Anthropic says Claude — specifically its Fable 5.2 model — has solved two open problems in theoretical physics, both concerning black holes and the search for a theory of quantum gravity. The company published the results on its research blog on November 17, and the work reached the front page of Hacker News on Thursday as physicists and AI researchers debated what it says about the state of AI-assisted science.

The first result completes a calculation that physicists set up more than two decades ago but never finished. The second pins down the exact limit of a technique that physicists have used since 2009 to rule out incorrect candidate theories of quantum gravity. Neither result describes our own universe directly, Anthropic notes, and neither has practical applications — but both feed a long-running effort to understand how quantum mechanics and general relativity fit together. For more on how frontier labs are positioning their models as research tools, see our ongoing latest AI developments.

Why black holes are quantum gravity's proving ground

Modern physics rests on two frameworks that are mathematically incompatible: general relativity, which describes the very large, and quantum mechanics, which describes the very small. Black holes are one of the few places where both matter at once, which makes them a natural testing ground for any would-be unification.

Physicists studying quantum gravity often rely on holographic duality: the idea that certain theoretical model universes can be described two ways — as a universe with gravity and spacetime, or as a gravity-free quantum theory "written" on the universe's outer boundary, in one dimension fewer. The trick turns hard questions about gravity into quantum questions that are sometimes easier to answer.

Problem one: the Hawking-Page transition in a shapeless universe

The first problem concerns what happens when you heat one of these theoretical universes. Below a certain temperature it fills with a gas of particles; above it, a black hole forms instead. In 1983, Stephen Hawking and Don Page showed that this switch is abrupt — a true phase transition, like water boiling at 100°C. But their calculation relies on general relativity, which only holds when the universe in question is vast compared with the scale at which that description breaks down.

As the theoretical universe shrinks toward that scale, spacetime loses any definite shape, and the gravity-side calculation stops working. Physicists found in 1999 that a phase transition still exists in that regime, but nobody knew whether it was abrupt or smooth. In 2003, a group of physicists set out the calculation that would settle the question. It was so complex that it was never completed — until now. Anthropic says Claude showed the transition is abrupt: the same sudden switch appears at both extremes, suggesting black hole formation is built into the underlying quantum theory rather than being an artifact of relativity.

Problem two: the exact limit of the modular bootstrap

The second result concerns a simplified universe with two dimensions of space instead of three. Physicists cannot yet build a full theory of quantum gravity even there, so they work in reverse: instead of finding the correct theory, they rule out wrong ones by showing they are internally inconsistent. This method is called bootstrapping.

Since 2009, a version of the method known as the modular bootstrap 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, and every lowering of the ceiling narrows the search for the true theory. Numerical calculations hinted at where the best possible ceiling lay, but they could not prove it, and the best proven ceiling fell well short of what the numerics suggested.

According to Anthropic, Claude Fable 5.2 noticed a pattern in the numerical results that physicists had missed, borrowed a method from elsewhere in mathematics and physics, and used it to calculate the precise ceiling for the simplest version of the technique — then proved that it cannot be lowered. The result does not identify the correct theory of quantum gravity, but it may point physicists in the right direction.

How the work was done: an agent swarm running for six weeks

The methods behind the results may matter as much as the results themselves. In August, Anthropic researcher Hongbin Chen was testing Claude's capabilities as a side project and asked it to draw up a list of interesting open problems in theoretical physics, focused on black holes and the bootstrap method. Claude returned hours later with 22 candidates. Chen set up separate agents to work on each problem; the agents could read each other's progress posts and collaborate on related questions, running in the background with no special setup for about six weeks.

The Hawking-Page solution came from one of those agents. Weeks after it reached a result, an independent physicist who had separately used Claude on the same problem shared his own solution — reached by a different method but matching the agent's answer, a useful cross-check. The modular bootstrap result came from several agents working together: one spotted the pattern in the literature, and another, originally tasked with a mathematically related problem about packing spheres into high-dimensional spaces, turned that pattern into a method. That method reproduced a sphere-packing result recently proved by an OpenAI model, Anthropic says. An earlier Claude agent had tried the bootstrap problem and failed across 61 parallel attempts, and two other agents ultimately arrived at the same answer independently.

In September, the results were formally verified in Lean, a computerized proof checker — a meaningful guard against the hallucination problems that plague AI-generated mathematics. Claude then wrote the paper with minimal guidance, and a physicist checked every main result. Preprints of both papers are available, per Anthropic.

A pattern, not a one-off

This is not Anthropic's first foray into AI-driven discovery. Earlier this month, on November 4, the company published work in which Claude investigated hypothetical "vault particles." The black hole results, though, are a step further: not analyzing a proposed phenomenon, but closing out questions that had resisted human mathematicians for years.

Caveats remain. The problems were theoretical, the universe models are toy models, and Anthropic — as the company selling Claude subscriptions — has an obvious interest in showcasing its model's capabilities. The independently verified Lean proofs and the physicist's review mitigate the most serious concern, that the results are simply wrong. Whether the broader physics community accepts the methods, and whether agent swarms running for weeks in the background become a standard tool of theoretical research, will be the real test of what this moment means.

---

Stay Ahead of AI

Get the latest AI news, analysis, and breakthroughs — all in one place.

Read more AI news →