It's not just onions that it's illegal to have futures contracts on; it's movie tickets too.

Robin Hanson has talked about this before:




Or, here's a more recent, longer article about it: https://www.theringer.com/movies/2018/11/15/18091620/box-office-futures-dodd-frank-mpaa-recession

So, Kalshi also won't be opening markets on what movies will do well, I guess.

(Yes, this is stupid.)

Expand full comment

On the AI GDP questions, you seem to have left/right swapped.

Expand full comment
Feb 8, 2021Liked by Scott Alexander

Huh, I've heard of "onion futures" being illegal before and didn't realise it was literally onions. I always assumed it was a type of future named metaphorically after onions - like futures on futures on futures on something - something with layers. TIL.

Expand full comment

Another prediction market site that right now only appears to have two things, but I vaguely remember being pretty liquid for the Trump election: https://ftx.com/markets/prediction (I have never used it)

Expand full comment

What do you think Metaculus would predict about its own demise? E.g. something like "predict the year in which Metaculus's monthly active users drops below 100 for 6 consecutive months." And how do you think this prediction effects people's other predictions? Do people care less about being right if they'll be right only after the site is no longer operating?

Expand full comment

On the AI-written book question - would a volume of short stories count? I would bet that GPT-4 could write a reasonably coherent short story, especially if you prune the more outre outputs .

Expand full comment

I think the most straightforward explanation for why the Polymarket Trump contract is still trading at 6% is that 1) people who aren't familiar with crypto are put off by the complexity and inconvenience of getting their money into USDC so that they can trade there, and 2) people who are familiar with crypto have no interest in locking up their capital for two months for a 6% return, because there are much better opportunities for yield currently available in that space.

Expand full comment

The May/June 2020 upward deviation in predictions (Will AI progress surprise us? / Date first AGI is publically known) coincides with the release of GPT-3.

Expand full comment

For "Date First AGI is publicly known" and similar:

It seems like this is a sort of problem where the time value of money really matters. I.e., a bet that yields $100 in 2025 is worth more than a bet that yields $100 in 2055 (even after adjusting for inflation), so there is the potential of having a huge thumb on the scale that encourages people to bet on sooner dates rather than later dates (or else encourages the "soon" predicters to bet and the "later" predicters not to).

How do metaculus or real-money sites account for this? Or do they even need to from a business perspective?

Expand full comment

Disclaimer: non-expert opinion; I got some USDC onto Polymarket last week and have read their docs and lurked their discord, but don't understand every detail of how it works.

Polymarket's transfer issues are recent, and the Trump-related markets have always been unreasonably pro-Trump (in my opinion).

The issue with the relayers is this: due to high trading volumes for ETH, it costs around $50 in gas fees to transfer USDC to Polymarket. Polymarket itself is picking up the tab for these fees, but they don't want to pay more than necessary. Therefore, they're only filling the gas tank when gas prices are relatively low, and they're not putting in enough for everyone who wants to process a transfer to do so.

Unfortunately, there is no queueing system, so your USDC will not be moved from the first relayer through to your Polymarket account until you click the button to do so when there is money in the gas tank.

There is no way to retrieve your USDC from the relayer except moving it through to your Polymarket account and then back out.

You can add money to the gas tank yourself to try to hurry things along, but it will cost at least $50 or so, and you should read up on the discord before doing so.

The Polymarket team is working on switching to a much cheaper transfer system, but it's unclear how long this will take.

Expand full comment

Why does Trump to be President trade at 6%? The question has two parts: (1) why is anybody crazy enough to bet yes on that, and (2) why arent people eager enough to offer it at a lower price.

Subquestion 1 is really an anomaly, but one that clearly exists in any prediction market and in the real world.

Subquestion 2 has an easy answer -- Polymarket might not have the value-at-risk cap and trader cap that PredictIt has, but it does have some real risk of being shut down or declared illegal at any moment. PredictIt has a no action letter from the feds, Polymarket does not. Some combination of being a pain to get started with, being crypto-based, not having any assurance from US authorities that the site will not find itself in some state of illegality, and other operational risks mean that nobody is willing to put a large amount of money there without some minimum expected return to compensate for those risks.

Expand full comment

As someone currently ranked in the top 50 on Metaculus having joined 6 months ago, and who has made >$100k on Polymarket having joined around 1.5 months ago, I have some thoughts.

First, on "But most of Metaculus is bread-and-butter questions about when such-and-such a system will meet such-and-such a benchmark, or how many parameters it will take, or things like that." Those questions are almost all from a specific contest run by the Open Philanthropy Foundation with $50k in prizes. See https://www.metaculus.com/ai-progress-tournament/

They do not actually represent the bulk of Metaculus questions, but it may appear that way if you sort in certain ways.

Second, Metaculus heavily rewards predicting on lots and lots of questions. If you just copy the community on every question, you gain net points in expectation, and do better if you can beat them, which you can do pretty easily just by looking at older questions where most predictors haven't updated recently. For two examples, see this thread https://twitter.com/avi_eisen/status/1357385286985121792. Often in practice, someone happens to see an older question and the community prediction looks wrong to them, they write a comment, which bumps up the question so a bunch of other people now update. Because of this dynamic, you need to be very careful if you're going to try to use Metaculus numbers as a prior/input into your model/etc. There are often fairly uninformed predictions simply because nobody has done a deep dive on the question for two months. I myself have predicted on 881 questions to date, and it's nearly impossible to update them all regularly. I try every other month or so and my brain just goes numb after scrolling through a few hundred and thinking which might have changed odds.

Third, as you can probably guess from my intro, it is absolutely worth it to figure out the kinks with Polymarket. The market is far from efficient. However, it is harder to beat than Metaculus, unless you're betting on the odds of a Trump coup.

Expand full comment

Scott, if it interests you, the site Papers With code has a nice plot that shows the state of the art on these problems, and explains the datasets. I think you might find at least some of the problems to be exciting.

The SQuAD dataset is a dataset of questions and answers based on wikipedia articles, and is described here: https://paperswithcode.com/dataset/squad

The Mini Imagenet is a subset of Imagenet, a dataset of a thousand classes of images. One shot learning is "learning to recognize X from a single image of X", which is pretty cool and non trivial thing (in my opinnion) for a neural network to do. It is described here:


Metaculus doesn't seem to expect any big changes in these domain in the next year, which is a bit unfortunate.

Expand full comment

"Presumably some AI-related advance short of AGI would cause an unprecedented world economic boom. I am pretty skeptical of this because there have been a lot of weird unprecedented things in economic history and they have never caused that severe a discontinuity."

We've never had one advance that affected every industry almost all at once; they've always had to diffuse out at the rate of human learning, human decision making processes and institutions, manufacturing and/or building infrastructure and logistics networks, and so on. That will still be somewhat true in a less-than-FOOM-takeoff, I think, but it doesn't mean there can't be a huge shock. Part of the value of AI is automating decision making, probably reducing those limitations. Wars and depressions do cause huge negative shocks, often followed by recovering faster than trend. In the Great Depression, US GDP fell by 30%. WWII was followed by a 30 year economic boom.

I will say I'm also quite confused about what I should expect to happen to prices and nominal GDP in a world where almost everything gets automated. My current understanding is that however indirectly, prices today are mostly about how fundamentally scarce the resources are to make something, and how much (and what kind of) labor goes into it. I expect asymptotically AI pushes the value of (non-human) labor to zero, and costs depend only on material composition and energy for production and maintenance, but the intermediate stages could be super weird in a slower takeoff depending on which labor components get zeroed out in what order and what prices of still-scarce-things we choose to bid up in response.

Expand full comment

I'm not sure why polymarket is so off but the Augur based prediction market is still fighting to resolve the Trump bet as well, Trump supporters seem to be throwing money at the dispute function of augur, which is only really able to delay the inevitable and waste their money as the cost to delay ramps up. If I had to guess why people are selling the nTrump tokens at 92 cents on the dollar it's because people want to re-enter the crypto market and may be willing to eat an 8% loss to get out of limbo. I'm really wishing I had sold my nTrump tokens when they were worth 95 cents after the election as I've missed out on nearly $10k in opportunity cost because the funds used to be in ethereum. which was worth less than $400 when I made the bet and is worth $1700 now.

There also may be some fear that the Trump supporters can actually break Augur, although I doubt they can.

The whole of the ethereum ecosystem is going through some very significant pain right now as the gas price is insanely high and the upgrades that will enable scaling are months or years away. It's an unfortunate time to be looking at ethereum based prediction markets, huge fees just for moving coins around let alone interacting with smart contracts that multiply fees makes the whole practice difficult and probably contributes to the current bad price prediction.

Expand full comment

Just wondering, did you find a way to do Etherium staking?

Expand full comment

"There will be an ambiguous resolution if there is no 4 year (nor 1 year) doubling interval by 2050, to isolate specifically the takeoff speed from other things like the chances of no takeoff occurring at all or human extinction." (from the "world output doubling" question)

I love the scrupulous thoroughness of specifying how the bet will resolve in the event of human extinction.

Expand full comment

"Always bet as though the market will keep existing" is a standard piece of investment advice. Markets suck at pricing in nuclear wars that will make the markets stop trading or make the traded shares worthless.

In a similar way, there's an incentive to predict that AGI (or nuclear war, or super-plague, or asteroid impact, or supervolcano eruption, or whatever other catastrophic or existential threat) is a long way off, because every year with no AGI is a year in which you get to boast about your insightful cyncism, whereas the year you're wrong you have much bigger things to worry about than Fake Internet Points.

Expand full comment

Once again I am a̶s̶k̶i̶n̶g̶ ̶y̶o̶u̶ ̶f̶o̶r̶ going to mention my suspicion that the eerily straight lines that economists draw through the last 300 years of economic progress are due to the economists having carefully chosen a lot of tuning parameters so that the lines would be straight, and not at all proof that discontinuities and rapid surges in economic productivity/wellbeing don't happen. I am going to keep mentioning this until somebody replies to my objection in a convincing way. Until then, I am going to assume "GDP over time" charts are a hilarious in-joke amongst economists.

Expand full comment

I'm extremely skeptical of AGI (meaning that I think the currently projected timelines are ridiculously optimistic), but sadly I'd have to bet on the "AI NYT bestseller by 2030" claim. Unfortunately, this problem is being solved from both ends: AI is getting better, and bestsellers are getting worse. It is not inconceivable that, by 2030, just having a semi-coherent book with mostly correct grammar would be sufficient. That's just my subjective opinion, admittedly.

Expand full comment

I wish there were a working combinatorial prediction market, real-money or not.

It is of course computationally intractable in theory but I think people can make it work? There seems to be some academic research in this area (seemingly from a few people at Microsoft), but no currently-active public implementation.


This would solve the obvious problem, very apparent on PredictIt, where one combination of events is equivalent to some other combination of events, yet they have different prices.

It would also just be very cool, and probably fun to trade on (although of course less opportunity for arbitrage).

Expand full comment

The Trump markets (and related markets such as the John James market) really shook my belief in prediction markets. I defended them when they got the 2016 presidential election wrong because almost everyone else also got it wrong. But I couldn't believe it when people continued to bet on events that had already happened, elections that had been certified well beyond the threshold for recount.

Expand full comment

So, the tragic fundamental issue with prediction markets is that they require ETH smart contracts. Ethereum gas fees are incredibly expensive. There are many different approaches to fixing the scaling problem here, with plasma and optimistic rollups being some of the more promising ones in development. Ethereum v2, which is apparently going to start rolling out this quarter (!?!), should also greatly improve scaling in a way that should trickle down to smart contract prices. (On the flip side, demand could increase sufficiently that unplatable fees would still be an issue).

Polymarket in particular is completely centralized and has broken prices in a few different ways, with many markets being illiquid or otherwise stuck (though it might use being centralized to lower some fees? not sure). The much more interesting alternatives are Omen and Augur, with Augur being the only completely decentralized resolution market. https://omen.eth.link/ used to work but doesn't come up for me right now; last I checked augur was more or less unusable on its own, but clever people managed to set up a Balancer market to do this with relatively lower fees at catnip.exchange.

I expect Augur to become more impactful in the next year or two as Ethereum gets its shit together.

Expand full comment

I've been using Polymarket since summer and have earned a significant amount of money on the election's market.

Regarding 6% chance of Trump's comeback, there's a general reason why prediction markets can't be very precise. This 6% market translates into actual 3-4% of profit, and for a market that runs for any length of time this might be worse than just investing your money in S&P. This is not the case for this particular market though, since it will be resolved fairly soon and 3-4% profit in a month is quite a good deal.

Expand full comment

I’m curious about something regarding the idea of taking prediction markets seriously, hoping someone more well-versed in this will indulge me—

Doesn’t the theory of prediction markets kind of rest on the notion that people will be rational actors placing their bets in such a way that it’s their sincere best guess at whether the prediction will come true, so they can make the maximum amount of money - and doesn’t this make them very vulnerable the moment they have consequences beyond who wins the bet? Suppose a government decides to implement COVID policy based on a prediction market - won’t that immediately just cause people to flock to the markets to “predict” that whatever policy they want enacted will save more lives? At that point, isn’t it just kind of a roundabout way of voting on the policy, except now only people with money to burn get a significant vote?

Expand full comment

TL;DR: market price does not map that easily to probability for some contracts, including all AGI/GDP related questions. If we have a simple binary contract paying 0 or 1 on a given date depending on whether a certain event occurs, price of this contract in a liquid market may not reflect market perception of event probability for several reasons. The main reason is that if the outcome is negatively correlated to the value of money, market price will underestimate the probability and vice versa. For example, if we had a contract on the event "Earth is wiped out by hostile aliens before 31/12/2021", a rational actor with unlimited free collateral would be ready to sell this contract at any price irrespective of perceived probability, as there would be no state of the world in which they would have to pay out anything. A more subtle example closer to the AGI questions would be a contract on "Inflation in USD is above 100% per annum in 2021". If the contract settles at 1, the value of a dollar won by a long contract holder is less than 50c in today money; if the contract settles at 0, the value of a dollar lost by a long contract holder is likely to be much higher. Thus, a rational market participant would bid for this contract far less than the perceived probability. Now, today value of $1 in a year of 100% GDP growth caused by AGI is pretty uncertain as nominal prices will be all over the place in such a year. Inflation for some consumption baskets may be -99% and for other consumptions baskets it may be +5000%. So, at best, contract price reflects the joint distribution of likelihood of the contract event and marginal market participant predicted inflation for their consumption baskets. At worst, the price does not reflect anything as people as not very good in understanding this convexity.

Expand full comment

Am I reading it right on the 'Will AI surprise us' question that it's predicting a 75% chance it will? This is based on whether other prediction market bets will move? If the market is 75% sure that those other bets are off, why haven't they corrected themselves to reflect that? Is this just an artifact of the low volume/market inefficiency nature of this and other prediction markets that could be corrected if they were made generally legal and widespread?

Expand full comment

There's also the Foresight Exchange , which is still fake internet points but at least has been around a bit longer. E.g. http://www.ideosphere.com/fx-bin/Claim?claim=GoCh

Expand full comment

GDP growth is far more likely to be limited by the rate of change of human behaviors (slow) and human values (even slower) than by technology or knowledge. You'll probably have to wait for a few generations to die out before you can possibly see any kind of real transformational change.

For example, the growth of lab-grown meat is going to be heavily limited by acceptance, not price or quality. There is nevertheless a stable equilibrium where lab-grown meat is all anybody has ever known and at that point, the prospect of taking a cow, growing in a way that is hugely destructive to the environment before slaughtering and dismembering it, only to obtain something that is indistinguishable from what you already have but far more expensive, will seem entirely preposterous. It will take several generations to get there, however. Superintelligence will not change this.

Similarly, approximately half of all wealth in the US is housing wealth. Think about this for a second! Half of all the stored value in our society relates to where and how we live. It is not the lack of a superintelligence that is preventing this from growing. Growing the amount of housing is going to hit hard caps in terms of land availability. Growing the value of housing at constant size is going to hit caps on the amount of value a person can get from a small living space. Growing the non-housing share of GDP is going to involve getting people to care about things other than their small, crappy house, but this is going to be a huge and slow shift in human values. There may be higher value equilibrium states that we could reach, but getting to them will involve reorganizing extremely large numbers of people in terms of where and how they live. This will not happen quickly, no matter how many supercomputers you have.

Expand full comment

Oh wow. This Cyberpunk novel on the Onion Futures Ban writes itself. According to wiki, In 2010 the motion picture association of America successfully had box office receipts added to the futures ban. Apparently some financial firm called Cantor Fitzgerald planned to start a real money Box Office receipts futures market in the 1990s, but the firm's corporate office was destroyed on 9/11. They were just two floors up from where one of the planes hit. Later this guy named "Robert Swagger" tried to start Media Derivatives Inc in 2007, but then the financial crisis hit and the MPAA got box office futures banned as part of the Dodd-Frank Act. What are they trying to hide?


Expand full comment