VibeusAnthropic’s STEM Fellows Program put Claude Mythos 5 to work on... - Vibeus
Anthropic’s STEM Fellows Program put Claude Mythos 5 to work on quantum-circuit design with plain-language specifications. In collaboration with Harvard researchers Susanne Yelin and Mikhail Lukin, fellow Andi Gu and mentor Alek Kemeny asked the model to design eight circuits spanning chemistry, optimization, physics simulation, number theory, and numerical computing. An automated evaluator checked correctness and scored each design.
Mythos 5 worked in a closed loop: propose a circuit, receive feedback and a score, then revise it. On some tasks, the resulting circuits improved on the best published designs by using fewer qubits and gates—sometimes by roughly an order of magnitude.
One elliptic-curve point-addition circuit took 53 hours of mostly autonomous work, with people providing the initial brief and occasional prompts. The process drew on mathematical identities, simulation, and a Lean proof.
For a decoder used in quantum error correction, the team reports an approximately 12× reduction after Mythos 5 replaced the algorithm and reconstructed temporary values. They sampled 403 million decodings to choose a memory reservation; the selected setting can fail in about one case per 10,000, within the task’s allowed error rate.
A separate FeMoco example illustrates why attribution matters: an estimate fell from 50 trillion gates to 340 million through years of progress across the field. That roughly 150,000-fold change is not a reduction produced by Claude alone.
These are research results for particular circuit-design tasks, not a solution to quantum hardware constraints or proof that practical quantum applications are already within reach.
