GPT-6 Astra 物理测试 50 题解出 49 道

Alex Veremeyenko @alex_verem原推发布于 2026.10.16

在 X 上打开原推

GPT-6 Astra disagreed with expert physicists on several research problems. It insisted the expert results were wrong and its calculations were right. The physicists checked their work again. And yes Astra was right. Its final score was 98%. Physicists at universities like Harvard, Duke and the University of Tokyo wrote the test. It has 50 problems taken from their current research, and none of the answers were public when it ran. A world expert needs about a day to solve one of them. Astra solved 49 with no human hints, at an average cost under $3 per problem. The authors call it a significant milestone for physics research and teaching. A day of expert work now costs a few dollars. They say problems like these shouldn't go to new researchers as first projects anymore, because an agent can already produce the answer. The other agents scored lower. Claude Opus 5.5 solved 46. Gemini 3.8 Flash solved 34. The best open-weight model, DeepSeek-V4.1-Flash, solved 27. The agents could search the web and write their own code, the same way a researcher works. With search turned off, Astra and Opus each solved 46. The authors say this suggests the models have memorized large parts of the physics literature. The physicists say humans still need to lead. The agents calculate as well as world experts, but in the authors' experience they don't know which problems are worth solving. When the authors asked them to invent new problems, the agents focused on details a human would find uninteresting. So the calculating part of physics research is cheap now. Choosing the problem is still a job for people. The paper is in the reply 👇