This reminds me of the patient I heard about from a colleague who got cured after one of the voices in his head recommended the correct antipsychotic medication.
> talking about immigrant crime can lead to longer jail terms than the immigrant crime itself
I wonder which jail terms did you have in mind. The most publicised one was over this tweet, sent and widely shared in the middle of riots that included literally setting hotels on fire:
> Mass deportation now. Set fire to all the fucking hotels full of the bastards for all I care. While you’re at it, take the treacherous government and politicians with them. I feel physically sick knowing what these families will now have to endure. If that makes me racist, so be it.
It's not exactly "talking about immigrant crime", is it, so presumably there are other jail terms that I'm not aware of.
Yes, there are other jail terms you're not aware of. For instance, Lee Dunn was jailed for eight weeks for reposting three memes on Facebook at the same time as the Southport riots (in which he did not participate), one of which showed a mocked up image of a migrant holding a knife near the palace of Westminster. Compare that eight weeks in jail to Tariku Hadgu (as one example), who did not get a jail sentence despite assaulting two female police officers in Bournemouth.
> Prosecutor George Shelley said Dunn had posted three separate images. The first one showed a group of men, Asian in appearance, at Egremont crab fair 2025, with the caption: “Coming to a town near you.”
> The second also showed a group of men, Asian in appearance leaving a boat on to Whitehaven beach. This, said Mr Shelley, had the caption: “When it’s on your turf, then what?”
> A final image showed a group of men, again Asian in appearance, wielding knives in front of the Palace of Westminster. There was also a crying white child in a Union flag T-shirt. This was also captioned, said Mr Shelley, with the wording: “Coming to a town near you.”
This is by no stretch of imagination "talking about immigrant crime", it's promoting racial fear and hatred.
They're controversial tweets that I'm certainly no fan of, but I'm not sure it's inaccurate to refer to them as 'talking about immigrant crime', insofar as posting a meme is talking about something. They're clearly intended to draw attention to crime committed by immigrants (even if in a controversial and offensive way), and quite different from the Connolly tweets you refer to.
I genuinely don't know what to say to this. I guess some might call the blood libel "talking about crime" too, but to me it's just that: stoking fear of an ethnic group. There is nothing inherently connecting this
> …image showed a group of men, again Asian in appearance, wielding knives in front of the Palace of Westminster
to immigration, except if it's also combined with an assumption that "men, Asian in appearance" have to be immigrants, which is entirely false, racist, and socially corrosive in itself.
I was thinking more of things like this person who was arrested for posting a "Fuck Hamas" meme ( https://archive.is/OBLrZ#selection-5647.7-5647.12 ) , but you're right that they weren't jailed and I was conflating it with some of the others. I'll edit that description.
EDIT: See UnabasedWatershed below, the meme also included "Fuck Islam"
[To start off: I think it's outrageous to arrest anyone for posts like this.]
From the article, the post was: "F--- Palestine. F--- Hamas. F--- Islam. Want to protest? F--- off to Muslim country and protest."
Based on the article and other anecdotes about punished speech in the UK, it seems more likely that he was arrested for the "Fuck Islam" part, not "Fuck Hamas." I'd expect posting "Fuck Islam" on its own would get you in trouble, while "Fuck Hamas" wouldn't.
I think it's important to get details like this right! Arresting people for posting "Fuck Islam" is an extremely bad policy that I do not support. Arresting people for posting "Fuck Hamas" would still be *much* worse (and bordering on incoherent, given that they're defined as a terrorist organization whom it's illegal to *support*).
This is a neat illustration of a disinformation campaign landing incredibly well across the political spectrum. Somehow Americans are primed to believe that people end up in jail for minor transgressions here, even people like you with plenty of intellectual integrity to issue corrections. I don't know why this particular campaign landed so well, and I haven't seen this phenomenon analysed yet.
They aren't false per se, they are misleading in the context. "Malicious communications" is not a good law, but even the anecdotes in that very article highlight that it's not about political speech, but about private disputes.
> …their child’s primary school objected to the volume of emails they sent and “disparaging” comments made in a WhatsApp group… were questioned on suspicion of harassment, malicious communications and causing a nuisance on school property
That's quite far removed from opposing immigration or disagreeing with the government, and yet that's the lens through which it's seen in the US.
Hell, we just had a massive march in central London led by outright racists, with Elon Musk telling the crowd they have to overthrow the government to survive. Far cry from the image of 1984 that seems to be very prevalent among online Americans.
I wouldn't attribute this to a disinformation campaign, just sloppiness. I suspect that most North American free speech advocates still remember Count Dankula, and I think they're incorrectly pattern-matching new stories of jail-time-for-speech with that incident.
(Yes, I know that Count Dankula was Scottish, not English--but I suspect that Europeans occasionally get confused about whether something occurred in the U.S. or Canada, too. Either way, an occurrence in one country often does indicate how likely a similar occurrence is in a neighboring, culturally similar country.)
Something similar happens with MAID (euthanasia) in Canada. The program is broadly popular here but there is a strong anti-MAID disinformation campaign that appears to be primarily in the US. They spread false stories like that teenagers can be killed for depression. I'm not sure why Americans decided to pick up this issue.
> Arresting people for posting "Fuck Hamas" would still be *much* worse (and bordering on incoherent, given that they're defined as a terrorist organization whom it's illegal to *support*).
I can easily imagine the UK state arresting people both for pro-Hamas speech and for anti-Hamas speech. Yes that would be incoherent, but it's never stopped them before.
This comment beat me to it...."talking about immigrant crime" is not criminalized in the UK and no one has been charged with let alone convicted of any crime for doing so.
(Being pretty close to a free-speech absolutist, still not thrilled about that one nitwit being sent to actual prison let alone for as long as 31 months. But that's a policy disagreement and not a basis for twisting the facts beyond recognition.)
I really like this because if these tools can get more people to "bet" on prediction markets, then the markets will probably become more accurate ("wisdom of the crowd", "bias", and all that).
May I suggest something like a dedicated 2026 FIFA World Cup site?
I know that American's usually don't care about soccer, but outside America this is *the greatest* sporting event. Even an Octupus went viral for "predicting" the winners! (https://en.wikipedia.org/wiki/Paul_the_Octopus)
I can’t tell if you read my mind or just read my substack! I’m a big soccer fan! I think the WC is so big that even the existing prediction market platforms will probably do a good job designing normie-friendly pages for basic stuff like win probabilities.
But that doesn’t mean there aren’t other interesting things to do. For example, something people love to speculate about, (even in the US) is who will be on the World Cup roster. A while back, I made a Manifold market about the US roster (there’s also one for Germany I believe). And then later I made a custom “normie-friendly” page that organizes those probabilities a bit better:
World Cup markets are probably outside the scope of my grant, but I used the roster website above in my ACX grant application to illustrate what it means to me to make prediction markets more “normie-friendly”. Essentially stripping out the betting UI and thinking of the probabilities themselves as a consumable news product.
For every market, there is probably a “perfect” design that almost certainly looks nothing like how it is presented on its host platform, and there are many markets that are important enough to deserve that design treatment.
> Markus Englund, $50K, for software to detect data fabrication.
As an ACX Grantee 2024 doing an EEG Entrainment study replication, I am wondering if it'd be helpful to submit my perfectly fine, hand-collected, artisanal, organic, grass-fed EEG data to this project that I'm currently collecting in London: https://forms.gle/X37zyTV3KhbSb3Ze9.
Congrats to all the new ACX Grantees! Good luck with your projects!
> He wants to create an introspection benchmark, and to see what happens when you train AIs to succeed on that benchmark.
Isn't this a terrible idea? This is just funding capabilities research. Sure, it could help with interpretability too, but so can every "make AIs smarter" goal.
Hi Isaac! Grantee here. So we've thought a bunch about this, and we don't think that training models to have the ability/propensity for accurate introspection will do all that much to increase capabilities (beyond what the AI companies are already doing with, e.g., self-correcting errors), while it could have significant upside for interp, evals / auditing, and safety in general. But if you disagree, we'd love to hear what kind of capabilities you're worried about this advancing.
More generally, I tend to think that it's impossible to do good safety work that has zero chance of advancing any capabilities (see, e.g., Neel Nanda's take on this: https://x.com/robertwiblin/status/1967668166773104973). Despite this, I think it's good to keep doing as long as it has a sufficiently high safety:capabilities ratio.
Hi Adam, thanks for the reply. I think we have a very poor understanding of what "general intelligence" means, and what humans have that current AIs don't, so all discussion here will necessarily be quite speculative and vibes-based. My vibes are that introspection is a big part of what makes AGI able to do what it does. The ability to notice that we're making an error, think about the cause of the error, and come up with a mitigation, seems crucial to me.
e.g. when I try vibecoding agents, they reliably stop working above a certain complexity, and a lot of that seems to be because they lack any drive to create organizational systems, streamline their process, take notes, etc. It just doesn't occur to them that they can modify their own process.
IIRC Eliezer has said something similar.
I agree that to some extent all safety work advances capabilities, since it requires us getting a better understanding of how AI works. But why is this one a high safety:capabilities ratio? That seems backwards to me; making AIs more introspective will make them more able to do things like give fake outputs in CoT traces when they're trying to deceive the human, without necessarily giving the *humans* any more insight into the AI's thought processes.
Like, if I'm in charge of a country and there's a powerful foreign country that may be friendly to my country and may be hostile, I'm not sure, which intervention makes more sense to you: my country should try to get access to more information channels from them so I can hear what they're discussing internally? Or my country should give them better communications technology so their officials can better coordinate among each other?
Hi Scott! I'm one of the grantees (Adam Morris), and I think you forgot to add the sentence I requested at the end of the blurb about requests for potential collaborators. It was: "Adam is excited to chat with potential collaborators who have experience in technical AI safety work (especially in interpretability, CoT faithfulness, and fine-tuning frontier open models); reach out to him at thatadammorris@gmail.com." Just letting you know in case you can still add that in.
Also, just to signal boost here: I'm excited to chat with potential collaborators! So if anyone with an ML background is interested in studying introspection in LLMs, reach out to me :).
I suspect this isn't true in the sense you're thinking - in every poor country I've been to, most medications are available over the counter. Prescriptions are helpful either as a guide so patients know what to take, or to access a small subcategory of controlled substances, or as some kind of interface with aid programs so that the programs know what people need. That having been said I don't know for sure how this works in Congo.
Some interesting research by CrowdStrike on bias in DeepSeek. Not sure if you’ve seen it. Depending on contextual modifier, output code can contain more vulnerabilities.
Good question! Maybe sourcing from a more ethical part of the food chain, like barely edible pig parts or insects, put into an artificial mouse package? Some types of snakes only eat mice and rats and some will even eat mealworms. I suspect transitioning those mice eaters to something more ethical will take research and development and scaling - ie. will these bits of chicken fool my pet ball python, and how.
There's an entire question here about should snakes even be pets? I don't know how well a python kept in a tank lives compared to its life in the wild, and I tend to think that the really, uh, invested in their hobby types tend towards the crazy (yeah I know it's Chinese Robber Fallacy, but the stories about "and please be on the lookout for an escaped Burmese python which might be crawling in your window" don't reassure me*)
*Though that story had a happy ending, as one neighbour was equally a snake nut and didn't mind the snake at all, just picked it up and brought it back to the owners:
"Jenny Warwick said she looked out of the window at 05:15 BST and spotted the huge yellow snake on her neighbour's roof but did not realise what it was until she saw it trying to get in the upstairs window.
She said: "I saw them trying to poke it out and it fell on their car. It was massive.
"People passing by were staring and couldn't believe their eyes."
Linda Elmer, who recognised the snake, said she was woken at 07:00 by worried neighbours banging on her door, trying to find the owner.
She said: "Everyone was panicking.
"I picked him up - no one wanted to help. It was very difficult because he's a big snake and I managed to hobble down the road with this 18ft python wrapped round me and knocked on the owner's door.
"I think it was a shock for all of us.
"He's beautiful and Burmese are very docile. They're not aggressive snakes anyway and I had one myself so I was comfortable picking him up."
The RSPCA previously urged owners to keep snakes securely contained during hot weather as the warmer temperatures make them very active and more likely to escape."
I still think pythons etc. would be better off in their natural, native habitats.
I came up with an idea similar to Dan Elton's metascience observatory a year or two ago, but I knew I didn't have the skills to lead a project like that. Even so, I'd be happy to contribute to it. I tried to contact Dan Elton by following the link, but I couldn't find an email address. I'd be happy if someone shared it with me.
My idea was to develop an index of how badly we're over-relying on various scientific conclusions. To calculate that for a given conclusion, I'd aggregate the impact scores of all studies that claim to support that conclusion, then "divide" it by an aggregate statistical power of all studies that analyzed that research question (regardless of whether they supported the conclusion). I think the most difficult part would be reliably classifying studies by their research question (i.e. figuring out which studies were attempting to replicate which others), but LLMs could speed up the process a lot.
My idea seems far beyond the scope of Elton's grant, but I wouldn't be surprised if he had similar ambitions. Either way, I respect his efforts to assess the reliability of scientific results.
> Aaron Silverbook, $5K, for approximately five thousand novels about AI going well.
I'm not sure how this is supposed to work. Presumably, these novels would need to be just shoddy enough to bias the next round of AI training, without biasing any humans who read them. And the novels do need to be available for humans to read, otherwise they'd be unlikely to end up in the next round's training corpus. But if the LLMs are already as intelligent as humans (as most proponents of this approach tend to claim), then wouldn't the shoddy propaganda fail to "persuade" them too ? Even if the LLMs are nowhere near human levels of comprehension, wouldn't they still be able to extrapolate from all the negative human-written reviews of the shoddy novels ? The obvious answer is "yes which is why we'll fake the reviews", but it would be hard to fake the lack of reader participation -- which is also not impossible to fake, but realistic fake users aren't free. The more I think about this project, the more expensive it appears to get...
If the shoddy AI slop novels work a little bit, in the next round of ACX grants, Scott can fund $50k for a bunch of unemployed creative writing MFAs to write slightly less bad sci-fi novels about AI going well. Then if *that* works better, the 2027 edition can grant $500k to a crack team of talented authors to churn out a bunch of really good novels, which should arrive just in time to save us.
It needs to be pointed out that Palestine Action were banned because they broke into a base and damaged British military aircraft not for a free speech related reason.
That is correct, UK has proscribed organizations which are illegal to advocate in favor of, unlike in America where you can eg proudly support Al qaeda
The UVC lamps sound neat, but won't diseases just adapt to resist the light after a bit of time? Can someone who knows much more about biology and this technology explain to me why I'm wrong?
That's the cool part. The wavelength (222 nm) doesn't penetrate very far in tissue and doesn't damage skin or eyes. However because air droplets containing bacteria and viruses are small, they get fully "cooked".
My only potential concern is the light generating ozone, which could end up being toxic if the light is run a lot in spaces without good ventilation. So you'll end up needing ventilation anyway.
Our immune systems have been killing pathogens by heating up into fevers for a long time. "Evolve to survive extremely high temperatures" is possible, as evidenced by extremophiles that live near hot vents, but it's easier said than done, so not really found much elsewhere. Similarly, resisting UV light would be hard to do, and anything spending the metabolic cost to do so would be at a disadvantage in any other environment.
I was a bit confused whether my application was received or not. Were we supposed to get a confirmation email after sending out the form? I didn't receive any, and when I tried again, there was none either.
I dare you to read through this list, or even just scan it, and tell me who in the world is a better person, dollar for dollar, than Scott Alexander. And this is why you should consider paying for this newsletter, if you can afford it.
Hyperstition AI lists The Moon Is a Harsh Mistress as a story where "humanity faces its demise at the hands of its own creation: artificial intelligence." This is an error. The only AI in the story is trying to improve the living conditions of the people of Luna, to the point of sacrificing itself to protect them from attack. It's not "aligned to humanity" the way we describe it today, since it's written as a character with free will, an internal world, and multiple desires, but it's definitely working towards a happily-ever-after.
Re germicidal (far-UVC) lamps, they may be a great thing for specific purposes and places (hospitals) but beware of more general deployment.
Our immune system is metabolically expensive. If we make large parts of our world safe from pathogens, over time our immune systems will adapt and get weaker.
The "hygiene hypothesis" is the leading explanation for the origin of the epidemic of asthma and severe allergies (peanuts, latex, etc.) in the first world. People who lived in filthy peasant conditions didn't have those (particular) problems, and children who grow up on farms still don't.
Children who grow up in cities where people believe "cleanliness is next to godliness", do.
Congrats to all the new ACX Grantees! Pretty excited about the low-cost ultrasound project. The tech has made so much progress that it no longer makes sense for medical ultrasound scanner to be so expensive. Subhash, please let me know if I can help you with anything (FPGA, DSP, electronics, manufacturing, etc) (https://alextoussaint.com)
Well, I must have been stunned by a passing comet or something, because this year I think the vast majority of the grants are reasonable and sensible.
I'm neutral on the AI ones (I can't see that writing 5,000 cheery fanfics about 'and then the AI and the humans lived happily ever after' is going to make much difference one way or the other, but hey it's harmless) and as for the African urbanism one, I venture to suggest "Not enough YIMBYism" may not be *the* major problem behind "why can't we get proper infrastructure projects off the ground?" there. But that's why we need research, I suppose!
The shrimp-stunning one makes me laugh, in an evil way. Now I'm visualising tanks of liddle shrimpies getting individually stunned at varying levels of shockeration in order to find out how high do you need to go to stun a shrimp sufficiently, and how can you tell if a shrimp is stunned?
The Christians for Impact one puzzles me. How is this different from the usual charitable programmes around churches and denominations, and there are a metric ton of those already out there, including ones which started off with the "unite the believers and the non-believers who want to do good in action", e.g. Oxfam, or which have broadened out from their original religious inspiration (e.g. Depaul, which seems to have spun off from the original Society of St. Vincent de Paul)?
Looking at the website, I'm not entirely sure what they want or intend to achieve (and some of the language is a turn-off even for me, a believer: "Discuss evidence-based ways to tackle important, nonpolarized problems near God’s heart." Oh, so you can read God's heart, can you? Nice to know you have a hotline to the Throne!)
Are they trying to get non-believers hooked up with church charities, or are they trying to get believers into the kind of secular, political, high-powered jobs like 80,000 Hours? I mean, if I look at the page about the "problems near the heart of God" it's climate change, global poverty, disease, etc. and organisations like Concern, Trocáire, etc. have those covered. The "Christian software engineer tackling climate change" - what makes it different that he's a Christian? Is the idea that as a Christian he wouldn't have thought about undertaking this? How is his Christianity an influence on his career decision there, because there's a lot of the secular EA folks going down that path already?
The best sense I can make of this is that it's outreach to Christians to persuade them "hey, if you want to do good and get involved in charitable causes, don't go the traditional routes through your local church or denomination like going on the missions or whatever, instead get a high-paying, high-powered job and work on Big Thinky Thoughts projects" very much in the EA mould.
And I don't quite know how I feel about that.
EDIT: Actually, now I do know how I feel about that, and it's "Oh boy, here we go". Because I hadn't read the pertinent part of the website, and it's just dressing up EA aims and goals and procedures in a light coating of "speaking to your heart, American non-denominational church style":
"You have 80,000 hours in your career—40 hours a week, 50 weeks a year, for 40 years.
That makes it one of your primary resources to love and serve others and honor God.
We’re here to help you make those hours count.
We provide prayerfully-researched advice, tools, and mentorship that empower you to tackle the world’s most pressing problems.
Our advice is especially focused on students and graduates aged 18–35 in the US or UK who are blessed to be able to choose a career based on impact and are keen to tackle problems we find most pressing."
I took a look at the Aerolamp prototype and maybe this is a dumbass question, but how does a tiny lamp like this have the necessary range and dispersion to effectively cover even a medium sized room?
This is Jacob Witten, sorry to nitpick but I'd say more precisely "his work is still in stealth" as there's no startup (yet! working on it! everybody should feel to reach out if they're interested in curing horrible lung diseases)
"The number of urban planners per 100,000 people in the UK is about 37, and in the United States is about 12. For OECD countries, the average is 21.5."
Is that actually an insight on the UK's housing market?
"Kasey Markel, $10K, for genetically engineered corn. Kasey and his team at Semilla Nueva use prime editing, a new genetic technology, to create corn which is rich in zinc, iron, essential amino acids, and other nutrients frequently deficient in corn-heavy poor country diets. Our grant helps fund greenhouse space, enzymes, DNA synthesis, and scientist time, and will let them expand faster into new regions that require corn with different genetic backgrounds."
Let's say this works perfectly. Will this be used even then if it doesn't offer financial benefits, especially with Greens fighting it tooth and claw every step of the way?
You think it's crazy that 5% of papers fabricate data? I'm pleased if it isn't *far* worse!
This looks like a fantastic tool, though, agree. I imagine people will just train LLMs against it for undetectable fabrication, but it will still catch tons and tons of old ones.
I hope this has nothing to do with the solutions proposed in the lesswrong or eaforums article you shared previously about snake feeding which frankly showed no subject matter knowledge and was totally off base. If it's artificial or plant based snake feed, that won't work biologically, way more than 50k has been sunk trying to make it work but I don't think it's possible.
In the comments to that article I suggested that an ethically raised rodent supplier would be a good idea and probably a good business, and would be a cool acx grant, if this is that then I'm very excited and would be a customer. So much of the price of feeder rodents is the shipping (express, dry ice etc) that even doubling the cost of raising the rodents is pretty minimal on the cost to consumer.
> Sam Glover, $60K, to fight for free speech in the UK. These are dark times for UK speech on both sides of the aisle: the left is upset that speaking in support of Palestine Action is now considered an act of terrorism, and the right is upset about arrests for racist tweets. So far, pushback has been siloed by cause and partisan affiliation. Sam and his two co-founders are early-career bloggers and aspiring public intellectuals who want to build a united nonpartisan free speech movement. They’re still in stealth, but I’ll promote their website as soon as it becomes public.
I very much support this. Far too many people object when it's speech by "their" side that gets suppressed but not when it's the "other" side.
New grant idea: A "psychotherapist" LLM who specializes in treating LLM psychosis.
This reminds me of the patient I heard about from a colleague who got cured after one of the voices in his head recommended the correct antipsychotic medication.
A robopsychologist who specialises in treating LLM psychosis (the term with a different meaning than the currently used one)
>We received 654 applications this year, and were able to fund 42. To the other 608: sorry!
The real question is what happened to the remaining four.
Aargh, why does this always happen?
It's reassuring that you are as terrible at simple arithmetic as I am.
Wait, maybe I mean terrifying?
they were eaten
They were so bad he isn't sorry about not funding them?
Probably re-writing the application and trying again. Or lying in wait to mug the successful grantees. One of those. 😁
The opposite of funding. They were so bad he had to charge them instead.
I'm a bit disappointed by this bit:
> talking about immigrant crime can lead to longer jail terms than the immigrant crime itself
I wonder which jail terms did you have in mind. The most publicised one was over this tweet, sent and widely shared in the middle of riots that included literally setting hotels on fire:
> Mass deportation now. Set fire to all the fucking hotels full of the bastards for all I care. While you’re at it, take the treacherous government and politicians with them. I feel physically sick knowing what these families will now have to endure. If that makes me racist, so be it.
It's not exactly "talking about immigrant crime", is it, so presumably there are other jail terms that I'm not aware of.
Edit: I believe sentencing remarks for that case can also be quite useful: https://www.judiciary.uk/wp-content/uploads/2024/10/R-v-Lucy-Connolly.pdf
Yes, there are other jail terms you're not aware of. For instance, Lee Dunn was jailed for eight weeks for reposting three memes on Facebook at the same time as the Southport riots (in which he did not participate), one of which showed a mocked up image of a migrant holding a knife near the palace of Westminster. Compare that eight weeks in jail to Tariku Hadgu (as one example), who did not get a jail sentence despite assaulting two female police officers in Bournemouth.
https://www.whitehavennews.co.uk/news/24513379.sellafield-worker-jailed-sharing-offensive-facebook-posts/
https://www.telegraph.co.uk/news/2025/04/30/small-boat-migrant-spared-jail-punching-female-police-offic/
Here are the memes in question (per https://www.whitehavennews.co.uk/news/24513379.sellafield-worker-jailed-sharing-offensive-facebook-posts/ ):
> Prosecutor George Shelley said Dunn had posted three separate images. The first one showed a group of men, Asian in appearance, at Egremont crab fair 2025, with the caption: “Coming to a town near you.”
> The second also showed a group of men, Asian in appearance leaving a boat on to Whitehaven beach. This, said Mr Shelley, had the caption: “When it’s on your turf, then what?”
> A final image showed a group of men, again Asian in appearance, wielding knives in front of the Palace of Westminster. There was also a crying white child in a Union flag T-shirt. This was also captioned, said Mr Shelley, with the wording: “Coming to a town near you.”
This is by no stretch of imagination "talking about immigrant crime", it's promoting racial fear and hatred.
They're controversial tweets that I'm certainly no fan of, but I'm not sure it's inaccurate to refer to them as 'talking about immigrant crime', insofar as posting a meme is talking about something. They're clearly intended to draw attention to crime committed by immigrants (even if in a controversial and offensive way), and quite different from the Connolly tweets you refer to.
I genuinely don't know what to say to this. I guess some might call the blood libel "talking about crime" too, but to me it's just that: stoking fear of an ethnic group. There is nothing inherently connecting this
> …image showed a group of men, again Asian in appearance, wielding knives in front of the Palace of Westminster
to immigration, except if it's also combined with an assumption that "men, Asian in appearance" have to be immigrants, which is entirely false, racist, and socially corrosive in itself.
Connolly's tweet was bad but not deserving of a multi-year jail sentence.
There were others jailed during the aftermath of Southport, including one for saying the murderer was a Muslim.
There were also the recent protestors against the Epping migrant who sexually assaulted a 14yo girl, who got longer sentences than he did.
I do agree that 31 months is a lot, but the multi-year jail sentence was pretty much guaranteed once Connolly pleaded guilty to the charge. It's explained pretty clearly here: https://davidallengreen.com/2025/05/explaining-a-31-month-sentence-for-a-tweet/
I was thinking more of things like this person who was arrested for posting a "Fuck Hamas" meme ( https://archive.is/OBLrZ#selection-5647.7-5647.12 ) , but you're right that they weren't jailed and I was conflating it with some of the others. I'll edit that description.
EDIT: See UnabasedWatershed below, the meme also included "Fuck Islam"
Thank you!
[To start off: I think it's outrageous to arrest anyone for posts like this.]
From the article, the post was: "F--- Palestine. F--- Hamas. F--- Islam. Want to protest? F--- off to Muslim country and protest."
Based on the article and other anecdotes about punished speech in the UK, it seems more likely that he was arrested for the "Fuck Islam" part, not "Fuck Hamas." I'd expect posting "Fuck Islam" on its own would get you in trouble, while "Fuck Hamas" wouldn't.
I think it's important to get details like this right! Arresting people for posting "Fuck Islam" is an extremely bad policy that I do not support. Arresting people for posting "Fuck Hamas" would still be *much* worse (and bordering on incoherent, given that they're defined as a terrorist organization whom it's illegal to *support*).
Thanks, I've edited this in.
This is a neat illustration of a disinformation campaign landing incredibly well across the political spectrum. Somehow Americans are primed to believe that people end up in jail for minor transgressions here, even people like you with plenty of intellectual integrity to issue corrections. I don't know why this particular campaign landed so well, and I haven't seen this phenomenon analysed yet.
Do you think claims like https://www.thetimes.com/uk/crime/article/police-make-30-arrests-a-day-for-offensive-online-messages-zbv886tqf are false?
They aren't false per se, they are misleading in the context. "Malicious communications" is not a good law, but even the anecdotes in that very article highlight that it's not about political speech, but about private disputes.
> …their child’s primary school objected to the volume of emails they sent and “disparaging” comments made in a WhatsApp group… were questioned on suspicion of harassment, malicious communications and causing a nuisance on school property
That's quite far removed from opposing immigration or disagreeing with the government, and yet that's the lens through which it's seen in the US.
Hell, we just had a massive march in central London led by outright racists, with Elon Musk telling the crowd they have to overthrow the government to survive. Far cry from the image of 1984 that seems to be very prevalent among online Americans.
I wouldn't attribute this to a disinformation campaign, just sloppiness. I suspect that most North American free speech advocates still remember Count Dankula, and I think they're incorrectly pattern-matching new stories of jail-time-for-speech with that incident.
(Yes, I know that Count Dankula was Scottish, not English--but I suspect that Europeans occasionally get confused about whether something occurred in the U.S. or Canada, too. Either way, an occurrence in one country often does indicate how likely a similar occurrence is in a neighboring, culturally similar country.)
Something similar happens with MAID (euthanasia) in Canada. The program is broadly popular here but there is a strong anti-MAID disinformation campaign that appears to be primarily in the US. They spread false stories like that teenagers can be killed for depression. I'm not sure why Americans decided to pick up this issue.
> Arresting people for posting "Fuck Hamas" would still be *much* worse (and bordering on incoherent, given that they're defined as a terrorist organization whom it's illegal to *support*).
I can easily imagine the UK state arresting people both for pro-Hamas speech and for anti-Hamas speech. Yes that would be incoherent, but it's never stopped them before.
This comment beat me to it...."talking about immigrant crime" is not criminalized in the UK and no one has been charged with let alone convicted of any crime for doing so.
(Being pretty close to a free-speech absolutist, still not thrilled about that one nitwit being sent to actual prison let alone for as long as 31 months. But that's a policy disagreement and not a basis for twisting the facts beyond recognition.)
To be more precise, that's 31 month with 40% served. If it were full 31 months, she couldn't have spoken at Reform's Conference.
Fair enough.
Just wanted to say thank you again for the opportunity!
If it’s okay to post this in the comments, I will try to share my ideas/progress on “normie-friendly prediction market interfaces” here:
https://substack.com/@charliemol?utm_source=user-menu
If anyone wants to follow along.
Awesome!
I really like this because if these tools can get more people to "bet" on prediction markets, then the markets will probably become more accurate ("wisdom of the crowd", "bias", and all that).
May I suggest something like a dedicated 2026 FIFA World Cup site?
I know that American's usually don't care about soccer, but outside America this is *the greatest* sporting event. Even an Octupus went viral for "predicting" the winners! (https://en.wikipedia.org/wiki/Paul_the_Octopus)
I can’t tell if you read my mind or just read my substack! I’m a big soccer fan! I think the WC is so big that even the existing prediction market platforms will probably do a good job designing normie-friendly pages for basic stuff like win probabilities.
But that doesn’t mean there aren’t other interesting things to do. For example, something people love to speculate about, (even in the US) is who will be on the World Cup roster. A while back, I made a Manifold market about the US roster (there’s also one for Germany I believe). And then later I made a custom “normie-friendly” page that organizes those probabilities a bit better:
https://usmnt-wc-roster.val.run/
World Cup markets are probably outside the scope of my grant, but I used the roster website above in my ACX grant application to illustrate what it means to me to make prediction markets more “normie-friendly”. Essentially stripping out the betting UI and thinking of the probabilities themselves as a consumable news product.
For every market, there is probably a “perfect” design that almost certainly looks nothing like how it is presented on its host platform, and there are many markets that are important enough to deserve that design treatment.
Let me get my prediction in here first, so: the winner will be... Brazil!
Or Germany.
(There you go, now it is all but a dead certainty those two won't win because I've never won a bet in my life).
> Markus Englund, $50K, for software to detect data fabrication.
As an ACX Grantee 2024 doing an EEG Entrainment study replication, I am wondering if it'd be helpful to submit my perfectly fine, hand-collected, artisanal, organic, grass-fed EEG data to this project that I'm currently collecting in London: https://forms.gle/X37zyTV3KhbSb3Ze9.
Congrats to all the new ACX Grantees! Good luck with your projects!
> He wants to create an introspection benchmark, and to see what happens when you train AIs to succeed on that benchmark.
Isn't this a terrible idea? This is just funding capabilities research. Sure, it could help with interpretability too, but so can every "make AIs smarter" goal.
Hi Isaac! Grantee here. So we've thought a bunch about this, and we don't think that training models to have the ability/propensity for accurate introspection will do all that much to increase capabilities (beyond what the AI companies are already doing with, e.g., self-correcting errors), while it could have significant upside for interp, evals / auditing, and safety in general. But if you disagree, we'd love to hear what kind of capabilities you're worried about this advancing.
More generally, I tend to think that it's impossible to do good safety work that has zero chance of advancing any capabilities (see, e.g., Neel Nanda's take on this: https://x.com/robertwiblin/status/1967668166773104973). Despite this, I think it's good to keep doing as long as it has a sufficiently high safety:capabilities ratio.
Hi Adam, thanks for the reply. I think we have a very poor understanding of what "general intelligence" means, and what humans have that current AIs don't, so all discussion here will necessarily be quite speculative and vibes-based. My vibes are that introspection is a big part of what makes AGI able to do what it does. The ability to notice that we're making an error, think about the cause of the error, and come up with a mitigation, seems crucial to me.
e.g. when I try vibecoding agents, they reliably stop working above a certain complexity, and a lot of that seems to be because they lack any drive to create organizational systems, streamline their process, take notes, etc. It just doesn't occur to them that they can modify their own process.
IIRC Eliezer has said something similar.
I agree that to some extent all safety work advances capabilities, since it requires us getting a better understanding of how AI works. But why is this one a high safety:capabilities ratio? That seems backwards to me; making AIs more introspective will make them more able to do things like give fake outputs in CoT traces when they're trying to deceive the human, without necessarily giving the *humans* any more insight into the AI's thought processes.
Like, if I'm in charge of a country and there's a powerful foreign country that may be friendly to my country and may be hostile, I'm not sure, which intervention makes more sense to you: my country should try to get access to more information channels from them so I can hear what they're discussing internally? Or my country should give them better communications technology so their officials can better coordinate among each other?
Great list.
LOLed at "after we start talking about becoming bodiless immortal machine-gods".
Also my new word that will be deployed until household members become irritated enough to request that I stop is, "corrigenda".
Hi Scott! I'm one of the grantees (Adam Morris), and I think you forgot to add the sentence I requested at the end of the blurb about requests for potential collaborators. It was: "Adam is excited to chat with potential collaborators who have experience in technical AI safety work (especially in interpretability, CoT faithfulness, and fine-tuning frontier open models); reach out to him at thatadammorris@gmail.com." Just letting you know in case you can still add that in.
Also, just to signal boost here: I'm excited to chat with potential collaborators! So if anyone with an ML background is interested in studying introspection in LLMs, reach out to me :).
Sorry, have fixed.
Thank you! (And thank you again for running the grant program! Super excited to keep working on this.)
It's truly horrific that in the Congo, even after armed groups have taken over, you still need prescriptions to be able to get medicine.
I suspect this isn't true in the sense you're thinking - in every poor country I've been to, most medications are available over the counter. Prescriptions are helpful either as a guide so patients know what to take, or to access a small subcategory of controlled substances, or as some kind of interface with aid programs so that the programs know what people need. That having been said I don't know for sure how this works in Congo.
> study truth-seeking and bias in LLMs
Some interesting research by CrowdStrike on bias in DeepSeek. Not sure if you’ve seen it. Depending on contextual modifier, output code can contain more vulnerabilities.
ACX Grants 2027 will be wild.
I look forward to reading about the snake project when it stops being secret. Edible robotic mice?
Good question! Maybe sourcing from a more ethical part of the food chain, like barely edible pig parts or insects, put into an artificial mouse package? Some types of snakes only eat mice and rats and some will even eat mealworms. I suspect transitioning those mice eaters to something more ethical will take research and development and scaling - ie. will these bits of chicken fool my pet ball python, and how.
There's an entire question here about should snakes even be pets? I don't know how well a python kept in a tank lives compared to its life in the wild, and I tend to think that the really, uh, invested in their hobby types tend towards the crazy (yeah I know it's Chinese Robber Fallacy, but the stories about "and please be on the lookout for an escaped Burmese python which might be crawling in your window" don't reassure me*)
*Though that story had a happy ending, as one neighbour was equally a snake nut and didn't mind the snake at all, just picked it up and brought it back to the owners:
https://www.bbc.com/news/uk-england-hampshire-62678259
"Jenny Warwick said she looked out of the window at 05:15 BST and spotted the huge yellow snake on her neighbour's roof but did not realise what it was until she saw it trying to get in the upstairs window.
She said: "I saw them trying to poke it out and it fell on their car. It was massive.
"People passing by were staring and couldn't believe their eyes."
Linda Elmer, who recognised the snake, said she was woken at 07:00 by worried neighbours banging on her door, trying to find the owner.
She said: "Everyone was panicking.
"I picked him up - no one wanted to help. It was very difficult because he's a big snake and I managed to hobble down the road with this 18ft python wrapped round me and knocked on the owner's door.
"I think it was a shock for all of us.
"He's beautiful and Burmese are very docile. They're not aggressive snakes anyway and I had one myself so I was comfortable picking him up."
The RSPCA previously urged owners to keep snakes securely contained during hot weather as the warmer temperatures make them very active and more likely to escape."
I still think pythons etc. would be better off in their natural, native habitats.
If you want to learn something about snake keeping, check out Clint's Reptiles. https://www.youtube.com/results?search_query=clint%27s+reptiles
Lots about knowing which snakes need what, and what it takes to supply it.
Also, snapping turtles are docile and friendly if they're supported from beneath. not that I'm trying it.
*Venomous* snakes sure as heck shouldn't - and snake-owners seem to be impressively good at letting them escape. I'm neutral on the rest,
I came up with an idea similar to Dan Elton's metascience observatory a year or two ago, but I knew I didn't have the skills to lead a project like that. Even so, I'd be happy to contribute to it. I tried to contact Dan Elton by following the link, but I couldn't find an email address. I'd be happy if someone shared it with me.
My idea was to develop an index of how badly we're over-relying on various scientific conclusions. To calculate that for a given conclusion, I'd aggregate the impact scores of all studies that claim to support that conclusion, then "divide" it by an aggregate statistical power of all studies that analyzed that research question (regardless of whether they supported the conclusion). I think the most difficult part would be reliably classifying studies by their research question (i.e. figuring out which studies were attempting to replicate which others), but LLMs could speed up the process a lot.
My idea seems far beyond the scope of Elton's grant, but I wouldn't be surprised if he had similar ambitions. Either way, I respect his efforts to assess the reliability of scientific results.
> Aaron Silverbook, $5K, for approximately five thousand novels about AI going well.
I'm not sure how this is supposed to work. Presumably, these novels would need to be just shoddy enough to bias the next round of AI training, without biasing any humans who read them. And the novels do need to be available for humans to read, otherwise they'd be unlikely to end up in the next round's training corpus. But if the LLMs are already as intelligent as humans (as most proponents of this approach tend to claim), then wouldn't the shoddy propaganda fail to "persuade" them too ? Even if the LLMs are nowhere near human levels of comprehension, wouldn't they still be able to extrapolate from all the negative human-written reviews of the shoddy novels ? The obvious answer is "yes which is why we'll fake the reviews", but it would be hard to fake the lack of reader participation -- which is also not impossible to fake, but realistic fake users aren't free. The more I think about this project, the more expensive it appears to get...
Aren't LLMs text-prediction systems? They don't read books, write a review, then forget about most of them.
They are, but presumably they would grant more training weight to documents that are often referenced and quoted, and reviews are a part of that.
"Presumably, these novels would need to be just shoddy enough to bias the next round of AI training"
That's easy to fix, just prompt your chatbot to write them for you 😀
If the shoddy AI slop novels work a little bit, in the next round of ACX grants, Scott can fund $50k for a bunch of unemployed creative writing MFAs to write slightly less bad sci-fi novels about AI going well. Then if *that* works better, the 2027 edition can grant $500k to a crack team of talented authors to churn out a bunch of really good novels, which should arrive just in time to save us.
It needs to be pointed out that Palestine Action were banned because they broke into a base and damaged British military aircraft not for a free speech related reason.
https://www.theguardian.com/uk-news/2025/jul/03/four-remanded-in-custody-aircraft-damaged-raf-brize-norton-allegedly-palestine-action
I understand this, but I was under the impression that it was not only illegal to join Palestine Action, but to say that you liked them.
That is correct, UK has proscribed organizations which are illegal to advocate in favor of, unlike in America where you can eg proudly support Al qaeda
The UVC lamps sound neat, but won't diseases just adapt to resist the light after a bit of time? Can someone who knows much more about biology and this technology explain to me why I'm wrong?
More importantly, if these lamps can burn out bacteria, wouldn't they also burn humans ? What happens if you look at one with the naked eye ?
Put them in the air vents.
That's the cool part. The wavelength (222 nm) doesn't penetrate very far in tissue and doesn't damage skin or eyes. However because air droplets containing bacteria and viruses are small, they get fully "cooked".
See: https://pubmed.ncbi.nlm.nih.gov/30947566/
My only potential concern is the light generating ozone, which could end up being toxic if the light is run a lot in spaces without good ventilation. So you'll end up needing ventilation anyway.
Our immune systems have been killing pathogens by heating up into fevers for a long time. "Evolve to survive extremely high temperatures" is possible, as evidenced by extremophiles that live near hot vents, but it's easier said than done, so not really found much elsewhere. Similarly, resisting UV light would be hard to do, and anything spending the metabolic cost to do so would be at a disadvantage in any other environment.
Thanks!
Patio11 did a great podcast episode on UVC research and potential.
Complex Systems with Patrick McKenzie (patio11): Killing viruses with light, with Jacob Swett
Episode webpage: https://www.complexsystemspodcast.com/killing-viruses-with-light-with-jacob-swett/
This is an excellent list. Thank you Scott for doing this work! I wish all the grant recipients much success!
I was a bit confused whether my application was received or not. Were we supposed to get a confirmation email after sending out the form? I didn't receive any, and when I tried again, there was none either.
A shame to see my pitch, "Weed for Dogs", rejected yet again. Maybe $100k was asking too much? There's always next year
Hang in there, Russ. Your time will come. lol
You need a rebrand, get marketing on it. Pitch it as "Canine Calming Aid" or the like 😁
I dare you to read through this list, or even just scan it, and tell me who in the world is a better person, dollar for dollar, than Scott Alexander. And this is why you should consider paying for this newsletter, if you can afford it.
Hyperstition AI lists The Moon Is a Harsh Mistress as a story where "humanity faces its demise at the hands of its own creation: artificial intelligence." This is an error. The only AI in the story is trying to improve the living conditions of the people of Luna, to the point of sacrificing itself to protect them from attack. It's not "aligned to humanity" the way we describe it today, since it's written as a character with free will, an internal world, and multiple desires, but it's definitely working towards a happily-ever-after.
Re germicidal (far-UVC) lamps, they may be a great thing for specific purposes and places (hospitals) but beware of more general deployment.
Our immune system is metabolically expensive. If we make large parts of our world safe from pathogens, over time our immune systems will adapt and get weaker.
The "hygiene hypothesis" is the leading explanation for the origin of the epidemic of asthma and severe allergies (peanuts, latex, etc.) in the first world. People who lived in filthy peasant conditions didn't have those (particular) problems, and children who grow up on farms still don't.
Children who grow up in cities where people believe "cleanliness is next to godliness", do.
the "epidemics" of asthma and severe allergies are a trivial fraction of human suffering compared to contagious disease.
Scott, is there anyway you might publish your list of great ideas that didn’t make the cut so that some of us readers might be able to find them?
Congrats to all the new ACX Grantees! Pretty excited about the low-cost ultrasound project. The tech has made so much progress that it no longer makes sense for medical ultrasound scanner to be so expensive. Subhash, please let me know if I can help you with anything (FPGA, DSP, electronics, manufacturing, etc) (https://alextoussaint.com)
Well, I must have been stunned by a passing comet or something, because this year I think the vast majority of the grants are reasonable and sensible.
I'm neutral on the AI ones (I can't see that writing 5,000 cheery fanfics about 'and then the AI and the humans lived happily ever after' is going to make much difference one way or the other, but hey it's harmless) and as for the African urbanism one, I venture to suggest "Not enough YIMBYism" may not be *the* major problem behind "why can't we get proper infrastructure projects off the ground?" there. But that's why we need research, I suppose!
The shrimp-stunning one makes me laugh, in an evil way. Now I'm visualising tanks of liddle shrimpies getting individually stunned at varying levels of shockeration in order to find out how high do you need to go to stun a shrimp sufficiently, and how can you tell if a shrimp is stunned?
The Christians for Impact one puzzles me. How is this different from the usual charitable programmes around churches and denominations, and there are a metric ton of those already out there, including ones which started off with the "unite the believers and the non-believers who want to do good in action", e.g. Oxfam, or which have broadened out from their original religious inspiration (e.g. Depaul, which seems to have spun off from the original Society of St. Vincent de Paul)?
Looking at the website, I'm not entirely sure what they want or intend to achieve (and some of the language is a turn-off even for me, a believer: "Discuss evidence-based ways to tackle important, nonpolarized problems near God’s heart." Oh, so you can read God's heart, can you? Nice to know you have a hotline to the Throne!)
Are they trying to get non-believers hooked up with church charities, or are they trying to get believers into the kind of secular, political, high-powered jobs like 80,000 Hours? I mean, if I look at the page about the "problems near the heart of God" it's climate change, global poverty, disease, etc. and organisations like Concern, Trocáire, etc. have those covered. The "Christian software engineer tackling climate change" - what makes it different that he's a Christian? Is the idea that as a Christian he wouldn't have thought about undertaking this? How is his Christianity an influence on his career decision there, because there's a lot of the secular EA folks going down that path already?
The best sense I can make of this is that it's outreach to Christians to persuade them "hey, if you want to do good and get involved in charitable causes, don't go the traditional routes through your local church or denomination like going on the missions or whatever, instead get a high-paying, high-powered job and work on Big Thinky Thoughts projects" very much in the EA mould.
And I don't quite know how I feel about that.
EDIT: Actually, now I do know how I feel about that, and it's "Oh boy, here we go". Because I hadn't read the pertinent part of the website, and it's just dressing up EA aims and goals and procedures in a light coating of "speaking to your heart, American non-denominational church style":
"You have 80,000 hours in your career—40 hours a week, 50 weeks a year, for 40 years.
That makes it one of your primary resources to love and serve others and honor God.
We’re here to help you make those hours count.
We provide prayerfully-researched advice, tools, and mentorship that empower you to tackle the world’s most pressing problems.
Our advice is especially focused on students and graduates aged 18–35 in the US or UK who are blessed to be able to choose a career based on impact and are keen to tackle problems we find most pressing."
80,000 Hours. Did I call it, or what?
I took a look at the Aerolamp prototype and maybe this is a dumbass question, but how does a tiny lamp like this have the necessary range and dispersion to effectively cover even a medium sized room?
This is Jacob Witten, sorry to nitpick but I'd say more precisely "his work is still in stealth" as there's no startup (yet! working on it! everybody should feel to reach out if they're interested in curing horrible lung diseases)
From Kurtis Lockhart's Asterisk article:
"The number of urban planners per 100,000 people in the UK is about 37, and in the United States is about 12. For OECD countries, the average is 21.5."
Is that actually an insight on the UK's housing market?
Hmm...
"Kasey Markel, $10K, for genetically engineered corn. Kasey and his team at Semilla Nueva use prime editing, a new genetic technology, to create corn which is rich in zinc, iron, essential amino acids, and other nutrients frequently deficient in corn-heavy poor country diets. Our grant helps fund greenhouse space, enzymes, DNA synthesis, and scientist time, and will let them expand faster into new regions that require corn with different genetic backgrounds."
Let's say this works perfectly. Will this be used even then if it doesn't offer financial benefits, especially with Greens fighting it tooth and claw every step of the way?
"Five out of ninety-two is a crazy result"
You think it's crazy that 5% of papers fabricate data? I'm pleased if it isn't *far* worse!
This looks like a fantastic tool, though, agree. I imagine people will just train LLMs against it for undetectable fabrication, but it will still catch tons and tons of old ones.
G 50k snakes.
I hope this has nothing to do with the solutions proposed in the lesswrong or eaforums article you shared previously about snake feeding which frankly showed no subject matter knowledge and was totally off base. If it's artificial or plant based snake feed, that won't work biologically, way more than 50k has been sunk trying to make it work but I don't think it's possible.
In the comments to that article I suggested that an ethically raised rodent supplier would be a good idea and probably a good business, and would be a cool acx grant, if this is that then I'm very excited and would be a customer. So much of the price of feeder rodents is the shipping (express, dry ice etc) that even doubling the cost of raising the rodents is pretty minimal on the cost to consumer.
> Sam Glover, $60K, to fight for free speech in the UK. These are dark times for UK speech on both sides of the aisle: the left is upset that speaking in support of Palestine Action is now considered an act of terrorism, and the right is upset about arrests for racist tweets. So far, pushback has been siloed by cause and partisan affiliation. Sam and his two co-founders are early-career bloggers and aspiring public intellectuals who want to build a united nonpartisan free speech movement. They’re still in stealth, but I’ll promote their website as soon as it becomes public.
I very much support this. Far too many people object when it's speech by "their" side that gets suppressed but not when it's the "other" side.