Dan Kagan-Kans interviews seven experts about language models changing theoretical physics. Calculations are becoming cheaper and faster, but choosing a worthwhile question, training new physicists, and testing a theory remain harder problems.
- A striking bounded result: the article reports that GPT-6 Astra answered 49 of 50 unpublished physics problems correctly after errors in human answer keys were examined. Correct answers averaged under $3 in compute.
- Different work, different impact: Andreas Karch describes near-continuous agent use. Daniel Harlow says most of his effort goes into deciding what to calculate in the first place.
- An apprenticeship at risk: calculations have helped students build intuition and reputations. Researchers worry about how students acquire that judgment when agents perform the work.
The reporting separates solving a well-defined problem from discovering a theory that explains nature. Even a mathematically consistent proposal can be beyond our ability to test experimentally. Predictions of imminent autonomous discovery remain predictions, and several enthusiastic interviewees disclose model-company support or consulting ties.
Faster output also puts pressure on reading and explanation. The article asks who will evaluate the growing pile of papers, and whether translating a discovery into something other physicists understand deserves equal credit.