Quantum computing: Claude redesigns eight circuits and reduces the number of gates by up to 17 times

Working with researchers from Harvard and QuEra, Anthropic had Claude optimise quantum computing building blocks; in a typical scenario, a Reed-Solomon decoder’s memory requirement is reduced from 665 to 19 qubits.

Anthropic detailed on October 16 an experiment in which Claude Mythos 5 designs quantum circuits in closed loop: he writes an implementation in Python, an automatic assessor checks that it is correct and the note according to the number of quadits and doors, then the iterian model. Eight primitives were processed, covering chemistry, optimization, quantum simulation, number theory and numerical calculation. The project was led by Andi Gu, coached by physicist Alek Kemeny, with Susanne Yelin and Mikhail Lukin (Harvard and QuEra Computing).

Two results highlighted

For the addition of points on an elliptical curve, a central operation in quantum attacks against current cryptography, one of the solutions uses 12% more qubits but 17 times less doors than the best known construction; its properties have been formalized in Lean. For a Reed-Solomon decoder, Claude replaced the extended Euclide algorithm with that of Berlekamp-Massey and recalculates the intermediate values instead of storing them: about 12 times fewer doors than Google's version, and a memory reduced from 665 to 19 qubits in the typical case. Memory exceedance remains possible about once in 10,000, which remains within the tolerance margin of error, verified on 403 million simulated decodings. The authors note that the model was initially cautious and had to be encouraged to explore truly new approaches.