Anthropic’s Andi Gu-led study gives Claude Mythos 5 a concrete engineering loop: write a quantum circuit as code, check it with an external program, and improve its cost score. The score combines qubits, the circuit’s storage, with gates, the operations it performs. Eight tasks test whether the model can find better designs under those constraints.

Two examples show what the loop produced:

  • A 53-hour run developed a way to recover intermediate computation history from the current state. One resulting circuit uses 12% more qubits than the smallest previous design, but 17 times fewer gates.
  • A decoder, which reconstructs errors from a compact summary, uses 12 times fewer gates than the previous construction. Part of its storage saving accepts a failure rate around one in 10,000, within the task’s stated allowance.
  • Claude sometimes stayed with familiar designs or repeatedly checked supported results. Requiring several different approaches in parallel, plus occasional human prompts to broaden the search, helped it pursue bigger changes.

Correctness checks and a measurable target make this an instructive agent workflow. The researchers still shaped the search, and the reported gains concern circuit designs and estimated resources, not a demonstration on a useful quantum computer.

The 32-comment thread on Hacker News adds context about the coding environment and challenges the rare-input trade-off.

What the thread adds

  • keeganryan — identifies the implementation environment: “Kickmix is an assembly language for classically simulatable quantum circuits” and argues that “LLMs benefit from the closed loop of verification, and can rapidly test and discard ideas.”
  • adastra22 — challenges the decision to omit costly handling for rare inputs: “So it’s not performing the same computation…”
  • perching_aix — pushes back on that objection: “It’s not like they’re sweeping this under the rug, it’s pretty explicitly stated for which optimizations this applies. It’s also a perfectly standard optimization strategy.”

Handles are pseudonymous; HN publishes no per-comment scores, and its ordering is HN’s own ranking. This is a slice of the thread, not a consensus.