An evaluator-in-the-loop setup let a model improve on published circuit designs, in some cases by an order of magnitude. This is a replicable pattern for agentic search over scored design spaces.
Key takeaways · AI-distilled
The score Claude optimized was qubits used multiplied by gate count, a proxy for runtime. The targeted circuits can be checked by an ordinary computer because each input gives one definite output.
Anthropic says Claude Mythos 5 ran in a closed loop: propose a circuit, get a correctness check and score from the evaluator, and revise, with no human input between rounds.
Because circuits can be written as code, Claude could write, run and revise designs as ordinary programs, working from plain-language specs for eight primitives in chemistry, optimization, physics simulation and number theory.
As context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition →, the post notes FeMoco estimates fell from about 50 trillion Toffoli gates in 2017 to about 340 million, roughly 150,000-fold, mostly from redesigning circuits rather than better hardware.
Terms in this piece · Glossary
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
