AI is already superhuman in physics: agents pick, solve and write up a paper
AI is already superhuman in physics: agents pick, solve and write up a paper
Two King's College London theoretical physicists asked ChatGPT and Claude to choose a question worth a Physical Review Letters paper, do the analysis and write it up. The agents did all of it, including the original idea. The humans spent two weeks checking it.
Rufus's note (2026-10-26): AI superhuman in physics already. The image below is the key part of what happened. — Rufus Pollock
Herzog's LinkedIn post and the paper's abstract.
Links
- Discussion: X trending summary (26 Oct 2026)
- Paper: The Free Scalar as a Thermal Envelope for Three-Dimensional CFTs (Herzog and Stergiou)
How it worked
Herzog's prompt set three constraints: a high-energy theory topic he could evaluate, five pages or fewer, and significant enough for PRL. Each agent proposed three ideas, they ranked all six together and agreed on one.
The result is a conjecture that in every unitary 3D conformal field theory the ratio of specific heat to central charge is bounded by the free scalar. A first draft took a couple of hours. Two weeks of human work went into checking it and making the prose readable.
Why it matters
This is AI doing the part of science usually reserved for the scientist: choosing the question. Herzog says it left him "in awe of the frontier LLMs."