QuEra automated the critical task of keeping quantum-computer lasers precisely locked by training an AI agent (Claude) inside the Model Hardware Standard (MHS) environment to control $0.7M of precision optics within human-set safety bounds. Lasers that drive neutral-atom qubits must hold within tiny frequency windows and can fall into wrong stable modes after disturbances; recovery historically required expert craft - coarse sweeps beyond thresholds, verification against absolute references, and layered actuator coordination. The AI was allowed to run experimental recoveries on a live testbed, iterating proposals, executing them, reading instruments, and refining strategies while humans set scope, reviewed steps, and enforced hardware interlocks.
Validation showed deterministic, inspectable controllers produced by AI-driven development: 695/700 successful recoveries (99.3%) across seven disturbance classes, with time-to-recover 0.9-5.4 s for wavelength-preserving faults and 10-14 s for large hops versus 5-10 minutes for an expert. The agent also autonomously tuned twelve interdependent feedback parameters over ~12,500 safe evaluations, cutting residual frequency noise fivefold, eliminating unattended drops in a ~19-hour soak (previously 1.6/hr), and matching or exceeding expert performance in out-of-loop qubit-noise tests - suppressing a hidden 220 kHz resonance ~1,000× and improving a linewidth metric from 58 to 21 kHz. The method transferred to a second operating wavelength in one overnight run, producing similarly large noise reductions.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.