Looking at how large the unsolved list is for Frontier Math open problems, I get the sense that despite the amazing successes AI models are not close to superhuman general mathematical problem solving.

The Frontier Math open problems list contains a lot of questions that are more attention constrained than truly unsolvable by humans. And current SOTA models cannot solve the vast majority of them, which shows how jagged the frontier is.

With that being said, I’m sure models will improve a lot more by 2028

@JaundicedBaboon doesn't sound like there's much "new mathematics" / "new paradigms" coming out of these models yet. Maybe some hard problems don't get solved until models reach that threshold.

How do you define "area"?

This seems inevitable. I think the main question is if a model will successfully become superhuman via reinforcement learning by the end of 2027.

A secondary question is if we're actually going to be able to do much given that the pace of discovery will likely be much faster than the pace of understanding. We will likely know if something is true or not but not know what it implies or how it could be used.

@DanielLittQCSn would you consider rephrasing the description slightly? it talks about outperforming humans in all mathematical research areas by 2028, but then this paragraph specifically makes the requirement for a yes resolution sound much weaker:

I will resolve this market YES if I think publicly available AI tools outperform most professional mathematicians (e.g. me) as of July 17, 2028, and NO otherwise. In particular, for a YES resolution, they should be able to create high-quality original research with very little human input.

I find it plausible that AI tools outperform most professional mathematicians on average or in terms of net utility or wtv but they are worse at many specific things. Also, that they can create high-quality original research in most domains but struggle in some niche domains that eg require some sufficient level of continual learning (like where you need to have spent many hundreds of hours competitively playing in a game to understand or think of a certain math-related dynamic) or good long term multimodal perception. Assuming that in those cases the market resolves NO (but lmk if that's not the case), I think this paragraph might be misleading? basically to rephrase, the vibe i get from the first sentence vs 4th parapragh seems pretty discordant

@Bayesian when I say “outperform” I mean in all areas, including domains requiring continual learning etc. I.e. if there is any math research capability I have that I do not think can be elicited from a publicly available AI tool the market resolves NO. Does this agree with your understanding?

@DanielLittQCSn Yes, ty!

@DanielLittQCSn If there is one area where they're better than mathematicians who don't specialize in it but worse than most mathematicians that do specialize in it, then AI could be better than most mathematicians in every area without being better than most of the specialists of each area in every area. Is that what you meant?

how do you want this to resolve in the even that you are unreachable?

@jim N/A I guess.

fill me

The anthropic shadow is too strong for me to want to participate, but it's a good question.

@JoshSnider Be assured that if I see a swarm of nanobots approaching, turning everything in their path into paperclips, I will resolve the market YES.

@DanielLittQCSn a swarm of nanobots approaching turning everything in their path into proofs of the Riemann hypothesis

@JoshuaTindall We will know whether P = NP by measuring how much of the world AI has to turn into data centers before it knows how to put oranges in a box.

@JoshuaTindall This is just viXra

Are you considering only the task of proving theorems, or also that of building theories and conjectures? In my opinion, while it's possible for a system like AlphaProof to become superhuman at proving theorems by 2028 (though it would be very difficult), I believe such systems are still very far from being able to create definitions, make conjectures, and build theories.

@Grothenfla The question is about "all areas of math research." This includes creating definitions, making conjectures, etc.

(commenting so i can repost)—good personal judgment market