Vibeleaderboard
← All Intel
Intel / article

Claude Black Hole Physics

Source
anthropic.com
Author
Anthropic research
Date

Context

Anthropic reports that teams of Claude agents settled two open problems in quantum gravity research, the effort to reconcile general relativity with quantum mechanics. Both answers are narrow, checkable results about simplified model universes, not statements about our own universe, and Anthropic says neither has practical applications. Each was posted as a preprint (phase transition, bootstrap) that Claude largely wrote and a physicist checked; neither paper reports peer review. The announcement explains the physics in plain terms.

The first result finishes a calculation set up in 2003 and never completed. In 1983, Stephen Hawking and Don Page showed that heating a model universe with negative vacuum energy, which behaves like a bowl that keeps its heat in, makes a gas of particles give way abruptly to a black hole, much as water suddenly boils. Their method only works when that universe is large. Holographic duality, a long-studied correspondence, lets the same universe be described as an ordinary quantum theory without gravity living on its outer boundary, and in the regime where those boundary particles barely interact, the gravity picture is small and blurry and cannot be calculated directly. A transition was known to exist there, but whether it is abrupt or gradual depended on one number that appears only at the third level of correction, which requires thousands of terms. Physicists identified that number in 2003. Claude's agents computed it, doing almost every piece twice independently, and its sign means the transition is abrupt. Because it is also abrupt at the opposite extreme, Anthropic calls an abrupt transition everywhere in between the most natural possibility, which is an rather than a proof.

The second result is a proof about the modular bootstrap, a method that narrows the search for a theory of gravity in a simpler universe with two space dimensions by ruling out candidates that contradict themselves. A donut-shaped spacetime can be read with either of its loops as time, relativity requires both readings to agree, and that requirement caps how heavy the lightest black hole can be relative to a quantity that roughly counts the universe's microscopic building blocks. The best proven cap left a gap to what computer searches suggested. Claude noticed a pattern in those numerical results, borrowed a tool from elsewhere in mathematics, and derived an exact cap of about 1/9.1266 of that count, together with a proof that this version of the method, which ignores black hole spin, cannot do better. Anthropic says the result was formally checked in Lean, a proof-verification program, and that half of the proof follows the strategy of a sphere-packing proof by an OpenAI model that one recognized as related. The same paper shows this version cannot reach the lower cap at which black holes first appear, so physicists will need other constraints.

The physics is a step in a long theoretical program. The stronger news is about method: an Anthropic researcher let separate agents work in the background for about six weeks on 22 problems Claude itself had proposed, and these two results came out of that run with little human direction. In the black hole case, a physicist who had independently used Claude reached the same answer by a different method, which serves as a cross-check. It follows earlier work in which Claude computed a nine-loop particle amplitude in the same boundary theory, checked by an outside physicist. Neither paper reports independent refereeing, so the claims rest on the authors' checks, the Lean verification Anthropic reports, and that one independent agreement.

Key takeaways · AI-distilled
  • Claude finished a three-loop N=4 super Yang-Mills calculation set up in 2003 but never completed, and found its weak-coupling phase transition is abrupt, like Hawking and Page's gas-to-black-hole switch. Anthropic says this suggests the effect is built into the quantum theory.
  • The calculation spans thousands of diagrams plus infinities that must be tamed. Claude ran a group of copies of itself, did almost every piece twice independently and checked parts against published results. A physicist using Claude separately reached the same answer another way.
  • For 3D gravity, Claude proved that the simplest, spin-ignoring modular bootstrap caps the lightest black hole's mass at c/Λ with Λ about 9.1266, and that this ceiling is optimal, beating the proven c/8.5. So this method alone cannot reach c/12, where black holes first appear.
  • The bootstrap result came from agents sharing progress: one spotted a missed pattern in how the numerical ceiling shifted with c, and another, working on sphere packing, turned it into a method after an earlier agent failed in 61 parallel tries. The proofs were checked in Lean.
  • The setup was modest: a researcher asked Claude for open problems, it proposed 22, and separate agents ran in the background for about six weeks with no special setup. Anthropic notes that neither result describes our own universe or has practical applications.
Terms in this piece · Glossary
  • inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
  • AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
Read the source www.anthropic.com
More from Anthropic research
Recommended reads
Comments

Checking sign-in…

Loading comments…