Claude Mythos 5 quantum circuits: build the evaluator first

Anthropic let Claude Mythos 5 design quantum circuits on its own, with a program grading every attempt. The model beat the best published designs. The research post from October 16 is about quantum computing, but the method is the part you can use this quarter: an AI agent works well without a human in the loop when a machine can check its work.

What actually happened

Per Anthropic's write-up:

• The tasks. Claude got plain-language specs for eight "computational primitives," the building blocks of larger quantum algorithms, across chemistry, optimization, simulation, number theory and numerical computing.
• The loop. Claude wrote each circuit in Python. An external evaluator checked it for correctness and scored it on qubits times gates, a proxy for runtime. Claude read the score and revised. Anthropic says there was no human input between rounds.
• The results. Across the tasks, Claude's designs used fewer qubits and fewer gates than the best published ones, in some cases by an order of magnitude or more. One point-addition design uses 12% more qubits but 17x fewer gates than the smallest earlier design. A decoder step came in about 12x cheaper than Google's previous construction.
• The run time. For point addition, Claude worked from a one-paragraph brief for 53 hours with minimal human input.
• The caveats. These are designs and resource estimates, not runs on quantum hardware. Anthropic also says Claude was too cautious early on and stuck to known approaches, so the team made it explore different approaches in parallel.

Why it matters for your business

Forget the qubits. Look at the shape of the system: a clear spec, an automatic grader, and a model allowed to iterate. That is why 53 unattended hours produced something useful and not 53 hours of confident noise.

Most small-business AI projects skip the grader. The agent writes the email, the product description or the price rule, and a person eyeballs it. That caps you at human review speed, and quality drifts.

Before you give an agent more autonomy, ask one question: can a program tell if the output is right?

• Yes: pricing rules checked against margin floors, SQL checked against known results, catalog data checked against a schema, code checked by tests. Let the agent loop.
• Partly: support replies checked against a policy list, ad copy checked for banned claims. Automate the checks you can and send the rest to a person.
• No: brand voice and judgment calls. Keep a human on those.

Anthropic also had to push the model out of its comfort zone. Budget for that. A grader with no pressure to explore finds the first "good enough" answer and stops.

Sources: Anthropic: Designing more efficient quantum circuits with Claude.