Signum
Feed
Useful signal19 Nov 2026high confidence

Anthropic reports Claude agents solved two open theoretical physics problems on black holes and quantum gravity, with preprints and Lean verification

Edited by · written with AI from the linked sources (how)

Anthropic published a post and two preprints claiming Claude agents (run by researcher Hongbin Chen over about six weeks) (1) completed a calculation set out in 2003 showing the Hawking-Page-like phase transition is abrupt even in the regime where spacetime has no definite shape, and (2) found and proved the exact limit of the simplest modular bootstrap ceiling on the lightest black hole in a 2D-space model universe. Results were formally verified in Lean (September), and a physicist checked every main result. One result was independently reached by an outside physicist using Claude, and the method reproduced an OpenAI-model sphere-packing proof.

Capability

Entities: Anthropic, Claude, Claude Fable 5.2, Hongbin Chen, OpenAI, Stephen Hawking

62Useful signal
1 source
0 primary
Was this useful?
01

What happened

Anthropic published a post and two preprints saying Claude agents, run by researcher Hongbin Chen over about six weeks, solved two open theoretical physics problems. The first completes a calculation set out in 2003 showing a Hawking-Page-like phase transition is abrupt even where spacetime has no definite shape. The second finds and proves the exact limit of the simplest modular bootstrap ceiling on the lightest black hole in a 2D-space model universe. The results were formally checked in Lean (September), a physicist reviewed every main result, and an outside physicist reached one result independently using Claude. The extraction lists no direct links to the preprints, so the papers themselves have not been checked here.

02

Why it matters

The main significance is for research practice: AI agents can now produce checkable, machine-verified results on narrow, well-posed theory problems, which is relevant to physicists, mathematicians and rival labs. Lean verification is the strongest part of the evidence, because it removes much of the doubt about whether the proofs are correct. The physics itself has limited reach, since Anthropic concedes neither result describes our universe directly or has practical applications. The wider decision it informs is whether research groups should budget for AI-assisted theory work, and that is still a modest, early-stage signal.

03

What is noise

"Claude solves" understates the human role: a researcher steered the work for six weeks and a physicist checked the results, so this is not autonomous discovery. Claims that black hole formation is "built into" quantum theory or that the search for quantum gravity is narrowed are interpretation of results from simplified model universes, and the post itself notes the limits. It is also Anthropic's own account of its own product, and the sphere-packing reproduction of an OpenAI-model result shows rivals are doing similar things, not that Claude is uniquely capable.

04

Watch next

  1. 01Independent physicist reaction to the two preprints, and whether they pass peer review or are challenged on whether the problems were truly open and whether the Lean formalisation matches the stated claims.
  2. 02Whether other researchers, not tied to Anthropic, reproduce these results or get new ones with Claude or rival models, and how much human steering each needed.
  3. 03Whether Anthropic releases the agent setup, prompts, logs and compute cost, so outsiders can judge how repeatable the six-week process is and whether it extends beyond narrow, formally checkable problems.

Coverage

1 story

Newsletter

Get the no-hype AI news straight to your inbox

Weekly. No spam, unsubscribe any time. Privacy

More capability signals

Full feed →