Anthropic details six weeks of Claude agents working on two black-hole theory problems

Text
iHOLYSHITVISUAL / TECHAI research breakthroughsiHOLYSHIT / ILLUSTRATION
iHolyShit illustration

Two black-hole theory preprints appeared on November 7. Ten days later, Anthropic described the workflow: models split the calculations, tried dead ends and borrowed ideas across problems, while physicists framed the work and checked key results. These are theoretical results, not new observations.

On November 17, Anthropic described how a group of Claude agents tackled two open problems in black-hole theory. Both preprints had already appeared on Hexagon on November 7 under the byline Claude (Anthropic); neither has been peer-reviewed. The new account explains how the work was split, where promising methods came from and which results physicists checked.

AI had already contributed to open mathematical research

AI had already contributed to open mathematical research. In May 2025, Google DeepMind reported that AlphaEvolve found a configuration in eleven dimensions with 593 equal spheres touching one central sphere, one more than the previous record. A construction proves that many can fit, not that no more can fit. Claude’s results also have precise limits: they settle questions inside particular models or methods, not quantum gravity as a whole.

The first problem asks whether the transition associated with black-hole formation remains abrupt at weak coupling, where there is no classical spacetime geometry. The preprint finds a negative sign for the coefficient that decides the issue. In the model studied, the transition is first order: the dominant state changes abruptly at a critical temperature. A physicist checked known special cases. The other uses a technique called the modular bootstrap: rule out inconsistent answers using symmetry, then find an upper bound on the lowest non-vacuum primary state. The preprint finds the optimal leading coefficient at large central charge. This defines what the chosen method can constrain under its assumptions; it does not weigh any observed black hole.

Anthropic says a researcher defined the area of inquiry, Claude proposed 22 candidate problems, and agents ran for about six weeks. An earlier agent’s 61 parallel attempts on the second problem failed; another group later adapted an idea from sphere packing. Anthropic reports Lean formal verification for the second result. Its provenance note credits an internal Claude model with the initial result and says the public Opus 5.5 later obtained it independently. People still decided which findings warranted a paper, checked the reasoning and assessed its scope. Agreement between model runs is not peer review. Running more routes at once could shift the bottleneck to expert checking, but the account does not give a total compute bill. It cannot show that the whole process cost less than conventional research.

iHolyshit review Over the next six to twelve months, I would look for outside physicists to reproduce the first calculation from public materials and check the second result’s optimal bound and the scope of formal verification. If later projects also publish unsolved attempts and the time required for expert checks, we can judge where this workflow actually saves researchers time.

Share this article

What do you think of this story?

Anonymous feedback, up to 5 reactions per article.

Reader comments

Comment without signing in, or sign in to display your member name.

Discuss the technology and the story. Hate, politics and war, threats, abuse, personal accusations, sexual content, ads, links and personal information are not allowed.

0/500

No comments yet. Be the first to share.

Sources

Anthropic details six weeks of Claude agents working on two black-hole theory problems|iHolyShit