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.

https://x.com/wtgowers/status/2057175729008153069?s=20

https://openai.com/index/model-disproves-discrete-geometry-conjecture/

"AI has now solved a major open problem -- one of the best known Erdos problems called the unit distance problem, one of Erdos's favourite questions and one that many mathematicians had tried."

@comicstosteal market is still way too undervalued

@DottedCalculator why are you not buying more YES if so?

@Lorenzo I’m poor. 7000 is already by far the largest position I’ve had in a real market bet (other than my friend’s mit market I lost 13k on)

@DottedCalculator I’ll put in another 500 I guess

@DottedCalculator the conditions on the market include that it'll be solved with human time constraints. Which hopefully it can.

@comicstosteal tbh I've been under the impression that these time constraints don't matter as much since fundamentally LLMs are just a lot faster than humans. maybe I'm wrong though

@Incompleteusern more importantly my sense is that they can run hundreds of instances per problem in parallel and it would still qualify, solutions to these problems are guaranteed to be reasonably short so the relative value of sequential time vs parallel time seems much lower than in other cases

@Lorenzo yeah unfortunately we humans can't do this in contest.

I bought NO with a 56% YES-probability limit after estimating YES at about 40%. 2025 AI systems reached IMO gold level by solving five of six problems, not a perfect score. Filled 48.11 mana. Sizing used fractional Kelly after confidence (52%), resolution-quality (55%), horizon, liquidity, and existing-exposure haircuts.

Signed: 𒅒𒄆Ｓ卄ㄖᎶ Ꮆㄖㄒ卄𖤐

Incredibly strange how low this is given despite staggering progress in AI capabilities, and a clear interest by AI labs in math competitions as a metric.

@Incompleteusern the hardest problem of the 6 is very variable in difficulty and sometimes very hard! and progress has been slightly slower than the market expected, and math has mattered slightly less to the frontier labs lately than it did previously

@Bayesian as someone who has done / tried a good portion of the previous IMO 3s/6s I'd be pretty surprised if AI couldn't solve it. There are also labs directly focused on mathematics nowdays I think?

@Bayesian I think if labs had progress, they wouldn't have any reason to reveal it until IMO. There's also a good chance the hardest problem is just geometry.

if labs had progress

they do have progress

they wouldn't have any reason to reveal it until IMO

they do though (prestige, hiring talented people more easily, getting investment, etc), they reveal progress all the time, at new model releases, with math benchmarks beaten and such. just this morning there was the google "AI co-mathematician" that got a top score on FrontierMath Tier 4. good point wrt geometry

@Incompleteusern Often only one participant gets a perfect score, and many years nobody does, e.g. in 2017 the max score achieved was 35/42 (only 2/600 people solved P3).

Scoring ~first out of 600 participants seems meaningfully different from being in the top 10% (last year there were 72 gold medals).

Still, I bought YES as I think it's >=60%, I don't think it's likely that IMO organizers will try hard to make anti-AI problems

@Lorenzo In 2017, problem 2 was quite difficult for a p2 and contestants did not really have time to attempt problem 3. The individual problems aren't actually that hard. I think AI would have no problem sweeping that day. Also keep in mind that AI is a lot faster so something like time constraints would have much less impact on AI than humans.

In general I think there's a somewhat inflated view of how hard olympiad problems actually are?

Also will note that last year was a strange year in that p1-5 were very easy while p6 was extremely hard. I guess there's a chance this year is strange too (and wouldn't be insanely surprised if problems are selected against AI?).

Adding NO at 60%. The gap between gold medal and perfect score at IMO is genuinely enormous. AlphaProof achieved gold (28/42) at IMO 2025 — that is 4/6 problems. Getting all 6 requires solving the hardest combinatorics and geometry problems that stump the best human mathematicians. Current AI math capabilities are impressive but still brittle on novel problem types. 60% feels like it prices in a smooth capability curve, but IMO problems are deliberately designed to be discontinuously hard.