Anthropic says Claude solved two open black hole physics problems

Anthropic says Claude solved two open black hole physics problems

Anthropic says Claude has solved two open problems in theoretical physics, both tied to black holes and the search for a theory of quantum gravity. The company's post links two preprints and explains the findings first in plain terms, then in more technical detail. The copy available for this retelling is cut off partway through the technical explanation of the first problem.

The first problem concerns the Hawking-Page transition. In certain model universes, a hot universe is filled with a gas of particles at low temperature, and above a certain temperature a black hole forms instead. In 1983, Stephen Hawking and Don Page showed that this switch is abrupt, a phase transition similar to water suddenly boiling at 100°C. Their calculation relies on general relativity, which holds only when the universe is vast compared with the tiny scale at which that description breaks down. As the universe shrinks toward that scale, spacetime loses any definite shape and becomes blurry, and nobody knows how to do the calculation on the gravity side. On the quantum side of holographic duality, the same situation can be calculated in principle. In 1999, physicists found a phase transition in that regime too, but did not know whether it was abrupt or smooth. In 2003, a group of physicists set out the calculation that would settle it, but it was extremely complex and was never completed. Claude did it by running a small research group of copies of itself, which worked out almost every piece of the calculation independently and checked its math against known results. It showed the transition is abrupt. Because the same sudden switch appears at both extremes, the post says this suggests black hole formation is not just a feature of relativity but is built into the underlying quantum theory.

The second problem concerns a simplified theoretical universe with two dimensions of space instead of three, which still has black holes and other mysteries similar to those of ours. Physicists cannot yet build a full theory of quantum gravity even there, so they try to rule out wrong theories by showing they are internally inconsistent. This method is called bootstrapping. Since 2009, a version of it has been used to set a ceiling on how heavy the lightest black hole in such a universe could be; any theory that breaks the ceiling is ruled out. Numerical calculations hinted at where the best possible ceiling lay but could not prove it, and the best proven ceiling was well short of what they suggested. Claude Fable 5.2 noticed a pattern in the numerical results that physicists had missed, borrowed a method from elsewhere in math and physics, calculated the precise ceiling for the simplest version of the technique and proved it cannot be lowered. The result does not identify the correct theory, but it may point physicists in the right direction.

Anthropic is explicit about scope: neither result describes our own universe directly, and neither has practical applications. Both feed a long-running theoretical effort to understand how quantum mechanics and relativity fit together.

The post also describes how the work happened. In August, Anthropic researcher Hongbin Chen was testing Claude's capabilities as a side project. He asked it to list interesting open problems in theoretical physics, focused on black holes and the bootstrap method. Claude returned hours later with 22 candidates. Chen set up a separate agent for each; the agents could read each other's progress posts and collaborate on related problems, and they ran in the background, with no special setup, for about six weeks.

One agent solved the Hawking-Page problem. Weeks later, a physicist who had independently used Claude on the same problem shared his result, and a colleague realised Chen's agent had reached the same answer earlier. The two solutions used different methods, so they served as a check on each other.

The modular bootstrap result came from several agents. The one assigned the problem spotted the pattern while surveying the bootstrap literature. Another, working on a related problem about how tightly spheres can be packed in high-dimensional space, turned the pattern into a method, and found that it reproduced a sphere-packing result recently proved by an OpenAI model. Claude then applied the method to the bootstrap problem, which an earlier Claude agent had tried and failed to solve in 61 parallel attempts. Two other agents independently reached the same answer. One half of the final proof, showing the ceiling cannot be lowered, follows the strategy of the OpenAI proof. In September the results were formally verified in Lean, a computerised proof checker. Claude then wrote the paper with minimal guidance, and a physicist checked every main result.

The visible part of the technical section starts with an everyday analogy for the sum over histories, in which a quantum particle takes every possible path. For gravity, what gets summed is the shape of spacetime itself. Hawking and Page found a case simple enough to handle: in anti-de Sitter spacetime, a bowl-like universe with negative energy of empty space, two histories dominate at any temperature, one a gas and one a black hole, and whichever contributes more wins. The text then turns to the holographic principle, from Bekenstein and Hawking's 1970s finding that a black hole's information capacity scales with its surface area, through the early-1990s proposal by Gerard 't Hooft and Leonard Susskind, to Juan Maldacena's 1997 formalisation, and breaks off there.

Key facts

  • Anthropic says Claude completed a calculation set out in 2003 that was never finished, showing the black hole phase transition stays abrupt even where spacetime has no definite shape.
  • In a two-space-dimension model universe, Claude Fable 5.2 spotted a pattern in numerical results and proved the exact ceiling for the simplest version of the modular bootstrap cannot be lowered.
  • Researcher Hongbin Chen asked Claude for open problems; it listed 22 within hours, and separate agents worked on them in the background for about six weeks.
  • The bootstrap proof drew on a sphere-packing method that reproduced a result recently proved by an OpenAI model, and one half of the final proof follows that proof's strategy.
  • The results were formally verified in Lean, Claude wrote the paper with minimal guidance, and a physicist checked every main result.

Why it matters

Anthropic presents this as Claude solving two genuinely open theoretical physics problems rather than reproducing known results. The first was a calculation physicists set up more than 20 years ago and never completed. The second closes a gap between what numerical bootstrap results suggested and what had been proven. The post also stresses the method: agents running in the background for about six weeks with no special setup, a different agent for each of 22 self-proposed problems, and cross-pollination between them. The bootstrap result leaned on a method that reproduced an OpenAI model's sphere-packing proof, so work by models from two labs fed into one result. The post is equally clear that neither result describes our own universe directly and neither has practical applications.

Who it affects

Most directly, theoretical physicists working on quantum gravity, holographic duality and the modular bootstrap, whose open questions are the target. The post says the second result may point physicists in the right direction, though it does not identify the correct theory. It also matters to AI researchers and labs interested in long-running autonomous research agents, since the post lays out how Chen's agents were set up, shared progress and were checked.

How to use it

This is a research report, not a product. The post links two preprints for the full results. The workflow it describes is the reusable part: ask Claude to draw up a list of open problems in a narrow area, start a separate agent for each, let the agents read each other's progress posts and collaborate on related problems, and leave them running in the background. In this case the results were then formally verified in Lean and a physicist checked every main result before the paper was written up.

How solid is it

The account is Anthropic's own, from a first-party post, and the claims rest on its description of the work. Several checks are reported: the Hawking-Page answer was matched, by a different method, by a physicist who had independently used Claude on the same problem; two other agents independently reached the same bootstrap answer; the results were formally verified in Lean in September; and a physicist checked every main result. The source does not say the preprints have been peer reviewed or accepted by a journal. The copy of the post used here is truncated in the technical section, so the detailed derivations are not covered.

Risks and caveats

The post itself limits the claims: neither result describes our own universe directly, neither has practical applications, and the bootstrap result does not identify the correct theory, it only may point physicists in the right direction. The abrupt-transition finding is described as suggesting, not establishing, that black hole formation is built into the underlying quantum theory. The post is Anthropic's account of its own model's work. It does not give compute cost, the number of agents on the first problem, or token usage, so the effort behind the results cannot be gauged from the text.

“Neither result describes our own universe directly, and neither has practical applications, but both contribute to a long-running theoretical effort to understand how quantum mechanics and relativity fit together.”

— Anthropic, research post on Claude and black hole physics