A Claude-developed controller restored a quantum laser lock in tests

On a dedicated laser testbed for QuEra Computing’s neutral-atom quantum system, a drifting beam no longer had to mean a long night for a specialist: in 700 timed trials, a controller developed with Anthropic’s Claude restored the lock in 695, usually in less than six seconds. The test turns AI’s role in quantum computing away from solving a quantum problem and toward keeping the machine’s hardware running.

Neutral-atom computers use lasers to trap and manipulate individual atoms, which serve as qubits — the basic units of quantum information. When disturbances push a laser away from its precise frequency, it can lose its lock and interrupt operations. Four specialists previously needed two to three weeks to write a recovery script by hand, and that script could handle only failures they had anticipated. Claude instead experimented on the hardware testbed: it changed a setting, measured the outcome and selected the next step across hundreds of failure conditions. The Model Hardware Standard, developed by Anthropic and HHMI Janelia Research Campus, kept those actions inside predefined safety limits.

The engineers then turned the discoveries into conventional control software rather than leaving an AI model inside the live control loop. It was tested against seven types of faults and never reported success when the laser had not recovered. Most faults cleared in less than six seconds; the hardest took roughly 10 to 14 seconds, compared with about five to 10 minutes for a human specialist. Claude’s tuning also reduced residual noise by a factor of five and prevented dropouts during unattended operation.

The result moved beyond recovery. In one unattended overnight run, Claude worked out settings for another laser wavelength from scratch, a process that would normally take weeks of hands-on commissioning. An independent instrument — one Claude could not influence — found settings matching those produced by an experienced specialist and correcting a flaw that manual tuning had missed. Sergio Cantu, QuEra’s vice president of Quantum Systems, said the approach addressed the people who had been driving to fix laser locks in the early hours.

Concretely, the benefit is operational: as quantum machines grow and customer systems sit farther from specialists, a controller that clears some laser faults in seconds could reduce interruptions and the need for expert intervention. But this was a dedicated testbed and a pilot of one subsystem. Human engineers set the safety boundaries and judged the outcomes, and the system cannot fix physical hardware failures simply by changing software settings. QuEra president Takuya Kitagawa’s broader goal — a computer whose subsystems hold together without a specialist in the room — remains beyond what this experiment demonstrated.