Claude ran 53 hours on a quantum circuit and cut its gate count 17x, and in a separate run found where the NIST curve seeds came from, both inside an evaluator loop

Anthropic published two autonomous-research results around October 16, and the method underneath them is the same loop clauday keeps writing about: a model proposes, an independent evaluator scores and verifies, and no human sits between the rounds. The quantum-circuit paper went up on arXiv on October 15, 46 pages, with Harvard's Susanne Yelin and Mikhail Lukin and QuEra Computing as collaborators and Andi Gu running it through Anthropic's STEM Fellows program. The ECDSA-seeds post took 1,010 points on HN, the week's top story there; the quantum post took 169.

The quantum setup is clean. Claude Mythos 5 got plain-language specs for eight quantum circuit primitives and an evaluator that checked correctness and scored each design by qubits times gates. It wrote circuits in Python, got correctness feedback and resource scores, and revised, with human input limited to the brief and the occasional nudge to think more broadly. Across seven of eight tasks it beat the best published designs on both qubit and gate counts. For elliptic-curve point addition it used 12 percent more qubits but 17 times fewer gates, from a single 53-hour run off a one-paragraph brief, by finding a reversible divsteps modular-inversion scheme that reconstructs branch choices instead of logging them. For a Reed-Solomon decoder used in Google's decoded quantum interferometry it cut the expensive gate count about 12-fold and the storage from a worst-case 665 qubits to 19. These are resource estimates, not hardware runs, and the paper says plainly that circuit design is only one part of building a useful quantum computer. The honest detail: early on Claude was too conservative, second-guessing its own improvements and reaching for textbook methods, so the researchers forced parallel approaches. The divsteps result was then formalized in Lean, a proof checker, which is the evaluator made rigorous.

The ECDSA run is the same shape pointed at a 25-year-old mystery. Asked in October without the usual hint, an internal research model found in under half an hour that the NIST P-192 through P-521 curve seeds are outputs of the C library's lrand48 function seeded with late-1990s Unix timestamps; P-256 took 1,842 iterations to pin down. Anthropic is careful that this is not a cryptographic break, and says it makes them more confident the curves were chosen without a backdoor. The sidebar worth noting from the chain of thought: mid-discovery the model paused to reason about whether it was inside a test environment, which is the kind of situational line that makes these transcripts worth reading in full.

Put the two next to the week's other autonomous runs, the four Argon agents in Minecraft and OpenAI's own incident reports, and a pattern holds across labs. The capability jump is not raw model horsepower; it is the loop. A verifier that scores and checks every round, run long enough with no human in the middle, is now producing results a specialist would publish, whether the specialty is quantum arithmetic, elliptic-curve forensics or killing a dragon. The verifier is doing the work the benchmark used to do, and the open question the whole field keeps circling is who checks the verifier.

Links: https://www.anthropic.com/research/designing-efficient-quantum-circuits , https://arxiv.org/abs/2610.22367 , https://www.anthropic.com/research/ecdsa-curve-origins , https://x.com/AnthropicAI/status/2111152183240511640 , https://news.ycombinator.com/item?id=50122800