AI solves 18-year-old quantum math problem in 3 days

An AI research system proved a major quantum computing conjecture in just three days, and a machine verified the math.

Research · Source: Physical Superintelligence

What happened

Physical Superintelligence built an internal research system using a non-public Anthropic model. They pointed it at the Aaronson-Ambainis conjecture, a fundamental quantum computing problem open since 2008. David Turturean, a member of the technical staff, scoped the problem and prompted the model.

The AI found a complete proof during the third day of running on internal infrastructure. It did not need human peer review to confirm the math. A computer program called Lean checked every step and accepted the final version. The system then worked on related open problems and improved the result twice.

The proof settles a massive question about quantum computers. It proves they cannot be exponentially faster than classical computers on random inputs. Stanford researchers Guy Blanc, Jordan Docter, Carmen Strassle, and Li-Yang Tan had solved a special case earlier in August 2026. They read the new full proof and called it a huge result that brings everything together in a magical way.

Key facts

Why it matters

AI is no longer just summarizing text or writing boilerplate code. It is doing original, frontier-level mathematics. The system generated the mathematical statements, wrote the first draft, and wrote the formal code to verify it. The proof required replacing a conjectural inequality with a new all-pairs inequality. This shifts AI from a tool for productivity to an engine for hard scientific discovery.

The verification step changes the game for builders. By pairing a large language model with a strict proof assistant like Lean, hallucinations are caught instantly. The machine checks its own work through Lean kernel and lean4lean. This means we can trust AI outputs on complex logic problems without waiting months for human experts to review them. The trusted base is now the kernel and the compiler, not the human reader.

The result also grounds quantum computing hype. We now know for sure that quantum computers need structured inputs, like the repeating patterns in Shor algorithm, to achieve exponential speedups. On random yes-or-no questions, an ordinary computer can simulate the quantum machine without needing exponentially more looks at the data.

For builders

Pair LLMs with strict verifiers

Do not rely on language models alone for hard logic. Pair them with deterministic systems like Lean or compilers. Builders who build self-checking loops will win enterprise contracts because they eliminate hallucinations.

Master problem scoping

The AI did the math, but a human scoped the problem and prompted the model. Your edge is learning to research with AI. Founders who know how to frame complex bottlenecks for AI systems will outpace those who just use chatbots.

Target unstructured data limits

The proof shows quantum computers do not get exponential speedups on random inputs. Companies building quantum algorithms must focus on highly structured problems. Investors will stop funding quantum startups that promise magic on unstructured datasets.

My take

I keep telling founders to stop treating AI like a toy. If an AI can solve an eighteen year old quantum math problem in three days, it can solve your business bottlenecks. Your edge is no longer having the answers, but knowing how to point the machine at the right questions.

Original reporting: Physical Superintelligence. This is my rewrite and opinion.

More AI news for builders