This is pretty remarkable. OpenAI just released 722 mathematical manuscripts generated by an internal AI model that hasn't even been made publicly available. These aren't homework problems. The manuscripts cover 372 groups of related results across mathematics, theoretical computer science and mathematical physics, including attempts to resolve problems that have challenged mathematicians for years.

According to OpenAI, the model was given roughly 4,000 research problems and used an average of just three hours of ChatGPT Pro-equivalent computing time per result.

Think about that, problems that could occupy mathematicians for years are now being tackled by AI in hours. Of course, there's a pretty important catch. Not all 722 manuscripts have been independently verified. Some include computer-checkable proofs, while others still need review and could contain mistakes.

But that's actually what makes this so interesting to me. We may be approaching a point where AI can generate potential scientific discoveries faster than humans can verify and understand them. And if this technology keeps improving, the implications go well beyond mathematics. Drug discovery, engineering, materials science, physics. Entire fields could eventually face the same situation.

The question might no longer be whether AI can help us discover something new, but whether we can keep up with what it's discovering. Are we watching the beginning of a new scientific revolution, or getting ahead of ourselves before these results are verified?

https://openai.com/index/sharing-ai-progress-in-mathematics/