Anthropic reports that Claude Fable 5.2 calculated a discontinuous transition to a black hole even in a very small universe and found the exact limit of a method for ruling out quantum gravity theories (preprints without peer review)
Anthropic reports that Claude Fable 5.2 calculated a discontinuous transition to a black hole even in a very small universe and found the exact limit of a method for ruling out quantum gravity theories (preprints without peer review)
Anthropic announced the results on November 9. Both papers appeared on November 7 on HEXAGON, a server that accepts papers without review by other scientists; Claude is listed as the author.
Both papers rely on holographic duality: certain universes with gravity can be described exactly by a quantum theory without gravity. In 1983, physicists Stephen Hawking and Don Page showed that heating a sufficiently large universe of this kind produces a discontinuous transition to a black hole, much like water boiling. In 1998, physicist Edward Witten identified the corresponding transition in the dual theory: as the temperature rises, particles cease to be bound by the strong interaction.
The weakly interacting case remained unresolved: in the gravitational description, the universe is so small that it has no classical geometry. A useful way to picture the system's state is as a ball on a landscape: the ball is stable only in a valley. The dual theory already has a transition when there are no interactions, but the landscape is flat at the transition point, so even a weak interaction determines the nature of the transition. In 2003, physicist Ofer Aharony and colleagues showed that there are two possibilities, determined by the sign of a number b that appears when interactions are introduced. When b is negative, a second valley appears nearby before the original valley disappears, and the ball jumps into it; when b is positive, the valley gradually becomes shallower, and two continuous transitions replace the jump. In 2005, the same physicists calculated b for a simplified version of the theory; the version dual to gravity has more than ten times as many terms.
In the first paper, Claude obtained b = −1,48·10⁻³·λ² (λ is the interaction strength). The value is negative, which means that at weak interaction strength there is a single, discontinuous transition. The landscape develops a valley, a hill, and a second valley; these appear, replace one another, and disappear in the same sequence as the universe without a black hole, the small black hole, and the large black hole in Hawking and Page's calculation.
Almost every numerical result in the paper was obtained using two independent programs or checked against an exact answer. In one instance, both programs agreed with each other but disagreed with the exact answer; as the paper puts it, “both were faithfully computing the wrong theory.”
The modular bootstrap, the subject of the second paper, helps physicists rule out candidate theories of quantum gravity in a simplified universe. It uses a symmetry of a two-dimensional quantum theory: the sum over all its states stays the same when the temperature T is replaced by 1/T. This symmetry gives an upper bound: in a theory with a large value of the size parameter c, the lightest state that cannot be reduced to the vacuum has a mass no greater than approximately c/Λ, where Λ depends on how the bound is derived. A larger Λ gives a lower bound on the maximum allowed mass; at Λ = 12, it would reach the threshold above which states correspond to black holes. Previously, Λ was known only to lie between 8,503 and 12. According to the paper, no method within a broad class can yield Λ greater than 9,1266, and one explicit method attains that value (established by a computer-assisted proof); reaching the c/12 threshold will require additional information. The authors adopted the approach to proving optimality from an OpenAI model's August result on sphere packing.
According to Anthropic, Fable 5.2 solved the first problem “largely independently.” According to the preprints, an internal Claude model obtained the first result in response to a physicist's prompt; the physicist checked it against special cases. The model obtained the second result autonomously, Claude Opus 5.5 independently reproduced it, and a physicist checked the main results.