Two curves govern whether quantum computing becomes useful: how capable the hardware gets, and how many resources the algorithms demand. A new collaboration between QuEra Computing, Harvard University and Anthropic attacks the second curve — using an AI model to design quantum circuits that need dramatically fewer gates and qubits.

The preprint, "Discovering state-of-the-art quantum circuits with large language models" (arXiv:2610.22367), pairs Anthropic's Claude with an automated evaluator that checked each proposed circuit for correctness and scored its resource use. Claude used that feedback to search for better designs. Across eight computational building blocks, AI-designed circuits reduced both gate counts and qubit requirements in seven of them; the eighth surfaced a new trade-off between the two.

The gains are large on paper. For a Reed-Solomon decoder — the costliest step in decoded quantum interferometry — the AI cut Toffoli gates roughly 12-fold, from 5.1 million to about 440,000, while also using fewer qubits. The Coulomb potential used in molecular simulation saw a 9.9x cut in a combined resource metric, and a fixed-point Gaussian function improved 39x, the paper's largest result, with gates falling from about 24,000 to 1,583. The authors stress these are resource estimates for individual blocks, not hardware runs, and that the combined metric is not a speedup.

Claude also discovered a new reversible approach to modular inversion, an arithmetic operation central to elliptic-curve cryptography. In one benchmark it needed about 17x fewer gates than the previous lowest-qubit design, at the cost of about 12% more qubits — new choices for balancing the two resources.

The work builds on an earlier QuEra-Anthropic pilot in which Claude wrote laser-control software that recovered a correct laser lock in 695 of 700 trials, in seconds rather than the 5-10 minutes a human expert needs. QuEra targets its fault-tolerant Libra system for availability on Amazon Braket in 2028. The broader claim: AI-assisted circuit discovery could make future quantum applications less demanding — and bring useful machines closer.