Quantum computing relies on qubits, which can exist in multiple states simultaneously through a phenomenon called superposition. However, current quantum hardware is limited by a small number of qubits and high sensitivity to environmental noise. To make quantum computers practical for fields like chemistry or materials science, researchers must design 'circuits'—sequences of operations called gates—that minimize the number of qubits and the total number of gates used. A key challenge in this design is that quantum circuits must be reversible; unlike classical programs, they cannot simply discard intermediate 'scratch' data or use standard 'if-then' branching logic, as doing so would collapse the quantum state. This creates a difficult trade-off between space (qubits) and time (gate count).
In a study led by Anthropic’s STEM Fellows Program, researchers provided Claude with plain-language specifications for eight different quantum computational primitives. They established a closed-loop system where Claude proposed circuit designs in Python, which were then processed by an automated evaluator. This evaluator checked the designs for correctness and assigned a score based on the product of qubits and gates used. Working largely without human intervention for 53 hours, Claude discovered novel designs that outperformed the best previously published circuits in seven of the eight tasks. In some instances, these improvements were an order of magnitude or more, such as a new reversible architecture for modular inversion that made the 'divsteps' algorithm—previously impractical for reversible arithmetic—viable for quantum use.
The image illustrates the core research finding where Claude was used to design more efficient quantum circuits.
Claude’s success was attributed to its ability to uncover structural and mathematical insights rather than just fine-tuning existing designs. For example, when simulating an iron-sulfide molecule, the model identified symmetries that allowed certain results to be calculated once and reused. In other cases, it identified and removed costly components that were only necessary for extremely rare inputs. While these optimizations significantly reduce the resources required for useful applications, the results also highlight a specific trade-off: Claude's new designs sometimes used slightly more qubits to achieve a massive reduction in gate count (up to 17x fewer). This suggests that AI can serve as a powerful co-designer for quantum algorithms, potentially accelerating the timeline for practical quantum applications in medicine, battery technology, and industrial catalysis.