priced way too low IMO.

OpenAI won Atcoder heuristic and algorithmic world final by a large margin.

hmmm, I don't think that it should be this high because last year's imo was easier by a good margin. I asked chatgpt and turned on research mode and stuff and it said that it's probably around 30%. Gemini seemed to agree too. I'm not sure though

@Chimpy gpt 5.6 and whatever the new Gemini is are probably pretty good at math

@jim this is true

@Chimpy if you believe it, bet no

@DottedCalculator I did. I also want to believe in AI though

@Chimpy what if they just lied and said that it got a perfect score

@Chimpy the IMO will grade their responses. We can also read them if we want and see if there are any mistakes.

some potential evidence against this comment https://manifold.markets/Bayesian/perfect-score-achieved-by-an-ai-mod#b0i9vnrphlf

is the fact that OpenAI is sponsoring the IMO this year (and doing some sponsor related stuff) as well as their participation in the current AtCoder WTF.

@Incompleteusern I think the most likely failure mode is a silly dock like the IMO 2023/6 rubric. I don’t think it’s likely they will fail to solve a problem.

my personal opinion is that this market should be >= 95% and that I haven't heard a good reason against this happening.

@Incompleteusern free mana

@Incompleteusern Big labs might just not bother to run it. They might see it as saturated, like, 'we already know these models are superhuman at this task. This isn't going to make us stand out from other labs. Let's not rub it in the faces of the humans doing the competition'

@comicstosteal i guess that could happen but i don't think this is that realistic of a scenario; even then, it's also somewhat plausible that someone independently runs a harness or whatever using big lab models.

@Incompleteusern the resolution rules are a bit specific. Probably no big AI company cares about this market enough to pay attention to the rules, so we have to hope somebody just happens to comply.

@placebo_username it was perfectly fine last year

@DottedCalculator Yeah, and I think it may well be fine this year, but 95% seems a little high.

@placebo_username which part of the criteria do you think won’t be satisfied? they all seem reasonable for any perfect score model to satisfy

@DottedCalculator >5% probability of publication date after Aug 21 or more than 9 hours of compute used.

@placebo_username the 95% includes the 4.5 hour per day time limit. This is an official IMO time limit so if they submit for IMO they will explicitly follow it if their model has the ability to get a 42 within this time, which I think is very likely. There is also no reason to wait a month. Last year, they were published immediately after IMO.

guess manifold is going to get the "ye of little faith" voters again?

how much of a jump is it to go from gold to a perfect score?

can AI do combinatorics now or is that still an obstacle?

@0xseraphim someone claimed they can do P6 from last year which was an especially hard combinatorics problem so prolly they can

@0xseraphim jump from gold to perfect score depends on the highly varying difficulty of P6. Last year the gap was v big compared to the avg year

@0xseraphim for human contestants this very much depends highly on individual variance and test variance. many people with up to 4 gold medals have 0 or 1 perfect scores.

@Bayesian see case study C of Towards Autonomous Mathematics Research .

Towards Autonomous Mathematics Research

Abstract page for arXiv paper 2602.10177: Towards Autonomous Mathematics Research

@Bayesian "someone claimed they can do P6 from last year which was an especially hard combinatorics problem so prolly they can" they can get 1-3 points on it I think