Ground Truth.
AI, checked against the source.

News · 2026-11-20

Anthropic reports two Claude-assisted results in black-hole physics

Anthropic reports that Claude-assisted researchers produced two theoretical-physics results: a calculation of a weak-coupling phase transition and a limit on a particular mathematical method used to study quantum gravity. Both results appear in unreviewed preprints, and their broader significance awaits independent scrutiny.

Key facts

The two results address different questions. One calculates which of two possible phase-transition scenarios occurs in a particular model. The other asks how far a particular mathematical technique can tighten a bound. Both relate to ideas used to connect quantum theories and gravity, but neither supplies a complete theory of quantum gravity.

The first preprint, titled “The Hawking-Page transition at weak coupling,” studies a gauge theory on a three-dimensional sphere in a regime where its coupling is weak. Earlier work by Aharony and collaborators showed that the relevant transition could take more than one form. The new calculation reports a coefficient whose negative sign selects a first-order transition, meaning the preferred phase changes abruptly rather than evolving smoothly.

A familiar analogy is water changing phase: under suitable conditions liquid and vapor do not blend smoothly into one another as temperature crosses the transition. The analogy illustrates an abrupt change, not the detailed physics of this gauge theory. The calculation concerns a boundary theory related to gravitational ideas through a proposed duality; it is not a direct computation of a black-hole gravitational path integral.

The authors report calculating nearly every number independently twice and checking known special cases. Anthropic says a separate Claude calculation using another method reached the same answer. Agreement and checks against known cases can catch mistakes, particularly in a long calculation, but multiple model runs are not the same as an outside group publishing an independent derivation.

The second preprint, “The spinless modular bootstrap at large central charge,” studies limits on a mathematical procedure for constraining possible theories. The method uses a symmetry of a theory's partition function to bound the lightest non-vacuum state under specified assumptions. Its reported optimum has a constant of about 9.1266 in the denominator of the asymptotic bound.

Imagine trying to rule out impossible buildings using only a measuring tape and one structural rule. Finding the best possible measurement using that rule tells you how powerful the method is. It does not establish that a building at the boundary can actually be constructed, and a different rule might rule out more possibilities. Similarly, the result limits what the stated class of bootstrap functionals can achieve; it does not show that a physical theory saturates the bound or identify a unique quantum-gravity theory.

Anthropic says its researcher Hongbin Chen asked Claude to suggest open problems, and agents pursued 22 candidates for about six weeks. The company describes a numerical pattern, a spectral-curve approach and additional agent checks behind the bootstrap result. It also says a physicist checked main results and that the proof was formalized in Lean, a proof-checking system. The reviewed material does not publish the Lean project or identify that physicist. A checker can validate the formalized logical steps, while the connection between the formal definitions and the intended physics still requires examination.

There is also a source-identification trap. The HexagonMath identifiers resemble arXiv numbers, but arXiv:2611.00175 and arXiv:2611.00174 are unrelated papers. The preprint host says it does not perform peer review. Discussion on Hacker News ranges from enthusiasm to requests for independent review and reproducible artifacts; it is not a substitute for specialist scrutiny.

If the results hold and prove useful, they illustrate two roles for AI in science: organizing a demanding calculation and proposing an approach to a mathematical constraint. The remaining tests are public artifacts, independent calculations, peer review and whether other researchers build on the work.

The work also illustrates why source identification is part of scientific verification. A number that looks familiar can refer to different repositories; checking the title and host avoids sending readers to an entirely unrelated article. The underlying mathematical claims require their own separate review.


Primary source, verified: read the paper →

Key questions

Did Claude solve quantum gravity?

No. The reported results concern a specific gauge-theory transition and a limit on one mathematical bounding method, not a complete theory of quantum gravity.

Have the two results passed peer review?

The texts are posted as HexagonMath preprints, and the host says it does not perform peer review; no journal acceptance was established in the reviewed material.

Are the physics preprints on arXiv under the same numerical identifiers?

No. The corresponding arXiv identifiers point to unrelated papers; the physics papers are linked on HexagonMath.
Cite this

APA

Ground Truth. (2026, November 20). Anthropic reports two Claude-assisted results in black-hole physics. Ground Truth. https://groundtruth.day/news/claude-black-hole-physics-preprints.html

BibTeX

@misc{groundtruth:claude-black-hole-physics-preprints,
  title  = {Anthropic reports two Claude-assisted results in black-hole physics},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {nov},
  url    = {https://groundtruth.day/news/claude-black-hole-physics-preprints.html}
}

Topics: research · physics · anthropic · ai-for-science

Comments are replies to this story on Bluesky — reply with any Bluesky account to join in.