Anthropic says Claude produced quantum circuit designs using fewer qubits and gates, including a point-addition option with 17 times fewer gates.
Anthropic says Claude Mythos 5 designed more resource-efficient circuits for eight computational primitives spanning chemistry, optimization, quantum simulation, number theory and numerical computing. According to the company, the designs used fewer qubits and fewer gates than the best previously published algorithms, potentially lowering the hardware and runtime requirements of larger quantum computations.
The model worked in a closed optimization loop. It wrote circuit implementations in Python, submitted them to an automated evaluator that checked correctness, received a score based on qubit count multiplied by gate count, and revised its next attempt. Anthropic says no human input was required between rounds, allowing Claude to explore alternatives against a consistent, objective measure.
The most detailed test involved point addition on an elliptic curve, a core and costly operation in quantum algorithms for elliptic-curve discrete logarithms. After 53 hours with minimal human input, Claude developed a reversible implementation of divsteps for modular inversion. Rather than retaining the complete history of conditional branches, the method reconstructs that information from the changing numerical state.
Anthropic reports that Claude first tested candidate decoding rules, then derived a formula using algebraic identities. The formula was checked through large-scale numerical simulations and subsequently formalized in Lean. Because decoding branch history saves memory but takes additional time, Claude also generated a family of circuits that vary how much information is stored versus reconstructed.
One design in that family uses 12% more qubits than the smallest previous circuit but requires 17 times fewer gates. Anthropic presents this not as a single universally superior design, but as an expansion of the available trade-off frontier between qubit use and runtime.
For Decoded Quantum Interferometry, a quantum optimization method, Claude replaced an extended-Euclidean-algorithm approach to Reed–Solomon decoding with a design based on Berlekamp–Massey decoding. Anthropic says this made the most expensive stage of the algorithm about 12 times cheaper by better matching the fixed sequence of gates required in a quantum circuit.
The results are algorithmic circuit-design findings, not a demonstration of these workloads running on a practical fault-tolerant quantum computer. Their significance lies in reducing estimated resource requirements: current systems have only hundreds to thousands of noise-sensitive qubits, while many useful algorithms would need substantially more. The work also highlights a suitable setting for scientific AI agents, where proposed code can be checked automatically for correctness and scored against a defined cost metric.
Sources
Originally published by AI Feed: Claude Mythos 5 redesigns circuits for eight quantum tasks.
Top comments (0)