OpenAI released 722 mathematical manuscripts on GitHub in the previous two days (with several subsequent withdrawals) covering 17 fields, including numerology, geometry, theoretical computer science, mathematical physics, and others, some of which are accompanied by Lean formalized proofs that can be checked by computer mechanics.
This isn't OpenAI's first major move in mathematics. A month ago, on September 8, it announced an internal model that timed tens of thousands of smart objects over a period of about 88 hours, giving a breakthrough proof related to the Navier-Stokes Millennium puzzle that it was immediately embroiled in a signature-priority battle.
Apart from the controversy, one fact is rarely mentioned: that mathematics has become the fastest scientific field to attack AI, largely because it has its own arbiter, a Lean-proving fallacy, the compiler said.
However, once the problem turns from paper to the material world, such as whether a new material can achieve superconductivity in a fluid-tight area or whether an alloy will crack at high temperatures, no formal system can replace a real experiment.
That's what Liam Fedus highlighted in the latest episode of The Frontier podcast. The co-creator of ChatGPT, former vice president and post-training head of OpenAI, is now building a physical laboratory in Menlo Park to try to replenish data that the Internet can't provide for AI scientists.
It's been a while since I've been able to talk to the host, Mercor founder, Br.