Clarification: From what I now understand, competitors can *try* FrontierMath, but they cannot access the full problem set and their solutions. OpenAI can, and this give them a large and unfair advantage. And very much raises the question of whether OpenAI trained on those

→ View original post on X — @garymarcus