Lots of people have talked about this before, but I'm putting it here to gather my thoughts:
What's the dream with prediction markets and AI? Here's a vision:
(1) Train LLM agents to be like superforecasters but better and cheaper. E.g. whereas it would cost a million dollars a year to have a team of top forecasters like Samotsvety weighing in on dozens of markets a week, with AIs maybe it costs only a thousand dollars a year to get a similar level of intellectual labor.
(2) Then, play-money prediction markets like Manifold can get turbocharged by AI bots. If someone is willing to pay a million / yr for AI compute, then there can be a vast liquid market with thousands of questions live at any given time. (For comparison Manifold has something like 10k open markets I think). So basically people can pay the AI company a million/yr to have their own Manifold on topics of their choosing, with higher-quality answers too.
(3) How would the incentives work for the AIs? Ideally these would be special fine-tuned versions of the flagship models, that are trained purely to maximize their Manifold mana / prediction market play-money. This hopefully prevents the AIs from being biased in various ways & also just continually trains them to be more accurate. You could use AIs from multiple companies (or better, open-weights models) to reduce the risk of one company secretly manipulating a market, not that that's a particularly big risk probably.
(4) How would the incentives work for the humans? Well, you probably want something like: Humans pay real money to create a market and subsidize it with play-money; AIs then compete to win that play-money by predicting how the market will resolve; the human who paid, resolves the market. (What if they are biased or terrible at resolving markets? Well, then they'll develop a reputation as such, and AIs will be more cautious interacting with their markets in the future, and they'll learn less from them therefore. Also, it doesn't benefit them in any way to spitefully do poor resolutions to your markets, all you are doing is hurting some poor AIs, you won't get your own money back.)
(5) The upshot is that it seems like this would allow individuals and organizations to efficiently ask questions about the future, of their choosing, to the AIs, for cheap. Like, you could spend $100 and get an amount of intellectual labor equivalent to Samotsvety working for more than a month trying to answer your question.
(6) Why is this better than just paying for API credits and having the AIs research your question for you?
--Well, it's more public. Positive externality of having the market that anyone in the public can view.
--OTOH that's a downside as far as some orgs and some questions are concerned; maybe for competitive reasons they want to keep secrets.
--Another advantage is that the aggregated wisdom of a bunch of prediction AIs specifically incentivized to get it right (since they are in a training loop) is probably more accurate than the individual wisdom of a group of AIs you ask to do it 'for free' (i.e. they aren't going to be trained on the accuracy of their response.) Three things going on here: Aggregation improves capabilities, training improves capabilities, and training decreases bias.
(7) I think this could maybe be a viable business model. Charge like a 5% fee or something, so if the answers are even slightly better in expectation then it is superior to using API credits. And they'll probably be better for the reasons above. And then a nice positive externality will be created, since the results of this knowledge creation will be public.
"make better brier/loss score"
Has anyone really been far even as decided to use even go want to do Garrabrant induction? IDK what this would look like, just generally asking. I assume there'd be significant issues, maybe e.g. dealing with a non-fixed total money pool.