· via dev.to (home feed)
dev.to post claims Anthropic's Claude Fable 5.2 solved two open black-hole problems
A dev.to blog post says Anthropic's Claude Fable 5.2 produced solutions to the black-hole information-loss paradox and cosmic censorship, but the claims rest on a single unverified source.

What the post claims
A post on dev.to's community feed reports that Anthropic unveiled a large language model called Claude Fable 5.2 on 19 November 2026, and that the model produced what the author describes as the first AI-generated answers to two long-standing open problems in black-hole physics, written up so that trained physicists could submit them for peer review. The two problems are the information-loss paradox for evaporating micro black holes, and the question of whether a regular rotating solution to the semiclassical Einstein equations can respect the weak cosmic-censorship conjecture.
According to the post, a technical note on Anthropic's research portal contains the equations, code and validation data needed to reproduce every result, and independent groups at the Institute for Advanced Study and the Max Planck Institute for Gravitational Physics cross-checked the constructions. The post also says the announcement reached the top of Hacker News with 252 upvotes and spread through Google News feeds. None of these claims can be corroborated from the available material: the dev.to post is the sole source, and no Anthropic statement, physics preprint or institutional confirmation accompanies it.
The claimed results
On information loss, the post says the model built a tensor-network representation of a black hole's interior, structured like a multi-scale entanglement renormalization ansatz, that yields a unitary map from interior states to radiation states. In a simulation of a black hole of roughly ten Planck masses, the model reportedly tracked the entropy of emitted radiation onto the Page curve to numerical precision, with the trace distance between exact and approximate density matrices below one part in a million. A back-reaction term is said to keep the effective Hamiltonian Hermitian, avoiding the non-unitary shortcuts of earlier proposals.
On the rotating solution, the post says the model derived an analytic metric that adds a type-A trace-anomaly correction, with the anomaly coefficient set to about 0.03, the Standard Model value. The claimed solution stays regular up to a spin parameter of 1.12, beyond the classical Kerr bound of 1, while a back-reaction analysis reportedly shows that infalling matter cannot push the geometry past the point where censorship would fail. The post also reports a shift in the fundamental quasi-normal-mode frequency of roughly 1.8 thousandths scaled by the anomaly coefficient, relative to Kerr, and describes that deviation as within reach of next-generation gravitational-wave observatories.
Problems the post itself acknowledges
Even taken at face value, the post lists reproducibility gaps. The Python solver library the model relied on, named aishoot, is not open source, and its heuristics are undocumented, so independent groups had to re-implement the algorithm from published pseudocode and obtained slightly different numbers. The tensor-network construction depends on the chosen bond dimension, and no systematic study of that dependence exists. The post also concedes that experts dispute whether a spin parameter above the Kerr bound reflects genuine physics or merely the presence of higher-derivative terms in the equations.
Why the skepticism is warranted
Beyond those internal caveats, the story has the shape of an unverified rumor rather than a confirmed result. It rests on a single community blog post. Parts of the write-up are explicitly hypothetical, walking through an imagined graduate student's workflow, and the narrative drifts between that illustration and claims of real replication, which makes it hard to tell what is documented. A result of this magnitude, on questions open for roughly half a century, would normally arrive as a preprint with named authors and detailed institutional announcements, not as a feed post.
Why it matters
If the claims survived verification, it would mark a real shift: language models moving from assisting with code and literature to generating symbolic derivations and candidate solutions at the frontier of theoretical physics, with human experts doing the checking rather than the creating. The bottlenecks the post itself admits, from closed-source solver libraries to parameter sensitivity, would then define the next problem: how to peer-review research whose provenance is a model. If the claims dissolve instead, the episode still illustrates how quickly extraordinary AI stories travel through community feeds, and why technical readers should wait for primary sources before amplifying them.
- #anthropic
- #claude
- #machine-learning
- #physics
- #research