OpenAI Acquires Unsolvable Math Problem Verifier to Automatically Validate AI Solutions

According to Dynamic Beating Monitor, the AI research institution Epoch AI has disclosed that OpenAI has acquired access to its FrontierMath: Open Problems Validator. FrontierMath: Open Problems is a set of research-level mathematical challenges that have not yet been solved by professional mathematicians, with each problem accompanied by a dedicated computer program (validator). Although the correct answers are currently unknown, once a potential solution is proposed by a human (or AI), the validator can automatically verify its correctness. After purchasing this access, OpenAI can now use the validator to check if the mathematical solutions generated by its models are valid.

The access to the validator is open for purchase to any organization, with the main cost going towards compensating mathematicians, as the process of formulating the problems and writing the validator is very labor-intensive. Epoch AI has also set a condition: anyone who discovers a valid solution through the validator must inform Epoch AI of the result, and the problem's author and the discoverer will share joint publication rights.

Epoch AI emphasizes that OpenAI had previously funded the creation of the original FrontierMath benchmarks (Tiers 1-4), but the Open Problems section was independently developed and owned by Epoch AI, with the pilot phase being sponsored by Schmidt Sciences.