AI & ComputeLOT 261122 · 3 min read

Claude Agents Claim Solutions to Two Open Black Hole Physics Problems

Anthropic says Claude agents proved the Hawking-Page transition stays first order at weak coupling and pinned the modular bootstrap slope at 9.1266. Neither preprint is peer reviewed.

By Grace Kim544 words

Key takeaways

  1. Anthropic claimed Nov. 17 that Claude agents solved two open black hole physics problems; preprints posted Nov. 7 on Hexagon.
  2. The modular bootstrap slope was pinned at about 9.1266, previously known only to lie between 8.503 and 12.
  3. The result was formally verified in the Lean proof checker in September.
  4. Agents ran for about six weeks on 22 candidate problems starting in August.
  5. Neither preprint has been peer reviewed.
Anthropic says Claude agents solved two open problems in black hole physics
Anthropic says Claude agents solved two open problems in black hole physics — BoliviaInteligente / unsplash

Anthropic says its Claude AI agents solved two open problems in theoretical physics, both concerning black holes and the search for a theory of quantum gravity. The company published the claim in a research post dated Nov. 17. Two preprints followed on Nov. 7 on Hexagon, a preprint site that labels both papers as "primarily AI-generated text." Mashable and other outlets picked up the story on Nov. 20.

The results remain claims from Anthropic and its own papers. Neither preprint has passed peer review. Anthropic states that neither result describes our universe directly nor has practical applications.

What did Claude reportedly prove?

The first result addresses the Hawking-Page transition. Stephen Hawking and Don Page found this abrupt switch between a hot gas and a black hole in a model universe in 1983. That calculation holds only where general relativity applies. Physicists found a transition in the opposite, weak-coupling regime in 1999. They did not know whether it was abrupt or smooth. A 2003 paper set out the calculation that would decide the question.

According to the preprint, the answer turns on the sign of one coefficient that first appears at three loops. Physicists had not computed it in the two decades since. Claude found it negative, meaning the transition stays abrupt — first order — at weak coupling. The paper says almost every number was obtained in two independent ways. A physicist checked the result against known special cases.

The second paper concerns the modular bootstrap. This method rules out candidate theories of three-dimensional quantum gravity by showing they are inconsistent. It sets a ceiling on how heavy the lightest black hole can be. The best slope of that bound was known only to lie between 8.503 and 12. The preprint pins it at exactly 12π²/(π² + 4 arcsinh²1), about 9.1266. It also states the simplest version of the method cannot do better.

How did the agents work?

Anthropic researcher Hongbin Chen started the project in August as a side test. He asked Claude to list open problems in black hole physics and the bootstrap method. Claude returned 22 candidates. Chen set up separate agents for each. They could read each other's progress posts and ran in the background for about six weeks.

The company credits Claude Fable 5.2 with spotting a pattern in published numerical results that physicists had missed. Anthropic also says the bootstrap method was first developed by an agent working on a related sphere-packing problem, where it reproduced a result recently proved by an OpenAI model. Half of the final proof follows the strategy of that OpenAI proof. An earlier Claude agent had failed the same bootstrap problem in 61 parallel attempts.

How were the results verified?

Anthropic says the bootstrap result was formally verified in the Lean proof checker in September. Claude wrote the paper with minimal guidance. A physicist checked every main result. On the Hawking-Page problem, a physicist who had separately used Claude reached the same answer weeks later by a different method.

The post follows a claimed AI-assisted proof of the Riemann hypothesis by Anthropic mathematician Levent Alpoge, reported earlier this week. Peer review will now determine whether these agent-produced physics results hold up as claimed.

mlq.ai (Source)

Filed under

Share this article:

More from Grace Kim

Grace Kim

Show full bio

Market editor covering people, launches and day-to-day practice in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.

1,213 articles

Read also

« Previous article