Anthropic details Claude's solutions to two black-hole physics problems
Anthropic has described two black-hole physics results obtained with Claude: a calculation left unfinished for more than 20 years, and an exact limit on a method for testing candidate theories of quantum gravity. The accompanying preprints also show how the work was checked. In one case, two independently written programs agreed with each other and were both wrong.
We mentioned the results when Fable 5.2 launched. Anthropic's November 17 account now gives a fuller explanation; the two preprints were posted on Hexagon on November 7.
The calculation that waited
The first paper concerns a mathematical model in which heating a system produces the counterpart of a black hole forming. Hawking and Page established an abrupt transition in the gravitational description. The unresolved question was whether the corresponding quantum theory also changes abruptly when its particles interact only weakly, where the usual picture of smooth spacetime breaks down.
A calculation proposed in 2003 could settle it. The answer depended on the sign of one coefficient, which first appears at the third level of a perturbative expansion. The preprint reports a negative value, giving a first-order, or abrupt, transition.
Its account of the computation describes several Claude instances working through files under a coordinator. Quantities normally needed either agreement between independently written programs or a check against an exact result. Code also had to reproduce known cases before being used.
That last requirement caught an error that agreement alone missed. Two calculation engines produced the same answer but disagreed with a known result from another method. Further testing showed that both had computed the wrong theory. The paper records an exception to its double-checking rule for some coefficients in its series expansion, rather than claiming that every number received two independent checks.
Finding the limit of a method
The second paper studies the spinless modular bootstrap. This method constrains possible quantum theories by demanding mathematical consistency. Through the connection between quantum theories and gravity, those constraints help narrow the possible descriptions of black holes.
For large central charge, a quantity denoted by c that characterizes the quantum theory, the paper bounds the lightest non-vacuum excitation's scaling dimension at roughly c/9.1266. The previous analytical ceiling was c/8.503. A larger denominator makes the ceiling lower, ruling out more possibilities. The paper also establishes that this version of the method cannot reach the tighter target of c/12.
The provenance statements are more specific about model identity than the announcement's Fable headline: both credit an internal Claude research model with the original result. The bootstrap paper says publicly available Opus 5.5 later obtained that result independently, and that a physicist checked every main result.
Anthropic says the broader experiment began with 22 candidate problems and ran for roughly six weeks. These are results in simplified theoretical settings. The bootstrap paper's limit concerns its spinless formulation and specified class of mathematical functionals; it does not settle the full theory of quantum gravity.
Comments
No comments yet. Be the first.