Claude solves two black hole physics problems

A glowing, fragmented sphere suspended in a dark void, its surface splitting into two distinct textures—one smooth and curved like spacetime, the other pixelated and crystalline like quantum data—conn

Anthropic’s Claude has solved two open problems in theoretical physics related to black holes and quantum gravity. The findings, while not directly applicable to our universe, mark a significant step in understanding how general relativity and quantum mechanics might coexist. Both results emerged from an internal research project at Anthropic and have since been formally verified using Lean, a computerised proof-checking system.

The work highlights AI’s growing role in fundamental science, particularly in domains requiring long chains of abstract reasoning. For researchers and technologists, especially in the UK, UAE, US, and Kazakhstan—where investment in AI-driven science is rising—this signals a shift in how theoretical breakthroughs may be achieved.

A phase transition that survives quantum uncertainty

The first problem concerns the Hawking-Page phase transition, a theoretical shift in a model universe where, upon heating, a gas of particles suddenly collapses into a black hole. Originally described in 1983 using general relativity, this abrupt transition was well understood in large, classical spacetimes. But in smaller, quantum-dominated regimes, spacetime loses its definite geometry, making the gravity-side calculation intractable.

In 1999, physicists found evidence of a similar transition on the quantum side of holographic duality—a framework where a gravitational system in higher dimensions is equivalent to a quantum system in one fewer dimension. However, they couldn’t determine whether the transition remained sudden or became smooth. In 2003, a method to resolve this was proposed, but the calculation was too complex to complete.

Claude completed it. By running a swarm of agent versions of itself—each capable of reading progress updates and collaborating with others—it independently worked through nearly every piece of the mathematical derivation and cross-checked its results against known physics. It confirmed the transition remains abrupt, even when spacetime is no longer well-defined. This suggests that black hole formation is not merely a feature of classical gravity but is embedded in the underlying quantum description.

Tightening the bounds on quantum gravity models

The second result involves the modular bootstrap—a technique used to test candidate theories of quantum gravity by ruling out inconsistent ones. In a simplified two-dimensional space model, physicists have long sought to determine the maximum possible mass of the lightest black hole. Any theory predicting a heavier one is invalid.

Since 2009, numerical methods have progressively lowered the theoretical upper bound. While simulations suggested an optimal limit, no proof existed. Claude 3 Opus (not Claude 5.2) identified a hidden pattern in the numerical data and applied a mathematical method borrowed from sphere-packing theory. This allowed it to derive the exact limit for the simplest version of the bootstrap and prove it cannot be lowered further.

Notably, one half of the proof mirrors a strategy recently used by an OpenAI model to solve a high-dimensional sphere-packing problem. The connection suggests cross-pollination between AI-driven advances in physics and pure mathematics. The result was later verified in Lean, a formal proof assistant, and independently replicated by multiple agent instances of Claude. A physicist reviewed all main claims, and Claude drafted the paper with minimal human input.

What it means

These findings do not change engineering or applied physics. They operate in highly abstract, non-physical models. Yet they contribute to the foundational effort to unify quantum mechanics and gravity—a challenge central to modern theoretical physics.

For businesses and institutions investing in AI for research, the implications are concrete. Claude didn’t just assist; it led a self-directed investigation, identified promising problems, collaborated across internal agents, and produced formally verifiable results. This level of autonomy goes beyond current tools used in academia or industry.

In the UK and US, where AI research is tightly linked to national science strategy, such capabilities could accelerate discovery in fields like materials science or quantum computing. In the UAE and Kazakhstan, where AI adoption in research is still emerging, this case offers a benchmark for what’s possible with strategic investment.

The fact that two independent researchers—human and AI—reached the same conclusion via different paths strengthens confidence in the results. Anthropic reports minimal human guidance; a physicist later verified all main claims.

AI is not replacing physicists. But it is becoming a co-researcher—one that can persist through complex, tedious calculations and spot patterns invisible to humans. The next challenge will be integrating such systems into collaborative scientific workflows without over-relying on opaque reasoning processes.

Sources

Leave a Reply

Your email address will not be published. Required fields are marked *