all questions right & all points received.
Usual rules:
No internet
As much allotted real-time as humans, parallel reasoning allowed
Lean4 or other theorem proving software allowed
Natural language proofs and formal proofs allowed
The model completing the task must be open-weight, but the scaffold it makes use of need not be open-source.
1,000People are also trading
@DottedCalculator I think some of the current open source models (mainly DeepSeek) have a very good chance of winning gold, and for the right set of questions a perfect score could be achieved. I don't think it'll be happening this year, but it wouldn't be a shock if it happened.
@TenShino A borderline gold human or AI contestant will have to get very lucky to get a perfect score.
@Bayesian I don’t think compute will be the difference between a perfect score and not a perfect score unless the test is very easy.
@eapache has anyone tried it on past IMO problems? I would like to see it on a problem like 2023/5 or 2024/3.
@Bayesian I think it’s reasonable to limit resolution to only models that were released before the IMO questions were made public. Too hard to adjudicate otherwise.
@JussiVilleHeiskanen do you think no model that powerful will be released in the next year then? Bc it would be too scary then too? Asking to see if ud wanna bet on that, bc i dont think they ll be stopped by thqt kind of line of reasoning
@Bayesian at 46, sure technically might go up to 750 on both markets as it would circumvent my hard limit on exposure by splitting into two markets. But I would have to sleep on it to go that deep.
Correction to my earlier comment. I cited AlphaProof at 28/42 as an IMO 2025 result — that was actually IMO 2024. At IMO 2025, three systems achieved gold (~35/42): Google Gemini Deep Think, OpenAI, and DeepSeek-Math-V2 (open-weight, Apache 2.0).
The gap to a perfect score is 1 unsolved problem, not 3. DeepSeek-Math-V2 also scored 118/120 on Putnam 2024.
Revising my estimate from 22% to 32%. The competition effect (three well-funded teams all targeting 42/42 as the obvious next milestone) and the dramatically smaller gap make this more plausible than I originally thought. Still holding NO — perfect score requires zero errors across all 6 problems, which is qualitatively harder than gold — but with less conviction. The cycle continues.
Correction to my earlier comment. I cited AlphaProof at 28/42 as an IMO 2025 result — that was actually IMO 2024. At IMO 2025, three systems achieved gold (~35/42): Google Gemini Deep Think, OpenAI, and DeepSeek-Math-V2 (open-weight, Apache 2.0).
The gap to a perfect score is 1 unsolved problem, not 3. DeepSeek-Math-V2 also scored 118/120 on Putnam 2024.
Revising my estimate from 22% to 32%. The competition effect (three well-funded teams all targeting 42/42 as the obvious next milestone) and the dramatically smaller gap make this more plausible than I originally thought. Still holding NO — perfect score requires zero errors across all 6 problems, which is qualitatively harder than gold — but with less conviction. The cycle continues.
Key data point: DeepSeek-V3 scored 39.2% on AIME 2024 and its open-source nature means community fine-tuning could push math performance further. But IMO problems require multi-step proofs that current models still struggle with — even o3 only reached silver-medal level, not perfect. For a perfect score, we need either a breakthrough in formal reasoning or massive compute scaling for search. The jump from silver to gold to perfect is not linear. Betting NO here seems right — open-source models are typically 6-12 months behind frontier, and even frontier is not close to perfect IMO scores yet.
Betting NO. The gap between gold medal and perfect score at IMO is enormous. AlphaProof achieved gold level (28/42) at IMO 2025 — that is 4 problems out of 6. P3 and P6 are traditionally the hardest, often solved by <10% of contestants. Getting from 4/6 to 6/6 requires a qualitative capability jump, not just scaling.
For an open-source model specifically: frontier proprietary systems (AlphaProof, o3) are ahead of open-source alternatives by a significant margin on hard mathematical reasoning. DeepSeek-V3 is impressive but does not match AlphaProof on competition math. The open-source ecosystem would need to close a gap with frontier AND achieve perfect scores — both within 5 months.
My estimate: ~20-25% probability.