研究称 GPT-6 Astra 解出 98% 未发表物理难题
帖子称 30 多位物理学家用 50 道未发表的研究级理论物理题测试 AI,GPT-6 Astra 解出 98%、每题成本低于 3 美元,部分「失败」实为专家答案有误。

Physics is solved. 30+ physicists from Harvard, Google DeepMind and top research institutes just admitted AI now matches world experts at physics calculations. They gave AI agents 50 unpublished, research-level theoretical physics problems. The kind a world expert needs about a day to solve. GPT-6-Astra solved 98% of them, for under $3 per problem. And when it "failed" a few times, it turned out the experts own answers were wrong. Even with no internet, the top agents still hit 92%, which means they've basically memorized the physics literature. So calculation stops being the human's job. You become the one who picks the right question, sets the direction and judges what's worth solving. Calculation becomes cheap, and taste becomes the job. That's exactly why i wrote the guide below on how to build your own agent system from scratch. You can also find the full paper below.
这条是引用着另一条写的:x.com/anatolikopadze/status/2090418682534961283