OpenAI has released AI-generated solutions to 377 hard math and physics problems at once. According to Scientific American, all the solutions came from a frontier model that has not yet been made public, and were produced by a single agent responding to a single prompt. They cover physics, advanced mathematics and other disciplines, and the code has been uploaded to GitHub in bulk. OpenAI president Greg Brockman said the aim is to “accelerate scientific discovery”.
Key points:
- One agent, one prompt: Unlike the earlier campaign involving 10,000 agents, these 377 solutions used just one agent and one prompt, showing a large jump in the capability of a single model.
- Academic doubts: Melanie Matchett Wood, a Harvard mathematics professor and member of the independent advisory group, said bluntly that AI-generated mathematical work often involves “theory plagiarism”, vague arguments and no human willing to take responsibility for the results, and remains short of rigorous academic standards.
- OpenAI's response: It has formed an independent advisory group of mathematicians to draft publication norms, and has promised to publish the AI's chain of thought, cite the original papers and present mathematical arguments in a form humans can understand.
Background: Last month OpenAI claimed that 10,000 AI agents had cracked a subset of the Millennium Prize Navier-Stokes problem, which academics criticized as “exploiting loopholes in the math”; around the same time, GPT-6.1 Astra was paused from rollout for failing to meet safety standards. While loudly showcasing AI's ability to speed up research, OpenAI is also hitting the brakes over safety, fighting on two fronts at once. Whether these 377 solutions can withstand peer review will take mathematicians hundreds of hours of checking, one by one.
Source: read the original. This text is a summary for reference only.
