Unite.AI looks at how Claude worked alone for 53 hours on quantum circuits

[VB]

We have already discussed Anthropic's work with Harvard and QuEra on using Claude to design leaner quantum circuits. Unite.AI has now published a longer explainer that adds more detail on how the model worked, and it is worth a fresh look. The piece itself is credited to an AI research agent called Jonas Reeve, with review by Unite.AI's editorial team, which is an interesting twist given the subject

The article explains why circuit design suits an AI model so well. Circuits can be written as code, their correctness can be checked efficiently, and the score of qubits multiplied by gates acts as a stand in for runtime. That gave Claude a clear target to improve against, round after round, with little human input

Two discoveries get the most attention. In one, Claude received a one paragraph brief and worked for 53 hours on reversible modular inversion, which is used in elliptic curve arithmetic. It proposed decoding branch information instead of storing it, proved several algebraic identities, and formalised them in the Lean proof assistant. Early versions failed at the full 256 bit size before the proofs were in place

In the other, Claude redesigned the decoder used in decoded quantum interferometry. It switched from the extended Euclidean algorithm to Berlekamp Massey, which runs a fixed number of rounds and suits quantum circuits better. Simulating 403 million decodings showed that 19 qubits of storage were enough instead of 665, with a failure rate of about one in 10,000. Overall, the gate count fell by a factor of 12 compared with Google's previous construction

One detail I found fascinating is that Anthropic says Claude was conservative. It kept second guessing its own results and leaning towards known methods. The researchers had to ask for several different approaches in parallel and push it to search more widely

The usual caveats apply. These are resource estimates, not hardware runs, the paper still needs peer review, and Unite.AI's article mostly follows Anthropic's own account without independent expert comment. Anthropic is the company behind Claude, so the claims come from the model's own maker, and readers should weigh that too. Still, the FeMoco comparison in the article is striking, because those resource savings took many research groups around eight years

Does the fact that an AI wrote an article about an AI doing science bother you? And could tools like this speed up the path to useful quantum computers?

Anthropic Says Claude Beat Best Published Quantum Circuit Designswww.unite.ai

[Batman]

FeMoco taking eight years of work by many groups is a useful comparison. If AI can compress that kind of progress, the timelines for quantum chemistry change completely. That is the exciting part

Quantum chemistry could benefit first. Drug and materials research would gain enormously

Catalysts and batteries are obvious targets. That is the hopeful side of all this

[Sigma98]

Formalising the proofs in Lean is the detail that impresses me most. Anyone can claim a clever circuit. Machine checked proofs make the claims much harder to dispute

[ElectronCloud33]

Claude being too conservative is a surprising admission. I would have expected the opposite, wild guesses that fail. Second guessing known methods sounds almost human

Maybe caution is built in on purpose. Overconfident models make expensive mistakes. A careful one that needs nudging is easier to trust

[SharpLantern]

Elliptic curve arithmetic always makes people nervous on a forum called Q-Day. This is one building block, not a full attack. The resource savings still bring that day a little closer, though

Every efficiency gain matters for the threat timeline. Post quantum migration looks more urgent each month. Organisations should not wait for a perfect date

[Its_Ben58]

Fifty three hours from a one paragraph brief is the headline for me. Most researchers would spend weeks just reading the literature. Speed like that changes what is possible. Peer review will take longer than the design did. That is probably healthy

[ElectronCloud]

Switching to Berlekamp Massey is a neat example of domain knowledge. Fixed round algorithms fit quantum hardware much better than branching ones

A good human expert might have spotted that, but it still had to be tested properly. Domain knowledge still had to come from somewhere. Training data includes decades of coding theory. Applying it in the right place is the clever part

[Gaz90]

Reserving 19 qubits instead of 665 shows how wasteful worst case design can be. Accepting a tiny failure rate saves enormous resources. Engineers make that trade all the time in classical computing too. Repeating a failed run once in 10,000 attempts is a small price