Epoch AI made a new benchmark of open math problems: https://epoch.ai/frontiermath/open-problems.

Resolves yes if an AI model solves one of the problems before 2027. Some caveats:

1. The AI's raw output does not to be publishable by itself, but the core mathematical insights must be interesting enough to motivate a publishable paper. There was already a case where an LLM solved one of the problems but the result was not deemed interesting enough to be publishable: https://epoch.ai/frontiermath/open-problems/explicit-deformations

2. Human collaboration is allowed, but the core insights involved in solving the problem must be produced by an AI model. Since I am not a math expert I will look towards judgements from EpochAI in particular, but also other from mathematicians in determining whether this is the case for a particular solution.