OpenAI says its Astra model solved 10 hard math problems

In short: OpenAI says an unreleased AI model called Astra produced proofs for 10 major problems in mathematics and related areas, and the claims are shaking up the math community.

What's going on

OpenAI recently published a blog post and several hundred pages of technical documents that it says back up “10 advances in mathematics and theoretical computer science.” The Verge’s Decoder podcast covered the response with AI reporter Robert Hart, who said many mathematicians were impressed, and also unsettled.

A key point is the mismatch in what today’s AI can do. Hart said these systems can still struggle with simple tasks like basic arithmetic, calendars, and time, but can do better with some high-level, abstract math. One way to picture it is a student who sometimes misses easy questions, but can also write an advanced essay that surprises the teacher.

The OpenAI results reportedly touched several specialized areas. Examples mentioned include sphere packing in higher dimensions and quantum game theory. Hart said researchers told him that if a human solved even one of these problems, it could be enough to build a strong academic career.

The announcement also raised questions about credit and messaging. Hart said OpenAI initially suggested there had been little progress on the problems for a decade, then later changed wording after people noted that some work built on named researchers’ earlier progress.

What to watch

Mathematicians told Hart that the biggest unknown is whether this is repeatable at scale, and how much human guidance was involved in steering the AI. Another issue is access. Running top AI systems can be expensive, which could leave smaller math departments behind even if the tools become important.

Source: The Verge AI