Anthropic STEM Fellows gave Claude plain-language specs and automated evaluators for eight quantum circuit primitives spanning chemistry, optimization, simulation, and number theory. In a closed loop, Claude Mythos 5 proposed Python circuit designs, received correctness checks and qubit×gate scores, and iteratively improved without human intervention between rounds.
Claude beat published baselines on all eight tasks, sometimes by an order of magnitude, including a new reversible modular-inversion approach for elliptic-curve point addition and ~12× cheaper decoding for decoded quantum interferometry. The work suggests AI-assisted circuit design could reduce resources needed for practical quantum algorithms (e.g., FeMoco chemistry simulations). Full paper on arXiv.