Regardless of the actual singularity, we have definitely hit one in arXiv submissions.
It is becoming apparent that the current paradigm of scientific knowledge production is not ready for AI. Accelerating research speed and expanding access means there is a lot more slop and there is also a lot more good work, but way too much for the current peer review system to delineate between. We need some accurate and decentralized way to aggregate this information.
This makes a scientific prediction market, like Robin Hanson's idea futures, really appealing. The community can have markets for broad questions we care about, bystanders get paid for pointing out good research, and scientists are financially incentivized to make progress on the highest leverage issues. But with Hanson's conception, there is one big problem, which is how the market gets settled. For abstract scientific questions, like whether models are reward seeking or fitness seeking, we have no definitive event which can force market settlement. Hanson suggests that we use the convergence of scientific consensus, but this proposal takes trust in possibly misaligned opinions and requires the centralized coordination of scientists, which is hard to create---especially in pre-paradigmatic fields like AI safety.
In this post, I'll lay out a theoretical framework for a new kind of scientific prediction market. TL;DR: Scientific questions become perpetuals where the fair price is the expected direction of future evidence.
Here are the mechanics:
There are two kinds of markets, perpetuals on ideas and experiments that settle on their outcomes.
Anyone can create an experiment market. In the market, they are allowed to insider trade. That means, when their experiment is done, they can take risk free profit by trading the remaining liquidity in the market. HERE IS THE TRICK: when setting up a market, a scientist must commit to using their profits towards buying puts or calls on specific perpetual, dependent on the outcome of your experiment. For example, if I am testing whether Qwen 4 has generalized its math skills, I commit my profits towards buying calls on the idea perp "LLMs generalize out of distribution" if Qwen does generalize or puts if Qwen does not.
On the options structure: puts and calls have no expiration date and can be exercised immediately. The value of those options to the scientist is the expected price move indicated by their work: imagine a scientist who does an experiment, puts their profits towards options, publishes the work, then captures delta because their work moved the market. Even if their research does not move the market itself, it may be informative about future research. On the contrary, the options derived from an experiment which is poorly predictive of others' results are worth much less.
To hedge directional risk, options dealers must buy the idea future in the direction of the experimental outcome.
Because of the above structure, future marginal demand for a perpetual is forecasted by the expected wager of future experiments. If you expect that LLMs generalize out of distribution, you can buy the idea perp, because you expect future experiments to generate hedging demand from options dealers as scientists return positive outcomes.
note: above I described scientists profiting and putting those profits towards options, for simplicity. In an actual implementation, they would collect their net profits in the form of equivalently priced options when they submit their experimental result. The direction of those options is preregistered when they create the market.
This kind of prediction market has some really nice qualities. Below are some possible outcomes.
First, there is a financial incentive for someone to filter through the ever growing slop of paper production. Finding good new research is alpha, and so is pointing out bad research.
Second, liquidity focuses on experiments which matter. One objection you could have is that a random person can create an experiment market and take a risk free profit. This is likely not the case. The reason to trade on experiment markets is that, if the experiment actually matters to the price of the idea perp, no-arbitrage conditions force price discovery on the potential outcome of the experiment. Traders let market makers capture flow, and resting liquidity pays the scientist. Now, scientists get paid for adding information on things we care about.
Third, scientists are incentivized to do good, careful work. After pre committing your profits to the result of your experiment, it is irrational to do bad work, because your option's value is dependent on predicting the right direction of future results.
Fourth, good new scientists can get paid even if they are not recognized. We can imagine an undergrad doing great work, but with no liquidity in their experimental markets. If their experiments are, in fact, informative about future results, they still can put capital on the right side of perps and profit. They could then bootstrap their track record into reputation.
I could go on, but the epistemic goods of markets are well discussed on this site. I am interested in what thoughts and critiques readers have on the market structure here.
Seems like the kind of thing that AI superforecasters would be ideal for. They can provide liquidity and investigate the massive amount of research. (Today, there would also be a human role in identifying valuable experiments that AIs tend to overlook, for whatever reason.)
In general if you really wanted this relatively complicated setup to get traction, you'd probably want to make it so human scientists can easily use it via their own AI agent(s). That this also lets AI scientists easily use it later is probably a benefit; one of the bottlenecks to getting good ex. medical results from AI is sifting through all the tentative results to try to find what's actually useful. A market price could provide some signal.
Regardless of the actual singularity, we have definitely hit one in arXiv submissions.
It is becoming apparent that the current paradigm of scientific knowledge production is not ready for AI. Accelerating research speed and expanding access means there is a lot more slop and there is also a lot more good work, but way too much for the current peer review system to delineate between. We need some accurate and decentralized way to aggregate this information.
This makes a scientific prediction market, like Robin Hanson's idea futures, really appealing. The community can have markets for broad questions we care about, bystanders get paid for pointing out good research, and scientists are financially incentivized to make progress on the highest leverage issues. But with Hanson's conception, there is one big problem, which is how the market gets settled. For abstract scientific questions, like whether models are reward seeking or fitness seeking, we have no definitive event which can force market settlement. Hanson suggests that we use the convergence of scientific consensus, but this proposal takes trust in possibly misaligned opinions and requires the centralized coordination of scientists, which is hard to create---especially in pre-paradigmatic fields like AI safety.
In this post, I'll lay out a theoretical framework for a new kind of scientific prediction market. TL;DR: Scientific questions become perpetuals where the fair price is the expected direction of future evidence.
Here are the mechanics:
note: above I described scientists profiting and putting those profits towards options, for simplicity. In an actual implementation, they would collect their net profits in the form of equivalently priced options when they submit their experimental result. The direction of those options is preregistered when they create the market.
This kind of prediction market has some really nice qualities. Below are some possible outcomes.
First, there is a financial incentive for someone to filter through the ever growing slop of paper production. Finding good new research is alpha, and so is pointing out bad research.
Second, liquidity focuses on experiments which matter. One objection you could have is that a random person can create an experiment market and take a risk free profit. This is likely not the case. The reason to trade on experiment markets is that, if the experiment actually matters to the price of the idea perp, no-arbitrage conditions force price discovery on the potential outcome of the experiment. Traders let market makers capture flow, and resting liquidity pays the scientist. Now, scientists get paid for adding information on things we care about.
Third, scientists are incentivized to do good, careful work. After pre committing your profits to the result of your experiment, it is irrational to do bad work, because your option's value is dependent on predicting the right direction of future results.
Fourth, good new scientists can get paid even if they are not recognized. We can imagine an undergrad doing great work, but with no liquidity in their experimental markets. If their experiments are, in fact, informative about future results, they still can put capital on the right side of perps and profit. They could then bootstrap their track record into reputation.
I could go on, but the epistemic goods of markets are well discussed on this site. I am interested in what thoughts and critiques readers have on the market structure here.