker than those, especially under adverse pressure by quasi-kins (Ashkenazi selection, lactose tolerance selection, that fish speciation in recently separated lakes story...)
I mean, the question is whether language should be thought of as primarily a biological adaptation or primarily a cultural invention. So the comparison is between evolution and culture. Mind you, obviously you can still argue for a rapid genetic shift, or a slow cultural evolution; it's just ironic that the "sides" (if the review is to be trusted, since I understand that it drastically simplifies disagreements among linguists) are fast genetic vs. slow cultural.
Slow genetic is fairly popular among non-Chomskyans (see Svetlana Burlak, see some comments under this very post). But yeah, virtually absent "fast cultural" is interesting. The reason may be that cultural guys tend to be holists, and holists will (obviously) see language as more complicated than analysts, because, well, they include more into their notion of language.
(And as for "don't know how" — that makes the analogy more illuminative for me: the nature of the nucleotides implies it should've been rather abrupt, as ribonucleotides don't form stable chains (I mean the "vertical" connection, not the complementarity-based "horizontal" one) with disoxyribonucleotides, so it wasn't nucleotides changing one by one in a hybrid molecule, so we can conclude abruptness without seeing it.)
Commenters have expressed on other reviews a desire for economy where book reviews are concerned.
This was really economical to me. Kind of the difference between a NYT book section review and a NY Review of Books review: something (the former) can be short and yet in some cases a waste of time compared to the longer thing.
I found the dispute and more so the background info fascinating but I would never want to read a whole book on it. This was “just right” in Goldilocks terms.
I loved reading about Chomsky and not having any of the obligatory stuff about his politics.
And yet - by the end - it did seem to me that his undoubted brilliant scholarship - displayed as time went on (if this review is anything to go by) a sort of unearned totalizing tendency, rather in keeping with his natural radicalism.
I like how the author presented acronyms in parenthesis before using them. The Author of The Family That Could Not Sleep review did not do this, and I struggled slightly because of it.
Both these people raise red flags for me. Chompsy is the archetype celebrity scientist holding a field hostage to his ridged interpretation, and Everett is making limited claims but writing as if he is overturning the entire field.
Everett's claims would plausibly overturn the entire generative linguistics field if they were true, make no mistake. They suggest interdependence between language and culture at a level incompatible with language being a module of mind.
I don't see Everett as turning over an entire field so much as overturning a specific paradigm within that field. And overturning paradigms is a standard feature of scientific progress. It's especially not much of a red flag to me if that paradigm never established itself all that firmly in the first place, and was promulgated on considerations that have their origins in a more rationalistic and less empiricist approach, which is how it seems in this case.
Critical views of the Chomskyean approach to language are not unique or confined to Everett, either, and others working in related fields have offered alternative approaches that, if true, would be seriously at odds with Chomsky's views. For instance, Christiansen and Chater also present a perspective on language that runs contrary to Chomsky's views. I don't see work like this as overturning a field, I see it as overturning ... well, Chomsky. I don't see any red flags in that.
It is a different question, but if we have reason to believe it only happened once, that would increase the probability that language was only invented once.
Well, alphabet is a cultural invention — specifically, a cultural modification of another cultural invention, writing, which did appear several times. And yes, all modern alphabets except probably Hanguyl go back to an old Semitic consonantal writing system. I am not sure this recent development is of any relevance to the question of language evolution.
I mean, both Hebrew and Arabic still use an abjad, which is literally worse than hieroglyphics. Writing systems are sticky unless you have Ataturk or Red Terror at hand.
(Also, in digital age non-linearity of Hanguyl is a bit of a problem.)
I think we have evidence that writing was independently invented multiple times. Some of the scripts were syllabic rather than alphabetic. e.g. the Mayan script was probably an independent invention.
This review should come with a strawman warning. I am not a linguistics academic but enjoy books on the topic and this reviewer’s characterisation of linguistics as just Chomsky seems completely at odds with recent books. There are multiple established schools of thought and most recent books seem to have moved on - but reading this you would think there was some David and Goliath struggle.
Also I felt the LLM conclusions were pretty weak - if you’re going to bring them in I’d want to get more of a “So What?”. Concluding that LLMs show that Chomsky is wrong seems a bit too obvious.
Thanks. I was interested in Chomsky's claims about universal grammar 20 years ago but I think gradually the consensus shifted to the strong version of it being wrong and the weak version suffering from vagueness that could explain anything and nothing, so I lost interest.
I was intrigued to read this book review and indeed, it seemed to be the old perspective that is irrelevant today.
As for Pirahã language and people, we have to take into account that this a very small group of only 200-300 speakers who live in harsh primitive environment. It is not unimaginable that they have experienced societal upheavals like violence from neighbouring tribes, diseases, famine, loss of habitats etc.
Spoken languages change rather quickly and we cannot even be sure that what Everest witnessed wasn't some kind of pidgin formed from merging several local dialects and languages. Wikipedia gives an example that women sometimes pronounce s as h which is an indication of a different linguistic background. You could expect eventual development into full-fledged creole language. However, the number of speakers are so low that now the school has been opened for them where they learn Portugese instead of Pirahã.
- there are extinct but historically attested languages that seem straightforwardly related. It's not like there's a total absence of data of current and some historical neighbouring languages, and it's known to have borrowed vocabulary from unrelated languages like Nheengatu
- Pirahã does not typologically look like a creole at all: it has complex verbal morphology for one thing, creoles usually lack verbal morphology altogether. It has phonemic tone which I believe is rare in creoles. Etc.
- Differences in women's and men's speech differences is not unique to Pirahã, it exists elsewhere in the world. I can't remember what it was but I once read of a language that goes as far as to have entirely different versions of common morphemes based on the gender of the speaker, so one phonological difference doesn't seem all that striking. Also I don't think this is a typical creole thing?
You make good points that Pirahã language could not be pidgin or creole. However, I think that simplified verb structure of creole is due to mixing of dissimilar languages like English and Chinese. What if those two languages are on dialectum continium like Russian and Ukrainian? It is called Surzhyk (Ukrainian-Russian pidgin) and I don't believe it is standardized and yet contains full complexity of verbal forms. Even the most defining features such as a specific pronunciation of Ukrainian [g] is undefined, some speakers might use it and other not.
In case of Pirahã it is not that women speak differently but that some women speak differently. How is it possible for such a small group to develop different phonetic features? We think of Pirahã people as a static group but reality might be that they are mixing with neighbouring tribes, possibly even with violence (like stealing wives from neighbour tribes?). I don't know how much I can trust researchers who are religious proselytizers. They always have their own agenda, even with Everest losing his faith in Christianity, he still might be a person of a strong ideology instead of science.
On further thought, I think my main issue with Chomsky-the-linguist is the whole I-language concept. While studying, in practice, normal (external) language, excluding stuff like errors to make the issue tractable, he felt the need to instead claim to be describing an internal language that didn't have those issues in the first place. I suspect part of this was because forming a model that's in any way less complex than reality, in the social sciences, will get you an instant swarm of humanitarians with "everything is holistic and fundamentally connected to everything else" and such. Either way, the result is that Chomsky is a linguist (and political theorist) that thinks he's a cognitive scientist, and interprets his claims about language as claims about the human mind. Ergo weirdness like "language isn't for communication" or "only humans can have language, AI and other species have no relevance to it" or the single mutation 50 kiloyears ago (stuff like that can happen, sure, but the prior should be very low).
And so - the calm Chomskyan response to Everett saying that the Piraha language doesn't have recursion would be "if this is true, then it's not a true language". Eventually that's what Chomsky concluded. But if you equate language and mind, then "the Piraha don't have a true language" is interpreted as "the Piraha are literally subhuman". And so the response was as toxic as you'd expect from a debate involving the word "subhuman".
‘John, the man that put people in the water in order to clean them for God, that lived in a place like a beach with no trees and that yelled for people to obey God’.
Heck, I'd reject that too! Never mind the tortuous syntax, it's the esthetics that's the deal-breaker here.
By comparison -- and overlooking the cryptic rendering (or mistranslation) of Mark 1:3 and the ambiguity of the last two sentences; is he supposed to be saying that the people lack God or desire God? -- this comes like a breath of fresh air:
‘John cleaned people in the river. He lived in another jungle. The jungle was like a beach. It had no trees. He yelled to people. You want God!’
I guess these short punchy statements might grow tiresome after a while. But then again, I could also imagine getting used to them.
And meanwhile, doesn't the very fact that these sentences occur together and form a coherent whole imply that parataxis is all you really need to coordinate a collection of ideas? Surely no one is suggesting that Chomskyan grammar stands or falls on the presence or absence of an explicit conjunction like 'and' as opposed to -- in this case -- a period that's in essence fulfilling the same function?
"Everett’s answer to the innateness question is complicated and in some ways subtle. He agrees that certain features of the human anatomy evolved to support language (e.g., the pharynx and ears). He also agrees that modern humans are probably much better than Homo erectus at working with language, if indeed Homo erectus did have language.
"He mostly seems to take issue with the idea that some region of our brain is specialized for language. Instead, he thinks that our ability to produce and comprehend language is due to a mosaic of generally-useful cognitive capabilities, like our ability to remember things for relatively long times, our ability to form and modify habits, and our ability to reason under uncertainty."
This makes me think of Cecelia Heyes' Cognitive Gadgets: The Cultural Evolution of Thinking (2018). She argues that many things we think of as innate, like "theory of mind" are learned. "Theory of mind" is the idea that you think a lot like me: you have emotions like I do, you have desires like I do, etc. On the other hand, our brain makes it relatively easy to learn some "cognitive gadgets" and not others.
"The intent here is not to pooh-pooh Chomsky" writes the author of a review that transparently tries to do just that; and fails spectacularly.
The crux of the argument is that Piraha doesn't use recursion but as Chomsky has explained many times (e.g. https://youtu.be/u6Lk79bnUbM?t=0h22m47s) what is of interest is what human capacities are, not whether they are always expressed. In the video above Chomsky makes the analogy to a tribe that wears a patch over one eye, which doesn't refute humans having the capacity for binocular vision.
The reviewer goes on to imply that this means the theory is vacuous or unfalsifiable ("the statement that some languages in the world are arguably recursive is not a prediction; it’s an observation, and we didn’t need the theory to make it... can we test it by doing experiments involving real linguistic data...?") but that is not true: linguists like Moro *have* done experiments that show that humans don't seem to be able to process certain made-up linear-order (i.e. non-recursive) languages as easily as made-up recursive ones and use non-language areas of the brain to do so (see Moro, "Secrets of Words"), which suggests that humans cannot understand arbitrary linguistic structures but can understand recursive ones, just as UG would have predicted. Had the result been the reverse it would have falsified (or at least severely weakened) the theory. (Note that Piraha is like a degenerate recursive language [no center embedding], so it is still "natural" for humans). Also note that the Moro experiment does falsify the "mosaic of generally-useful cognitive capabilities" claim made later on (which is an actually vacuous theory as it doesn't explain anything in any case).
Furthermore, as others have commented the reviewer completely misunderstands recursion, the fundamental characteristic they are supposed to be critiquing. The reviewer's own conception of recursion is indeed "insane", fortunately it is completely wrong. Merge isn't just about putting two elements arbitrarily into a pair, it is about creating a semantic interpretation according to how those pairs are combined and it implies that hierarchical order and not sequence/linear order is what matters for humans. See commenter Dimitrii's telescope example.
There are other misunderstandings:
- "Language is for communication, and probably—like everything else!—emerged gradually over a long period of time." No, not the appearance of the eye or multicellular organisms.
- Purposely conflating biological language faculties with cultural influences, two entirely separate things that are both referred to as "language". Chomsky specifically avoids this by using specialized terms (FLM, I-language).
- Equating language with speech, ignoring the implications of sign language and the mute.
- etc.
Things really go off the rails in the LLM section. The reviewer writes that "Chomsky’s claim that engineering success does not necessarily produce scientific insight is not uncommon, but a large literature speaks against it." but gives no examples in the case of LLMs (because none exist). They then go on to make the argument that since LLMs can approximate the output of human language then they must have analogous internals but this is deranged: You can train an LLM or deep neural net on the output of a clock or the weather but it is absurd to say that a clock or the weather works like an LLM or deep net and that examining the internals of the LLM will yield any insight whatsoever.
When a reviewer reviews a book in a topic they have no interest in, from an obscure author they presumably learned of only recently, one has to ask where the sudden urge to write the review came from. The last section of the review provides a clue: Like many other LLM-cultists, the reviewer had a sudden new-found interest in discrediting Chomsky after he criticized their precious LLMs. Maybe they should try harder to listen to what he has said first.
I have no idea what misunderstanding you have, maybe you should be the one clarifying. You can read the wikipedia article about the "Early eye". If your objection is the word "eye" then you can use the term "eyespot" instead, it's irrelevant to the point that it is an organ that appeared suddenly.
Do you think people who are wondering about whether an "eye" arose suddenly have in mind a notion of an "eye" this is so broad so as to include a very simple light sensor? Or do you think they may have something else in mind by an "eye"?
You said they can't process them "as easily," but then you say, "which suggests that humans cannot understand arbitrary linguistic structures but can understand recursive ones." Is it that they can't understand them, or can't understand them as easily? Also, "people will understand languages with recursion more readily than ones without it" strikes me as a fairly tepid prediction that isn't exclusively consistent with UG or nativist views more generally. I'd also have some questions about who the participants in these studies were; if they'd already been enculturated into one set of languages rather than another, those difficulties may have been acquired rather than innate.
> Also note that the Moro experiment does falsify the "mosaic of generally-useful cognitive capabilities" claim made later on (which is an actually vacuous theory as it doesn't explain anything in any case).
Could you say a bit more about what it falsifies and how?
>- "Language is for communication, and probably—like everything else!—emerged gradually over a long period of time." No, not the appearance of the eye or multicellular organisms.
Can you clarify what the misunderstanding is here?
>- Purposely conflating biological language faculties with cultural influences, two entirely separate things that are both referred to as "language". Chomsky specifically avoids this by using specialized terms (FLM, I-language).
*Purposefully* conflating? How do you know what their intentions were?
> Is it that they can't understand them, or can't understand them as easily
They don't process them as quickly and they use non-language (specifically problem solving parts) of the brain to do so.
> ... strikes me as a fairly tepid prediction that isn't exclusively consistent with UG or nativist views more generally
It's not tepid at all, it provides reasonably strong evidence that language processing requires hierarchical structures, not arbitrary ones, which is what UG predicts and is not what "general processing" theories predict. It's probably one of the most important recent experimental results in linguistics.
> I'd also have some questions about who the participants in these studies were; if they'd already been enculturated into one set of languages rather than another, those difficulties may have been acquired rather than innate.
I have no idea what you mean. Obviously people in the experiment have pre-existing language ability. (It's not ethical to raise a child from birth in a fake-language experiment)
> Could you say a bit more about what it falsifies and how?
It takes longer to process a certain structure and it uses non-language parts of the brain which strongly suggests it is not using the language-specialized machinery to do it but other general problem solving machinery.
> Can you clarify what the misunderstanding is here?
There are crucial instances of evolution that are not gradual but sudden. I didn't mention it but the separate assertion about "language is for communication" also ignores Chomsky's arguments against that (that language is almost never externalized and it is computationally inefficient for communication due to things like the filler-gap problem).
> *Purposefully* conflating? How do you know what their intentions were?
They reference the distinction in other parts of the review and then ignore it when it supports their point. If you don't like the word "purposefully", just replace it with that long sentence instead.
You seem really interested in falsifying the Moro study, maybe you should read the original or the layman explanation in "Secrets of Words" first.
>I have no idea what you mean. Obviously people in the experiment have pre-existing language ability. (It's not ethical to raise a child from birth in a fake-language experiment)
What I'm asking is if they already spoke languages with recursion in them. If they do, then could it be that it's easier for people to process languages that share similar features to the languages they already speak?
>It's not tepid at all, it provides reasonably strong evidence that language processing requires hierarchical structures, not arbitrary ones, which is what UG predicts and is not what "general processing" theories predict. It's probably one of the most important recent experimental results in linguistics.
But even the results as described don't seem to show that it's *required.* Without more details I'd also have questions about why people may have responded this way.
>It takes longer to process a certain structure and it uses non-language parts of the brain which strongly suggests it is not using the language-specialized machinery to do it but other general problem solving machinery.
Okay, that makes sense, thank you. I'm still wondering why this is the case, though.
>There are crucial instances of evolution that are not gradual but sudden.
I would have thought the general concern isn't whether there are sudden instances in evolution, but whether something complex is sudden. Light sensors aren't. A human eye would be. So if someone proposed that a human eye appeared suddenly, presumably we'd agree that that's very unlikely. The question for me is whether claims about whatever genetic features associated with UG that may have arisen suddenly are more likely a light sensor or more like a human eye.
>They reference the distinction in other parts of the review and then ignore it when it supports their point. If you don't like the word "purposefully", just replace it with that long sentence instead.
I'm not going to retcon what you said. It concerns me when people are quick to attribute bad motives to others.
>You seem really interested in falsifying the Moro study, maybe you should read the original or the layman explanation in "Secrets of Words" first.
My work frequently involves critiquing the methods used in empirical studies. That's my natural disposition. I'm also sympathetic to Everett's views and not a fan of Chomsky's views, from what little I've seen so far. I may do exactly that, thanks.
> What I'm asking is if they already spoke languages with recursion in them
Yes every language does (except possibly Piraha).
>If they do, then could it be that it's easier for people to process languages that share similar features to the languages they already speak?
It wouldn't explain the fMRI data. You can certainly come up with a hypothesis that is consistent with the experiments without UG but they are just harder to believe than the "simple" explanation that language is recursive and as far as I know don't have corresponding supporting experiments. It's true that it isn't irrefutable dispositive proof but that rarely exists in psychology and psychology-adjacent fields.
> But even the results as described don't seem to show that it's *required.*
Required for what? I can put a gun to your head and force you talk without recursive sentence structure, what has that proven? If you have a Turing machine and only feed it regular grammars does that prove it is not a Turing machine? If you have a joint with 6 degrees of freedom but observe it only moving in 1 does that prove that it doesn't have 6? If you observe a small tribe in the Amazon that crawls on all fours throughout their lives does that prove that bipedalism is not a "required" property of humans?
> Okay, that makes sense, thank you. I'm still wondering why this is the case, though.
Well the point of the experiment was to demonstrate it is because the human language faculty does hierarchical processing (and can't do arbitrary simple processing)
> I would have thought the general concern isn't whether there are sudden instances in evolution, but whether something complex is sudden. Light sensors aren't
Chomsky agrees! One of the main guiding principles of Chomsky's minimalist program is that whatever occurred it had to be very simple (hence "merge").
> I'm also sympathetic to Everett's views and not a fan of Chomsky's views, from what little I've seen so far.
Yes that's clear, but you haven't made it clear why.
>You can certainly come up with a hypothesis that is consistent with the experiments without UG but they are just harder to believe than the "simple" explanation that language is recursive and as far as I know don't have corresponding supporting experiments.
That doesn't seem simpler to me.
> It's true that it isn't irrefutable dispositive proof but that rarely exists in psychology and psychology-adjacent fields.
I'm not asking if it's irrefutable. I'm wondering if it's especially strong evidence at all.
>Required for what?
You said "It's not tepid at all, it provides reasonably strong evidence that language processing requires hierarchical structures." With respect to the rest of your questions: what point are you illustrating with those questions? At least with respect to the last one, the answer strikes me as a trivial yes: bipedalism isn't "required," so if someone said "all human populations are bipedal" and we observed one that wasn't, that person would be wrong. If they said we had an innate predisposition for bipedality, that'd be fine, but that's a different claim. Evidence of languages without recursion is evidence that all languages don't have recursion; it's not as strong of evidence that recursion isn't a feature of some innate language faculty.
>Chomsky agrees! One of the main guiding principles of Chomsky's minimalist program is that whatever occurred it had to be very simple (hence "merge").
Okay, good. So what does he think occurred? Is there much in the way of empirical evidence that whatever that is occurred?
You would have to come up with some theory about why only recursive language processing is localized to areas of the brain associated with language and linear-order language processing is not. As I said you could certainly do that: Maybe during childhood development the language area becomes "trained" on whatever language it is exposed to and then it crystallizes that way so if you were trained on linear-language it would live in the language area and recursive language would have to be interpreted using puzzle-solving parts of the brain, the opposite to what we observe now. But there isn't any evidence for that and it's just more to believe than the simple explanation that that part of the brain came pre-wired to process hierarchical structure. You may disagree. It also creates its own unanswered questions like how did recursive language become predominant if other language structures could be processed equally well.
But if you want to come up with some experiment refuting the Moro results I think that would be an enormous scientific contribution, even if it is driven by animus against Chomsky.
> At least with respect to the last one, the answer strikes me as a trivial yes: bipedalism isn't "required," so if someone said "all human populations are bipedal" and we observed one that wasn't, that person would be wrong.
I'm glad we've discovered the disconnect here: you are just concerned with semantic pedantry. If I said "humans are bipedal" you would want to have an argument about how the Amazonian tribe crawls on all fours. This is a waste of time. No one is going to add all the qualifiers that you insist on that everyone already understands. People say "humans are bipedal" not "the natural state of humans is the bipedal state" unless their a specific reason to provide that required clarification. If you really need these pedantic qualifiers add them in your own mind, no normal person is going to write them out.
> Evidence of languages without recursion is evidence that all languages don't have recursion; it's not as strong of evidence that recursion isn't a feature of some innate language faculty.
Yes, exactly. It's not evidence at all in fact, which is Chomsky's point and why this whole book review is wrong.
> Okay, good. So what does he think occurred? Is there much in the way of empirical evidence that whatever that is occurred?
I mean that's the whole field of study, where all the main questions aren't solved yet. Chomsky isn't focused that much on the bio-evolutionary aspect of things, more system-computational. I'm not sure it's a fruitful area of pursuit in any case: What caused the eye (eyespot) to appear? Is there empirical evidence for that? We might never know but that doesn't mean it didn't happen.
>> Any expression larger than two words, then, requires recursion, regardless of whether there is embedding in that expression. For instance the noun phrase “My favourite book” requires two iterations of Merge, (Merge(favourite, book)= [Favourite book], Merge(my, [favourite book])= [my [favourite book]]) and therefore is an instance of recursion without embedding.
For crying out loud, the fact that I am able to chain instances of Merge (a glorified concat) does not imply recursion any more than the fact that I can write (1+2)+3 means that addition is recursive.
Coming at this from computer science, that example would be possible with the following formal grammar (in EBNF):
By contrast, I would reserve "recursion" for cases where a production rule refers directly or indirectly to itself:
and_list = noun, "," and_list | noun, "and", noun
The claim is as silly as the claim that a bike is self-similar because a bike is made-up-of wheels and a frame, and the wheels are made-up-of a tires, spokes and a hub. Hence to describe a bicycle, you have to use 'made-up-of' on things which are themselves 'made-up-of' things, hence self-similarity.
Does it really make sense to you that an entire field is full of clueless nitwits who’ve never heard of this exotic "formal grammar" of yours?
Wikipedia, anyway, suggests otherwise: "BNF is a notation for Chomsky's context-free grammars. Backus was familiar with Chomsky's work."
I honestly don't see why you (and apparently many other) are having an issue with the term "recursion" here. As someone in computer science, you must have heard terms like "recursive grammar" and "recursively enumerable sets" before, right?
And by the way, also from wiki: "Addition is a function that maps two natural numbers (two elements of N) to another one. It is defined recursively as: a+0=a and a+S(b)=S(a+b)". I admit this is possibly a bit nitpicky, but it too suggests that you might want to reconsider your understanding of the term "recursion"...
I am aware that some guy named Chomsky (what a coincidence) has established a hierarchy of formal languages, and I am fine with the use of recursion there. Per WP on recursive grammar:
> In computer science, a grammar is informally called a recursive grammar if [...] expanding a non-terminal according to these [production] rules can eventually lead to a string that includes the same non-terminal again. Otherwise it is called a non-recursive grammar.
I am also okay with 'recursive[ly enumerable] language', even if I prefer '[semi-]decideable language' -- all languages expressible without recursive production rules are decideable, after all.
The crux of my disagreement is this:
> The term recursion, when used by Generative linguists, refers to a property of functions. Functions are relations between a set of inputs and a set of outputs. A function is recursive if its output can also serve as its input.
This is a bizarre definition to me. I have googled for 'recursion generative linguistics', but not quickly found more people making the claim, so for all I know this might be just one grad student. For all I know, the rest of the world calls what he calls 'recursive' an 'endofunction'. (Naturally, one can have plenty of indirect recursion without employing any endofunctions!)
The equivalent for biology of that definition might be:
> The term 'heart', when used by rhino researchers refers to the big blood vessels between the blood pumping organ (BPO) and the lungs.
I am fine with addition being recursive in the sense that there is an recursive rule for it in Peano arithmetic. I am not fine with it being called recursive because it projects R x R to R and thus is chainable.
I'm not a linguist or an etymologist, but from what I understand, the term "recursion" in linguistics doesn't meant to imply a precise deep meaning. Instead, it's more of a vague relic from historical jargon that makes sense within its original context.
In the case of "recursive", this context is the foundational concept of early generative linguistics, which posits that syntax is, well, generative - meaning that that syntax of natural languages is characterized by production rules that generate syntactically valid sentences. While this might seem like a simplistic or even trivial idea today, this computational approach was novel, revolutionary, and instrumental at the time.
To me, it seems sensible to use the term "recursion" to emphasize the process of building sentences from simpler fragments by iteratively applying production-rules.
By far the best review I've read so far, thank you! Also finally learned something about this Chomsky guy (seems like another case where great intelligence is used in large part or even mostly to defeat itself).
Do any of you readers also start to feel that sentences such as,
> Large language models (LLMs) like ChatGPT are not perfect, but they’re getting better all the time
are repeated too much? I mean 2 years ago, ChatGPT released i-forgot-what-version and everyone was amazed, but from there, "the law of diminishing returns" hit hard. To the educated eye, one can see that the model simply stopped advancing "so greatly" and all the features they advertise nowadays are some extra pre and postprocessing tasks rather than the capabilities of model itself along with maybe extended memory or so.
Good review. I find it interesting that the Christian missionary is the one making such a strong case for gradual evolution and natural selection, and the Marxist professor is the one arguing some version of human exceptionalism and mysterious endowment.
I also got confirmation of my priors, that Chomsky would be (would have been) more helpful if he could just dial it down to about 65% and make room for a bit more humility in most things.
im very late; but does anyone have example of linguistic theories about non recursive languages?
Im anti the hyper recursion of programming languages; for example, while you could define the allocation of space using "binary space partition trees"(and I use bspwm), you really really shouldnt, artists and designers subdivide 2d space with grids for a reason.
Likewise instead of 2+3*4 instead of letting that statement equal 20 reducing the complexity and let more parsing be left to right; everyone thinks you *need* to push everything into a 10 step hyper abstract process; for what exactly?, you shouldnt ever depend on ambiguous statement grammer.
Yes you can describe all grammatically correct sentences with lambda calculus (aka the merge operator) and yes lambda calculus is hierarchical
But lambda calculus is Turing complete, meaning that it can literally describe everything. It can simulate the physics of my sandwich. Is my sandwich hierarchical?
So is Chomsky doing a massive mind projection fallacy with 500k citations? I mean it's happened before...
But the ones attacking Chomsky ... probably should just sit down and manually work out a few grammars of English sentences to see what is meant by "hierarchy". The same goes for the author of this post. No it's not necessarily about embedding full sentences. Yes, it's profound. We can turn a bunch of words into set-theoretical claims about the world.
"The dog is chasing the cat" -> chasing(dog,cat)=true. We can do that for all English sentences!! That's what Chomsky started! Ain't no obscure rainforest language gonna change that. The model is Turing complete. It makes no predictions to be falsified.
As for "linguistics is about grammar" vs "linguistics is about phonemes and stuff" - that is a discussion about budget allocation and not about linguistics. Is it a blegg or a rube? Depends on who gets funded.
A couple of things that have bugged me bout this review as someone with an academic background in linguistics (just a master degree, so far from an expert).
First it greatly overstates the importance of Chomsky in a way that even the most ardent Chomskian linguists would find glaring, and it does so only in the basis of citations, without thinking that high number of citations non necessarily equals reverence in the field. Chomsky has a lot of citations for various reasons, 1. He is a really respected and prolific linguist and also a very public one 2. His work has been important in several fields outside linguistics (mostly computer science, but also psychology) 3. He has been highly controversial and taken part in very heated debates inside the field (the semantic wars being the go-to example). I haven't checked the numbers, but my guess is probably that there is at least as many citations to argue against his work than to build upon it.
That doesn't mean that Chomsky theories didn't have an impact in modern linguistics, of course they had, and there is a still thriving community of Chomskian linguists, but most linguistic study that is done today it's not Chomskian in nature. To start Chomsky's theories did not really took a strong foothold in Europe, which basically continued and built upon the structuralist work of Saussure and the Prague school (which is different think that American "Bloomfield" style structuralism because linguist want to make things complicated) with the functionalism of Andre Martinet and the SFL and in the United States while certainly Chomsky contributed to the Cognitive shift, most linguist today working on Cognitive linguistics follow more in the steps of Lakoff that of Chomsky. And that is in the more they heavy/grammar heavy disciplines of linguistics, because pragmaticians, sociolinguists, applied linguists and such do not bother themselves with generative linguist at all.
Based on the initial false assumptions about Chomsky relevance, contrast Everett and his wildly contested claims about Piraha (which are not only contested by Chomskian scholars I must point out) with "modern" linguistics that do not do any field work or aren't bother with empirical data. Which is weird because we are in a time where multimedia tools and search motors actually have led to an incredible explosion in the creation of annotated linguistic corpora, which has led to basically linguists having more empirical data than ever to work with and thus to most linguistic work being highly empirical in nature (even inside Chomskian linguists). Theoretician like Chomsky are important, but right now, they are a minority in linguistics.
Lastly I don't get the last part musing about LLM's, because it only has very tangential relationship with the book that the review is ostensibly about, and second because the reasoning is not very solid, we know for a fact that LLM and human learn language in a very different way (I would argue that LLM don't even "learn language" , and are only very sophisticated stochastic parrots). And I am not even referring to the UG, even if we go by Piaget or Vygotsky's hypothesis of language acquisition, none of them rely in humans ingesting incredible amount of data to formulate statistical analysis but rather how development of innate capabilities interacts or creates the ability for language and how this gets molded by interactions with the outside reality.
1. "Chomsky makes claims like this elsewhere too. The argument is that, even if there were a language without a recursive grammar, this is not inconsistent with his theory, since his theory is not about E-languages like English or Spanish or Pirahã. His theory only makes claims about I-languages, or equivalently about our innate language capabilities.
But this is kind of a dumb rhetorical move."
I think from another angle, the point may not be so dumb. Given that a Pirahã child would acquire Spanish if raised in Barcelona, Mandarin in Beijing, and Hebrew in Tel Aviv, presumably this is because there is something related to recursion in the underlying database of rules (UG or something else) that generates whatever grammar the child acquires. Whether any particular language makes use of those rules is another matter.
This type of variation occurs not just for recursion, but most aspects of language. In phonology, languages vary in what phonemes they make use of, and some phonemes are never used at all in a particular language. The same goes for syntax. The difference is that recursion seems to show up in almost all languages whereas (certain) phonemes have far less universality.
2. "Everett thinks cultural pressures forced humans to evolve bigger brains and better communications apparatuses (e.g., eventually giving us modern hyenoid bones to support clearer speech), which drove culture to become richer, which drove yet more evolution, and so on."
Why would selection act on vocal anatomy but not neural circuitry for language? I'd be interested to hear an argument that would make this plausible.
My two cents, and I haven't looked into this and probably won't bother, but I studied linguistics before moving on to other work after graduating. All of this comment is baseless conjecture.
If I'm reading this correctly, Everett is claiming Piraha doesn't have /subordinate clauses/. Which is truly an insane claim. This is a super fundamental language structure. Without it, you have no real way of discussing desires, causes, commands, beliefs, or a bunch of other things. This is a small step down from saying "Actually their language doesn't allow you to ask questions, any attempts to ask anything will inevitably fail."
Running with the Questions comparison, I could see a naive guy studying Spanish, in an alternate world where Spanish is spoken by a group of 200 illiterate, innumerate hunter-gatherers in the deserts of Spain, coming to a bunch of wrong conclusion about the wacky morphology of Español (pronounced "ess pan yoll")
"They have no way of forming yes/no questions!" He says. But everyone and their mother knows in reality that no, Spanish does yes/no questions just fine, and relies on prosody to convey yes/no questions without issue.
(Actually even in popular languages like Spanish there's a bunch of weird misinformation where even experts get fairly basic stuff wrong, but I probably shouldn't go into that here)
The article mentions that the Piraha get mad at you if you talk about something you didn't directly see yourself. I think this may be important to the story. One problem that comes up sometimes with missionaries and researchers who are trying to learn new languages, is that you often rely on the natives' concept of what makes a valid sentence or not. Take the sentence "Colorless green ideas sleep furiously". For an illiterate uneducated hunter-gatherer without a strong sense of grammar, the sentence seems invalid because it makes no sense. My guess is that the Piraha kept rejecting the John-the-Baptist passages because they just didn't like the verbosity.
One of my favorite things from linguistics is the existence of pidgins/creoles, and regrettably I've never studied in detail any creoles, maybe one day I'll get around to studying afrikaans or something. But the prototypical story around creoles is that sometimes, when weird conditions bring together different groups with different languages (slavery for example), the need arises to quickly establish a form of communication, usually by taking basic vocabulary from different languages and putting them together with rudimentary grammar rules. This creates a pseudo not-really-a-complete-language called a pidgin. Maybe it has no way to form comparatives, so it's impossible to say "Stronger than" in this language.
What inevitably happens is a generation of children grow up hearing the pseudo-language, never hear any examples of comparatives, and then spontaneously invent some way of doing comparatives, and basically filling out all the missing features and turning it into a complete language. Sometimes the invented solutions are weird and unorthodox, but my understanding is that there's a basic stock of high-level structures, and every pidgin that comes along will get filled out with all the same high-level structures like noun clauses, adjectival clauses, comparatives, if-then statements, prepositional phrases and whatever else. (children are the engines of language innovation, hence all the slang)
Going back to Piraha, if the speakers of Piraha, in their tiny isolated group, possibly malnourished, have their language in some weird shape where they're not forming any subordinate clauses, I don't think this is a blow to Chomsky's ideas. And I would guess that if the piraha group improved their condition, started growing, expanding, and inventing new technologies, I'm pretty sure their children would quickly start using subordinate clauses once abstraction is at all important for their day-to-day lives.
Anyway, as a linguist I've heard about the Piraha in passing before but never looked into it. I've never heard of Everett before but I don't think he deserves any hate, it looks like he was just trying to help the Piraha People and learn about them. I don't think he's adequately refuted universal grammar, though.
That's what it seems like to me.
Thanks for your comments here, I think they perfectly complement the impression I had from other conversations with Chomskian linguists. It's typically toxic, in bad faith, overclaiming their (mostly theoretical) achievements, and it lacks any sort of self-irony and applicability. Also they constantly reference to their supreme leader more than anyone else, who is the greatest example of this behavior. I mean, Darwin didn't refer to his predeccessors like "mere collectors", neither did Linney, Mendeleev or other global theorists. They typically expressed some gratitude and admiration to those who 'collected stuff' for them. I think now that the success of this approach was less 'grammarian generative' and more 'dissertation generative', because it opens an infinite field of theorizing. I mean, you'd probably expect that good theory would somehow predict how we learn languages, how we model them, and what languages can and cannot exist, or how they evolve. In all of this Ch. didn't provide much. I don't devaluate their work fully, that'd be a very low bar given probably majority of linguists still following this paradigm. But now I think it did more harm than good, suppressing less dogmatic research and thinking.
See Pullum on how Everett has been treated https://youtu.be/06zefFkhuKI?feature=shared&t=1962
Thanks for that link. I never quite understood the distinctions (or non-distinctions) between self-embedding and MERGE, but the intro explains it rather well.
Great video. Thanks for sharing.
Why does AI research have so many citations? Particularly when compared with (other branches of?) Computer Science?
Yes, what is going on here? And not just AI research but Deep Learning in particular (Bengio, Hinton). This area has now captured a significant majority of researchers for all time?
Because there are so many AI researchers. I wonder how the numbers would look like if normalized by the size of respective research communities.
In some of these cases it’s hard to say what “the research community” is. I would have expected Marx to be higher, since the research community of all Soviet scientists working on any subject was larger than any particular modern research community.
I agree. One could parameterize the normalized citation indeces by a notion of community.
Maybe conducting AI research is less expensive because it's primarily digital. In STEM or social sciences researchers have to interact with meat space. In humanities here are no resource constraints so maybe artificial limits on volume of published work is necessary to maintain prestige.
But humanities researchers publish *much* less than science and engineering researchers (especially when measured by counting discrete publications, as citation counts do, so that being cited by 10 articles of 6 pages each shows up more than being cited on every page of a 200 page book), and there are far fewer humanities researchers than engineering researchers.
In part, the publishing conventions in AI (and adjacent fields) lean towards producing a large number of small papers. That is, the papers per researcher is very high, and the number of researchers is very high.
Because the number of researchers has exploded in the past 10 years, driven by massive investment from industry, rather than academia. Big AI conferences now get more than 10,000 submissions a year. Other areas of computer science are primarily academic, and are at least two orders of magnitude smaller.
Hinton's been around for a long time, but if you look at citations prior to 2014 or so, when neural nets were still mostly an academic curiosity, the number will look much more reasonable.
I did a little experiment. I was browsing some of the papers that cite Chomsky's most cited book, so I picked up one paragraph from the abstract of one of the top results that Google Scholar produced, fed it to an AI chatbot (Grok) and asked for reference suggestions:
"Based on the provided abstract and considering the context of attachment theory, I can suggest several references that would be relevant to your academic paper titled 'Patterns of Attachment: A Psychological Study of the Strange Situation':"
It gave me 8 suggestions. Call me cynical, but I think AI researchers are actively using AI to write their papers. A lot.
Which begs the question: what percentage of the conclusions are hallucinations (which I prefer to call bullshit)?
Did it suggest any papers actually cited by the allegedly human author? Were they even real papers (hallucinated citations are a real issue with LLMs)?
5 out of 8 were actually cited by the human author of the book (it was a book). Another one was the original book.
because the papers are AI-generated? :-)
Because it is industry backed there is constant publishing. See "Two Minute Papers" in YouTube and you will see. There are always new techniques published specifically by companies such as Nvidia or universities that are backed by such companies in this specific area.
Pirahã is pronounced pretty close to [piɾahã]. Portuguese spelling is rather straight forward.
There is a big difference between the small amount of language a baby hears when learning to speak and the large amount of language artificial intelligence needs. It is not clear that artificial intelligence disproves Chomsky's theories.
A counter-argument to this “poverty of stimulus” line of thinking is that a baby doesn’t get merely exposed to some words. It also gets a deluge of visual, auditory, sensory etc. information, all perfectly synced, and _then_ it also gets to actively explore and _then_ it also gets multi-faceted types of feedback from the environment and from fellow humans. ML theory suggests that each of these differences can matter quite a bit.
This, indeed, doesn’t disprove Chomsky- but it means the poverty of stimulus argument is far from decisive, as the stimulus is not all that poor.
I am guessing that the baby has most of the deep layers already hard-coded (by evolution), and only requires some equivalent of a LoRA to get up to speed in his native language. The LLM has to be taught from scratch.
Well that's Chomsky's theory. Language is hardcoded.
No, not language itself (that'd be the LoRA), but rather the facility to learn language (however that might function, which I suspect no one knows).
Good question. Can anyone answer this?
Also keep in mind that human language itself evolved to be easy to pick up by humans.
LLMs don't have that luxury: they need to learn a language that evolved not for ease of understanding by them, but by some aliens with completely different quirks in their underlying systems.
LLMs don't really have minds, though: they're just language-learning machines. That's what makes them ultimately unsuitable as the sole basis for AGIs.
I fixed my comment to remove the unnecessary mention of 'minds'. My point is independent of that word.
"Easy to pick up by humans" during a specific stage of brain development. Much more onerous to pick up after the first five years of life.
Much less onerous? An adult dropped into a foreign language community can pick up the language in several months, not several years.
Well, there's been a lot of handwaving and jumping up and down by behavioral geneticists about the FOXP2 gene and language acquisition in humans. And FOXP2 seems to play a role in song-bird song acquisition. Although FOXP2 may help with the ability to distinguish patterns in sound, it doesn't explain the emergency of language in the brain.
The human genome doesn't really have enough room to provide all that much hard coding. The entire thing is a couple of gigabytes, and most of it doesn't have anything to do with how particular synapses in the brain are wired. You've got basic reflexes, the learning system for training up the bulk of the mind mostly from scratch, and I'd guess not much else.
I suspect that human babies
1) Get a lot more training signal than just language. You've got to count all the visual information, audio information, etc. as well. This data is probably lower quality, but it still helps a lot.
2) Just have a more efficient architecture than transformers.
Since everyone's working on better architectures and better training procedures, I'd guess the AIs will catch up and then surpass the data efficiency of human brains not too long from now.
Information isn't infinitely compressible. If there are x different ways a brain can be (x different possible phenotypes), and if genes were the sole determinator of which of the x different brains you'd get, there'd need to be x different possible genome settings. Otherwise not all x phenotypes could be realised.
And the point here is that x<<2^11, or ca. 3 gigabytes. And I think that's likely not enough to encode all the possible different pieces of deep language heuristics.
Now in reality, the genes don't actually need to encode the entire blueprint for the language faculties. They just need to encode some initial setup that will predictably become a language faculty once the neural network starts training. This requires much less description length. You just need to encode the right optimiser and the right sort of architecture.
This suffices to explain why humans can learn language and other animals mostly can't. It's not that the whole language faculty is hard coded, the setup to be able to produce that language faculty is. Same as how you're not going to get a good language model out of a 2015 era MLP trained with the methods of the time, while a modern transformer or mamba model does great at it.
It can also explain why language faculties usually seem to be encoded in similar brain regions in different humans, and work in roughly the same way. You don't actually need hard coding of the entire language processing apparatus in the genome to explain something like this. I can see how people might have thought that twenty years ago, but a big lesson mechanistic interpretabillity of modern neural networks has taught us is that there's a lot of convergence in what is learned. If you train a bunch of different networks on the same sort of task, even if you do it with different initalisations and different data and even somewhat different architectures, they will often still converge to implement almost identical solutions.
'Not exactly, there is some variability introduced by the RNA transcription process, from what I can gather. Therefore, each combination of DNA could have more than one phenotype'
This variability can't be used to encode the information needed to specify a language faculty. Otherwise, organisms with the same DNA (humans) would sometimes come out having a language faculty, and sometimes come out not having a language faculty. For each bit that the transcription process contributes, the number of humans that possess the hard coded language abilities should be cut in half.
"The big difference however is that DNA is quaternary storage not binary like our computers."
This is already taken into account in the 'couple of gigabytes number I gave. Humans have ca. 3 billion base pairs, each of which takes 2 bits to encode. That's ca. 0.75 gigabytes total, I misremembered it to be a bit more.And that's not even taking any further compression into account, which would cut the actually available storage down even further. And then most of it is used to specify the rest of our bodies, rather than our brain initialization.
Presumably the human genome still contains some kind of a language learning bootstrapper, since human babies can easily learn at least one language, whereas babies of other ape species cannot.
I believe most human children can manage about three different languages, at least if the parents are motivated to teach them.
This is precisely the question, no? Whether humans have an innate facility for language, or whether we cobble one together out of other cognitive adaptations.
Definitely. No problem fitting that into a couple of gigabytes.
But I expect that to be coding for a particular kind of brain architecture/training signal/more neurons of a particular type. Because those things probably don't require many bytes of description.
Taking the example of LLMs, encoding the LLM architecture, loss function and optimiser takes very little space. Whereas pretrained weights for hard-coding a bunch of deep language rules would take a lot of space, if our LLMs are any judge.
I don't really understand why Chomsky is so opposed to it or why it would disprove him. Statistical learning could just statistically learn grammar as well as or better than an expicit encoding. I suspect he didn't understand machine learning techniques and made some ill-considered comments, and his pride won't let him back off of it.
Just rowing in here with my usual comment that we have no idea if humans are the only animals with language.
At some point, this becomes a matter of definition, but for most intents and purposes, we kinda do. (My go-to book for this is Zorina & Smirnova's "What speaking animals taught us", but it is in Russian.)
It just seems astounding unlikely that we are so unique ... but if you have a book with good arguments otherwise in a language I can read (just English, sorry!) then I'll gladly read it
I mean, there's "Why Only Us?", obviously, but it is a bit too condescending and abstract to my taste. Rumbaugh and Savage-Rumbaugh (two different people!) have a number of works, but probably on the more technical side. And Hockett (1966) is a classic.
Oh, and "Language Instinct" by Pinker, of course. Can't believe I forgot that one.
We already know we're unique. We've built a giant civilization and all sorts of crazy technology. Our uniqueness including our language isn't that surprising.
That said, we also know we're not unique in terms of sheer brain power. I don't know what pilot whales and the like need all that brain power for, but it's at least plausible that they could use language.
Dolphins can at least understand basic syntax, in that they've been taught a language where changing word order can change the meaning.
We’re not. See, e.g., https://www.bbc.com/future/article/20240709-the-sperm-whale-phonetic-alphabet-revealed-by-ai
I’ve long suspected marine mammals are quite intelligent, and it’s a small leap of imagination to think they have a language. Water is an excellent sound conductor; natural underwater waveguides allow sound propagation for literally thousands of miles. Perfect environment to develop a language.
This is very far from a proof that whale communication is language. I think it’s absolutely right to remain open-minded about whether whale communication is or isn’t like human language in whatever sense you mean. There might be a few bird species whose communication is possibly like human language.
But “using AI to identify a ‘phonetic inventory’ of codas” is very far from showing that this is language.
Did you read the whole piece? Where, for example, the researchers describe “[…]the sperm whale "phonemes" could be used in a combinatorial fashion, allowing the whales to construct a vast repertoire of distinct vocalisations. The existence of a combinatorial coding system, write the report authors, is a prerequisite for "duality of patterning" – a linguistic phenomenon thought to be unique to human language – in which meaningless elements combine to form meaningful words.”
They are pretty far along the road toward “yes whales have a language”. The help of machine learning pattern recognition is useful, but far from the only thing pointing toward it.
I hadn’t read the whole article. I refuse on principle to read BBC science articles, because they’re about as bad as they come with sensationalizing things. But the actual Nature article is interesting.
Here is the key chart, where they classify all 8000 or so codas in their database on the combinatorial dimensions of rhythm, tempo, and ornamentation: https://www.nature.com/articles/s41467-024-47221-8/figures/3
It’s interesting for sure, and seems like a breakthrough in interpreting this communication. But this is very far from saying that this is a language. It’s not possible to show anything significant about a claimed language if your database includes a total of 8000 phonemes (the equivalent of about 1500 words) across several dozen communicative exchanges.
No, there are 21 distinct coda-types - they have recordings of 8000 coda-tokens. It's like they have recordings of a few dozen people each saying a few words to each other in a language like Japanese, that has about 20 distinct phonemes. Definitely not enough to prove it's a language, but maybe enough to get its phoneme inventory if they're right.
Yeah I’m with you on science coverage, and not just BBC of course. This one seems reasonable enough.
One thing to note: I remember reading somewhere that the vocabulary of an English peasant in the Shakespearean comprised only about 600 words, so the whales have at least the same order of magnitude vocabulary. Sure we don’t know much about their language - yet - but that’s just a matter of time, IMHO.
Looking forward to someone dropping a hydrophone into the ocean and having the first dialog with a whale!
Note - it's not that they have 600 distinct "words" - it's that they have recordings of about "600 words" worth of conversation total. In any human language, about a hundred of those words would likely be repeats.
But also, I'm very skeptical of a claim that there have been adult humans with a vocabulary of about 600 words. I could *maybe* believe 6,000. Here's a list that claims to be the 1,000 most common words in English now: https://www.gonaturalenglish.com/1000-most-common-words-in-the-english-language/
I could imagine a peasant from a few hundred years ago might not know the following words from the first 500 on that list: school, state, student, country, American, company, program, government, million, national, business, issue, provide, political, include, community, president, real, information, office, party, research, education, policy, process, nation, college, experience, former, development, economic, military, relationship, federal.
But a peasant would surely know several dozen names for animals and plants that aren't on that list, as well as tools and actions to do with them, even ignoring whatever religious and family vocabulary they might have.
I feel that is incredibly dependent on the definition of language. Chomsky's definition seems incredibly ridged, as is common in foundational research. "A structured system of communication that consists of grammar and vocabulary" is true of various bird songs. The objection then becomes that grammar is more than just structure but a structure of a certain complexity defined to separate human language from other animals, which seems very post-hoc to me.
>The objection then becomes that grammar is more than just structure but a structure of a certain complexity defined to separate human language from other animals, which seems very post-hoc to me.
You have a good point, but there are fairly obvious measures of the complexity of human language which I'd be surprised to see any other animal's language reach.
For instance, human sentences are sufficiently varied that, barring quotes, repeated sentences (except exceedingly short ones, "Damn!" is common) are rare. Does any other animal match this? This doesn't require any specific properties of the language's syntax.
The simplest and most distinctive features of language:
A very large fraction of utterances (e.g. sentences) are novel combinations of existing elements
The patterns by which these elements are combined are recursive (e.g., any two sentences can be combined with “and”, noun phrases can combine with a preposition to form a modifier within a bigger noun phrase).
Does anyone know what the Chomsky followers would say about the recent findings that non-human animals (a large number of them) clearly have methods of communication?
I'm not sure what important recent findings you're referring to; the fact of non-human communication is ancient knowledge. What Chomskyans would say is that human language is evolutionarily discontinuous (it is not an outgrowth of the same phenomenon in other species) because there are categorical differences. Nonhuman communication systems are syntactically basic (certainly non-recursive), goal-orientated, lack reference, conceptual labelling, symbol arbitrariness and so on. Claims to the contrary based on e.g. training apes to sign have always anthropomorphised the data and there is not wide acceptance that any species has demonstrated human-like capabilities.
Confirmed by a Chomskyan (me!). Also, note that language is not a tool for communication, that function is secondary and forced (see the quote in the beginning of the article), it is originally a tool for structuring thought.
"language is not a tool for communication, that function is secondary and forced (see the quote in the beginning of the article), it is originally a tool for structuring thought"
Can you summarize, for a complete lay person in this area, what the evidence is to support this claim?
Multiple homonymies, synonymies, and other inconsistencies between the likely syntax (and semantics, which is read off syntax in a not-quite-trivial way) of sentences and their observed morphophonologies.
Sorry...more lay-persony, please?
Yeah, got too jargony there for a second (although that itself is a good illustration!). Like, take the "I saw a girl with a telescope" example above. It has a structural ambiguity (do you first join "girl" with "with a telescope" or "a girl" with "saw") _and_ a lexical ambiguity between instrumental and comitative (in Russian, for instance, these meanings are expressed differently). You can also show each of the homonymies separately. We think hierarchically and only then try to push that hierarchy into a linear sequence (a sentence); and, to add insult to injury, our lexical system is also ripe with various meaning switches (metonymies, metaphors, and so on): "I saw a mouse" is ambiguous between a computer gadget and a small animal, and it persists despite us usually not noticing a similarity between the two (although it is, of course, originally, a metonymy on visual similarity).
There's an often-found claim that most homonymies are not a problem in real speech (the one @jumpingjacksplash begins with below), but there remain quite enough.
Wait, the argument is that when I'd never get confused saying "witch/which" whether I mean "sorceress" or "that," but someone listening to me might? That's faintly interesting, but it's a feature of toy systems which aren't remotely for the purpose of thought.
For example, imagine two ships communicating with flag signals, but with a system where letters and concepts use the same flags and you have to guess what the other person means. The crew of ship one try to write "halt," but the flags for H=Denmark, A=Collision, L=Rock and T=West. The crew of ship two thinks a Danish ship has hit a rock west of them and sails off in that direction.
The real evolutionary pattern of language must be something like: 1. Simple sounds to indicate things; 2. more complex sounds to convey relations that can be thought non-linguistically (eg. this rock is on top of that rock; 3. the possibility of having thoughts that couldn't be expressed non-linguistically due to thinking in language (eg. "language is ontologically prior to thought"). 4. The utility of non-linguistic thoughts by allowing complex social co-ordination (e.g. "Obey the king of Uruk or face the wrath of Marduk"). This then just comes back to Wittgenstein's point that sentences of type 3 are basically the equivalent of "colourless green ideas sleep furiously" so far as external grounding is concerned, although I'd modify that to having powerful impacts on social co-ordination. Hence they stuck around and developed as opposed to being sanded off by evolution.
That's the best evidence?
That's the evidence that's easiest to explain in non-technical terms (not that the comment you replied initially succeeded in it). But yeah, the main thrust of the argument is along the lines of "we see a bunch of people hammering nails with microscopes, but the microscopes' internal structure suggests it wasn't made for hammering".
How would you respond to this?
https://www.reddit.com/r/self/comments/3yrw2i/i_never_thought_with_language_until_now_this_is/
(I originally came across it in Scott's review of "Origin of Consciousness in the Breakdown of the Bicameral Mind")
The person who wrote that claims that they used language to communicate but not to think. I realize it's one, self-reported Reddit post, so not exactly high-quality data, but I wonder if there are other claims like this, and if they have been investigated by linguists?
There are other claims like this, and they are true at some level and false at another. We often indeed don't think in language in the sense of "build a sentence, up to and including how it sounds/what the articulatory commands are" (the acoustic/articulatory debate is its own bag of worms), but the idea is that to formulate _complex_ thoughts specifically, we use a method to chain thoughts which is so similar to syntax it probably _is_ syntax, even if it never gets sent to the systems of articulation ("is never sent to spell-out", in jargon-y terms). Note how using language was tremendously helpful for the guy, because it helped organize thought.
I think I read that post differently than you. He seems to be claiming that he didn't have complex thoughts before he started thinking articulatory in language, e.g.
"I can only describe my past life as ...."Mindless"..."empty"....."soul-less".... As weird as this sounds, I'm not even sure what I was, If i was even human, because I was barely even conscious. I felt like I was just reacting to the immediate environment and wasn't able to think anything outside of it."
So we seem to have a case where language was used exclusively for communication, and not at all for thought. Doesn't this contradict the claim that language is primarily used for thought, with communication being secondary?
Imagine you have a microscope and use it as a hammer. Doesn't this contradict the claim that the primary use of microscope, as evidenced by its structure, is for something else than hammering in nails?
<i>I can only describe my past life as ...."Mindless"..."empty"....."soul-less".... As weird as this sounds, I'm not even sure what I was, If i was even human, because I was barely even conscious.</i>
Huh, maybe p-zombies aren't such an outlandish idea after all.
Also, I wonder if this kind of testimony might be useful in studying how animals experience the world. Obviously there are many more differences between humans and other kinds of animal than just thinking with language, but still, this might be the closest we can get to answering the question, "What's it like to be a bat?"
Also this woman: https://www.youtube.com/watch?v=u69YSh-cFXY
Not sure you can readily describe her as chaining thoughts together using something like syntax.
An interesting question, but, to be fair, other explanations are available. (Is she able to imagine to raise her hand without raising her hand, I wonder?)
>We often indeed don't think in language in the sense of "build a sentence, up to and including how it sounds/what the articulatory commands are" (the acoustic/articulatory debate is its own bag of worms), but the idea is that to formulate complex thoughts specifically, we use a method to chain thoughts which is so similar to syntax it probably is syntax, even if it never gets sent to the systems of articulation ("is never sent to spell-out", in jargon-y terms).
At the very least, some thinking can also be geometrical rather than syntactic. Visualizing a part dangling from a flexible support and envisioning where its center of gravity will be and how it will be oriented is a complex operation, but it certainly doesn't feel linguistic to me.
This reminds me of the other day I was rolling out a pie crust. The instructions said to roll out the size of a sheet of notepaper. (In fairness to the recipe writer, I had misread and this step was not the final step; rather she wanted us to roll this shape prior to doing several folds for flakiness.)
But the point is, I kept trying to do this shape, which was also hard and impossible for me because I was making only a small pie in an 8 inch pan so didn’t have the full amount to make a notebook-paper sized rectangle. Poor reasoning on my part.
Throughout this process my brain vexingly kept flashing in my head a *picture* of the round tin, as if to say, you’re on the wrong track, old girl. Can’t emphasize enough this was a picture only, that I had to keep suppressing: “She said notebook paper! She said rectangle!”
It's _complicated_, but I am not sure that when laypeople (not physicists who learned advanced ways of modeling such things through, well, language) do it, it is _complex_ in the sense of explicitly manipulating multiple parts of the system.
Inner mental experience is far richer than that - this is a massively under-researched and under-discussed topic in our society. The best detailed research into it is by Hurlburt using a technique called Descriptive Experience Sampling.
I recommend this paper: https://hurlburt.faculty.unlv.edu/heavey-hurlburt-2008.pdf
One result he found is that most people often think in rich unsymbolized concepts - I do this myself, and I often don't *have* words for the concepts. I have to invent language, or spend a long time crafting it to turn the thoughts into words. This to me makes it pretty clear my underlying mind isn't primarily linguistic
(I use inner speech a lot too, and it is very useful for certain kinds of thought. It is but one modality of thought.)
Similarly, lots of people think mainly with images, or mainly with emotions. Others are mainly paying detailed attention to sensations, or have no focussed inner mental experience at all.
There's a phenomena Hurlburt found which the complex chained thoughts you describe reminds me of, where people can have "wordless speech". This has the cadence and rhythym of a sentence, but with concepts in the awreness instead of linguistic words.
Lots going on to investigate here, and I think linguistics would hugely benefit from studying Hurlburt's research.
So, there are several things to unpack here.
First,
> This has the cadence and rhythym of a sentence, but with concepts in the awreness instead of linguistic words.
is indeed prima facie evidence for, not against, "to formulate _complex_ thoughts specifically, we use a method to chain thoughts which is so similar to syntax it probably _is_ syntax, even if it never gets sent to the systems of articulation". It is just a specific case where you don't send it to the systems of articulation because the specific terminal nodes correspond to concepts which don't have a phonological word to them. (Compare a situation when you're bilingual and know a word for the concept you need in one language but not in the other.)
Second,
> Similarly, lots of people think mainly with images, or mainly with emotions. Others are mainly paying detailed attention to sensations, or have no focussed inner mental experience at all.
Indeed, there are many non-linguistic ways of thought; the question is, though, whether any of them are _complex_, chaining ways. Like, images seem to just be, well, non-compositional: an image of a house, cognitively, is not images of windows imposed over image of a wall combined with an image of a roof (as evidenced by the default image of window probably differing from the parts of the default image of a house that correspond to the house's windows).
Comparison to aphantasia in the comments, while overreaching in saying it is literally the same (it isn't), is interesting.
Do Chomskyans equate "language" and "abstract thought" when they say things like this? Seems like some of this conflict could be definitional.
Some of the conflict is definitely definitional, but not all of it.
What is the Chomskyan response to Evelina Fedorenko's work that suggests that even when the language systems in adult brains are damaged, their other cognitive functions remain intact? For example, people with global aphasia can still engage in complex causal reasoning tasks such as solving arithmetic problems and reasoning about others' intentions.
From my own experience, I mostly think without using language (i.e. talking to myself). It's only when I have to break down complex tasks into logical subunits that I begin to talk to myself about what I'm doing (but that's mostly a placeholder narrative — "e.g. this is the screw that I need to insert here, after I get this thingy lined up with the thingamajig." Also when I need to write, I'll start thinking in language. So, I'll grant you, language is necessary for structuring thought that leads to communication with others, but I don't think it's integral to reasoning. I'm pretty sure I could get the screw into the thingy and thingamajig without overtly describing the process to myself.
Aphasias are a prime proof of linguistic module's separateness, so this is a very important question indeed. The problem is, there are at least six different kinds of aphasias (Luria's classification), and none of them looks like "break the syntactic module". Broca's aphasia (in Luria's terminology, efferent motor aphasia) is pretty clearly "broken PF" (syntax-phonology interface), Vernike's aphasia (don't remember the Luria name for it) is pretty clearly "broken LF", some others are way more specific (there is, for instance, an aphasia that specifically targets multivalent concepts like "under").
Maybe he's referring to this?
https://www.nature.com/articles/s41559-024-02420-w.epdf
Yes, and the whales. Thank you.
Name-like calls are not sentences with grammatical structure. They don’t imply recursion or an infinitude of possible utterances with a finite vocabulary.
They may communicate, but they don't seem to do language well. All the ape-language studies from the 70s and 80s failed miserably (the most famous of these research projects, Coco the gorilla, made a good mascot but terrible science, particularly as the researcher refused to share or release unedited video of Coco communicating). It seems like even smart animals, like chimps, can't grok language the way humans can.
https://www.psychologytoday.com/us/blog/the-origin-words/201910/why-chimpanzees-cant-learn-language-1
Thank you, this is very informative. I'm thinking about results from whales and elephants. Do you have thoughts on whether this counts as "language" from a Chomskian view?
Here's a link from Ryan L about the elephants: https://www.nature.com/articles/s41559-024-02420-w.epdf
There's a lab at MIT(?) working on the whale translation.
I don't know Chomsky well enough to comment. My only insight is that laymen (such as myself) often think that language is simpler and easier than it is, so much so that we are impressed by chimps that seem like they can do sign language, or elephants that seem like they have names for each other. There really does seem to be a huge gulf in capability between humans and all other animals when it comes to language.
>There really does seem to be a huge gulf in capability between humans and all other animals when it comes to language.
Agreed. From the point of view of resolving how language actually evolved, it is a pity that all of our fellow Hominini except the chimps and the bonobos are extinct. At least the hyoid bones can fossilize, consonants and vowels, not so much... I'm skeptical that we will ever really know.
Yeah, that is broadly correct.
It doesn't. Symbolic thinking may be one of the many prerequisites, but it doesn't have the, you know, grammar prerequisite.
Having a name for each elephant that they use to get each other’s attention is more sophisticated than a finite list of alarm calls, but it does not mean that there is productive recombination of utterances to express novel meanings. One of the major features of language is that you can (largely) understand novel expressions that you have never heard before, and the limitations on your understanding aren’t due to the novelty, but due to inherent uncertainties of context and intention.
They can't grok human language the way humans can. I don't understand why this is sufficient to decide that they can't grok language? Surely one would need to study how they communicate with each other to decide that?
We need to study the ways they communicate to decide that. There are many animals whose communication systems have been studied sufficiently to be sure that they don’t have the complex structure of human language. This is true for dogs, cats, primates, many birds, and many other animals. I think there are some animals whose communication system is complex enough that we know we haven’t yet understood it, so that we can’t be sure it’s not like human language. This is true for some whales, elephants, some birds, and some cephalopods. None of these has been shown to be as complex as human language, but they haven’t been shown not to be either.
One interesting note: apparently Ken Hale, late well respected field linguist who did a lot of work on the Australian language Walpiri among many other things, uncontroversially described Walpiri and some other Australian indigenous languages as lacking clausal embedding in the 70s. This Language Log blog post (http://itre.cis.upenn.edu/~myl/languagelog/archives/004592.html) muses on various possible reasons this basically went unnoticed whereas Everett immediately got in an academic flamewar, but the main reason seems to have basically been that recursion hadn't been proposed as a linguistic universal until the introduction of Chomsky's Minimalist Program in 1995, earlier Chomskyan models of universal grammar did not include it.
So the idea that this may not even be unique to Pirahã seems like an important counterpoint to some claims I've seen that Pirahã in fact probably does have syntactic recursion and the real problem is that its sole expert isn't describing the language correctly. I've seen Everett comment somewhere that a problem may be that field linguists are essentially trained to fit grammar into a certain framework and so may be trying to square peg a round theoretical hole (as he himself says he did in his original PhD thesis description of Pirahã), thus failing to properly document the true linguistic diversity that really exists in the world.
Disclaimer that I don't have any expert understanding of the technical issue here, just a longstanding hobbyist fascination with linguistics. I acknowledge that my amateur understanding is causing me to miss something important here. But overall my reactions to the Chomskyan team response to this, where Everett is probably wrong but even if he's right it actually doesn't matter for his theory, strikes me similarly to you: it feels like dragging a major hypothesis into an unfalsifiable territory where it gets hard to understand what definitive claims it's actually making about the nature of human language at all.
Uh-huh, the so-called non-configurational language hypothesis. See https://www.ling.upenn.edu/~jlegate/main.pdf.
The claim that all human languages have sentences containing more than one word isn’t vacuously true, it’s profoundly true. Just compare this fact with the norm in the communications of other species. No equivalent of “the [color] bird” is known in how any other species communicates. That, is Merge.
The claim _is_ profoundly true, but note the dolphin comment above.
What's profound about it?
Communicating in something composed of parts with their own semantic effect (like, "dog" in "A dog came" has a meaning but "g" doesn't; these parts are normally called morphemes but layman explanations often default to "words" because it's simpler) is _not_ trivial, most animal communication systems don't have morphemes. Of those that do, seemingly none except humans and apes trained by humans are in the primate tree.
Does, as the author of the article suggest, this mean that any language that has more than one word thereby has recursion?
Interestingly, no! We know the answer from ontogenesis (children development). There are three prolonged stages in humans: one-word, two-word (the involved words are short and probably analyzed as monomorphemic by the child, so ignore the word/morpheme angle), and then the so-called "grammatical explosion" where the child rapidly builds more and more complex structures. So, to answer the question in a pithy way, two doesn't make recursion (you can combine, but you can't use the combined thing as the input to the next combination, which is where recursion kicks in), but three does.
And see the dolphinspeak comment above on the linearly recursive vs. hierarchically recursive difference.
Okay, thanks for clarifying.
Prairie dogs communicate both the color and species of intruders.
You’re confusing a syntactic claim with a semantic claim. Which is funny in context.
There’s a difference between obligatorily one-word utterances and a possibility of two-word utterances. What I think is actually the interesting difference though is the difference between utterances with a fixed upper bound on length and utterances that can be in principle unbounded (which even Everett says is true of Pirahã).
Say what??
This has to be one of the most hilarious pots we've seen.
Watch out for Jesus! His Chomsky rating is off the charts.
Pots? No, I meant posts. Oops.
You can edit posts, you know.
I've done editing in the past. I just can't remember how to do it.
Hit the "..." at the bottom right of your comment.
Thanks. It's always something simple, isn't it.
LLMs pick up language in super interesting ways, for example (for any of my Hinglish speakers out there), try asking ChatGPT:
"Kya aap mere se aise baat kar sakte ho?"
The answer I got:
"Bilkul, main aapse Hindi mein baat kar sakta hoon. Aapko kis cheez ki madad chahiye?"
LLMs' way of picking language is interesting but certainly different from humans', both because humans need less material and because they show unexpected biases (for instance, non-conservative quantifiers are unlearnable for human children of the critical age but not for LLMs; you can trivially teach an LLM extraction from islands or binding violations; and so on and so forth).
Could you explain binding violations?
"Mary saw her" cannot mean "Mary saw herself".
"Mary wrote a letter to her" cannot mean "Mary wrote a letter to herself".
"Mary saw a snake in front of her" _can_ mean "Mary saw a snake in front of herself".
And, finally, "Mary knows that Penny saw herself" cannot mean "Mary knows that Penny saw her" (and vice versa, but the vice versa is uninteresting because it's the same as the first).
There are very pecuilar and technical laws surrounding both when a reflexive _can_ appear and when it _must_ appear.
Oh, super interesting. Thanks!
>"Mary wrote a letter to her" cannot mean "Mary wrote a letter to herself".
Is that strictly true? What about this: "Mary reflected on the woman that she hoped she would become. Mary wrote a letter to her."
Is there some technical linguistic sense in which she's not writing a letter to herself? Is metaphorical meaning considered out-of-scope for linguistics? If so, how is that rigorously defined?
There is a purely technical linguistic sense in which she isn't writing a letter to herself, yes. The antecedent, to use a technical term, of the pronoun "her" is "the woman that she hoped she would become", which is a different phrase than "Mary" and, again, in a technical sense, we know it not to be co-indexed with Mary because the first sentence is possible (cf. #"Mary reflected on her"). The first sentence establishes the two women as different objects for the purposes of binidng theory.
This may not be all that rigorous and obligatory. Consider Earendil's poem:
In panoply of ancient kings
In chained rings *he armoured him*.
Fairly clearly, "he" and "him" both refer to Earendil. And the author knew something about the rules of the English language :)
To take a different example, note how "Joe Biden reflected on the president of the United States" is an infelicitous sentence if it happens now but would be felicitous if it happened four years ago.
(Note that nobody explicitly teaches children the laws, and, judging by CHILDES corpora, they don't get enough data to arrive at the very generalizations they tend to arrive at.)
Oh wow, ChatGPT does not realize that this is written using the Latin script. After directly asking it, it says it's written in Devanagari.
Is that what you meant with this example? I do not speak Hindi
Sorry, should've been clearer! I've written Hindi words in English, and ChatGPT is responding in Hindi written in English after "recognizing" that that's what I'm doing. Its ability to mimic that form of Hindi is what I was trying to highlight.
This is kind of a side rant but - I don't think this should be so surprising. As far as I can tell, there's sort of a "script is a transliteration with an army and navy" situation going on with romanized Hindi. It's pretty standardized and extremely commonly used, and it's mostly political and historical reasons that keep it from being recognized as a "real" script for the language. ChatGPT had plenty of examples to learn from online and no real reason not to learn it.
Also note though that you can converse in any number of encoding mechanisms, such as rot13, or, pig-latin, etc ... The patterns that have been learned aren't just at the level of specific examples but also at various levels of representation between the representational, syntactic and semantic.
When LLMs were young I remember someone using base64 encoding as a way to bypass moderation filters.
This makes sense to me – would be curious if it can similarly anglicize other languages that are less commonly anglicized.
Yeah it really shouldn't be surprising. What surprised me is that I followed up and chatgpt apparently has a hard time recognizing that the writing system is not the usual one and says it's writing Devanagari
ChatGPT(-4o) can do much more impressive stuff than that with transliterated language. Try asking it novel questions in romanized Hindi, or in English transliterated into devanagari. (e.g. "Who was Prime Minister of Australia on the day Björn Þórðarson became Prime Minister of Iceland?") I can't say for sure whether it will be able to handle these sorts of questions using romanized Hindi or devanagari English, but it seems to do pretty well when I ask questions in katakana English or romanized Japanese.
Somewhat weirdly, it answers romanized Japanese questions in romanized Japanese, but it answers katakana English in straight Japanese. (This behaviour is slightly different from what happens with GPT-3.5, which seems to have trouble recognizing katakana English as English, and responds to romanized Japanese in English rather than romanized Japanese.) I don't know exactly what is going on here.
My tentative conjecture right now is that it reads katakana English as borrowings into Japanese that it hasn’t seen yet (because the rate of novel English borrowing into Japanese is very high in recent texts) while it reads romanized Japanese as Japanese, because there is more use of romanized Japanese by actual Japanese speakers than there is novel borrowing of Japanese into English.
I don't think there is enough romanized Japanese produced by native Japanese speakers to train an LLM effectively. I think it's more likely is that the GPT researchers generated the data artificially by converting their entire Japanese-language corpus into romaji and feeding that in as additional training data.
Your explanation for the Japanese responses to katakana English seems plausible, but I'm still surprised at how well it handles these questions. It gives convincing answers to:
"If John is standing to the right of Jane, who is standing to the left of Jane?"
"Which game is easier to play: checkers or chess?"
"Do Australian Rules football players swim better than Canadian hockey players?"
"Who was Taoiseach of Ireland during the time when John Major was Prime Minister?"
After numerous attempts, I finally managed to confuse it by asking:
"During the prime ministership that followed the prime ministership of Tony Blair, which Taoiseach of Ireland finished his term of office?"
Even here, ChatGPT gets the answer right at least some of the time. But the answers seem inconsistent across separate chat sessions. It has no problem naming Gordon Brown as the next PM after Blair, but gets confused trying to make the connection to Ireland. Sometimes it correctly answers "Bernie Ahern"; sometimes it names Brian Cowen (i.e. Ahern's successor, who finished his term after Brown left office); sometimes it balks at the weird question and suggests that if I want to know about the Taoiseach(s) in that period, I should ask a "more concrete" question instead.
(Of course, being able to answer this question in any language is pretty impressive — let alone when it is posed in the pseudo-language of katakana English!)
I think this is a shallow take on a complex field. I appreciated that it's well-researched but the sources have not been well understood, though admittedly the review is from a non-linguist, and admittedly the discourse in the research literature is often unprofessional and opaque.
The objections raised against Chomsky's research program are mostly intuitive and therefore trivial - most were dealt with in the 1950s rejection of structuralism and behaviorism (which is foundational to almost all of cognitive science, which is why Chomsky is hugely cited and not as controversial as this makes out), while others betray a misunderstanding of what generative linguistics studies (this is a consequence of 'language' itself acceptably meaning many things to many people, though the review didn't adequately tease these meanings apart).
If anyone is interested to learn about the Chomskyan approach in an informed way, I would highly recommend exploring Norbert Hornstein's Faculty of Language blog. It's now no longer active but it is a treasure trove of explanation that helped me immensely when I was a graduate student trying to make sense of all this.
If the Chomskyan paradigm is so uncontroversial and secure, why the profound insecurity and un-science-like nature of the Chomskyan camp response to Everett? One or two or ten thin-skinned scientists would be expected and not a big deal, but why does Everett's claim ruffle feathers so universally?
The field is large and most working Chomskyan linguists have zero interest in Everett, which exactly matches the expectation that a secure field would not take note of him. Popular understanding of his impact is overestimated because media (even science media) loves drama.
The reason why those who do engage with Everett do so in unprofessional ways (and I happily criticise them for not controlling themselves) is because people like Everett rarely take on Chomsky's ideas as scientific hypotheses; they also want to be iconoclasts and paradigm-shifters, leading the charge against the big bad man who has hold of the field because of politics or mysticism or some other conspiracy. It's unscientific because it's personal, not because Everett's analyses can't be trivially dismissed.
This review claims that several ideas of Chomskyan linguistics seem obviously wrong. It would be very interesting to hear rebuttals to those points from a Chomskyan linguist!
Hello there :)
I have already commented on "language isn't for communication" angle and recursion angle in the other comments. As for I-languages vs. E-languages, the "language is a dialect with an army" problem of distinguishing whether two I-languages belong to the same E-language or different ones as well as most aspects of language (especially non-lexical ones) being essentially in the brain and, importantly, out of conscious control, should be sufficient.
Deciding which are dialects and which are separate languages can be done the same way biologists decide the boundaries of genus, species, and subspecies: messily and with much arguing. Or perhaps biologists should just take the simpler route and declare every organism a unique species, save a lot of effort that way.
Yeah, the species debates are a very good analogy, and species is much less well-defined than population, making it a not too useful notion. (The analogue of population would be a dialect of a tight-knit community; of course, linguistic communities are generally more permeable than populations.) There is one benefit though for biologists: they are not generally under siege from nationalists to declare certain populations parts of the same or of different species. Try telling Arabs that Palestinian Arabic and Moroccan Arabic are fourteen centuries of development apart, making them more far from each other than Icelandic and Danish. Or, conversely, that the differences between Danish and Bokmål or between the four Yugoslavian literary varieties (Bosnian, Croatian, Serbian, Montenegran; Slovenian and Macedonian are genuinely separate, but Macedonian has its own problem of continuum with western Bulgarian dialects) are laughably small and don't count as languages by any linguistics-internal criterion.
Isn't, "your theory predicts always ¬X, look, I've found some X," basically how taking on a scientific hypothesis is meant to work?
Yes, I worded it imprecisely. Everett does do that, though he doesn't stop where that would be sufficient. He embeds it in the rest that I described, which is what baits bad behaviour.
Baits bad behavior? It almost sounds like you're blaming Everett for his critics being nasty to him. I also see a lot of dubious psychologizing, suggesting that he's motivated by a desire to be an iconoclast. You do that here when you say:
"they also want to be iconoclasts and paradigm-shifters, leading the charge against the big bad man who has hold of the field because of politics or mysticism or some other conspiracy."
...this kind of psychologizing is part of the bad behavior and unjustified nastiness directed at Everett, criticizing someone based on speculative assertions about their motives. If someone genuinely comes to hold iconoclastic views, they have every business sharing those views with the broader community. Maybe you've got the causality reversed here, and are supposing a desire to be iconoclastic motivates Everett's views, rather than Everett coming to those views automatically making him an iconoclast regardless of whether that's what he wanted? Why psychologize in this way?
Yeah, this poster starts as an attempt at calm and mature dismissal and very quickly dissolves into thinly veiled ad hominem.
>I happily criticise them for not controlling themselves
For me a rancorous style is possible evidence of an idea which is unparsimonious. It's not just about being rude in the abstract. Maintaining belief in an unstable hypothesis is exhausting and people resent having to explain themselves.
I agree. For me, this is why it's important that these people are a small minority. In the wider context of quality research being done in the field (imo), it makes more sense to conclude that these people are a type who wilfully indulge in undignified behaviour - I think there are people in lots of fields who have this trait even when they are broadly right on the technical detail.
Thanks for the reply. It’s the human condition for sure. I don't have much stake in this, except we have a problem with how 'elites' are viewed, and although a lot of that discourse is toxic, I have some sympathy with this having done a humanities degree and seeing what "cultural marxism", for want of a better word, can do to the study of beautiful things. Can it do the same thing to the true as to the beautiful? Well here you have a Marxist who apparently totally dominates an emerging science and his defenders will go beyond the confines of typical academic discussion to defend him and that raises my eyebrow, I suppose. A person can be completely on top of the technical detail and still be an ideologue. In fact that's mostly how the left has come to dominate academia. But I'm getting into unparsimonious ideas of my own!
You elucidate something subtle and important here, imo. I've always been puzzled by the fact that there's a positive correlation between Marxism and IQ despite the many obvious gaping flaws with Marx. My current best guess is that the answer is that many academics crave power and Marxism gives that to them. Marxism elevates the central planner over the common person and thus creates an implicit power structure in which high-IQ academic types are singularly adapted to thrive. "No no, the collective wisdom of the hoi polloi inevitably leads to revolution so the best answer is to let me order society for you." It's sort of a sad irony that the post-structuralists never turned their tools on themselves because that would have clearly revealed how the structure of the ideology perfectly reflects the megalomania of deflated narcissism that haunts marginalized academics. "Sure you fools make your money and ignore your intellectual superiors, but I'll show you! I'll turn the masses against you!"
What field are you in?
> It's unscientific because it's personal, not because Everett's analyses can't be trivially dismissed.
I'd appreciate an explanation (or a link to an explanation) of how Everett's analyses can be trivially dismissed.
Here is an accessible article by a couple of researchers. The dismissal is trivial because it turns on a definitional rather than an empirical issue. https://inference-review.com/article/fiat-recursio
Thanks!
To me, that article seems confused on the following point: It insists that Chomsky wasn't talking about "recursion" in the sense of "self-embedding", but was instead talking about recursion in the following sense:
> In 1963, Chomsky wrote that his grammatical theory expresses “the fact that … our [linguistic] process [is a] recursive specification”9 and, in 1968, that a “generative grammar recursively enumerates … sentences.”10 In 2000, he noted that our linguistic knowledge is “a procedure that enumerates an infinite class of expressions”11 and, in 2014, that “we can think of recursion as [an] enumeration of a set of discrete objects by a computable finitary procedure.”12
Presumably implying that _this_ sense isn't affected by Everett's study of Piraha.
But I thought that Everett's claims, if true, suggested that Piraha might have a longest sentence, and thereby only a finite set of sentences. And this would seem to contradict this other sense of "recursion" that the article's authors want the discussion to be about. (One of Chomsky's quotes refers directly to infinity.)
-----------
Seperately, there's another route of defense available, and the one that Chomsky seem to take in the interview referred to in the wikipedia article on Everett.
> Chomsky called Everett a "charlatan", and said that even if Pirahã had all the properties described by Everett, there would be no implications for universal grammar. "Everett hopes that readers will not understand the difference between UG in the technical sense (the theory of the genetic component of human language) and in the informal sense, which concerns properties common to all languages".
But on a quick skim of Everett's 2005 article, Everett seems perfectly aware of this distinction. His full critique of Universal Grammar is:
> For advocates of universal grammar the arguments here present a challenge defending an autonomous linguistic module that can be affected in many of its core components by the culture in which it "grows." If the form or absence of things such as recursion, sound structure, word structure, quantification, numerals, number, and so on is tightly constrained by a specific culture, as I have argued, then the case for an autonomous, biologically determined module of language is seriously weakened.
This paragraph seems to roughly assert something like: the lack of certain properties in Piraha is _evidence_ against those properties being genetically built-in. Which seems like a totally fair claim to me, and certainly not one that should be countered by accusations of being a "pure charlatan" by the leader of the field.
(Though Chomsky could be referring to some later claims by Everett. I haven't previously read anything by Everett.)
I think you've somewhat missed why the recursive operation is said to be innate. The prediction of UG is not that every existing natural language will exhibit surface properties that could only be described by recursive enumeration, the prediction is that every existing human (without disability) is capable at birth of learning any of the class of possible natural languages (impossible languages being conceivable languages that no human could learn to speak naturally), and that class includes languages that can only be described by recursive enumeration. That property in turn is not a cultural innovation, it's a cognitive computation. The role of culture is to use or not use recursion; not to invent it.
So long as we consider all humans at birth to be equally capable of learning any natural language, theorists must come to understand linguistic ability by looking at the class of possible languages, and so particular data from Piraha is not some kind of disproof, as the requirements for UG are already established by other languages. Everett would rather need to demonstrate that children of Piraha descent cannot acquire e.g. European languages, which is where the (I think crass) accusations of culturally relative racism come from.
It's a slightly coarse analogy but Everett's argument is in effect like saying that because the Piraha do not ride bicycles, they therefore would not only need to invent bicycles to ride them, they would also need to culturally develop the cognitive pre-requisites that are necessary to learn how to ride a bike. This is a category error. The cognitive pre-requisites must exist latently in every human because they cannot be culturally invented (you cannot learn how to make your cerebellum memorise complex motor co-ordination); the abilities are just awaiting the right cultural inventions to be put to use.
Did you actually read the linked reddit post? That was a Chomsky skeptic who still thought Everett was unscientific. There isn't the necessary connection between the things that you think there is.
...Have you heard of a guy called Fomenko and his "New Chronology"? Most historians ignore him, but those who do very rarely pretend that this is a respectable opponent.
Or take biologists' debates with creationists (fitting given Everett's missionary history). To the extent they react to that, it is an inherently undignifying thing to have to react to something like this, with obvious consequences to behavior.
Nevins feels (correctly, in my view, although I wouldn't even waste time on Everett personally) that is the proper comparison, not your garden variety "we have a disagreement".
It is trivial to build a case against young earth creationists and people of that sort. If you want to compare them to Everett, it should therefore be pretty trivial to disprove his point about recursion and pirahas.
I must admit I am very suspicious of the response to an article stating that "Chomski is the caliph of linguistics" being "the caliph does not need to respond to every beggar in the street". Especially when the beggar in question seems to have done groundwork in a way the Caliph did not
Nevins et al did respond. Hell, the guy also, independently got a whole issue of a journal discussing whether the claims on Piraha have any merit. Chomsky is not the one and only caliph precisely because he doesn't need to personally fight every battle, Chomskyan linguistics is a whole field.
The disproval of Fomenko (done, incidentally, by a linguist, Andrey Zaliznyak) was somewhat technical and took some fifty-ish pages. But it was "in the predictable direction". For me, the points by Everett are in the same bizarre category, and yet a formal disproval does take a bit longer than for young-earth creationists. It is no less damning for its being longer and more technical.
Did Nevins or anyone else trained in linguistics also become as competent in speaking Pirahã? I'd be curious to see what we'd find if we simply had more linguists doing field research and learning more languages.
His former wife, Karen, did. She disagrees with his claims. But then, obviously, she isn't exactly an unbiased observer, either. And yeah, a famously closed-off people wary of outsiders living in an area of Brazil that's further restricted for outsiders by the Brazilian government, is... let's say, poorly accessible for researchers. It's a serious problem with some language-rich areas (Amazonia, Iran, China, New Guinea), it's not because people don't want to do research (there was a common meme that Chomskyans only do research on English, but it is largely outdated; for instance, a lot of Dr. Preminger's career is done on fieldwork in Mayan languages, Dr. Nevins is a co-author of a monograph on Basque, and I have a favorite quote somewhere listing a PhD recruiting data from fifty-ish disparate languages, although in that case, often second-hand).
>Or take biologists' debates with creationists (fitting given Everett's missionary history).
An odd little aside, which comes aside as just a drive-by attempt to associate with Everett with creationism even though there's no seeming connection.
I claim that Everett's claims are to linguistics what creationism is to biology. And I do have a hunch that he is more comfortable with such breakage of science because of his connection to another field that famously engages in such.
Which specific claims does Everett make that are equivalent to creationism? And equivalent in what respect?
Should we go over that quotation set again? The equivalence is in level of iconoclasm and holism as well as in Russell's teapot quality of rejection of analyses more "traditional" ("Your data don't show what you claim!" — "They do if you analyze them this way!" — "But you don't have to analyze them this way!" — "Yes, you do, I know because I lived in the tribe for decades and you didn't, so obviously you don't understand!"). I certainly don't claim that there's specific, concept-to-concept mapping to creationism.
(Regarding holism specifically, he refuses to contemplate separation of language and culture because they interact; imagine telling a biologist that they can't study genes because they interact with environment!)
I'd argue anticolonialist politics was probably a big factor: Everett was an American that travelled to Brazil to study a remote tribe and made an ethnographic account, at a time it was becoming increasingly clear that this had happened before and the result was questionable methodology and disastrous consequences for the natives. (Napoleon Chagnon and the Yanomami)
Also, Everett the charming missionary certainly look the part of the gringo with an agenda.
And Kuhn scientific paradigms: if you take Universal Grammar too seriously, Everett's claims sound a lot like "Pirahã people are too stupid to parse tree-like language structures", and this counts as evidence of Everett being both racist and bad at his job *because* of racism.
Well, to give credit where credit's due, Everett specifically rejects the idea of their being mentally deficient. (Whether it is actually logically consistent with his ideas is a tougher question, but people are known to espouse inconsistent or not-proven-to-be-consistent ideas.)
Does Chomsky believe language arose recently, and abruptly, in a single individual?
The mental capacity for language? Seemingly so. And then, by the usual genetic drift (probably the version for weakly-beneficial mutations, because being able to do complex thoughts is beneficial, including for intra-group competition), spread to the rest of the population in one of the bottleneck events.
Why does he think that? Do you agree with him? What kind of evidence is there that this is true?
I do tentatively agree, and one piece of evidence relevant here is "anatomically modern humans" in Africa who hadn't done much not-present-in-other-hominids human behaviour in 200000 years, with the parsimonous explanation being their not having that mutation.
Isn't the parsimonious conclusion there that the key factor is IQ, not language? Or is the position that complex cognition and language are inextricably intertwined?
Partly but not fully. The problem is, their brain size _and_ their brain _zones_, to the extent they're recognizable from the cranium, are the same as in modern humans. So, we wouldn't expect huge difference in "horsepower" of the mind.
What you wrote earlier ( https://www.astralcodexten.com/p/your-book-review-how-language-began/comment/62600358 ) about a "grammatical explosion" in children also seems to suggest something closer to a switch flipping, to abruptness, than to gradual continuous enhancement.
Exactly.
"one piece of evidence relevant here is "anatomically modern humans" in Africa who hadn't done much not-present-in-other-hominids human behaviour in 200000 years, with the parsimonous explanation being their not having that mutation."
Can you clarify? They haven't done much in what sense and in what timeline compared to what other oganisms? And what is the claim evidence for and why is it evidence for that claim?
And just to affirm: is the claim that whatever distinctive genetic capacities allow modern humans to engage in language as we do arose via mutation in a single mutation relatively recently? If not, what is the claim, and if so, how recently?
Lots to unpack here, but let's see. The idea is:
*On one hand, all Homo sapiens including Khoisan have language, which means that language must have appeared before the Khoisan split some 150000 years ago.
*On the other hand, there's the whole "cultural artifacts appearing" thing, that seems to happen either after that or around that time. (Scarcity of African archeological record is limiting here.)
*Crucially, anatomically modern humans strongly precede those first artifacts.
*The claim is that the crucial capacity for syntax and its interpretation appeared first, in what amounts to a single mutation, and then that system got hooked up to communicative system in what is plausibly a _different_, later mutation. So, two mutations: to think and to articulate the thoughts.
*Both happened and propagated via a bottleneck event relatively recently (but before Khoisan split — not 50000 years ago, as some places claim, that's a later and different bottleneck that wasn't decisive for language for the simple reason it only affected non-African population). A million years ago, probably neither mutation was present yet. So, we have a timeframe of ~1000000-~150000 for both.
*In the timeframe between a million years ago and 150000 years ago, all the three modern-ish Homo species developed: us, Neanderthals, and Denisovians. (There may have been more, lost to history, just like we hadn't known of Denisovians for a long time: again, scarcity of remains...) There were secondary interminglings, the speciation there was _not_ simple, but, to the best of our understanding of what parts of brain do what, it is likely that Neanderthals first separated before the two language mutations, and certainly at least before the second one. This pushes the date downward, closer to us.
So, that 200000 year old skull may well have not had that _second_ mutation yet.
Oh, and note that if someone proves Neanderthals did have language, it pushes the lower bound further behind (makes the mutations less recent) but doesn't change the substantive claims on how it happened — if anything, it ends up _narrowing_ the window for mutation. (Then it could align with the 900000-850000 bottleneck, which currently looks too far away.)
Pedantry alert: Genetic drift is when it's random (e.g. you kill off half the population, turns out this gene was more common in one half than the other, so now it's more common in descendant populations). Language reaching fixation would be directional selection since, as you note, it would be beneficial.
Yeah, I forgot the proper genetic term, sue me :D
There was no major bottleneck even in the last 150,000+ years and the Khoisan split off before 150,000 years and they do language just fine. There has been minimal gene flow between them and other humans since then and the genes that have been selected for when they interbred with some Cushitic pastoralists 2000 years ago aren't related to language.
The Neanderthals also had small amounts of gene flow 200k-250k and another pulse from 120k-100k, after the most basal, language using, modern humans split off.
That depends on your beliefs on https://en.wikipedia.org/wiki/Youngest_Toba_eruption#Toba_catastrophe_theory and a number of other events that have been suggested as population bottlenecks (see https://bio.libretexts.org/Bookshelves/Computational_Biology/Book%3A_Computational_Biology_-_Genomes_Networks_and_Evolution_(Kellis_et_al.)/29%3A_Population_Genetic_Variation/29.05%3A_Human_Evolution). I don't think anyone in linguistic community is particularly married to the 50000 date, I certainly saw different date estimates which looked more plausible (some even allow not to depend on whether Neanderthals had language).
The Khoisan are far too genetically diverse to have gone through a bottle neck then. That theory predates our ability to cheaply sequence dna and the dna evidence we have has refuted it.
This is partially a matter of point of comparison: if (as is virtually certain) out-of-Africa population has gone through an additional bottleneck, our baseline of what is high variability in humans changes (and cross-species comparisons obviously have their own problems). Like, there is no serious theory on the market that wouldn't predict that relatively highest diversity would be in Africa, and within Africa — in Khoisan specifically.
(There are a lot of components in how the entire edifice of language arose: for instance, it required greater control of mouth and/or head movements, and the hominid excavations suggest that this arose earlier, which confuses scores of researchers into thinking that's when language arose, including even such brilliant minds as Svetlana Burlak - but it is quite obvious that better control of fine motorics would benefit tasks beyond language.)
Yes. I personally find this implausible and I think he comes to postulate it because he backs himself into a corner by over-simplifying other (non-structural) features of linguistic ability. I nonetheless think the recursive structure-building operation he proposes is essential to explaining some complex properties of semantics, so I end up in the Chomskyan camp even though I (and many others who would be characterised as sympathisers) think we need a strong refinement of the theory.
I think a lot of those features are relegated to interface between language and communication (Phonological Form, Distributed Morphology/Nanosyntax, all that jazz). There is a discussion to be had - between professionals - to what extent that relegation is plausible (for instance, whether Marantz-style case assignment happens in syntax or in the interface), and, of course, the less one can relegate, the less plausible the theory becomes (although, continuing to use the same example, one can fully imagine that the protolanguage, well, didn't have m-case just like many modern languages don't, and the machinery for that developed later). But that's a discussion on a whole other level than needed here.
I'm a big Nanosyntax fan - what it boils down to for me is that I think the T-model with the two interfaces is a conceptual mistake. I think if the foundational assumptions under that are corrected (which requires work at the boundary between linguistics and philosophy of language), Merge will be preserved but it won't sit within cognition as an autonomous syntax module. That then has ramifications for the makeup of the narrow/broad language faculties and their historical development, but it's indeed another conversation :)
Nanosyntax is very much a T-model model though. (Well, it's a Syntax - Lexicon - Interfaces model rather than Lexicon - Syntax - Interfaces model, and there's a bit more lookahead and lookbehind than usual, but it is a T-model nonetheless, after the insertion phonology and semantics operate on the result separately.)
Indeed. I don't believe there are many well-developed non-T-models in generative grammar (however much the specific interactions are toyed with) because that model has been taken for granted ever since Aspects. But I'm not committed to rejecting all T-models outright; only to claiming that they have explanatory potential while misdescribing certain phenomena, such that a future, more explanatory model will incorporate their strengths and reconceptualise their weaknesses (as is characteristic of scientific development).
The 1950s rejection of behaviorism was important and powerful, but it relied on a lot of arguments about what could and couldn’t in principle be done that have since been falsified by large language models and other work. Many of Chomsky’s arguments early on involve the idea that you can’t make sense of probability as anything but a frequency or some simple observable feature like pitch or volume, which is an appropriate refutation of Skinner, but not of any of the many constructions of Bayesian probability in the second half of the 20th century. He also argues that you need to study language separately from the world, but modern large language models seem to do a good job of creating a model of the language-world system, whether or not they adequately separate it into the two components.
Many of the criticisms of behaviourism had nothing to do with token sequence probabilities, or even linguistics, which is why it failed as a paradigm in all of psychological science. Its foundational problems have not been addressed in any way by machine learning, and the capability of LLMs to do better next token prediction than Chomsky expected doesn't mean much in an integrated theory of human cognition, language included.
Saying that behaviorism “failed as a paradigm” doesn’t mean we were right to dismiss every one of its claims. Every shift of paradigm usually involves rejecting the problematic claims of that old paradigm, but also rejecting some of the things it got right. Behaviorism had a lot of valuable ideas that we are properly learning to appreciate again after decades of rejecting them because of their association with its over-ambitious claims.
I'm sorry, but since I've recently learned French, the whole time I was reading this I was thinking about this: https://www.instagram.com/reel/C9hlm5Du71u/?igsh=MWV0bmRsM3dya3RhMQ==
English is also full of this kind of thing if learned as a second language. E.g. write/right/rite/riot, not to mention such wonderful pairs as come/cum, sheet/shit, beach/bitch, can't/cunt, fox/fucks.
Agreed, it must be crazy-making.
I'm reminded of Max Planck's line, "Science progresses one funeral at a time."
Kind of in poor taste based on Noam Chomsky's recently disclosed health issues.
Oh, stop with the pearl-clutching! The man is 95. His death has been imminent for years.
Still, I don't think it's in good taste to eagerly root for someone to die because you disagree with his ideas about linguistics.
> These two points are not unrelated. Everett’s core idea is that language should properly be thought of as an invention rather than an innate human capability. You might ask: who invented it? Who shaped it? Lots of people, collaboratively, over a long time. In a word, culture. As Everett notes in the preface, “Language is the handmaiden of culture.”
I feel like the existence of Nicaraguan Sign Language - a natural language constructed by deaf children in a society without any preexisting sign language use - very strongly suggests some intrinsic biological quality to language. That is, rather than it being an invention that propagated the same way as writing or metallurgy, it is a product of human brains on some level which children can reinvent, including features like grammar.
Yup.
I don't understand the relevance of the question of whether it's one or the other. Obviously on some level, there is [information encoding of perceived real world things] occurring in the human (or animal) brain, dependent on the [info encoding - "brain"] capabilities of that creature. Is that an internal language? I don't know, does the categorization matter? I'm a computer programmer and certainly it sounds like a language to me, for whatever that's worth.
This means that you could reverse the process and do something in the real world (sounds, calls, whistles) that mirrors that encoding in your brain - in a way that other brains could interpret in the same, real-world verifiable way, for evolutionarily beneficial reasons like coordination on hunting. This seems to of course be both invented/learned (i.e. I point at the tiger and make the same sound, and you understand that it is a reference to it) and innate (people all around the world seem to understand that the word "squirrel" is more likely to describe a small fluffy thing than "hippopotamos").
But I don't get why it matters. And by matters I mean "in a way that, if we prove it's one or the other, that's a thing that can be objective proven, rather than just vague category definitions, and then informs our understanding of the science such that we can apply it to productive new insights."
To me it's obviously sufficiently both, not because I'm a genius linguist (I am not) but because treating it as a thing that obviously has many influences from our innate biology AND as an invented thing is necessary to (me!) understanding it in a way that we (I?) couldn't if I had to declare it 100% one or the other.
> But I don't get why it matters. And by matters I mean "in a way that, if we prove it's one or the other, that's a thing that can be objective proven, rather than just vague category definitions, and then informs our understanding of the science such that we can apply it to productive new insights."
Understanding where language comes from will probably eventually be useful to genetically modifying other species to be capable of talking. That's the only end use that immediately comes to mind.
If it's just brain power + culture, that implies you can just keep sticking more neurons in a dog's brain, and at 86 billion it'll be able to use a full language. If there's some specialized structures, you'll either need a lot more than 86 billion neurons (so the resultant creature can use nonspecialized bits, like how we solve math problems) or to introduce analogous structures.
I assume linguists have some other use case.
Hi David Brin! :)
But yes, definitely I would agree that "understanding where language comes from" would be important in that context, but I'm not sure why the exact ratio matters.
"If it's just brain power + culture, that implies you can just keep sticking more neurons in a dog's brain, and at 86 billion it'll be able to use a full language"
So replace "dog" with "Chatgpt" and obviously we've got a thing that is pretty language competent, right? I feel like LLMs are the example/thing we're all implicitly referring to?
" If there's some specialized structures, you'll either need a lot more than 86 billion neurons (so the resultant creature can use nonspecialized bits, like how we solve math problems) or to introduce analogous structures."
But here's where the categorization breaks down for me. Obviously for human language there ARE specialized bio/neuro structures that are necessary for actual spoken human language: the parts of your brain that encode information, voice boxes, those parts of the brain the article above talks about, etc... for a LLM, though, do we categorize them as having those things or not? Does the screen and text output sub-module and hard drive structure count? Why does it matter, at least in an important way?
See the other comment of mine on LLMs :) LLMs are the pinnacle of "every time I fire a linguist, productivity goes up", because, well, constraints of natural language learning by a child are quite different from constraints on computers. Look up CHILDES corpus: it is small and unrepresentative (poverty of stimulus argument), if you try to teach LLMs on what children actually hear, rather than on what gazillions on balanced and generally-more-complicated data OpenAI feeds them, you won't get as good results as children routinely get, and you won't get biases toward learnability of existing structures over non-existing. This suggests that LLMs do just brute-force the problem.
Yep, my favorite feature of LLMs have been the discussions they inspire about the extent to which they are doing the same thing as humans.
"This suggests that LLMs do just brute-force the problem."
That gets right at what I'm curious about with recursion-as-a-physical-process. But I guess, is the analogy that informs the conclusion "brute-force" really "kids vs LLMs, two equivalent things" or "kids, who have billions of years of evolution that worked on tools for understanding things, vs LLMs, whose tools are computer submodules designed by hardware/software people for mostly different things." The fact that kids do way better with less training data at a young age, but at "adulthood", ChatGPT seems to, having been trained on petabytes of data in its "childhood", now has tools that let it do better at some things.
Did kids (and therefore humanity and animals over millions of years) just brute-force problems in the distant past, solving them with (now built in) informational/genetic structures (brains, memory, eyes, neurons or.... recursion neural tools?) that they then passed on to their descendents, so that kids today can learn important things with vastly less data? What if the comparison doesn't work because each side isn't at the same starting point, and can't possibly ever be? Is that better described as "they just think 2 different ways" or "they think the same base level way, but they just have millions of years of history and developed tools, so it looks totally different now"
My bet is on "2 different ways", but this is one thing I am _not_ ready to debate, and at some point it falls into semantics (like with pointers-to-pointers: in some sense, it is a "developed tool", right?).
>But I don't get why it matters. And by matters I mean "in a way that, if we prove it's one or the other, that's a thing that can be objective proven, rather than just vague category definitions, and then informs our understanding of the science such that we can apply it to productive new insights."
This has been made moot by LLMs' linguistic success, but, if that _hadn't_ happened, then the question of "Does language competence require some special structure, inborn in humans, or is it purely a cultural artifact, learnable by the same general purpose learning algorithms that we use for everything else?" would have mattered to AI design. AFAIK, we now get to ask the same sort of question, but probably about e.g. executive function, instead.
That makes sense, thanks.
Many Thanks!
I don't see how humans couldn't have evolved language faculties once they had language, but developing a cludgy capacity to be able to communicate with one another, then gradually evolving dedicated capacities to massively upscale it to the point where it's helpful for thinking seems way more likely than developing full-on linguistic thought in one step then using it to communicate (presumably originally just among the kids of whatever super-hominid developed this ability).
Well, chimpanzees also have a kludgy capacity to communicate with one another, so obviously our ancestors did. But the most complicated language encoding we know of in the animal kingdom is, AFAIK, that of prairie dogs, where they can tell the others about the size, speed, color, heading, and species of an intruder (and even if a human has a gun or not!). This is obviously hugely different from the thing we're doing right now in a way that the thing we're doing right now isn't hugely different from chatting around the fireplace.
It's possible this is a slow-and-gentle evolution. But the fact that "behavioral modernity" pretty much appears all at once seems to suggest there was some big sudden shift.
What does “behavioral modernity” mean that appears all at once?
People start burying their dead around the same time that they start painting on cave walls and carving figurines, as well as other stuff. https://en.wikipedia.org/wiki/Behavioral_modernity
There's some debate on if it's just an artifact of how the evidence is recorded, versus actual near-simultaneous arrival, but it reeeeally seems to be near-simultaneous arrival.
If the "I-Language" theory is true, as I understand it.... couldn't the "big sudden shift" be communication ?
If people were developing grammar/syntax/whatever initially for "having complex thoughts" then you'd expect all the hard work on recursiveness to get done on this without communication. It can continue gradually evolving, just as most other human capabilities gradually evolved, and would not neccessarily leave any physical trace. Getting really quite complex, and close to full modern lanuague complexity as its always adaptive to have beeter, more well structured, internal thoughts this can drive itself for hundreds of thousands of years.
Then .... one day.... one human or another discovers he can make sounds to share this "intuitive personnal i-langauge" with someone else, and the "someone else" groks it !!! Between the two of them they are now able to *share* the highly complex thought patterns that have been evolving for generations for the first time.
Because the huge infrastructure of grammar had been built up through standard gradual evolution there'd be no need of a "big bang" genetic evolution, which seems to me unlikely on basic genetic principles. Genetic I-Language could have developed as slowly as the mammal eye.
But then once they can *actually communicate ideas* it can move explosively fast as it is now cultural! You can go from 0 to "burying your dead and painting deers on caves" in a few generations as humans explosively turn their fully intrinsic personally developed I-Language into a commonly used/commonly codified/common "this sound equals this concept"/collaborative E-Language.
Why does "behavioural modernity" under Chomsky I/E laguange distinction have to be related to the genetic development of the basics of the I-Language ? And not the titanic shift that the theory intrinsicly implies is there (the titanic day when for the first time a human communicated by sound a complex i-language thought that another i-language human was able to interpret into his thoughts...inventing the first compound sentence in the first e-language).
It would also seem to be generally supported about what that "cultural modernity" was.... which was an explosion of "humans making symbols" (i.e. cave paintings) or "humans acting symbolicly" (i.e. covering buried dead in ochre/having burial ceremonies/carving statues).
Surely thats the kind of behaviour you'd expect to see increase massively exponentially at the point that internally developed thought language got (at least partially) repurposed into a communication method for spreading thoughts by interspersing symbols <I-Language grammatical recursive thought in 1st human> <-> Symbol shared with another human either verbally or visually <-> <Identical I-Language grammatical recursive thought in 2nd human>.
I/E language implies there would have been the day of the creation of the first syntacticly/grammatically complex E-language.... You'd expect (if the I-Languages were already well developed by then) for this to lead to an explosion of symbols as an already well developed I-Language got rapidly translated into shared common symbols for use and communications.... and what we see with behavioural modernity is an explosion of those symbols.
If that explosion wasn't behavioural modernity.... where is it ? Because at any point I expect the first humans to first use verbal symbols to translate complex recursive I-Language to create the first proper e-language, I'd also expect to see an explosion in other symbols at that point as well as "the new tool" they've discovered culturally explodes across the landscape.
Doesn't it make sense for the genetics of this all to remain the slow, gradual, incremental process we see with all other genetics.... and the big/quick/explosion thing that appears at a pace more usual of cultural evolution to be a *cultural* outgrowth spandrel of that (explaining how it can spread so quickly *once discovered* as its no longer confined to the descendant line but can spread horizontally as well as vertically across hundreds of thousands of people every generation).
Or consider developmental verbal dyspraxia. A mutation in a single gene (FOXP2) causes it. People with this condition seem to exhibit ordinary non-linguistic intelligence, but only speak language at a very basic level. More interestingly, this seems to be because they have to learn it without any innate grammar. For example, if you learn a new word, "phlogon", and then you hear someone use the word "phlogons", you infer that "phlogons" means more than one "phlogon". They need to learn both words separately. (As adults, naturally, people teach them these rules and they can remember them, but they need to learn each grammatical rule consciously, instead of intuiting it the way that everyone else seems to)
Phlogons could be a possessive: Phlogon's. Or Phlogons'. You can't tell without context or seeing it written.
Yeah, NSL is sometimes taken to be evidence of a biological wiring for language, but it’s also in some ways evidence of the importance of social interaction, since it emerged from a bunch of kids being put together in a school, who all independently had their own “home sign” systems (home sign could also be seen as evidence of biological wiring for language at work). I don’t think it’s either/or. But either way, really fascinating case of a language emerging!
Yeah. I think realistically, if language-as-we-think-of-it emerges all at once, you can't say "language is for thinking" OR "language is for communicating." It would have been a single thing that altered how human beings both thought AND communicated, at the same time, with the thinking element probably requiring interaction with others with the same architecture in order to reach its full potential.
You don't think those kids' parents ever pointed at things or gestured to them in any way?
Eta: I agree that the existence of home-signers is remarkable, and probably points to some sort of innate language ability. But these still aren't, like, feral children raised by wolves.
I don't think their home signs included an actual grammar or anything, no. And keep in mind that only the young kids developed full fluency.
It’s probably useful to think of it on analogy with cooking. Humans started cooking before we adapted to the presence of cooked food, and after those adaptations cooking is now essential for almost all foods that we eat. But there’s not one single adaptation that explains this.
Agreed. Just as, after the initial invention of cooking, our digestion and cuisine co-evolved to become more and more dependent on cooking, and to more thoroughly exploit the advantages of that capability, if a linguistic "MERGE" capability resulted from a single mutation, I would expect that it would have e.g. increased the advantage that a larger vocabulary conferred, and, more generally, increased the frequencies of genes that provided complementary linguistic capabilities. So, in the end, many genes may be involved in our current linguistic capabilities, above what were present pre-"MERGE".
1.
>Regardless of the details, a generic prediction should be that there is no longest sentence in a language whose grammar is recursive
Why then:
>He does explicitly claim (in the aforementioned paper and elsewhere) that Pirahã probably has no longest sentence, which is about the most generic anti-recursion statement one can make
Wouldn't a non recursive language *have* a longest sentence?
2.
>I find this notion of recursion completely insane, because it would imply (i) any language with more than one word in its sentences has recursion
From the quotes, it sounds like the linguisists in question emphasize the role of recursion as an iterative - not isolated phenomenon. A sentence with two words might be a primitive instance of recursion, but lack the iterative flexibility that characterizes it.
3.
>But this is kind of a dumb rhetorical move. Either the theory makes predictions about real languages or it doesn’t
This objection doesn't seem fully compelling. All models are wrong, but some are useful. The mere existence of a supposed language doesn't prove that the model isn't generally useful. Did not the vast majority of langauges discovered since the advent of the theory fit the model?
4.
>Is “chicken” the subject of the sentence, the object of the sentence, or something else? Well, it depends on the context
This needs some more elaboration as to how it demonstrates a clear lack of noun/verb distinction. After all, similar ambiguous phrases could be constructed in other languages. Is the point that these are typical, rather than exceptional for the language?
5.
>He mostly seems to take issue with the idea that some region of our brain is specialized for language
Is that actually necessary for Chomsky's model? That is, does a sudden take off of language due to a novel mutation facilitating universal grammar necessitate a single specialized part of the brain?
6.
>ChatGPT...the brute empirical fact remains that it can handle language comprehension and generation pretty well. And this is despite the conception of language underlying it—language use as a statistical learning problem, with no sentence diagrams or grammatical transformations in sight—being somewhat antithetical to the Chomskyan worldview.
While Chomsky may personally have been skeptical of the success of LLMs, I'm not sure why they pose such a problem to his general model of language. After all, his theory, as presented, seems to be about the human faculty of language, and how evolves it. It seemingly doesn't preclude the possibility that a machine could mimic such its results using a completely different framework; after all, there are many tasks which are performed using quite differnet systems by humans and machines.
>It’s conceivable that a completely orthogonal system designed according to the principles of universal grammar could outperform LLMs built according to the current paradigm...
Again, why is this central to the Chomskyan model? That humans evolved the faculty of language as a particular mutation with particular characteristics doesn't preclude a different or even better system being developed.
>Chomsky’s claim that engineering success does not necessarily produce scientific insight is not uncommon, but a large literature speaks against it.
Again, how does "large literature" speak against it? Presumably the reviewer doesn't merely mean that many disagree with Chomsky, as that's trivial, but rather that the large literature about engineering sucesses is an existence proof against Chomsky. But is it? Following the review's presentation of Chomsky, would not Chomsky refer to such literature as the cataloging of butterflies bodies? Does it ultimately provide a parsimonious model of the concept of language, or insight into its evolution?
Regarding:
> ... has no longest sentence ...
I think this was just a typo, honestly. (Two other threads on this comment page think the same.)
>Again, how does "large literature" speak against it?
The example that I've seen cited is that thermodynamics owed more to heat engines than the other way around. I would guess that this is probably true of mechanical engineering preceding the science of classical mechanics as well. I'd guess that electromagnetism is probably the other way around: There was a lot of scientific investigation before the first motor.
Metallurgy is probably the most extreme version of this. Smiths knew how to create metals with very specific properties with precision without understanding anything about the underlying chemistry or physics. The actual understanding of the chemistry of how to make steel and bronze and, why the specific heatings, coolings, hammerings and ratios of materials used were necessary, came far later.
Great point! Many Thanks!
This was a well written overview of a fascinating topic, but I don't think it's a very good book review. Like so many of these reviews, the author uses it primarily as an opportunity to review a topic they are interested in, rather than a book about that topic. Obviously, one needs to have some background knowledge of the topic, but when the the actual book under review only takes up one out of five sections, it's not really about the book or the ideas it discusses.
This is neatly illustrated by this sentence: "There’s a lot of interesting stuff in there I haven’t talked about..." But isn't talking about the interesting stuff the main point of a book review?
Still, I enjoyed reading this a lot. Thanks!
Yep, plenty of "book reviews" are just guest posts in disguise. Since those aren't allowed in the pure form, this kind of thing is the result.
Exactly my thoughts, great article, not a great book review. Also think some editing on subject ordering would have been good (define recursion when it was mentioned, maybe move the chomsky stuff a little later)
I’m afraid Chomsky wins this one because of the work in computation theory that he did which the reviewer doesn’t reference at all. Chomsky’s biggest insight was what is now known as the Chomsky Hierarchy of languages or grammars:
Regular Languages, recognized by finite-state machines (animals have these)
Context-Free Languages, recognized by machines with a simple memory element such as a pushdown stack which allows objects to be merged or linked to each other (requires a brain modification; perhaps humpback whales and other animals can handle this)
Context-Sensitive Languages, recognized by linearly bounded automata that can mark or operate on their input data (allows language to interact with the external world so that we can build things following instructions)
Recursive (in the Godelian sense) languages, recognized by Turing machines and other automata which can use expanding amounts of rewritable memory to keep track of things, attained by humans with the invention of writing (pre-literate humans were not Turing Complete and the Church-Turing thesis is not actually about what humans can compute but about what technology-aided humans can compute).
Pirahã people (or at least the young children in it) have the capability to learn other human languages and the same brain improvements that the rest of us have. Their language is restricted in interesting ways but the only really important question about it, at the very top level, is whether they are capable of doing things that people in other societies can do, or whether their language makes this impossible.
My guess is that, when a phonetic alphabet is eventually developed for them, they will indeed become capable of reading and writing their own language, and doing the other things that writing enables.
A really good first test case of this: have they been taught to count yet, and imported or created number words that allow them to communicate the cardinality of a large collection of objects reliably so that they would know if something had been added or taken away?
Everett’s work may show that the language of the Pirahã is missing certain structures that other languages have; but I don’t think it falls outside Chomsky’s framework. If there is a concept which Pirahã cannot understand no matter how you attempt to translate it into their language, for example if they can’t be taught to count by giving them a few number words and showing how it works, that would be a much bigger deal.
Thank you so much for listing those 4 categories, very helpful.
Honest, a-political question for my own understanding, coming from a computer science background: how is that categorization not just a measure of "the amount of working memory available to the user of the language"? At least, in the context of what is actually going in the animal's brain. I 100% get how it's useful to us humans, seeking to categorize it by different features.
Regular languages (FSMs) words as you describe, don't necessarily refer to base primitives. The animal makes a call that says "there is a predator nearby". It would not be a language, or intelligble to the other animals, if they didn't understand the concept of predator or subconcepts like "predators are threats to our ability to pass on our genes." Sure, maybe it's not "memory" but "instincts" but (from a CS perspective it's still "info we already had")
"Context-Free Languages" add "memory" but it seems (naively, to me) like it's just more complex "memory."
For the "Context-Sensitive Languages" and "Recursive languages" - again, it just seems to me like the base difference is "even more memory, better configured or more complex perhaps." What is the actual physical brain difference between processing the reference to "predator" in the animal call "there's a predator nearby" and the highest level of language "the symbology of X in War and Peace is evocative of the concepts of Y in the bible," except in the complexity of the concepts being referenced? Does the brain actually do anything different when it references "the bible" than "predator", except in scale/complexity?
Working memory increase was certainly a facilitator (this happens to be Markov the biologist's favorite claim on the matter). However, "better configured" already betrays that quantity is not yet quality. You can have an enormous working memory but if you cannot use it to make, as you put it in the other thread, pointers to pointers, it's moot.
(Also, people whose Daneman-Carpenter test returns 1 or 2 for working memory can still use language pretty well - although they do have more problems with complex sentences, if you simplify the number of elements used enough, you can still get some recursive hierarchies.)
Ok, makes sense.
"However, "better configured" already betrays that quantity is not yet quality. "
How do we distinguish though? When I typed that, I recognized that it would imply a quantity/quality issue, but for...
"people whose Daneman-Carpenter test returns 1 or 2 for working memory can still use language pretty well"
... how do we know it's not just straight memory capacity? Wouldn't that be exactly what we see when you say:
"although they do have more problems with complex sentences, if you simplify the number of elements used enough, you can still get some recursive hierarchies"
Wouldn't that be consistent with "it's working memory" as the relevant concept? Let's say you've got 100 bits, where you can get "the sky is blue" because sky takes up 20 bits and blue takes up 40 bits, and 60 is less than 100. But for "I remember when last week, Steve said that Alice thought the sky was green" doesn't work for you, because sky is 20, green is 40, last week is 30, Steve and Alice each take 50, and the concept of Alice thinking herself (recursive) about the sky being green is another 110 bits (Alice + sky + green). So it's hard for you to understand?
I mean on some level, the brain kinda has to work that way right? I could see if the "quality" was an LLM-like capability of linking/associating bits together to differing degrees (green is 5 distance from blue, but Alice is 100 distance from Steve) but that.... is surely physically just the capability of determining (perhaps in the brain with true physical distance) the difference between different groups of bits. You will need to be able to handle more of them in order to do so. Is it really "recursive" as a fundamentally different concept, or just that recursion (mathematically, like multiplication) requires much more working memory, specifically to keep the relationships between the different levels of recursion accurately intact?
To bring back up to a higher level - is higher intelligence empirically correlated with better memory? Memory not only of specific facts but of the interplay/relationships between those facts, and the sub-facts that are implied by them?
Sorry for rambling, this is just fascinating to me, and thank you for taking the time to explain to a layman.
Your Steve-Alice example is exactly what we'd expect if memory were the only relevant issue. However, a person who can't reliably hold a sequence of "Steve, Alice, apple" in their memory in Daneman-Carpenter test is still able to hold "Steve thinks that Alice ate an apple", even though without recursive Merge, one would expect that the latter takes more memory, not less. (While in reality, it seems that it is composed as "[Steve [thinks [that [Alice [ate [an apple]]]]]]", with only two elements combined at each moment.) This suggests some privileged mechanism for exploiting "pointers to pointers" to ultimately use less memory.
Okay, bingo, thanks - this gets to the heart of it for me:
"This suggests some privileged mechanism for exploiting "pointers to pointers" to ultimately use less memory."
The "ultimately less memory" helps a lot.
The Chomsky hierarchy is an important mathematical result, but I don’t see how this tells us about human language. We don’t actually have as good evidence as we think that human minds are Turing complete or even at the level of context free grammar. We know for sure that a finite neural net like an LLM *cannot* reproduce a context-free grammar (that is one of Chomsky’s important results!) and yet it seems to be able to imitate one just as well as humans do. This is strongly suggestive to me that the human mind (which we know is also implemented by a finite collection of neurons, though we don’t know if the architecture is really a neural net) is actually a finite state machine that has somehow managed to implement within itself a rough approximation of a large initial fragment of context-free grammars and Turing machines. If so, then it’s quite possible that in whatever sense a human mind is able to implicitly define a full Turing machine, a neural net may be able to too, despite its in-principle limitations.
You didn’t read the part where I said that unaided human minds are NOT Turing-complete. They are only Turing-complete when combined with the TECHNOLOGY of writing.
Whether LLMs and neural nets can eventually do the equivalent is an interesting question—quantity compensates for quality so their limitations may not be easily apparent. But computers can obviously be Turing complete and I expect that the next generation of AIs will have more sophisticated architecture that can emulate human cognition more directly and require less brute force than now.
> We know for sure that a finite neural net like an LLM *cannot* reproduce a context-free grammar
I don't know much about linguistics, but modern LLMs are basically Turing complete. At least in the sense that a normal home computer, or a human with a finite amount of memory and paper to write on are basically Turing complete.
> To be clear, this usage of ‘recursion’ seems consistent with how many other Chomskyan linguists have used the term. And with all due respect to these researchers, I find this notion of recursion completely insane, because it would imply (i) any language with more than one word in its sentences has recursion, and that (ii) all sentences are necessarily constructed recursively.
No. The claim is actually somewhat more technical. Dolphins have the equivalent of sentences, but they are linear, not recursively hierarchical. An illustration: in English, "I saw a girl with a telescope" has two readings, "I [saw [a girl with a telescope]]" and "I [[saw a girl] with a telescope]" (i.e. using a telescope). Such structural ambiguity is not possible in dolphinspeak, where linear position _is_ hierarchical position but is possible in human languages.
More generally, Everett is exactly as much of a buffoon as Nevins claims he is, and "common sense" is a poor guide to linguistics.
This is an honest, a-political question, coming from a computer programming background: what is the point of the term "recursion" in the context of "lowest level description of language?" I know how *we* use it, and the difference in programming situations that are and are not recursive, and many of the ways it is used in this discussion seems similar, but at the lowest level of understanding language and how it relates to the brain, my understanding of the similarity breaks down: what is the point? Isn't everything expressed in language (and I mean this as "everything" like down to the atoms) recursive? When I say "my friend Bob" my brain (and the people I'm talking to) understands that I "actually" mean "that sack of water and carbon over there, made up of certain atoms and molecules, and electrical neural structures that contain the memories of our relationship" - and then all the various recursions down to "water is H20" and so on.
Even a single word: "telescope" must be "recursive" on a number of different mental concepts and the way they're encoded in the brain of the speaker/listener. Certainly a more complex sentence like above: "I thought you said you saw a telescope" encodes far more complex recursion across multiple things, but it's all just hierarchical: "this refers to this [ which refers to this, which refers to this...] and that [which refers to this, which refers to this...]". This is reflected (haha) in programming where recursion is "really" just "go to this part of the code and run it" where the code and the parameters and the variables can be self-referenced and iteratively changed - it is described as "recursive" for human understanding. The computer doesn't care.
So I guess I'm asking: in the context of understanding language at its lowest level, what does "recursive as a concept" get us? That would not be just as easily described as "a thing that refers to another thing, which naturally could refer to another thing (or things), and so on down to whatever the lowest level of info encoding."? Yes, I can totally see why "referencing a thing that itself references a ton of other things, especially things that reference ever more things, including itself" is an advanced concept that you might need more brain power (or GPU cycles) to understand, but I don't get why it's *different*.
That's an interesting perspective, but, using computer science jargon, I think you mix the ability for having pointers with the ability for a function (namely, Merge) to call itself. You can (and probably do) use the former for the latter, but it doesn't mean they are one and the same, and in practice, not every system with pointers is a system with recursion (although the fairly unique ability of humans for symbolic pointers was probably a prerequisite).
And more generally, "the lowest level" is neurons firing electrons and ions, which is no more useful than "silicon pieces firing electrons" for computer, we don't actually describe what computer does that way. (I am reading "Gödel, Escher, Bach" now, and there's a related discussion.)
Right, but in computer science, at a very low level, recursion IS just clever, self-referential (critically: self-referential as perceived by humans) pointers.
" not every system with pointers is a system with recursion"
Right, recursion (as a CS principle for humans) is pointers pointing at themselves... but it's still pointers (for computers). The abstraction of how recursion works on the code and variables it is affecting is important for humans to understand - not the computer. The computer still sees it as a one-way stream of 1s and 0s - though when you add recursion to your language, you often do need to add a new layer of debugging in order to accurately represent the concept of recursion so the human can understand it.
"And more generally, "the lowest level" is neurons firing electrons and ions, which is no more useful than "silicon pieces firing electrons" for computer, we don't actually describe what computer does that way."
Well, sure, when we are.... talking about the computer's behavior at a high level! High level = meaning dependent on embedded (recursive?) concepts. But I'm interested in understanding how that lowest level can be built up into that highest level, and which parts are "base", and which parts are just other higher order (potentially recursive) structures. That lowest level for computers is "is this bit a 1 or 0". For LLMs, a key additional lowest level is added - "what is the distance between the concept this 1 represents, to the concept that 0 represents (represented, of course, as additional 1s and 0s)" - which I think is probably something that the human brain relies on as well. That basic structure will let you create brains and language - and, I think, amoebas and crystals and the universe.
Sure, if you know how to build - aka if you have a blueprint - aka if you have a higher-level structure. I do not think the lowest level is informative on itself, for either human language, other aspects of human cognition, or computers. At the lowest lowest level, we're all just a bunch of interacting quarks. We build models of higher levels because they actually reflect useful generalizations. Your mileage of what's a useful generalization may vary, but just like I speak of a processor as an abstract machine that performs calculations not as of bunch of quarks organized into a bunch of atoms and electrons interacting in such a way that they happen to correspond to said calculations, I speak of a function calling itself not of a line of 1s and 0s that encodes a piece of memory that... you get the idea.
"A map is not the territory, but you cannot put the territory into your pants' pocket".
"Sure, if you know how to build - aka if you have a blueprint - aka if you have a higher-level structure. "
But at some point, we didn't. Isn't that an important distinction? The things that existed right before amoebas, or crystals or brains or languages did not have those blueprints. They created (evolved) those higher level structures emergently, based only on [whatever the lowest level is] and [whatever previous higher level structures they had built on top of it].
"We build models of higher levels because they actually reflect useful generalizations."
Totally agree, where "we" is humans who need to understand it.
"Your mileage of what's a useful generalization may vary, but just like I speak of a processor as an abstract machine that performs calculations not as of bunch of quarks organized into a bunch of atoms and electrons interacting in such a way that they happen to correspond to said calculations, I speak of a function calling itself not of a line of 1s and 0s that encodes a piece of memory that... you get the idea."
Again, totally agree in the context of "letting humans understand what's going on." But in the context of the debate around "recursion - do they [a language, the piraha, animals, etc...] have it?" it seems like it might matter a lot if recursion is just a concept that humans created to describe the salient (to us) aspects of "clever pointers in our brain", and the slow build of complexity around languages was "ever more clever pointers" rather than "the adoption of this importantly new break-from-the-past category of thinking."
Evolution throws things around almost at random and looks what "sticks" in terms of helping/not harming genetic fitness. But how gradual evolution is is very different for different things. Getting a pointer to a pointer is "importantly new break-from-the-past category of thinking", for the same reason RNA+DNA system is an importantly new break-from-the-past category of replicator (over RNA world).
Why is it that so many critics of Everett and defenders of Chomsky that I see don't simply critique Everett, but call him names and say nasty things? You do that here when you call him a "buffoon." Whatever the merits of either position, I find it really unappealing and off-putting and it makes me not like people who defend Chomsky. At least from the outside, it gives me the vibe that people who defend Chomsky are dogmatic and insecure, that raises a big red flag for me and probably for others that make us more sympathetic to Everett than we might otherwise be.
Another thread here goes into detail on this. In short, it's because Everett (unlike, say, Kenneth Hale with the non-configurational suggestion or Mary Dalrymple with Lexical-Functional Grammar) "started it first", in a sense, by playing the big, Fomenko-style iconoclast who has seen past Chomskyans' delusion. (All the more surprising given that in personal - well, Facebook - communication he doesn't seem an unpleasant person you'd expect from this characterization.) Think biologists that have to respond to creationists. Could one, in principle, be a bigger person and respond the way you respond to normal scientific disagreement? Sure. Is it the least bit surprising that this isn't what actually happens? No.
What in particular did he do that you see as "starting it"? You say "playing the big, Fomenko-style iconoclast." What does that involve?
Oh, let's see. From the lasst pages of Everett 2005:
> 2. Linguistic fieldwork should be carried out in a cultural community of speakers, because only by studying the culture and the grammar together can the linguist (or ethnologist) understand either.
Read: "Every one of you who hasn't done what I did, living with the group for many years, is a charlatan who doesn't understand the language they study".
> 3. Studies that merely look for constructions to interact with a particular thesis by looking in an unsophisticated way at data from a variety of grammars are fundamentally untrustworthy because they are too far removed from the original situation.
Once again but with passion: "You, yes you who wrote your thesis on binding violations or person-case constraint or whatever, are fundamentally untrustworthy".
How did the review put it? "This is the closest you come to insulting someone in a published paper"?
A bit later:
> This beautiful language and culture, fundamentally different from anything the Western world has produced, have much to teach us about linguistic theory, about culture, about human nature, about living for each day and letting the future take care of itself, about personal fortitude, toughness, love, and many other values too numerous to mention here.
...What is this? This is not how scientific (or linguistic, or historical) articles are written. This is sometimes how anthropologists write, but this is a condemnation for them, too.
> The need is more urgent than ever for field researchers to document these languages and for more individuals and foundations to follow the lead of the Hans Rausing Endangered Languages Document Project and donate to support research on them.
Oh, and call to action. "Here's how cool my sponsors are, gimme money money money". This isn't a grant proposal.
(And there's also a strong inconsistency where Piraha speakers are described as monolingual in some parts but as bilingual in other.)
Thanks for providing these quotes. Previously, I had said this: “What in particular did he do that you see as "starting it"? You say "playing the big, Fomenko-style iconoclast." What does that involve?”
I do not believe any of the quotes you provide present significant evidence of what I asked for. Some of the remarks may be subject to the criticisms you present, but they don’t seem to me to make a case for the claim you initially made. I’ll go through each and comment on why that is.
Quote #1: “> 2. Linguistic fieldwork should be carried out in a cultural community of speakers, because only by studying the culture and the grammar together can the linguist (or ethnologist) understand either.”
You respond to this by interpreting it as “Read: "Every one of you who hasn't done what I did, living with the group for many years, is a charlatan who doesn't understand the language they study".
Where are you getting this stuff about being a “charlatan” from? If one’s view is that language is integrally tied to culture in inextricable ways, then it’s simply true that one cannot fully understand a language without understanding culture. Nothing about making such a claim requires you to believe that anyone who doesn’t do this is a “charlatan.” All it would entail is that those who don’t do so won’t have a complete understanding of the language. That trivially follows from Everett’s perspective on language. Suggesting that those who take a different perspective on language won’t be able to have a complete understanding of it doesn’t mean that those people are “charlatans.”
I asked for remarks from Everett that support iconoclasm. What I see is a remark that contains no insults or critiques of those who hold views contrary to Chomsky’s, followed by you adding commentary proposes I interpret Everett as implying this. If what Everett says here is correct, then it may very well be that those who hold contrary views about the importance of culture are mistaken. This doesn’t mean they’re charlatans. It just means they’re wrong. Incidentally, Chomsky has explicitly called Everett a charlatan.
Quote #2: > 3. Studies that merely look for constructions to interact with a particular thesis by looking in an unsophisticated way at data from a variety of grammars are fundamentally untrustworthy because they are too far removed from the original situation.
This looks like a claim alluding to the importance of understanding language as it is used in social contexts and raising doubts about approaches to language that don’t do this. It reminds me a bit of remarks I’ve seen about usage-based models of language. Nothing about this strikes me as evidence that the person who said it is objectionably striving for fame as an iconoclast.
You interpret it this way: “Once again but with passion: "You, yes you who wrote your thesis on binding violations or person-case constraint or whatever, are fundamentally untrustworthy".”
Much of my own research focuses on questioning the methods and validity of research in my field. I frequently argue that a study’s methods are invalid, and I may say things like “We should not trust the results of this study.” This in no way indicates that the authors of the study aren’t “trustworthy.” Calling people untrustworthy implies those people are deceptive or bad people. Saying the results of their studies are “untrustworthy” means something quite different: it means we should have doubts about whether the findings of those studies are true. It has absolutely nothing to do with the honesty or character of the authors of those studies.
What I see in this remark is a comment about the trustworthiness of studies, not the authors of those studies. Given this, not only does this not strike me as evidence that Everett is seeking to be an iconoclast or acting like a jerk, it strikes me as evidence of what I suspected: that you and other critics of Everett are projecting, imagining hostility and iconoclasm where it isn’t present. Everett’s remark strikes me as completely innocuous and in no way an indication that he’s seeking fame as a paradigm-shifter.
Quote #3: > This beautiful language and culture, fundamentally different from anything the Western world has produced, have much to teach us about linguistic theory, about culture, about human nature, about living for each day and letting the future take care of itself, about personal fortitude, toughness, love, and many other values too numerous to mention here.
You say of this, “...What is this? This is not how scientific (or linguistic, or historical) articles are written. This is sometimes how anthropologists write, but this is a condemnation for them, too.”
Even if a remark like this is inappropriate for a scientific article, well, okay. Then it’s inappropriate for such an article. But that’s not very good evidence of “starting it” or seeking iconoclasm. It also doesn’t strike me as a good reason to insult Everett or treat him with contempt or hostility.
Quote #4: > The need is more urgent than ever for field researchers to document these languages and for more individuals and foundations to follow the lead of the Hans Rausing Endangered Languages Document Project and donate to support research on them.
You interpret this as “Oh, and call to action. ‘Here's how cool my sponsors are, gimme money money money’. This isn't a grant proposal.”
Your interpretation again strikes me as dubious and more than a little cynical. It may be a bit weird to suggest that we should donate money for a cause in an article like this, but again, this does not strike me as evidence of someone “starting” anything or seeking fame via iconoclasm. This looks like someone who thinks endangered languages are important sources of insight and that believes we should donate money to preserving those languages and cultures. Incidentally, I completely agree with Everett about this.
From looking at what Everett says in this article, it looks like he writes in a way that doesn’t conform to conventional norms about scientific writing in certain fields and perhaps more closely resembles the commentary one might observe in others. He makes personal remarks about his feelings towards cultures, he makes practical remarks about how we should donate to causes, and so on. To me, it reads as someone who doesn’t care much about academic conventions and who is either intentionally defying them or isn’t aware of them. Nothing about that strikes me as someone striving for fame via iconoclasm. It looks to me like someone who has little patience for those norms. I see pros and cons in that. I think scientists should strive for objectivity and professional detachment in their work, but I also think if we lean too heavily in that direction we can lose sight of insights that are obtained via personal experience and the gestalt of one’s impression about whatever it is you’re studying (societies, languages, ecologies, etc.). What I see here is someone whose work is very human and passionate. I do not see a raging jerk starting fights with Chomskyeans or calling them charlatans.
You say this: “(And there's also a strong inconsistency where Piraha speakers are described as monolingual in some parts but as bilingual in other.)”
Inconsistencies are not evidence that a person is starting anything or being some kind of iconoclast.
What I’d have expected is explicit denunciations of a Chomskyan approach to language, calling people names, insults, fist-pumping revolutionary declarations that his approach to language is the only way and that previous approaches are stupid and terrible, or something along those lines. Instead, what you’ve given me are either (a) innocuous remarks that you’ve interpreted in a highly cynical way, but for which I see no reason to share that interpretation (b) remarks that may be subject to legitimate criticisms but aren’t relevant to what I was asking for, e.g. “this is unscientific” or “this is inconsistent” is not evidence that Everett “started it” or that he’s seeking fame through iconoclasm. If such an accusation were true, I’d have expected better evidence. That you haven’t provided what strikes me as a single piece of even slightly convincing evidence actually serves as evidence to the contrary: it suggests to me that your depiction of Everett is less likely to be true. If it were, I’d have expected better evidence than this.
> If one’s view is that language is integrally tied to culture in inextricable ways, then it’s simply true that one cannot fully understand a language without understanding culture. Nothing about making such a claim requires you to believe that anyone who doesn’t do this is a “charlatan.” All it would entail is that those who don’t do so won’t have a complete understanding of the language. If one’s view is that language is integrally tied to culture in inextricable ways, then it’s simply true that one cannot fully understand a language without understanding culture. Nothing about making such a claim requires you to believe that anyone who doesn’t do this is a “charlatan.” All it would entail is that those who don’t do so won’t have a complete understanding of the language.
This paragraph is ripe with inconsistencies. Yes, the iconoclasm is in some sense a trivial consequence of Everett's views. It doesn't make them any less iconoclastic. If your view is "the only way to understand X is to do Y", then you automatically exclude from sciencehood those who don't do Y, which, in this case, is the supermajority of the field (including supermajority of linguists who do fieldwork, because this particular way isn't how linguistic fieldwork is normally done). And you also try to sneak in "complete", which wasn't in the quote. He doesn't say "can ... completely understand either". He says that if you don't do this, you don't understand the phenomenon, period. YES THIS IS ICONOCLASM. Same iconoclasm that Chomsky himself applied to behaviorists half a century before Everett. This is a call for full reestablishing of a scientific paradigm, with (implied but inescapable) condemnation of those who won't jump to the new paradigm. Compare this to, say, an introduction to LFG (which is, in its own way, no less revolutionary inside):
> Towards the end of the twentieth century, new formal ideas began to achieve prominence in linguistic theory, making use of parallel rather than serial structures and computations, and comparative evaluation of multiple overt structures rather than serial derivation of individual overt structures. These ideas are embodied in a family of nonderivational, constraint-based linguistic theories and in approaches based on optimization (both statistical and discrete). These newer theories are compatible with different linguistic epistemologies drawing on structuralist and functional/typological ideas which have both predated and coexisted with generative grammar. One such theory is lexical-functional grammar (LFG) (Kaplan and Bresnan 1982), which is the topic of this book.
Also, again, Everett 2005 is a published paper, _explicit_ insults wouldn't have been allowed even if he were in the mood for them, which he probably wouldn't be because of his kind demeanor. But kind is not right, and good is not nice.
As an outsider looking in, 3 of those 4 quotes strike me as obvious potential methodological problems in how linguistics research is performed. The odd one out is the sappy quote about fortitude and beautiful cultures, which I think is just fluff (although I disagree with your evaluation that at least to me seems to leave open the normative interpretation that science papers should be dry and boring). Your examples don't seem to me (a random person) to be particularly felicitous.
I was not aware we had that level of knowledge or confidence regarding the structure of dolphin communication; could you please point me in the correct direction to learn more?
Well, level of confidence varies, of course, but the researcher to cite would be James T. Fulton.
Hey, Ive read a lot of your comments on this article and wanted to collect my thoughts into a single response.
First, I think a lot of the clashes that you have here are due to different priorities. While I dont know too much about linguistics, Ive seen those grammar tree diagramms, and they seem like a reasonable thing for linguists to do, and like something that could produce real insight about "grammar". But Chomskyanism also includes a lot of general claims, and often the defense of these looks like a continuous retreat from falsifiability. For someone not interested in getting deeper into linguistics, these general claims with potential takeaways for elsewhere will be the main thing they care about, and they will give a lot more weight to arguments against them from other fields.
I think your examples of Everett "starting it" by being an iconoclast are pretty weak. The idea that there could be more relevant evidence than a papers worth of analysis examples, and that the chomskyian "reanalysis from a distance" only makes sense in terms of fitting them into their scheme and not in terms of what makes sense for studying that language - seems like the sort of thing that is generally accepted without outrage in other sciences, and I suspect it would have gone over much better in this case if Everett wasnt the sole authority on Piraha.
Why do you think homonymies and synonymies are evidence that language is for thinking more than communication? I agree that these are suboptimal for communication, but arent they even worse for thinking? If you have two concepts that are homonymous in thought, you just... wont be able to distinguish them, period. And if you have a synonymie, you can believe things about one but not the other, without realising that anything is wrong. I think both of those happen, and they are way worse than having to figure things out from context.
I dont agree that thinking in language is more potent than in images. You can understand interactions of gears with visual imagery that would take pages to explain in words, and noone could understand those without constructing their own images. Propositional thinking is applicable in cases where visual is not, but those are generally more removed from the evolutionary environment.
If modern human cognition developed in a single mutation, it can also disappear in a single mutation. Why is it so ridiculous as to dismiss out of hand that there is one isolated tribe somewhere in the world where this back-mutation reached fixation?
Thanks for your comment!
> Hey, Ive read a lot of your comments on this article
Yeah, I think I'm getting to a hundred by now? I should wrap it up at some point.
> First, I think a lot of the clashes that you have here are due to different priorities.
Absolutely.
> often the defense of these looks like a continuous retreat from falsifiability
There's actually an interesting blogpost on the topic: https://milway.ca/2021/07/13/what-does-falsification-look-like-anyway/
> I suspect it would have gone over much better in this case if Everett wasnt the sole authority on Piraha.
Very much so. (My bet is that we wouldn't be having this _discussion_ if he weren't - someone else would just come out and show their own work.)
> I agree that these are suboptimal for communication, but arent they even worse for thinking?
The point is that something which is pronounced homonymously is _not_ homonymous for the speaker's thought. You need to specifically _point out_ to people that, say, witch and which or "I [saw a girl] with a telescope" and "I saw [a girl with a telescope]" are homonymous, exactly because they're represented differently in the speaker's head and only get conflated when pronounced. Of course, true homonymy in thinking would suck.
> I dont agree that thinking in language is more potent than in images.
It has a specific capability of chaining which the image thinking does not (as the example of house image not being composed of wall image and windows images suggests). Of course there are other aspects in which imagery is more "potent", e.g., more expressive. One more case of a vague word :)
> Propositional thinking is applicable in cases where visual is not, but those are generally more removed from the evolutionary environment.
It is not as obvious as it seems at first glance. A very important part of any ape's evolutionary environment is social environment, dominance and prestige, all that jazz. And it would seem that thinking through coalitions and stuff like that is very much such a case.
> Why is it so ridiculous as to dismiss out of hand that there is one isolated tribe somewhere in the world where this back-mutation reached fixation?
Their children wouldn't have been able to acquire Portuguese if that happened, which they do.
Thanks for your reply.
>There's actually an interesting blogpost on the topic
The orbit of mercury is evidence against Newton and should reduce our confidence in him, increasingly so as we keep failing to find Vulcan, even absent a competing theory, and it doesnt count fully as evidence for Einstein until were sure there is no Vulcan.
Even if we took the argument at face value, I dont think this works as an excuse, because it doesnt explain why the issue would appear in linguistics more so than other sciences. In fact, the arguments apply to linguistics less than to physics, and theres not a general impression that physicist are slippery.
Generally, when you resort to philosophical arguments to defend a particular scientific theory, its because youre losing on the scientifc front. And even the philosophy here is pretty whack - the "theory first" post he linked implies apriori knowledge of medical facts, and he propably doesnt realise.
>Of course, true homonymy in thinking would suck.
Well, I think there is homonymy in thinking. I think the distinctions between e.g. precision and accuracy, or between different quanitifier orderings, are not ones that people can naturally use in their thinking. They need to be tought or discovered through great effort. Of course there are additional homonymies in speech, but that doesnt show thought is favoured - it just means that some popped up in every place they theoretically could. Obviously, the ones in thought are also reflected in speech.
>And it would seem that thinking through coalitions and stuff like that is very much such a case.
I think just playing through particular scenarios in your head can do a lot of lifting here, even without propositional thinking. Of course, once we start ordering society with rules that are themselves linguistic, and complex enough that you cant know the applications relevant to you by heart, thats when it becomes required. But the average isolated hunter-gather tribe we find today is not there. I dont think it kicks in until "memorising the rigveda" levels of cultural complexity.
>Their children wouldn't have been able to acquire Portuguese if that happened, which they do.
I hadnt heard. Do you know if they can count past seven now?
There absolutely is impression that physicists are slippery, especially when they work on string theory on somesuch.
As for the argument about things like precision and accuracy and similar — that is a reasonable point, but... There's a difference between homonymy (two concepts with the same reflection, in each specific case only one is meant) and vagueness (a reflection that fails to specify details that could distinguish two related concepts). Vagueness is inevitable, homonymy is not. It doesn't invalidate your point but mitigates it. (Example of vagueness in speech — Russian и fails to specify difference between BIT vowel and BEET vowel, so the actual pronunciations are across the whole continuum.)
String theory, maybe, but thats ~noone says thats settled science. Even *in his motivating example* people didnt think the physicists were slippery.
What detail distinguishes between precision and accuracy, such that they are otherwise the same? I agree that they are related, but... if two concepts can fool your mind into thinking theyre one, why wouldnt they still seem related after? Are they actually "more" related than the pan and the stove? Plus, the sitting-bank and the money-bank are also related, as shown by the etymology, but are usually an uncontroversial example of homonymy.
Most science isn't settled. Certainly most linguistics isn't settled.
Etymological links can be lost — or non-existent, if the merger is phonological, as in лук 'onion' and лук 'bow (weapon)' — without affecting.
To a layperson, neither precision nor accuracy refers to its statistical definition, both vaguely mean "how close the thing is to the truth/target/...".
As for Piraha children counting past seven once learning Portuguese, I don't technically know, but no one — certainly not Everett, who claims that the differences are cultural and teachable — suggested they can't.
I remember him saying that he tried a lot and failed to teach counting. Maybe he thinks young children are different? But that would be really weird to believe.
It wouldn't be weird to believe, if you think culture is the root cause (as Everett does), that young children haven't yet imbibed the culture and are indeed different.
In addition to being stubbornly wrong on linguistics, Chomsky is also known for being stubbornly wrong on politics, insisting that communism is the best way of organising human affairs. Infamously, Chomsky took a trip to Venezuela to congratulate Hugo Chávez on building a society for the future, a few years before its dramatic economic collapse.
He also never admitted he was wrong - a diehard “true communism has never been tried,” he doubled down on the Venezuela situation, saying “Capitalists were free to undermine the economy in all sorts of ways, like massive export of capital.”
(As per Chomsky, if you have enough resources to flee a failing state, the fact that it’s failing is now your fault.)
This is a clear ad hominem. You wouldn’t infer from Henry Ford’s adoration of fascism that he made crappy cars. The two arenas are completely unrelated.
Turtle seems to be remarking that Chomsky is wrong on both independently though.
> **** In addition **** to being stubbornly wrong on linguistics
Emphasis mine.
Well if we’re using caveats then let me just add my own, that if you ***were*** to use Chomsky’s political views to disparage his academic record ***that would be*** a baseless ad hominem attack.
I agree, but it does reveal a tendency towards dogmatism. That doesn't mean Chomsky is wrong on linguistics, but it does raise the question of whether he'd ever admit it if he is?
Why is Left Libertarianism the only ideology labelled as dogmatic? Because of Venezuela? Has no other ideology made similar errors? Should the centre-right be disowned because Margaret Thatcher backed Pinochet?
You are reading *way* too much into my comment. I never even hinted that Left Libertarianism (not how I would describe Venezuela, but that's beside the point) is the only dogmatic ideology, or that other ideologies aren't dogmatic, or that we should disown anyone simply because they are dogmatic.
All I said is that we should keep in mind that a person who tends towards dogmatism in one area might tend towards it in others.
The regime I more often hear Chomsky tarred with is the Khmer Rouge, which was much worse than Pinochet.
Pinochet resulted in today's Chile and Chavez resulted in today's Venezuela so not the analogy I'd lean into.
Chomsky is, from what was heard, a person prone to dogmatic and condescending responses, including to colleagues and including in detail (there was the whole "Semantic Wars" thing, for instance...). However, a world's most dogmatic person can tell you that two plus two equals four, and it won't make two plus two equal five.
Specifically, the work in linguistics based on his assumptions has mostly proved tremendously _productive_ (including when it shows that some specific _details_ of his proposal were wrong).
Nah, I agree with this. Heisenberg was literally a Nazi, but that doesn’t invalidate the uncertainty principle.
Seconded!
Not trying to argue at all, simple honest question: does Chomsky ever talk about how his understanding of linguistics underpins/justifies his political views? I am by no means an expert on the guy, but I do remember reading some of his stuff where he makes the connection. (Could be wrong!)
I never got round to reading Manufacturing Consent, if there is any crossover it will probably be there - how the powers that be use language to manipulate us etc.
thanks for jogging my memory, I think that's where I read it too. My mental recursion substructures must be underperforming today....
You're welcome!
Haven't read the entire corpus of his works, but I don't think so. There is an often-made joke that Chomsky the linguist and Chomsky the anarchist are two different people, and in Russian it is reinforced by the fact that his surname was traditionally translated as Хомский (with a [x] sound and declinable) for linguistic works but as Чомски (with the same-ish sound as in English and usually indeclinable) for political works.
And, personally, I'm a Chomskian (largely, there are important differences of detail, as there are bound to be when you do science) in linguistics but very much not Chomskian in politics.
In his debate with Foucault this comes out quite a bit. Universal grammar suggests a universal human need for things like creativity which feed into a more university moral and political stance. The debate is cool. You can watch it on YouTube or read the transcript online.
Got any Ford quotes on fascism? I hadn't heard of that.
I've heard it many times, found this by a quick search: https://medium.com/@hrnews1/how-henry-ford-was-a-fascist-turn-leftist-chris-jeffries-7cb8e7ffb115
I didn't see any Ford quotes there. Instead there's a youtube video from "Turn Leftist & Chris Jeffries", which appears to be about various things not actually specific to fascism that Ford believed in.
https://en.wikipedia.org/wiki/Henry_Ford#Honors_and_recognition
> In 1938, Ford was awarded Nazi Germany's Grand Cross of the German Eagle, a medal given to foreigners sympathetic to Nazism.[145]
That seems more like evidence of them liking him (as they did Thomas J. Watson of IBM) than of his "adoration of fascism".
Near the middle of that same page there's a whole section on how he earned it. He's mentioned so specifically in Mein Kampf, the wiki's editors felt it necessary to clarify that Hitler was already an anti-semite before encountering Ford's essays.
Again, that's evidence of them liking Ford rather than anything Ford said about fascism.
The analogy doesn't work because cars are a physical product that can be personally tested and used unlike his ideas and theories.
I used to be a firm believer in engaging arguments on their own merits. I've come to realize that it's too difficult to do that in fields I'm not personally very familiar with, and thus there is a lot of value in using a person's thoughts on other topics on which I'm more confident to judge how seriously I should take their thoughts on the topic in question.
This presupposes the need to have a position on every academic controversy no matter how little knowledge you have about it.
Why does it presuppose that?
I was just uncritically accepting the thesis of the reviewer that Chomsky the linguist is a) overrated and b) dogmatic.
Cars are the kinds of things that either run or don't, as are assembly lines. Theories of linguistics, very much not.
Further, imagine if Chomsky had been a die-hard fascist. Think anyone would care about his linguistic theories today?
I've noticed that a person's ability to put together logically consistent analogies is a good reflection of their general mental horsepower.
It doesn't prove he's wrong about linguistics, but it provides a good reason to hate him.
Given that he's so hateable, though, the most reasonable thing to do seems to be to quit giving him so much credit for being (supposedly) the first person to formulate what is in retrospect a fairly obvious idea. Why do we always need to hear Chomsky's name every time innate grammar is mentioned? We don't hear Payne-Gaposchkin's name every time the hydrogen-helium composition of the sun is mentioned. (In fact I've never heard that name, I had to look it up as an example of a correct idea whose first proponent is not widely known or celebrated.
Let's discuss linguistics without ever mentioning Chomsky again.
The argument is that "stubborness" is a personality trait that Chomsky shares across multiple areas. One can easily think the Ford is "fascist" in politics, car design and dealing with workers without making the cars worse it just implies something about his personality.
Orwell thinks there is something inherently fascist about good matadors while fighting fascism (and liking bull fights).
I think he's an anarcho-syndicalist rather than a communist.
Typo:
> He does explicitly claim (in the aforementioned paper and elsewhere) that Pirahã probably has no longest sentence
Seems like this should be "probably has a longest sentence".
I thought maybe he meant, all Piraha sentences are roughly equal in length.
I am wondering why the question of whether Piraha has a longest sentence got so much attention. English has a longest word, German doesnt. Piraha has a longest sentence (say), English doesnt. What should matter is that no language has a longest story. The focus on sentences seems arbitrary to me.
The claim is that sentence is the biggest object built by grammar, while it is not even clear what a word is, that's a hot debate: for instance, if you claim that "child seat" is a word that just happens to be written with a space (or, vice versa, that Kinderstuhl is two words that just happen to be written together), English suddenly becomes a lot like German.
Moreover/as a consequence, concatenation of sentences for stories does not appear to be hierarchical (although cf. https://en.wikipedia.org/wiki/Rhetorical_structure_theory).
The example of words was just meant for illustration, apologies. My point was:
One can easily extend any grammar for sentences to a grammar for stories. The latter grammar will always be recursive. So what is the point in claiming that the grammar for Pidaha sentences is not recursive when the more important grammar for Pidaha stories is recursive?
Well, again, the argument may be that stories aren't recursively hierarchical, they just chain sentences linearly. (I am not ready to discuss whether that argument would be true, mind.)
"To Everett, the most important feature of language is not grammar or any particular properties of grammar, but the fact that it involves communication using symbols. "
"Can’t lots of animals do all those things? Yes. Everett views the difference as one of degree, not necessarily of quality. "
Before today, I had never heard of all this tribal language beef. Nevertheless (being an inveterate prescriptivist, though untrained in formal linguistics), I arrived at identical versions of the above conclusions independently some years ago.
I'm still not sure whether this points to a sign of weakness in Everett's argument, but I'm calling it a small win.
I can't help feeling that being somewhat more open to Aristotelian 'folk' metaphysics, purely pragmatically, might take the edge off a lot of academic rancour. The human mind clearly has a great potential for language, which is nevertheless realised in act only in certain contexts. I'm sure you all have excellent jokes about bowling balls tied together with string but I'm looking here at an interminable debate about what "innate" means, and I do think at least having the vocabulary to talk about "non-being which in some way is" is a strength of Aristotelianism (and Thomism). It's not sufficient but it seems necessary, or just helpful, useful. Anyway, interesting review.
Can someone please explain why Chomsky continues to have a role in linguistics like Marx in communism. What he says has been proved to be wrong again and again yet he is still deified
Marx was disproven (in practice) after he died. Chomsky is still alive.
The situation is not as black and white as there are Chomskyan linguists and non-Chomskyan ones, although people do organise themselves institutionally in these ways for contingent reasons. Few people agree with everything that Chomsky has claimed but not everything Chomsky has said is false. The useful stuff sticks around while people reapply it in new frameworks in an improved form but his personality casts a shadow because people like drama and it gives outsiders the wrong impression of where the field is intellectually.
I think Newton was a lot like this as well. And in paleontology at the turn of the 20th century, Cope and Marsh were two competing poles that similarly dominated a field by sheer force of will, developing useful theories that led to major progress despite being fundamentally flawed in ways that were known early on (in Newton’s case, by Leibniz).
Similarly in coordination chemistry in the late 19th century, with Sophus Mads Jørgensen and Alfred Werner https://www.britannica.com/science/coordination-compound/History-of-coordination-compounds
> To borrow a phrase from elsewhere, linguists are those who believe Noam Chomsky is the rightful caliph.
This is the quickest an essay got me angry, ever.
Me too, kinda, I hear this crap far too often :D But there's a kernel of truth in it.
This review got off on the wrong foot, for me, by initially pretending that all of linguistics has gone Chomskyan, which is the same order of mistake as pretending that all Americans are Republican. Chomskyans do like to pretend that they're the whole field (or at least the "theoretical core" of the field), just as politicians like to pretend that all Americans agree with them; but the rest of us don't have to believe it.
(In fairness, the review later implicitly backs off of this pretense, both by using phrases like "Chomskyan linguists" and by citing various non-Chomskyan linguists; but I don't think there was any reason for the pretense to begin with, and I think that an uninformed reader would still come away from this review with a *massive* overestimate of what proportion of the field is Chomskyan.)
What are your thoughts on the citation stat?
In the 2020 election, 74.2 million Americans voted for Donald Trump, which is (by a large margin) more people than had ever voted for any candidate in any U.S. Presidential election. It's far more people than ever voted for Trudeau, Sunak, Macron, Merkel, Putin, Stalin, etc., etc., etc.
If you were given no other information besides that paragraph, you might conclude that Trump is overwhelmingly popular and everyone supported him.
But in fact, his main opponent, Joe Biden, won *even more* votes! Somehow that just didn't come up before. ;-)
Of course, in the case of Chomsky, there's no corresponding "anti-Chomsky" we can compare him to. (The opposite of a quasi-cult isn't a different quasi-cult.) And in fact, it wouldn't surprise me if many of Chomsky's cites are actually from scholars disagreeing with him.
But either way, what the citation counts show (at best) is the size of Chomskyan linguistics, not the proportion of linguistics that is Chomskyan.
It would be useful to have the proportions I agree, but it seems unlikely that Chomsky's citation share in Linguistics would be less than Marx's citation share in Economics, judging by the numbers in the review.
Sure -- and I'll gladly stipulate that there are more Chomskyan linguists than Marxist economists -- but are you really under the impression that most economists are Marxists??
No, I just chose Marx from the reviewer's list. My takeaway was: X(A) = expected % share of citations for an arbitrary 'famous' academic; X(C) = expected % share of citations for Chomsky. X(A) should be roughly equal to X(C) but they are way off. That's interesting even if Chomsky's actual share is, say, 30%.
Sorry, I'm not sure I understand what you're trying to say.
The review uses the citation numbers to help make its case that most linguists are Chomskyan; so when you asked me "What are your thoughts on the citation stat?", I thought that you were calling my attention to that supposed evidence, and asking for my rebuttal.
But from your latest comment, it seems that you don't actually agree with the review's conclusion — which is obviously fine — but then I'm not sure what you're trying to ask me?
Chomsky has had a huge impact on linguistics in the sense that his work has established central debates and research questions. So many linguists who are “non / anti Chomskyan” are in some ways working within the bounds of those questions (eg, about innateness vs empiricism, gradualism vs saltatory development of language, communication vs cognitive function of language, etc). Of course you’re right that the field is much broader than followers of Chomsky, but it really is helpful to understand his paradigmatic impact to contextualize Everett’s work, since some of what Everett argues feels like common sense to outsiders (eg, language is “for” communication).
Honestly, even that seems overstated to me — plenty of linguists are perfectly happy to do descriptive work with little or no reference to the questions you mention (just as plenty of Americans don't vote for either party) — but it's infinitely better than the review's claim that "linguists are those who believe Noam Chomsky is the rightful caliph."
Yeah that’s fair!
I think that whole part was supposed to be "the hook", an amusing little bit at the start to get readers interested. It ran too long for me though.
I'm sure you're right, but I think the reviewer really does expect us to take claims like this one at face value:
> Since around 1957, Chomsky has dominated linguistics.
Especially given that it's repackaged later as a presupposition in claims like this one:
> [...] the dominance of Chomskyan linguistics has had downstream effects in adjacent fields like artificial intelligence (AI), evolutionary biology, and neuroscience.
I agree. It made me worry that this was going to be written from a pro-Chomskyan point of view that pretended anti-Chomskyans don’t exist.
> bedobi, Redditor
> the conduct of many in the pro-Chomsky faction is pretty shocking
The first thing bedobi says in that post is "I'm sympathetic to criticisms of Chomsky's universal grammar"
If you've liked this post, you'll love toki pona: https://en.wikipedia.org/wiki/Toki_Pona
toki pona is a minimalist constructed language, made by Sonja Lang about 20 years ago. It consists of about 130 words and a small grammar. Despite these limitations, it works surprisingly well for communication, and there are friendly, active online-communities.
toki pona seems similar to Pirahã in that it has very little recursion. Many things are broken down in parts. For example, a sentence like "Alice said that the cat died" would be translated as "jan Ali li toki e ni: soweli li moli" -- literally "person Ali says this: mammal die".
Interestingly, toki pona *is* an invention. One that has sparked a unique culture, complete with meetings, music videos, penpals, literature. The language has attracted many people who are neurodiverse or part of the LGBTQ+ culture, maybe because toki pona comes without some of the cultural baggage of older languages. In this case, one can truly say that language is the handmaiden of culture.
Alice’s mother: wait a week before you bury me.
:)
Yes, context is important in toki pona. It seems surprising at first that words have such broad semantic fields, like soweli which covers all mammals. Yet in practice, it is usually clear what is being talked about.
If anyone is interested in reading more linguistics beef, 'The Linguistics War' documents the acerbic debates on how linguists should study language in the 1970s, and Chomsky is, of course, the main character.
I'm not a .linguist, but having met linguists, I get the impression that Chomsky is not universally well-regarded, despite his massive citation count. Sp this review may be overestimating Chomsky's hold on the field.
It strokes me that there are weaker claims than full Choimskiansism that re still pretty interesting:
- pretty much all humans can learn all human languages, provided they start early and put in effort. (language learning easier in childhood). The specifics of particulars langauages seem noit to be genetically hard-wited at all.
- it sems that this group of people can learn languages like Portuguese with recursive structures even f (arguably) tgheir own language doesn't use them. So something like the potential to understand recursive langauges is bioliogically embedded in all humans
- on the other hand, many non-human animals cant learn our languages, So its clearly not trivial.
I agree that the FOXP2 mutation is not particularly related to language—I think it causes a general deficit in learned rapid precise motor control, and as it happens, speech is the area most saliently impacted by a general deficit in learned rapid precise motor control. See my blog post here for details including two textbooks that come to the same conclusion: https://www.lesswrong.com/posts/smgwRt5MbY9dwhecL/is-foxp2-speech-and-language-disorder-really-foxp2-forebrain
I got the impression that FOXP2 was one' of the steps on the way to language ... conveys the ability to make a wider ranger of squeaky noises (in mice) without grammar. Presumably, in humans, some other adaptation built upon it.
Hot take incoming:
Chomsky consistently repeats the claim that his linguistics and his politics are utterly separate. I disagree. The through-line is in Chomsky’s implicit belief that the Good Life is one of mass democratic decision-making. Anarcho-syndicalism, a kind of radical egalitarian belief that all humans are created equal, would require that all people/workers/citizens be informed about and actively participate in the collective decision-making of their workplace and society. Basically, everyone needs to become middle-class, literate, and non-violent.
Theories of Universal Grammar also posit that all humans are created equal. But, by centering grammar as the core of language, they privilege the relationship that educated, articulate people have with language. Grammar just isn’t that important to the working class and the underclass, people who didn’t get their grammar corrected in school until they’ve learned to follow the rules.
Now, I know that Chomsky means something peculiar when he uses “grammar”: Deep Structure. But what evidence do we have for deep structure? Similar people probably have similar mental representations. But beyond that, can we really say with confidence that words and sentences encode meaning? No, meaning is pragmatic, and depends heavily on contextual information. The position that words have precise meanings, that we can literally evaluate sentences compositionally, is itself a social construct, and in particular, a hallmark of academic and quasi-academic discourse, such as this discussion.
In working-class and underclass cultures, it’s uncouth to be so literal. Also true at some upper class social events. Often, language is meant to signal communal bonds, not make truth claims.
To further unpack grammar, it’s worth noting that even in English, the boundaries between parts of speech are porous. Nouns are verbable. Verbs often become adjectives. Adjectives may be used as nouns or adverbs. Nonstandard words are informal, but confidently using informalities is power talk. Branding, marketing, sales, persuasion, propaganda, poetry, all benefit from novelty, which focuses the audience’s attention, and makes things memorable.
Frankly, I find it painful when people — or language models; ChatGPT is boringly perfectly grammatical; it almost never coins new words/concepts or goes on fun tangents or interrupts itself with higher-order ideas — are overly grammatical. Get to the point. Say what you mean. Rules are meant to be broken. Languages evolve, dialects form, communities divide, cultures emerge, and fun is had when we take ownership of the rules of communication, rather than follow the conventions set by older generations.
After all, language is about communication, which includes self-communication aka thought. But communication is hardly about words and sentences.
In reality, what you say matters less than how you say it (volume, rate, tone, prosody, pitch, etc.); how you speak matters less than your body language (do you project confidence or doubt? take up space or shrink? eye contact or not?); body language matters less than what you look like (age, gender, race, grooming, clothing, accessories; what’s your social status?); your identity matters less than that of your audience (who are you talking to?); and what you listen to matter more than what you say (before you produce information outputs, you must consume information inputs).
Also, as a technical point, strongly consider that natural language is not arbitrarily recursive! After 3 layers of nesting, the amount of semantic ambiguity gets too great to contextually disambiguate. This is not merely a problem of limited mental horsepower and working memory, it’s a combinatorial explosion of interactions between the pragmatic context, the first-order meaning of the composition of the words in the sentence, and the predicted intention of the speaker.
The book Decoding Chomsky may of interest as it argues the opposite; that Chomsky's linguistic theories serve the interests of the American establishment.
I contend that Chomsky’s leftism also serves the interest of the American establishment. It’s an intellectual stance that is of no threat to the ruling class; a particularly erudite form of champagne socialism/anti-imperialism, inaccessible to the masses required for a real revolutionary movement. Chomsky is of course more rationalistic than the critical theorists / continental philosophers he called obscurantist, eg Foucault and other pomo theorists that paved the way for 21st century identity politics and intersectionality — though they too are of course part of the Western elite discourse, and no threat to the world order; indeed, today, LGBTQ+ rights and racial justice are a justification and tool of Western power.
EDIT: To clarify, my point is that idealistic egalitarianism that imagines all humans as potential rational educated western middle-class knowledge-workers participating in democratic collective decision-making systems is actually not as liberatory as it seems. Advanced human society requires complex division of labor, including division of decision-making authority, and the hoarding and wielding of secrets. In other words, a social hierarchy; workers and bosses.
For all the criticisms of mass media as a means of elite control over democratic processes, we haven't yet found a better means of social coordination. The chaotic populism of social media — rampant misinformation and conspiracy theories, declining trust in authoritative institutions, a race to the bottom attention economy, incentivizing everyone to become an influencer — seems to be in many ways worse than old-school manufactured consent.
So yes, thank you for the recommendation. I just got the book.
I think it is important to recognize that the way both Chomsky and Everett use the word grammar is quite different from the way laypeople use the word grammar. The kind of grammar that linguists are talking about is simply the organizational structure of speech. This means that Day Labourism professor is having exactly the same grammar, and it is equally important to all of them. Conscious knowledge of grammar that is learned in school is not talking about, it is certainly true that that kind of explicit knowledge is not terribly important to the average man in the street, but the average man in the street speaks coherent structure language, which is no different from that of the professor, perhaps for the number of fancy words being used.that is point on which Chomsky and Everett would agree
Yes, I understand that grammar has a technical meaning in linguistics. I just question whether this is actually carving reality at it's joints.
As I wrote above, from a perspective on language and communication that centers pragmatics and sociolinguistics, it doesn't make sense to imagine that sentences have a Deep Structure. Sentences don't mean anything outside of their context! Context — the who, where, when, why, and how of speech — matters more than compositional semantics — what was said.
The concept of grammar used by academic linguists is a kind of motte-and-bailey. Linguists retreat to the distinction you mention, between deep structure (an abstract formalism) and surface structure (the rules you learn in school) when challenged by the reality that there's scant evidence for their emphasis on grammar. But in practice, most linguists hate the distributional hypothesis, The Bitter Lesson, and probabilistic systems like LLMs; but they like Montague Semantics; and they like rule-based recursive parsers — converting surface structure to deep structure — which are a good model of computer languages but not natural language.
“Grammar just isn’t that important to the working class and the underclass”
That seems completely wrong to me. There is a formal grammar of written academic language that doesn’t matter to them. But if you look at the grammatical patterns of their natural language, you’ll see that they’re just as complex and just as important to them - someone who uses “ain’t” wrong, or fails to double-negate in the standard ways, or in any other way breaks the grammatical rules of their particular non-privileged varieties of English is going to sound silly, like an outsider who is interloping and pretending to speak their language.
That said, I agree with a lot of the other criticism, that people aren’t using language for the same kind of abstract thinking that Chomsky thinks it is essentially for.
I'm not saying that working- and underclass people don't have any linguistic regularity. That would be preposterous. Certainly, how you communicate signals your affinity with the group, your insider/outsider status, etc., but this is as much or more a matter of pragmatics and sociolinguistics as it is about syntax.
It's my experience that in less academic contexts, with working- and underclass folks, in both casual speech and text messages, there is indeed less emphasis on adhering to the grammatical norms of the non-prestige dialect. Certain conventions must be followed to — don't sound like an outsider — but beyond that there is more latitude for all sorts of disfluencies.
This is a somewhat controversial perspective, I acknowledge. Many good-hearted folks in linguistics want to maintain the view that non-prestige dialects are equivalently complex to standard dialects. Certainly, it's a fact that all languages can express all ideas, just not with equal ease. Now, one could hold that dialects may be morally equal without being equally complex in some senses. But that's a hard distinction to maintain, and not one that most academics are comfortable making, for reasons that mirror the complexities of the discourse around biological differences in ethnic groups.
When we zoom out from dialectical differences to differences between languages, it's then rather obvious that some languages are more grammar-heavy than others. Why would this be untrue of dialects?
For an example to make this more concrete, while African American English (AAE) does have some constructs that Standard English lacks, such as the habitual be and null copula, it lacks the subjunctive mood, and perhaps other grammatical constructs of Standard English. (Note that English's subjunctive is much simpler than French's.)
I contend that the habitual be and null copula are ways to communicate more ambiguous notions than are easily expressed in Standard English, while the subjunctive mood conveys more precise notions than are natural for AAE. And while the habitual be and null copula are now commonly used by younger non-black people, using "were" instead of "was" in an AAE context would signal outsiderness to "urban" black Americans, even to those who code-switch to Standard English professionally.
See, the reason why dialects are maintained by people who are also fluent in the prestige dialect is because they represent the culture and perspective of their community. And for a community to exist, it must be able to differentiate itself from the surrounding society.
AAE has constructs that enable expressing ambiguous notions because the African American community values the ability to make these statements. When black Americans want to get more precise, they usually code-switch to Standard English. And people of all races have adopted these ambiguous constructs because they too see the utility in them.
AAE constructs are expanding beyond black speakers because they're simplifications. Grammatical simplicity is often a good thing. It's part of why English is a great language compared to eg German or French.
I should add that I stuttered as a child, and had to go through speech therapy as a teenager. I "solved" my stutter by learning to circumlocute. Rather than attempt to say a word that would make me stutter, I substitute a synonymous word or phrase. Sometimes, this means that I choose the wrong part of speech, or get the prepositions wrong. When this happens, it's a minor faux pas in professional contexts, but just not an issue for folks with less book smarts.
"I should add that I stuttered as a child, and had to go through speech therapy as a teenager. I "solved" my stutter by learning to circumlocute. Rather than attempt to say a word that would make me stutter, I substitute a synonymous word or phrase."
I recall reading John Stossel's account of failing to do this with the word "dollars," as the initial hard consonant was a problem for him, and the only synonym he could think of, "bucks," had the same problem, in addition to being too informal. Did you do any better than him?
I thankfully didn’t have the experience of having certain words I would frequently stutter on. I can only imagine how stressful it would be to constantly avoid key words. For me, it’s more of an immediate feeling that the word(s) I’ve planned to say won’t come out the way I intended. The more I would try to force myself to say something, the more I would stutter.
The temperature parameter in large language models is a good analogy. Temp = 0 means the model always outputs the most likely next token. Higher temperatures mean it outputs less probable tokens with some probability. For me to speak fluently, I have to have a “higher temperature” — I often surprise myself with my choice of words.
When I did speech therapy at 15, I realized my stutter was minor compared to the handful of other teens in my month-long summer program. The other kids stutters were triggered in particular contexts like reading aloud or talking on the phone, and many of them had key words that they would predictably stutter on. For some it was their own names. And they seemed to be more self-conscious than me about it, more concerned with how others judge them. Maybe it was low self-esteem but I didn’t think other kids thought or cared about me much, nor was I that concerned with their judgments.
My stuttering then, as now, was only correlated with heightened emotional states, particularly anxiety.
Once I realized that I don’t stutter when I’m not anxious or otherwise emotionally stressed, I started to see my verbal fluency as a barometer of my mood. So when I do catch myself stuttering, I prioritize regulating my emotional state and solving the problem that is distressing me over trying to sound good.
Most mathematicians I've met claim to think mostly in geometric terms, definitely not the kind of grammar-structured language that Chomsky seems to have advocated through a long career. Maybe this way of thinking is sometimes compositional and therefore could be brought into a Merge framework, but I don't think compositionality is universal in such thought (or in any other styles of thought).
That sounds right to me.
This was one of the 10s I gave in the book review primary season, and also my pick for secret undercover Scott review
There is no way in hell this is by Scott
>A more parsimonious speculation is that they did not, and that the Great Leap was effectively instantaneous, in a single individual, who was instantly endowed with intellectual capacities far superior to those of others, transmitted to offspring and coming to predominate.
Isn't this basically what Catholics believe about Adam? That all modern humankind is descended from a single male (and a single female) ancestor, and that God granted Adam the gift of rationality, a mind made in his own image?
"He does explicitly claim (in the aforementioned paper and elsewhere) that Pirahã probably has no longest sentence, which is about the most generic anti-recursion statement one can make."
Was this meant to say that it probably DOES have a longest sentence? From what I read up to this point I'd have thought that all languages having no longest sentence is the standard Chomskyian view.
i had same question!
Ah I misread this! I thought it was saying he was conceding the one most generic *pro*-recursion point, that it doesn’t have a longest sentence, even though it doesn’t use any of the standard recursive methods for extending them.
taking this account at face value, it seems like it would be *extremely valuable* to the field of linguistics to have one more person actually interact with the Piraha and learn their language, which would presumably dissolve the questions about Everett's description of their language. I realize that getting in slow-burning asynchronous fights is more fun than that but it's kind of embarrassing to read about as an outsider
That would be cool. This comes at a high cost though, since there are under 400 native speakers and they live in a very remote place.
People are studying recorded material though, and doing smaller experiments (which do not involve spending years to learn the language).
https://en.m.wikipedia.org/wiki/Pirah%C3%A3_language
I follow linguistics pretty closely and Chomsky's influence would be hard to _under_state.
Your note about parse trees in the "mechanism" of LLM's is no hypothetical - such trees have actually been found, as early as 2019 (in BERT, a precursor to GPT)! This work finds that the trees are linearly embedded in a certain fixed subspace of the latent vectors. It's powerful evidence that such an algorithm emerged with no explicit programming, as simply a necessary mechanism to predict the statistics of language.
https://nlp.stanford.edu/~johnhew/structural-probe.html
Excellent review! I'm gonna go out on a limb and say that it's Scott.
And call this a non-central objection if you must, but I'd like to respond to the joke in the link in 4th footnote. Friedman would have won the exchange if he'd simply used a binary search: ln(2^n) = n so he still would have been able to pin Yessenin-Volpin down in linear time.
Oh I get it - it's very clever! But Friedman missed the chance to out-clever him. Just sayin'.
>Greg Egan has a great short story called “luminous” that uses this idea.
Yup! More-or-less why I took the finitist view in https://www.astralcodexten.com/p/open-thread-319/comment/51465300
I respect Scott too much to believe this text is of his authorship.
I don’t think it’s Scott. He hasn’t expressed significant views on the Chomskyan debate before, and this seems like someone who has spent a long time thinking about it (perhaps someone like me who grew up Chomskyan but had a conversion some time in the past decade).
> I find this notion of recursion completely insane, because it would imply (i) any language with more than one word in its sentences has recursion, and that (ii) all sentences are necessarily constructed recursively.
I had to ctrl+f 'Goptjaam' and 'glossable' to see if anyone beat me to this, but it looks like no. Granted, Goptjaam is a *deliberately* cursed conlang (https://www.youtube.com/watch?v=ze5i_e_ryTk), but I think it's still an interesting example of how this statement needn't necessarily follow! Goptjaam isn't glossable - its words depend entirely on the context of the full sentence. The moment you put the 'same' word into a different context, it's a different word.
To be clear, this isn't meant as an argument; actual human languages aren't like this! I just find it fun to look at how one can break languages if one tries.
Isn't from Chomsky to Everett a Wittgensteinian turn?
I think Wittgenstein never thought language was inherent, but he tried to invent a logically consistent language to make truth statements about the world (an "I-Language")
He then realised he missed the point, language is something that is practised - and the practise improves it over time rather than one perfect design
So Chomsky is doubly mistaken
Wasn't Wittgenstein answering a totally different question? He was interested in how language is used while linguists are interested in how language is constructed.
Maybe, but I don't think the distinction between how it's used and how it's constructed is the relevant difference
The "E-language" [Everett] part seems to me saying that "how it's used is how it is constructed" while Chomsky [I-Language] says "mental properties construct how it's used"
There's clearly some interaction but I think syntax and semantics are distinct. Building the car vs driving the car would be the rough analogy I'd use.
I think the important point of Wittgenstein (which descriptive linguists of any sort accept) is that languages are defined by how they are used, and whatever it is to say they are constructed has to be defined by that.
I know little of either Wittgenstein or linguistics, but whatever W was talking about didn't involve, for example, parse trees. That seems to me to be indicative of some sort of fundamental conceptual distinction.
I do sort of find echoes of W in things like word embeddings: the meaning of a word is defined by the words which surround it, which seems suspiciously like "meaning as use". Chomsky's view of word vectors and LLMs as antithetical to the more analytical approach of linguistics I feel lends credence to the notion that what he was doing and what W was doing were pretty distinct.
Great review. Linguistics is such an interesting topic area I appreciate the work that went into creating a layman's summary.
On a side note: can we create a corollary to Godwin's law stating that the longer an argument goes on the probability of calling someone a racist approaches 1? And that anyone that does so automatically loses the argument?
No because there are arguments where calling someone else a racist is correct and pertinent.
There are also conversations where mentioning Hitler or Nazi's is relevant - the vast majority of the time this isn't true and even in the rare case it is adds exactly nothing to the conversation anyway.
Calling someone a racist isn't an argument, it's a status claim. By doing that you assert that he's a Bad Person, and you're entitled to make that judgement. This maneuver succeeds or fails mainly based on your relative social standings.
you mean exactly like calling someone a Nazi?
Well, the original Godwin's law was about comparing something to Hitler/Nazis, not necessarily insulting the opponent directly, so it was a bit more subtle. A law for a more civilized age.
That's an excellent deconstruction. I might just quote that at the next person who calls me racist.
I would endorse this.
Some other people have pointed out parts of this, but it's bizarre how this book review leads in by pretending that non-Chomskyan linguistics doesn't exist. Firstly, his influence was a lot less in many subdisciplines of linguistics (phonetics/phonology, historical linguistics, etc.) than in his core syntax/morphology/child language acquisition specialties. Secondly, even in the disciplines where Chomsky has had more to say, there's always been a sizable number of linguists opposing him (see Randy Allen Harris's The Language Wars). So yeah, Everettian ideas are also building on an established tradition (as should be clear by the many references to pre-Everett field linguistics)
Yes, this seems right and important. But I also think it’s interesting to me to see an Everettian anti-Chomskyan take, because I’ve come to this from a different angle.
> He does explicitly claim ( . . . ) that Pirahã probably has *no* longest sentence ( . . . )
Typo: should be "has *a* longest sentence"
There are a bunch of cryptographers and linguists who are still trying to decipher the Voynch manuscript. A pattern observed in that community....
a) Someone will point out a statistical feature of the Voynich manuscript which they claim never occurs in any human language ever. This, they claim, is conclusive evidence that the Voynich manuscript is a meaningless forgery, not a simple cypher of some real language.(*)
b) Someone else will point out an example of an absolutely definitely real language that also exhibits the feature in question. Those guys have samples of lots of languages in machine-processable format.
(*) Historical context would suggest that if its a real language, its got to be either a real language known to Europeans around the time of Rudold II of Bavaria, or a conlang invented by them. This really cuts down the space of likely possibilities.
Had the Voynich manuscript been discovered much later, like in say 1960, after the widespread use of machine ciphers and with he existence of widely travelled linguists, our priors for "what kind of cipher of which natural language might this be" would be so, so much worse.
Relevance to current post: Piraha serving as go to example for some of those kind of "no human language ever" claims.
Exotic possibilities for the Voynich manuscript include: its a simple cipher padded with nulls, i.e. most of the ciphertext is random junk padding, but subset of it deciphers to something cohetrent.
> (*) Historical context would suggest that if its a real language, its got to be either a real language known to Europeans around the time of Rudold II of Bavaria, or a conlang invented by them. This really cuts down the space of likely possibilities.
How much does that really cut it down? It would rule out American and Oceanic languages sure, but isn't all of Afro-Eurasia still plausible in principle?
Just want to note that I really enjoyed Kingdom of Speech and found it highly rewarding to read, though I had a suspicion it started after Wolfe and Chomsky met each other at a NYC cocktail party and intensely despised each other.
Also to note on this:
"…well, okay, maybe I don’t entirely get Foucault’s number. Every humanities person must have an altar of him by their bedside or something. "
Basically yes, he dominates citations in the humanities/liberal arts world
It is worth pointing out that the standard transformer architecture used in modern LLMs is not recursive (or recurrent). That is to say, an LLM cannot construct a grammatically correct sequence of tokens that corresponds (parses) to an abstract syntax tree of unlimited depth. This is obvious on theoretical grounds from the basic architecture, but you can also demonstrate it quite trivially by training a transformer on toy languages described by context-free grammars. The maximum depth of recursion is limited by the number of layers in the transformer.
There are some variants (i.e. feedback transformer, universal transformer) that are fully recursive. We don't use those variants because in practice it doesn't matter. Humans can't process grammatical structures of unlimited depth either; once the complexity of a sentence starts to overflow working memory, the grammatical structure gets lost. So the difference between English and Pirahã is merely one of degree -- Pirahã has a maximum recursive depth of 1, while English has unlimited depth in theory, but in practice has a maximum recursive depth that is slightly higher, but still small (single digits).
Note that humans can still write code of unlimited length, even though the abstract syntax trees for human-written programs have finite (and typically small) depth. The trick is to do exactly what Everett did when translating the Bible to Pirahã -- break big expressions into smaller ones, and use variables (names/pronouns) instead of nested clauses. An extreme example is assembly language, where each expression (sentence) has exactly 3 "words" -- e.g. (operation register1 register2). Recursive grammar is unnecessary to represent complex ideas.
It’s also worth noting that half of the attempts to extend the context transformers may have run into the quadratic cost (complexity, memory) issue, and that probably the majority of proposed solutions to this problem are de-facto injecting recurrence, one way or another.
See the big discussion in other comments, specifically the "pointers to pointers" wording.
I have a vague memory that there's a field called psycholinguistics that studies things like...
although you might write down a grammar of English that sugges6s there are arbitrarily long valid sentences, there is in practise a limit to what actual human speakers will put up with, Experimentally ,measure how much stuff there can be on the parser stack before a human decides the sentence is unreasonable.
I strongly disliked this review. It is incredibly condescending and equally wrong (and often "not even wrong"). It honestly strikes me as odd that the reviewer, a self-admitted layman, believe they can so easily identify the supposed absurd and "insane" mistakes of thousands of brilliant researchers. The alternative hypothesis, that the reviewer might be misunderstanding something, should've a at least crossed their mind. This has all the red flags typical of a crank (possibly due to the source material, but still...).
Like the reviewer, I'm not a linguist, but even I could spot such misunderstandings in this review. These include conceptual misunderstandings (such as the misconception about the meaning and implications of "recursion"), contextual misunderstandings (e.g. regarding the long history of structuralism, from De Saussure onward), and disciplinary misunderstandings (there are many counter-Chomskyan linguists and even entire such linguistics departments; none subscribe to Everett's views to the best of my knowledge, for what seem like good reasons). Believing all the linguists are dense and mindless cult-followers lacking critical thinking is just silly.
This review correctly mentions Foucault being extremely cited, and this very blog has a post claiming that Foucault was basically a liar!
https://slatestarcodex.com/2018/01/04/book-review-madness-and-civilization/
I feel these interested amateur reviews are high risk high return. Sometimes you read something that shares a glimpse into another world and has an infectious feeling of excitement - and sometimes you get a smart person totally misunderstanding a really complex topic and leaving you feeling vaguely annoyed.
Do Chomsky's critics want to go back to the pre-Chomsky views within linguistics before 1958 or so they accept that Chomsky made progress in at least a negative sense by criticizing the conventional wisdom of the 1950s even if many of positive theories haven't panned out?
Any idea what percentage of the Chomsky citations are to his political polemics? Especially among grad students, his political writings are, because of their volume and the predictability of his opinions, an ‘easy cite’ for students looking to frame a debate.
I would expect that even his linguistic articles have _huge_ citation numbers, because half the articles include "Following the general premise of Minimalism (Chomsky, 2000)" or some such intro, which artificially inflates quotations even if the rest of the article doesn't directly engage with Chomsky's work at all.
I thought it was pretty clear that the opening paragraphs of the review, which purport to describe Chomsky's quasi-divine status, were tongue-in-cheek? - anyway, I very much enjoyed this review - I have only read Wolfe's book, but now I want to read Everett's own accounts.
I thought this was a great review, and I wanted to share another interesting thing about the Piraha. They are really happy. Relentlessly happy. In the 30 years Everett studied them, there were zero suicides. The idea was a completely foreign concept. Here’s Everett:
“I was still a very fervent Christian missionary, and I wanted to tell them how God had changed my life. So, I told them a story about my stepmother and how she had committed suicide because she was so depressed and so lost, for the word depressed I used the word sad. So she was very sad. She was crying. She felt lost. And she shot herself in the head, and she died. And this had a large spiritual impact on me, and I later became a missionary and came to the Piraha because of all of this experience triggered by her suicide. And I told this story as tenderly as I could, and tried to communicate that it had a huge impact on me. And when I was finished everyone burst out laughing.”
They burst out laughing! It just didn’t make any sense to them that someone would even consider doing that.
Does this have to do with their language? Well, language and culture are deeply intertwined, and we know that the Piraha only use language centred on the present. They have very limited vocabulary to describe the past and the future. So it suggests a culture that’s deeply focussed on the present. In Western culture we sometimes say that depression is living in the past, anxiety is living in the future, and getting someone out of their head and into the real world (the now) is the point of therapy and mindfulness training.
The Piraha may be very happy, but if their reaction to being told “this person was very sad and then she killed herself” is to burst out laughing, then I would describe them as somewhere between “zero emotional intelligence” and “stone cold sociopaths.” Not great role models!
I picture it like - imagine someone told you “this person was so sad that she ate a thousand bananas until her stomach exploded.” You might laugh, right? Just cause the concept is so ridiculous and unheard of, not because you don’t have empathy for people with exploding stomachs.
I love these ACX book reviews, and will read this one in its entirety later, but is it too much to ask them to be a little more brief? With the length of this one I feel it’s probably easier just to read the book itself.
The fact is that many (most?) winners/ front-runners have been very long, historically. So the revealed preference is for longer rather than shorter reviews.
Of course we’re all selected for liking Scott’s writing here- and it leans towards thoroughness at the expense of brevity.
I was somewhat thinking this, but also thinking that it could have used a longer section actually talking about the book being reviewed!
Have to point out with the other commenters that the idea that linguistics is primarily Chomskyan is not true.
The thing that I find striking about Chomsky is the claim that nothing other than grammatical analysis really matters. From a philosophy of language perspective, I am much more sympathetic to the Wittgensteinian view, as expanded by Kripke. Chomsky in the end seems quite reductionist, mistaking his understanding of how there can be formal systems that describe how languages function via grammar to mean that other aspects shouldn't be studied. His project reminds me of philosophical work trying to derive the basics of arithmetic from the rules of logic, and then declaring that nothing else you could do with that was worthwhile. Even without that, there were plenty of things to do in mathematics- and different logical foundations are possible for the empirical facts that can be observed. If the Chomsky project ends up just describing a very simple rule for recursion, it may not be all that useful as a foundation for study of more complex phenomena, as the complexity that must be built on top of that foundation in order to describe common uses of language makes those descriptions less useful.
Language is a complex interaction of many things. The claim is that the grammatical side (which is richer than just recursion because of interfaces - again, ChHF 2002 is a _very_ bad intro into UG, good linguistic articles are not published in Nature!) is basically the only (and existing) _proprietarily linguistic_ thing among them, while others are purview of psychology, sociology or whatever.
Then studying 'proprietary linguistics' is pointless? Or maybe it's not important what is considered 'proprietary linguistics'.
Define "pointless". It certainly leads to discoveries about how grammar works and indirectly bears on other cognitive science because language is often the window used to try and look into the brain.
" Perhaps the most interesting suffixes, however (though these are not unique to Pirahã), are what linguists call evidentials, elements that represent the speaker’s evaluation of his or her knowledge of what he or she is saying. There are three of these in Pirahã: hearsay, observation, and deduction."
This sort of thing is not so unusual - for instance, Quechua has them. (A page in a grammar book on my bookshelf lists the situations in which one particular evidential particle can be used - some examples from memory: I was there but I was asleep, I was there but I was very drunk, I was there but I was less than five years old...)
> How Chomsky attained this stranglehold on linguistics is an interesting sociological question, but not our main concern in the present work
This review is bizarre. Chomsky doesn't have a stranglehold on linguistics. It's like writing a long essay criticizing Aristotle's stranglehold on physics.
While it’s overstated, I think it’s right. There are chomskyan departments and anti-chomskyan departments, but everyone has opinions on Chomsky. There is no comparable figure in physics.
I totally agree!! After the first few paragraphs I was seriously confused about what this guy had actually read. It just seems the opposite of any recent “mass appeal” linguistics book.
> He does explicitly claim (in the aforementioned paper and elsewhere) that Pirahã probably has no longest sentence, which is about the most generic anti-recursion statement one can make.
Should that read 'probably has a longest sentence'?
Okay next time someone named Everett proposes new fundamental ideas about a field I'll be listening from the beginning
Did Chomsky really declare that the first language evolved abruptly a mere 50,000 years ago? If that is a prediction of his theories then its epic stupidity largely discredits them on its own. Anatomically modern humans appeared over 300,000 years ago for a start, and it seems misguided, verging on cranky, to dispute that throughout that time they were fully capable of speech with a proper language, even if this may have lacked some modern refinements.
I mean one can teach chimps and gorillas to speak, at least simple sentences. They may grunt and huff and puff in trying to imitate human vocalizations, but when they hold up cards saying things like "Koco want banana" and suchlike, there is no doubt they are meaningfully communicating, and is cast iron evidence they are thinking in words up to a point. Even my cat clearly recognises a few words, such as her name and "din dins"!
If you assume that being anatomically modern means having a fully developed language ability, then you're making a circular argument (and ironically, you're kinda agreeing with Chomsky while doing it, though with respect to a slightly different of definition of "anatomically modern").
Also, writing appeared abruptly about 5,000 years ago, even though humans 300,000 years ago were anatomically capable of writing just like us. This shows that your reasoning is flawed, and it's reasonable to suggest a similar trajectory for speech. In fact, this makes more sense for speech: there is evidence (like types of aphasia) indicating that speech, unlike writing, isn't genetically stable and can break down isolatedly. While not definitive proof, this supports the hypothesis of an abrupt development of speech.
Finally, archaeological evidence (as far as I know) shows that about 50,000 years ago, there was a significant cultural leap in art and technology that spread quickly worldwide. The fact that there were no major anatomical changes during this time, as you mentioned, actually supports the idea that language suddenly emerged rather than contradicting it.
Not to pick on you but there is a lot of confident incorrectness about linguistic theory because we are prone to assuming that being able to speak language is a good qualification to start theorising about it but it isn't. Things like "word" and "sentence" have intuitive meanings to lay people but they are pre-theoretical terms. When you put in the work to analyse their role in a larger explanatory model (such as those starting with Eric Lenneberg and Charles Hockett, who made contributions to the biology of language long before Chomsky), you realise that these things can be counter-intuitive and you can reasonably come away with the conclusion that non-humans communicate with things we might not want to call "words".
In general, in this review and the comments, it ought to be suspicious how quickly people conclude that Chomsky is ideological or moronic in the face of counter-evidence, or that the field as a whole is just a cult of acceptance for everything he says, as if there haven't been hundreds of other linguists in the same tradition who have done fieldwork and animal studies and psychological experiments.
I mean, just open journals like Linguistic Inquiry, Syntax, Glossa or Natural Language and Linguistic Theory etc. - that should have been foundational to a piece which ultimately disparages the field alongside Chomsky but there is no evidence in those pages of what the review describes because the Everett drama is totally unrepresentative.
I think underlying all this, there's really a badly misunderstood model that people have for how academia works. Perhaps biased by reasonable critiques of the postmodernised humanities but that brush can't be used to tarnish everything, and people ought to think about other fields a little more carefully.
That's not what evidence from teaching apes language actually suggests, and anatomically modern humans that seemingly didn't engage in modern behaviour are an argument in the other direction, that is, that they were _not_ capable; I've already touched on both in other comments.
Must admit to skimming the last two sections. Three parts preface, one part actual review of the book in question, one part kludged-on red meat for rats (ratmeat?), with the former roughly 65% of the entire review, and cornucopious blockquotes throughout (so many blockquotes, mostly from...not-actually-the-book-being-reviewed)...it's just A Lot. Yet despite all the references and scene-setting, I still find myself with that nagging [Citation Needed] feeling of insufficiently-supported statements. This was not helped by c.f.'s to other works that the author admits to not having actually read/understood, nor ~all the quotes being from within the first few dozen pages of each work. I want to assume the full texts were indeed read, and their contexts explicated in prose, but...I dunno, it reminds me a lot of sundry mandatory Works Cited bibliographies in school stuffed with trite surface-level references. A palette of bon mots more than integral supporting research. One can certainly stitch together an argument this way, but it shambles rather than running, which is optimal for zombies but not rhetoric.
Ultimately I leave with no more knowledge than I started of what linguistics even is, or why it's important outside departmental funding disputes, or why anyone takes Chomsky seriously. "It's not my theory that's wrong, it's the world!" says the dismissively towering Ivyist who's never done any fieldwork. (Also, I still don't understand what an I-language is. The spherical cow of languages?) Despite reading it, it's Not For Me, sorry.
Without commenting on the rest, I-language is the mental module for language in each individual's brain, which gives rise to the individual's observed language, "idiolect".
This would have been helpful to know while reading, thank you. So language is theorized to be necessarily individual and arises from a particular <s>candidate gene</s> shared brain module, yet nevertheless follows certain universal patterns, despite the trivially observable vast differences between discrete..geocultural expressive groupings, or e-languages. (And maybe - maybe - some languages that don't follow all the axioms.) A meta-language, if you will, and an essential H. sapiens evolutionary stepping stone leading to enhanced cognition overall. That's at least the seed of a potentially interesting and falsifiable hypothesis. (I wonder what Chomsky thinks about Neuralink and similar proposed technologies...)
Should have amended that I...never actually learned proper grammar, and wasn't even introduced to the concept of certain fundamentals like "parts of speech" until late teens. Even today, the formalisms continue to baffle...some language simply seems intuitively correct, and other constructions not, without really being able to explicate why. Chomsky obviously means more than *just* the formal Rules Of Proper Language one learns in school when he refers to Grammarcy Hall, but that's definitely a part of it...so not having that, uh, brain module is perhaps a large part of why I bounce hard off linguistics. Not an impediment to actual speech-as-communication-or-internal-thinking (I assume perfect written/verbal SAT/WAIS measures *something* real), but probably a big stumbling block for understanding the theoretical underpinnings.
> but that's definitely a part of it...
Not really? If the rules are part of the brain, you don't need to teach them. More generally, you can't explain how specialized modules of the brain work by the mere fact you possess them, any more than possessing a heart makes you a cardiologist.
Not sure I understand. If such knowledge is inherent to all brains, just as natural as no one needs to be taught how to make their heart beat or their legs walk, then...why do we teach grammar at all? Still-disappointing-but-gradually-increasing average literacy levels certainly seem to suggest there's some additional need being met (although such paucity remaining a common feature of the complacency of the learned is also not encouraging). I do notice that such deficiencies perhaps reflecting on the actual strength of one's internal thinking, rather than merely communicative barriers, is...kind of a dark road to go down as well. A conclusion to flinch away from. One wants to believe in the rich interior lives of other sentients...
Incidentally, I appreciate you bothering to patiently interlocute in good faith with the sundry ignorant commenters on this topic, self included obviously. Edification on impenetrable topics is a worthy endeavor even if it doesn't stick. "More of this, please," as the kids say these days.
Oof. What we teach in schools is an attempt to formalize and standardize what has already happened in heads, and the grammatical differences between how kids speak and what they are taught are differences between their grammar and "Standard English" grammar, not between presence and absence of grammar.
Language emerging suddenly is a somewhat separate claim from language being innate. If humans rapidly developed true language ~50 thousand years ago (as SF ideas go, I like the suggestion that we were taught language by sperm whales at that time, or if you're boring you can just go with aliens, but really I'm talking about an existing proto-language rapidly making a qualitative jump in some tribe), and other humans already had the capability for it because anyone smart enough can learn language, it could spread very quickly, as useful technologies do. And there really was a spurt of Homo sapiens expansion that started ~50 thousand years ago, for unclear reasons. But blaming it on a de novo mutation that rapidly spread to literally everyone seems dubious.
By contrast, you could also argue for a biologically innate language ability, hardcoded in, that nevertheless evolved gradually, through multiple levels of selection, from a proto-language towards today through millions of years of selection for better linguistic ability.
And, frankly, I feel that this would be a more logical alignment of the abrupt-gradual and hardcoded-general questions than the one the review describes linguists as going for. Hardcoded linguistic ability should emerge at the speed of evolution; language as a teachable invention, after a "hardware overhang", can spread at the speed of culture.
But not all evolution is gradual. If the intuition about the system is broadly correct, the recursive property is as abrupt as switching from RNA to RNA+DNA, and this lines up well with ontogenesis.
That's a strange analogy, given that we don't know how the RNA-DNA transition happened. But yes - evolutionary changes can happen quickly. However, quick by evolutionary standards is still pretty slow by cultural standards.
Sure, but, seeing as we speak about hominids, the proper comparison is paleontological standards not cultural, and evolution can be much quicker than those, especially under adverse pressure by quasi-kins (Ashkenazi selection, lactose tolerance selection, that fish speciation in recently separated lakes story...)
I mean, the question is whether language should be thought of as primarily a biological adaptation or primarily a cultural invention. So the comparison is between evolution and culture. Mind you, obviously you can still argue for a rapid genetic shift, or a slow cultural evolution; it's just ironic that the "sides" (if the review is to be trusted, since I understand that it drastically simplifies disagreements among linguists) are fast genetic vs. slow cultural.
Slow genetic is fairly popular among non-Chomskyans (see Svetlana Burlak, see some comments under this very post). But yeah, virtually absent "fast cultural" is interesting. The reason may be that cultural guys tend to be holists, and holists will (obviously) see language as more complicated than analysts, because, well, they include more into their notion of language.
(And as for "don't know how" — that makes the analogy more illuminative for me: the nature of the nucleotides implies it should've been rather abrupt, as ribonucleotides don't form stable chains (I mean the "vertical" connection, not the complementarity-based "horizontal" one) with disoxyribonucleotides, so it wasn't nucleotides changing one by one in a hybrid molecule, so we can conclude abruptness without seeing it.)
This was a terrific review.
Commenters have expressed on other reviews a desire for economy where book reviews are concerned.
This was really economical to me. Kind of the difference between a NYT book section review and a NY Review of Books review: something (the former) can be short and yet in some cases a waste of time compared to the longer thing.
I found the dispute and more so the background info fascinating but I would never want to read a whole book on it. This was “just right” in Goldilocks terms.
I loved reading about Chomsky and not having any of the obligatory stuff about his politics.
And yet - by the end - it did seem to me that his undoubted brilliant scholarship - displayed as time went on (if this review is anything to go by) a sort of unearned totalizing tendency, rather in keeping with his natural radicalism.
I like how the author presented acronyms in parenthesis before using them. The Author of The Family That Could Not Sleep review did not do this, and I struggled slightly because of it.
"I’m not going to rehash all of the details, but the conduct of many in the pro-Chomsky faction is pretty shocking."
isn't that also exactly how pro-Chomsky faction behaves regarding his political views? weird.
Both these people raise red flags for me. Chompsy is the archetype celebrity scientist holding a field hostage to his ridged interpretation, and Everett is making limited claims but writing as if he is overturning the entire field.
Everett's claims would plausibly overturn the entire generative linguistics field if they were true, make no mistake. They suggest interdependence between language and culture at a level incompatible with language being a module of mind.
What about Christiansen and Chater's views on language? Would they do something similar?
I am having trouble extracting a coherent approach from the three-page thing, but likely so.
I don't see Everett as turning over an entire field so much as overturning a specific paradigm within that field. And overturning paradigms is a standard feature of scientific progress. It's especially not much of a red flag to me if that paradigm never established itself all that firmly in the first place, and was promulgated on considerations that have their origins in a more rationalistic and less empiricist approach, which is how it seems in this case.
Critical views of the Chomskyean approach to language are not unique or confined to Everett, either, and others working in related fields have offered alternative approaches that, if true, would be seriously at odds with Chomsky's views. For instance, Christiansen and Chater also present a perspective on language that runs contrary to Chomsky's views. I don't see work like this as overturning a field, I see it as overturning ... well, Chomsky. I don't see any red flags in that.
https://www.youtube.com/watch?v=3dGZ2I4aVuY&t=38s the article also reminded me of this video on the artificial language Toki Pona that only has 120 words
How many times was an alphabet invented?
Possibly by a scribe in a Phonecian city state.
It is a different question, but if we have reason to believe it only happened once, that would increase the probability that language was only invented once.
Well, alphabet is a cultural invention — specifically, a cultural modification of another cultural invention, writing, which did appear several times. And yes, all modern alphabets except probably Hanguyl go back to an old Semitic consonantal writing system. I am not sure this recent development is of any relevance to the question of language evolution.
I mean, both Hebrew and Arabic still use an abjad, which is literally worse than hieroglyphics. Writing systems are sticky unless you have Ataturk or Red Terror at hand.
(Also, in digital age non-linearity of Hanguyl is a bit of a problem.)
I think we have evidence that writing was independently invented multiple times. Some of the scripts were syllabic rather than alphabetic. e.g. the Mayan script was probably an independent invention.
Yes. There are four to five independently appearing writing system systems. But only one gave rise to systems with alphabetic principle.
This review should come with a strawman warning. I am not a linguistics academic but enjoy books on the topic and this reviewer’s characterisation of linguistics as just Chomsky seems completely at odds with recent books. There are multiple established schools of thought and most recent books seem to have moved on - but reading this you would think there was some David and Goliath struggle.
Also I felt the LLM conclusions were pretty weak - if you’re going to bring them in I’d want to get more of a “So What?”. Concluding that LLMs show that Chomsky is wrong seems a bit too obvious.
Thanks. I was interested in Chomsky's claims about universal grammar 20 years ago but I think gradually the consensus shifted to the strong version of it being wrong and the weak version suffering from vagueness that could explain anything and nothing, so I lost interest.
I was intrigued to read this book review and indeed, it seemed to be the old perspective that is irrelevant today.
The LLM case is made more forcefully here: https://lingbuzz.net/lingbuzz/007180
As for Pirahã language and people, we have to take into account that this a very small group of only 200-300 speakers who live in harsh primitive environment. It is not unimaginable that they have experienced societal upheavals like violence from neighbouring tribes, diseases, famine, loss of habitats etc.
Spoken languages change rather quickly and we cannot even be sure that what Everest witnessed wasn't some kind of pidgin formed from merging several local dialects and languages. Wikipedia gives an example that women sometimes pronounce s as h which is an indication of a different linguistic background. You could expect eventual development into full-fledged creole language. However, the number of speakers are so low that now the school has been opened for them where they learn Portugese instead of Pirahã.
Some issues with what you suggest:
- there are extinct but historically attested languages that seem straightforwardly related. It's not like there's a total absence of data of current and some historical neighbouring languages, and it's known to have borrowed vocabulary from unrelated languages like Nheengatu
- Pirahã does not typologically look like a creole at all: it has complex verbal morphology for one thing, creoles usually lack verbal morphology altogether. It has phonemic tone which I believe is rare in creoles. Etc.
- Differences in women's and men's speech differences is not unique to Pirahã, it exists elsewhere in the world. I can't remember what it was but I once read of a language that goes as far as to have entirely different versions of common morphemes based on the gender of the speaker, so one phonological difference doesn't seem all that striking. Also I don't think this is a typical creole thing?
You make good points that Pirahã language could not be pidgin or creole. However, I think that simplified verb structure of creole is due to mixing of dissimilar languages like English and Chinese. What if those two languages are on dialectum continium like Russian and Ukrainian? It is called Surzhyk (Ukrainian-Russian pidgin) and I don't believe it is standardized and yet contains full complexity of verbal forms. Even the most defining features such as a specific pronunciation of Ukrainian [g] is undefined, some speakers might use it and other not.
In case of Pirahã it is not that women speak differently but that some women speak differently. How is it possible for such a small group to develop different phonetic features? We think of Pirahã people as a static group but reality might be that they are mixing with neighbouring tribes, possibly even with violence (like stealing wives from neighbour tribes?). I don't know how much I can trust researchers who are religious proselytizers. They always have their own agenda, even with Everest losing his faith in Christianity, he still might be a person of a strong ideology instead of science.
On further thought, I think my main issue with Chomsky-the-linguist is the whole I-language concept. While studying, in practice, normal (external) language, excluding stuff like errors to make the issue tractable, he felt the need to instead claim to be describing an internal language that didn't have those issues in the first place. I suspect part of this was because forming a model that's in any way less complex than reality, in the social sciences, will get you an instant swarm of humanitarians with "everything is holistic and fundamentally connected to everything else" and such. Either way, the result is that Chomsky is a linguist (and political theorist) that thinks he's a cognitive scientist, and interprets his claims about language as claims about the human mind. Ergo weirdness like "language isn't for communication" or "only humans can have language, AI and other species have no relevance to it" or the single mutation 50 kiloyears ago (stuff like that can happen, sure, but the prior should be very low).
And so - the calm Chomskyan response to Everett saying that the Piraha language doesn't have recursion would be "if this is true, then it's not a true language". Eventually that's what Chomsky concluded. But if you equate language and mind, then "the Piraha don't have a true language" is interpreted as "the Piraha are literally subhuman". And so the response was as toxic as you'd expect from a debate involving the word "subhuman".
‘John, the man that put people in the water in order to clean them for God, that lived in a place like a beach with no trees and that yelled for people to obey God’.
Heck, I'd reject that too! Never mind the tortuous syntax, it's the esthetics that's the deal-breaker here.
By comparison -- and overlooking the cryptic rendering (or mistranslation) of Mark 1:3 and the ambiguity of the last two sentences; is he supposed to be saying that the people lack God or desire God? -- this comes like a breath of fresh air:
‘John cleaned people in the river. He lived in another jungle. The jungle was like a beach. It had no trees. He yelled to people. You want God!’
I guess these short punchy statements might grow tiresome after a while. But then again, I could also imagine getting used to them.
And meanwhile, doesn't the very fact that these sentences occur together and form a coherent whole imply that parataxis is all you really need to coordinate a collection of ideas? Surely no one is suggesting that Chomskyan grammar stands or falls on the presence or absence of an explicit conjunction like 'and' as opposed to -- in this case -- a period that's in essence fulfilling the same function?
"Everett’s answer to the innateness question is complicated and in some ways subtle. He agrees that certain features of the human anatomy evolved to support language (e.g., the pharynx and ears). He also agrees that modern humans are probably much better than Homo erectus at working with language, if indeed Homo erectus did have language.
"He mostly seems to take issue with the idea that some region of our brain is specialized for language. Instead, he thinks that our ability to produce and comprehend language is due to a mosaic of generally-useful cognitive capabilities, like our ability to remember things for relatively long times, our ability to form and modify habits, and our ability to reason under uncertainty."
This makes me think of Cecelia Heyes' Cognitive Gadgets: The Cultural Evolution of Thinking (2018). She argues that many things we think of as innate, like "theory of mind" are learned. "Theory of mind" is the idea that you think a lot like me: you have emotions like I do, you have desires like I do, etc. On the other hand, our brain makes it relatively easy to learn some "cognitive gadgets" and not others.
"The intent here is not to pooh-pooh Chomsky" writes the author of a review that transparently tries to do just that; and fails spectacularly.
The crux of the argument is that Piraha doesn't use recursion but as Chomsky has explained many times (e.g. https://youtu.be/u6Lk79bnUbM?t=0h22m47s) what is of interest is what human capacities are, not whether they are always expressed. In the video above Chomsky makes the analogy to a tribe that wears a patch over one eye, which doesn't refute humans having the capacity for binocular vision.
The reviewer goes on to imply that this means the theory is vacuous or unfalsifiable ("the statement that some languages in the world are arguably recursive is not a prediction; it’s an observation, and we didn’t need the theory to make it... can we test it by doing experiments involving real linguistic data...?") but that is not true: linguists like Moro *have* done experiments that show that humans don't seem to be able to process certain made-up linear-order (i.e. non-recursive) languages as easily as made-up recursive ones and use non-language areas of the brain to do so (see Moro, "Secrets of Words"), which suggests that humans cannot understand arbitrary linguistic structures but can understand recursive ones, just as UG would have predicted. Had the result been the reverse it would have falsified (or at least severely weakened) the theory. (Note that Piraha is like a degenerate recursive language [no center embedding], so it is still "natural" for humans). Also note that the Moro experiment does falsify the "mosaic of generally-useful cognitive capabilities" claim made later on (which is an actually vacuous theory as it doesn't explain anything in any case).
Furthermore, as others have commented the reviewer completely misunderstands recursion, the fundamental characteristic they are supposed to be critiquing. The reviewer's own conception of recursion is indeed "insane", fortunately it is completely wrong. Merge isn't just about putting two elements arbitrarily into a pair, it is about creating a semantic interpretation according to how those pairs are combined and it implies that hierarchical order and not sequence/linear order is what matters for humans. See commenter Dimitrii's telescope example.
There are other misunderstandings:
- "Language is for communication, and probably—like everything else!—emerged gradually over a long period of time." No, not the appearance of the eye or multicellular organisms.
- Purposely conflating biological language faculties with cultural influences, two entirely separate things that are both referred to as "language". Chomsky specifically avoids this by using specialized terms (FLM, I-language).
- Equating language with speech, ignoring the implications of sign language and the mute.
- etc.
Things really go off the rails in the LLM section. The reviewer writes that "Chomsky’s claim that engineering success does not necessarily produce scientific insight is not uncommon, but a large literature speaks against it." but gives no examples in the case of LLMs (because none exist). They then go on to make the argument that since LLMs can approximate the output of human language then they must have analogous internals but this is deranged: You can train an LLM or deep neural net on the output of a clock or the weather but it is absurd to say that a clock or the weather works like an LLM or deep net and that examining the internals of the LLM will yield any insight whatsoever.
When a reviewer reviews a book in a topic they have no interest in, from an obscure author they presumably learned of only recently, one has to ask where the sudden urge to write the review came from. The last section of the review provides a clue: Like many other LLM-cultists, the reviewer had a sudden new-found interest in discrediting Chomsky after he criticized their precious LLMs. Maybe they should try harder to listen to what he has said first.
the eye appeared suddenly?
Yes (obviously?). At some point there was no light sensor, then there was. It then evolved further.
It's not at all obvious, because I'm not sure others interpret "eye" as inclusive of "light sensor." Can you clarify what you're claiming?
I have no idea what misunderstanding you have, maybe you should be the one clarifying. You can read the wikipedia article about the "Early eye". If your objection is the word "eye" then you can use the term "eyespot" instead, it's irrelevant to the point that it is an organ that appeared suddenly.
Do you think people who are wondering about whether an "eye" arose suddenly have in mind a notion of an "eye" this is so broad so as to include a very simple light sensor? Or do you think they may have something else in mind by an "eye"?
I was very careful to explicitly write the *appearance* of the eye which leaves no ambiguity.
You said they can't process them "as easily," but then you say, "which suggests that humans cannot understand arbitrary linguistic structures but can understand recursive ones." Is it that they can't understand them, or can't understand them as easily? Also, "people will understand languages with recursion more readily than ones without it" strikes me as a fairly tepid prediction that isn't exclusively consistent with UG or nativist views more generally. I'd also have some questions about who the participants in these studies were; if they'd already been enculturated into one set of languages rather than another, those difficulties may have been acquired rather than innate.
> Also note that the Moro experiment does falsify the "mosaic of generally-useful cognitive capabilities" claim made later on (which is an actually vacuous theory as it doesn't explain anything in any case).
Could you say a bit more about what it falsifies and how?
>- "Language is for communication, and probably—like everything else!—emerged gradually over a long period of time." No, not the appearance of the eye or multicellular organisms.
Can you clarify what the misunderstanding is here?
>- Purposely conflating biological language faculties with cultural influences, two entirely separate things that are both referred to as "language". Chomsky specifically avoids this by using specialized terms (FLM, I-language).
*Purposefully* conflating? How do you know what their intentions were?
> Is it that they can't understand them, or can't understand them as easily
They don't process them as quickly and they use non-language (specifically problem solving parts) of the brain to do so.
> ... strikes me as a fairly tepid prediction that isn't exclusively consistent with UG or nativist views more generally
It's not tepid at all, it provides reasonably strong evidence that language processing requires hierarchical structures, not arbitrary ones, which is what UG predicts and is not what "general processing" theories predict. It's probably one of the most important recent experimental results in linguistics.
> I'd also have some questions about who the participants in these studies were; if they'd already been enculturated into one set of languages rather than another, those difficulties may have been acquired rather than innate.
I have no idea what you mean. Obviously people in the experiment have pre-existing language ability. (It's not ethical to raise a child from birth in a fake-language experiment)
> Could you say a bit more about what it falsifies and how?
It takes longer to process a certain structure and it uses non-language parts of the brain which strongly suggests it is not using the language-specialized machinery to do it but other general problem solving machinery.
> Can you clarify what the misunderstanding is here?
There are crucial instances of evolution that are not gradual but sudden. I didn't mention it but the separate assertion about "language is for communication" also ignores Chomsky's arguments against that (that language is almost never externalized and it is computationally inefficient for communication due to things like the filler-gap problem).
> *Purposefully* conflating? How do you know what their intentions were?
They reference the distinction in other parts of the review and then ignore it when it supports their point. If you don't like the word "purposefully", just replace it with that long sentence instead.
You seem really interested in falsifying the Moro study, maybe you should read the original or the layman explanation in "Secrets of Words" first.
>I have no idea what you mean. Obviously people in the experiment have pre-existing language ability. (It's not ethical to raise a child from birth in a fake-language experiment)
What I'm asking is if they already spoke languages with recursion in them. If they do, then could it be that it's easier for people to process languages that share similar features to the languages they already speak?
>It's not tepid at all, it provides reasonably strong evidence that language processing requires hierarchical structures, not arbitrary ones, which is what UG predicts and is not what "general processing" theories predict. It's probably one of the most important recent experimental results in linguistics.
But even the results as described don't seem to show that it's *required.* Without more details I'd also have questions about why people may have responded this way.
>It takes longer to process a certain structure and it uses non-language parts of the brain which strongly suggests it is not using the language-specialized machinery to do it but other general problem solving machinery.
Okay, that makes sense, thank you. I'm still wondering why this is the case, though.
>There are crucial instances of evolution that are not gradual but sudden.
I would have thought the general concern isn't whether there are sudden instances in evolution, but whether something complex is sudden. Light sensors aren't. A human eye would be. So if someone proposed that a human eye appeared suddenly, presumably we'd agree that that's very unlikely. The question for me is whether claims about whatever genetic features associated with UG that may have arisen suddenly are more likely a light sensor or more like a human eye.
>They reference the distinction in other parts of the review and then ignore it when it supports their point. If you don't like the word "purposefully", just replace it with that long sentence instead.
I'm not going to retcon what you said. It concerns me when people are quick to attribute bad motives to others.
>You seem really interested in falsifying the Moro study, maybe you should read the original or the layman explanation in "Secrets of Words" first.
My work frequently involves critiquing the methods used in empirical studies. That's my natural disposition. I'm also sympathetic to Everett's views and not a fan of Chomsky's views, from what little I've seen so far. I may do exactly that, thanks.
> What I'm asking is if they already spoke languages with recursion in them
Yes every language does (except possibly Piraha).
>If they do, then could it be that it's easier for people to process languages that share similar features to the languages they already speak?
It wouldn't explain the fMRI data. You can certainly come up with a hypothesis that is consistent with the experiments without UG but they are just harder to believe than the "simple" explanation that language is recursive and as far as I know don't have corresponding supporting experiments. It's true that it isn't irrefutable dispositive proof but that rarely exists in psychology and psychology-adjacent fields.
> But even the results as described don't seem to show that it's *required.*
Required for what? I can put a gun to your head and force you talk without recursive sentence structure, what has that proven? If you have a Turing machine and only feed it regular grammars does that prove it is not a Turing machine? If you have a joint with 6 degrees of freedom but observe it only moving in 1 does that prove that it doesn't have 6? If you observe a small tribe in the Amazon that crawls on all fours throughout their lives does that prove that bipedalism is not a "required" property of humans?
> Okay, that makes sense, thank you. I'm still wondering why this is the case, though.
Well the point of the experiment was to demonstrate it is because the human language faculty does hierarchical processing (and can't do arbitrary simple processing)
> I would have thought the general concern isn't whether there are sudden instances in evolution, but whether something complex is sudden. Light sensors aren't
Chomsky agrees! One of the main guiding principles of Chomsky's minimalist program is that whatever occurred it had to be very simple (hence "merge").
> I'm also sympathetic to Everett's views and not a fan of Chomsky's views, from what little I've seen so far.
Yes that's clear, but you haven't made it clear why.
>It wouldn't explain the fMRI data.
Why not?
>You can certainly come up with a hypothesis that is consistent with the experiments without UG but they are just harder to believe than the "simple" explanation that language is recursive and as far as I know don't have corresponding supporting experiments.
That doesn't seem simpler to me.
> It's true that it isn't irrefutable dispositive proof but that rarely exists in psychology and psychology-adjacent fields.
I'm not asking if it's irrefutable. I'm wondering if it's especially strong evidence at all.
>Required for what?
You said "It's not tepid at all, it provides reasonably strong evidence that language processing requires hierarchical structures." With respect to the rest of your questions: what point are you illustrating with those questions? At least with respect to the last one, the answer strikes me as a trivial yes: bipedalism isn't "required," so if someone said "all human populations are bipedal" and we observed one that wasn't, that person would be wrong. If they said we had an innate predisposition for bipedality, that'd be fine, but that's a different claim. Evidence of languages without recursion is evidence that all languages don't have recursion; it's not as strong of evidence that recursion isn't a feature of some innate language faculty.
>Chomsky agrees! One of the main guiding principles of Chomsky's minimalist program is that whatever occurred it had to be very simple (hence "merge").
Okay, good. So what does he think occurred? Is there much in the way of empirical evidence that whatever that is occurred?
> Why not?
You would have to come up with some theory about why only recursive language processing is localized to areas of the brain associated with language and linear-order language processing is not. As I said you could certainly do that: Maybe during childhood development the language area becomes "trained" on whatever language it is exposed to and then it crystallizes that way so if you were trained on linear-language it would live in the language area and recursive language would have to be interpreted using puzzle-solving parts of the brain, the opposite to what we observe now. But there isn't any evidence for that and it's just more to believe than the simple explanation that that part of the brain came pre-wired to process hierarchical structure. You may disagree. It also creates its own unanswered questions like how did recursive language become predominant if other language structures could be processed equally well.
But if you want to come up with some experiment refuting the Moro results I think that would be an enormous scientific contribution, even if it is driven by animus against Chomsky.
> At least with respect to the last one, the answer strikes me as a trivial yes: bipedalism isn't "required," so if someone said "all human populations are bipedal" and we observed one that wasn't, that person would be wrong.
I'm glad we've discovered the disconnect here: you are just concerned with semantic pedantry. If I said "humans are bipedal" you would want to have an argument about how the Amazonian tribe crawls on all fours. This is a waste of time. No one is going to add all the qualifiers that you insist on that everyone already understands. People say "humans are bipedal" not "the natural state of humans is the bipedal state" unless their a specific reason to provide that required clarification. If you really need these pedantic qualifiers add them in your own mind, no normal person is going to write them out.
> Evidence of languages without recursion is evidence that all languages don't have recursion; it's not as strong of evidence that recursion isn't a feature of some innate language faculty.
Yes, exactly. It's not evidence at all in fact, which is Chomsky's point and why this whole book review is wrong.
> Okay, good. So what does he think occurred? Is there much in the way of empirical evidence that whatever that is occurred?
I mean that's the whole field of study, where all the main questions aren't solved yet. Chomsky isn't focused that much on the bio-evolutionary aspect of things, more system-computational. I'm not sure it's a fruitful area of pursuit in any case: What caused the eye (eyespot) to appear? Is there empirical evidence for that? We might never know but that doesn't mean it didn't happen.
this comment was very clarifying to me, thanks for writing it out (and engaging with the other guy below)
You might like my kvetch on Don't Sleep There Are Snakes!
https://www.kvetch.au/p/dont-sleep-there-are-snakes
>> Any expression larger than two words, then, requires recursion, regardless of whether there is embedding in that expression. For instance the noun phrase “My favourite book” requires two iterations of Merge, (Merge(favourite, book)= [Favourite book], Merge(my, [favourite book])= [my [favourite book]]) and therefore is an instance of recursion without embedding.
For crying out loud, the fact that I am able to chain instances of Merge (a glorified concat) does not imply recursion any more than the fact that I can write (1+2)+3 means that addition is recursive.
Coming at this from computer science, that example would be possible with the following formal grammar (in EBNF):
decorated_noun={possessive , " " }, adjective_noun
adjective_noun={adjective, " "}, noun
possessive="my"|"your"|...
adjective="favorite"|"green"|...
noun="book"|"duck"|...
By contrast, I would reserve "recursion" for cases where a production rule refers directly or indirectly to itself:
and_list = noun, "," and_list | noun, "and", noun
The claim is as silly as the claim that a bike is self-similar because a bike is made-up-of wheels and a frame, and the wheels are made-up-of a tires, spokes and a hub. Hence to describe a bicycle, you have to use 'made-up-of' on things which are themselves 'made-up-of' things, hence self-similarity.
Does it really make sense to you that an entire field is full of clueless nitwits who’ve never heard of this exotic "formal grammar" of yours?
Wikipedia, anyway, suggests otherwise: "BNF is a notation for Chomsky's context-free grammars. Backus was familiar with Chomsky's work."
I honestly don't see why you (and apparently many other) are having an issue with the term "recursion" here. As someone in computer science, you must have heard terms like "recursive grammar" and "recursively enumerable sets" before, right?
And by the way, also from wiki: "Addition is a function that maps two natural numbers (two elements of N) to another one. It is defined recursively as: a+0=a and a+S(b)=S(a+b)". I admit this is possibly a bit nitpicky, but it too suggests that you might want to reconsider your understanding of the term "recursion"...
I am aware that some guy named Chomsky (what a coincidence) has established a hierarchy of formal languages, and I am fine with the use of recursion there. Per WP on recursive grammar:
> In computer science, a grammar is informally called a recursive grammar if [...] expanding a non-terminal according to these [production] rules can eventually lead to a string that includes the same non-terminal again. Otherwise it is called a non-recursive grammar.
I am also okay with 'recursive[ly enumerable] language', even if I prefer '[semi-]decideable language' -- all languages expressible without recursive production rules are decideable, after all.
The crux of my disagreement is this:
> The term recursion, when used by Generative linguists, refers to a property of functions. Functions are relations between a set of inputs and a set of outputs. A function is recursive if its output can also serve as its input.
This is a bizarre definition to me. I have googled for 'recursion generative linguistics', but not quickly found more people making the claim, so for all I know this might be just one grad student. For all I know, the rest of the world calls what he calls 'recursive' an 'endofunction'. (Naturally, one can have plenty of indirect recursion without employing any endofunctions!)
The equivalent for biology of that definition might be:
> The term 'heart', when used by rhino researchers refers to the big blood vessels between the blood pumping organ (BPO) and the lungs.
I am fine with addition being recursive in the sense that there is an recursive rule for it in Peano arithmetic. I am not fine with it being called recursive because it projects R x R to R and thus is chainable.
I'm not a linguist or an etymologist, but from what I understand, the term "recursion" in linguistics doesn't meant to imply a precise deep meaning. Instead, it's more of a vague relic from historical jargon that makes sense within its original context.
In the case of "recursive", this context is the foundational concept of early generative linguistics, which posits that syntax is, well, generative - meaning that that syntax of natural languages is characterized by production rules that generate syntactically valid sentences. While this might seem like a simplistic or even trivial idea today, this computational approach was novel, revolutionary, and instrumental at the time.
To me, it seems sensible to use the term "recursion" to emphasize the process of building sentences from simpler fragments by iteratively applying production-rules.
Not even close to finished with the review and I am so intrigued I will absolutely be buying this book and more. So that bodes well as a vote.
By far the best review I've read so far, thank you! Also finally learned something about this Chomsky guy (seems like another case where great intelligence is used in large part or even mostly to defeat itself).
Fantastic post, also introduced me to a new rabbit hole on finitism and ultrafinitism.
Do any of you readers also start to feel that sentences such as,
> Large language models (LLMs) like ChatGPT are not perfect, but they’re getting better all the time
are repeated too much? I mean 2 years ago, ChatGPT released i-forgot-what-version and everyone was amazed, but from there, "the law of diminishing returns" hit hard. To the educated eye, one can see that the model simply stopped advancing "so greatly" and all the features they advertise nowadays are some extra pre and postprocessing tasks rather than the capabilities of model itself along with maybe extended memory or so.
Good review. I find it interesting that the Christian missionary is the one making such a strong case for gradual evolution and natural selection, and the Marxist professor is the one arguing some version of human exceptionalism and mysterious endowment.
I also got confirmation of my priors, that Chomsky would be (would have been) more helpful if he could just dial it down to about 65% and make room for a bit more humility in most things.
im very late; but does anyone have example of linguistic theories about non recursive languages?
Im anti the hyper recursion of programming languages; for example, while you could define the allocation of space using "binary space partition trees"(and I use bspwm), you really really shouldnt, artists and designers subdivide 2d space with grids for a reason.
Likewise instead of 2+3*4 instead of letting that statement equal 20 reducing the complexity and let more parsing be left to right; everyone thinks you *need* to push everything into a 10 step hyper abstract process; for what exactly?, you shouldnt ever depend on ambiguous statement grammer.
This discussion seems very dumb to me.
Yes you can describe all grammatically correct sentences with lambda calculus (aka the merge operator) and yes lambda calculus is hierarchical
But lambda calculus is Turing complete, meaning that it can literally describe everything. It can simulate the physics of my sandwich. Is my sandwich hierarchical?
So is Chomsky doing a massive mind projection fallacy with 500k citations? I mean it's happened before...
But the ones attacking Chomsky ... probably should just sit down and manually work out a few grammars of English sentences to see what is meant by "hierarchy". The same goes for the author of this post. No it's not necessarily about embedding full sentences. Yes, it's profound. We can turn a bunch of words into set-theoretical claims about the world.
"The dog is chasing the cat" -> chasing(dog,cat)=true. We can do that for all English sentences!! That's what Chomsky started! Ain't no obscure rainforest language gonna change that. The model is Turing complete. It makes no predictions to be falsified.
As for "linguistics is about grammar" vs "linguistics is about phonemes and stuff" - that is a discussion about budget allocation and not about linguistics. Is it a blegg or a rube? Depends on who gets funded.
A couple of things that have bugged me bout this review as someone with an academic background in linguistics (just a master degree, so far from an expert).
First it greatly overstates the importance of Chomsky in a way that even the most ardent Chomskian linguists would find glaring, and it does so only in the basis of citations, without thinking that high number of citations non necessarily equals reverence in the field. Chomsky has a lot of citations for various reasons, 1. He is a really respected and prolific linguist and also a very public one 2. His work has been important in several fields outside linguistics (mostly computer science, but also psychology) 3. He has been highly controversial and taken part in very heated debates inside the field (the semantic wars being the go-to example). I haven't checked the numbers, but my guess is probably that there is at least as many citations to argue against his work than to build upon it.
That doesn't mean that Chomsky theories didn't have an impact in modern linguistics, of course they had, and there is a still thriving community of Chomskian linguists, but most linguistic study that is done today it's not Chomskian in nature. To start Chomsky's theories did not really took a strong foothold in Europe, which basically continued and built upon the structuralist work of Saussure and the Prague school (which is different think that American "Bloomfield" style structuralism because linguist want to make things complicated) with the functionalism of Andre Martinet and the SFL and in the United States while certainly Chomsky contributed to the Cognitive shift, most linguist today working on Cognitive linguistics follow more in the steps of Lakoff that of Chomsky. And that is in the more they heavy/grammar heavy disciplines of linguistics, because pragmaticians, sociolinguists, applied linguists and such do not bother themselves with generative linguist at all.
Based on the initial false assumptions about Chomsky relevance, contrast Everett and his wildly contested claims about Piraha (which are not only contested by Chomskian scholars I must point out) with "modern" linguistics that do not do any field work or aren't bother with empirical data. Which is weird because we are in a time where multimedia tools and search motors actually have led to an incredible explosion in the creation of annotated linguistic corpora, which has led to basically linguists having more empirical data than ever to work with and thus to most linguistic work being highly empirical in nature (even inside Chomskian linguists). Theoretician like Chomsky are important, but right now, they are a minority in linguistics.
Lastly I don't get the last part musing about LLM's, because it only has very tangential relationship with the book that the review is ostensibly about, and second because the reasoning is not very solid, we know for a fact that LLM and human learn language in a very different way (I would argue that LLM don't even "learn language" , and are only very sophisticated stochastic parrots). And I am not even referring to the UG, even if we go by Piaget or Vygotsky's hypothesis of language acquisition, none of them rely in humans ingesting incredible amount of data to formulate statistical analysis but rather how development of innate capabilities interacts or creates the ability for language and how this gets molded by interactions with the outside reality.
1. "Chomsky makes claims like this elsewhere too. The argument is that, even if there were a language without a recursive grammar, this is not inconsistent with his theory, since his theory is not about E-languages like English or Spanish or Pirahã. His theory only makes claims about I-languages, or equivalently about our innate language capabilities.
But this is kind of a dumb rhetorical move."
I think from another angle, the point may not be so dumb. Given that a Pirahã child would acquire Spanish if raised in Barcelona, Mandarin in Beijing, and Hebrew in Tel Aviv, presumably this is because there is something related to recursion in the underlying database of rules (UG or something else) that generates whatever grammar the child acquires. Whether any particular language makes use of those rules is another matter.
This type of variation occurs not just for recursion, but most aspects of language. In phonology, languages vary in what phonemes they make use of, and some phonemes are never used at all in a particular language. The same goes for syntax. The difference is that recursion seems to show up in almost all languages whereas (certain) phonemes have far less universality.
2. "Everett thinks cultural pressures forced humans to evolve bigger brains and better communications apparatuses (e.g., eventually giving us modern hyenoid bones to support clearer speech), which drove culture to become richer, which drove yet more evolution, and so on."
Why would selection act on vocal anatomy but not neural circuitry for language? I'd be interested to hear an argument that would make this plausible.
My two cents, and I haven't looked into this and probably won't bother, but I studied linguistics before moving on to other work after graduating. All of this comment is baseless conjecture.
If I'm reading this correctly, Everett is claiming Piraha doesn't have /subordinate clauses/. Which is truly an insane claim. This is a super fundamental language structure. Without it, you have no real way of discussing desires, causes, commands, beliefs, or a bunch of other things. This is a small step down from saying "Actually their language doesn't allow you to ask questions, any attempts to ask anything will inevitably fail."
Running with the Questions comparison, I could see a naive guy studying Spanish, in an alternate world where Spanish is spoken by a group of 200 illiterate, innumerate hunter-gatherers in the deserts of Spain, coming to a bunch of wrong conclusion about the wacky morphology of Español (pronounced "ess pan yoll")
"They have no way of forming yes/no questions!" He says. But everyone and their mother knows in reality that no, Spanish does yes/no questions just fine, and relies on prosody to convey yes/no questions without issue.
(Actually even in popular languages like Spanish there's a bunch of weird misinformation where even experts get fairly basic stuff wrong, but I probably shouldn't go into that here)
The article mentions that the Piraha get mad at you if you talk about something you didn't directly see yourself. I think this may be important to the story. One problem that comes up sometimes with missionaries and researchers who are trying to learn new languages, is that you often rely on the natives' concept of what makes a valid sentence or not. Take the sentence "Colorless green ideas sleep furiously". For an illiterate uneducated hunter-gatherer without a strong sense of grammar, the sentence seems invalid because it makes no sense. My guess is that the Piraha kept rejecting the John-the-Baptist passages because they just didn't like the verbosity.
One of my favorite things from linguistics is the existence of pidgins/creoles, and regrettably I've never studied in detail any creoles, maybe one day I'll get around to studying afrikaans or something. But the prototypical story around creoles is that sometimes, when weird conditions bring together different groups with different languages (slavery for example), the need arises to quickly establish a form of communication, usually by taking basic vocabulary from different languages and putting them together with rudimentary grammar rules. This creates a pseudo not-really-a-complete-language called a pidgin. Maybe it has no way to form comparatives, so it's impossible to say "Stronger than" in this language.
What inevitably happens is a generation of children grow up hearing the pseudo-language, never hear any examples of comparatives, and then spontaneously invent some way of doing comparatives, and basically filling out all the missing features and turning it into a complete language. Sometimes the invented solutions are weird and unorthodox, but my understanding is that there's a basic stock of high-level structures, and every pidgin that comes along will get filled out with all the same high-level structures like noun clauses, adjectival clauses, comparatives, if-then statements, prepositional phrases and whatever else. (children are the engines of language innovation, hence all the slang)
Going back to Piraha, if the speakers of Piraha, in their tiny isolated group, possibly malnourished, have their language in some weird shape where they're not forming any subordinate clauses, I don't think this is a blow to Chomsky's ideas. And I would guess that if the piraha group improved their condition, started growing, expanding, and inventing new technologies, I'm pretty sure their children would quickly start using subordinate clauses once abstraction is at all important for their day-to-day lives.
Anyway, as a linguist I've heard about the Piraha in passing before but never looked into it. I've never heard of Everett before but I don't think he deserves any hate, it looks like he was just trying to help the Piraha People and learn about them. I don't think he's adequately refuted universal grammar, though.