IMO 2025 is scheduled to take place between July 10 and July 20 this year. This market asks: will (at least some) members of the public be able, once the problems are posted online, to use an AI model to produce solutions that would win gold at the IMO?

"Publicly accessible" is here meant to be a relatively lenient condition: it is not required that everyone, or even most people be able to access it - it just needs to be known about by the public, and there must exist credible evidence that some people unaffiliated with the lab that trained the AI had been able to run it and produce the solutions.

If possible, I plan to defer to the Will an AI get gold on any International Math Olympiad by the end of 2025? market for the resolution - if it resolves YES based on some AI that satisfies the above requirements, then this resolves YES in turn; likewise for NO. In the case of some more complicated situation, where it might not be clear if some solution "counts", I'll try to seek community consensus on whether it does.

• Update 2025-02-04 (PST) (AI summary of creator comment): Publicly Accessible Clarification

  • Inclusions: External tests such as the o3-mini safety testing (e.g. mid-Jan) are intended to fall under the "publicly accessible" criterion.

  • Exclusions: Internal tools like Alphaproof (e.g. internal Google tests) are not considered publicly accessible.

  • Intent: The goal is to minimize the flow of detailed IMO information to the AI that is being used to generate solutions.