OpenAI says GPT-5.6 Sol handled routine measurements on an MIT quantum chip, but unclear signals still demanded expert oversight.
OpenAI reported on September 8 that MIT graduate student Beatriz Yankelevich had linked GPT-5.6 Sol, working through Codex, to software in a superconducting quantum-chip laboratory. The test examined whether an AI agent could take over some of the manual preparation and calibration work that precedes experiments with qubits.
The trial began with a six-qubit device that had not yet been calibrated. It followed a standard chip design that MIT’s Engineering Quantum Systems Group uses when evaluating manufacturing quality. OpenAI says the software connection allowed Codex to control the equipment, set measurement parameters, inspect the resulting data and determine whether a reading needed further refinement before the workflow advanced.
Yankelevich created measurement-specific skills that set out the procedures and evaluation rules for each type of test. With those instructions and sufficiently clear experimental signals, OpenAI says GPT-5.6 Sol required little human involvement while identifying qubit transition frequencies, tuning control and readout pulses, and determining how long a qubit preserved quantum information.
The company says the setup could keep routine measurements running for hours, including overnight, with the researcher watching its progress and changing course when needed. OpenAI also reports that the MIT group now uses agents regularly to characterize standard chips. Without such assistance, that routine work can occupy several days of a researcher’s time; preparation for quantum experiments may involve hundreds or thousands of preliminary measurements.
Performance deteriorated when readings were faint or affected by noise. In those conditions, the model could spend more time searching for usable parameters and sometimes required instructions from an experienced researcher. OpenAI further acknowledges that a specialist may reach the best calibration settings faster. For unfamiliar experiments, Yankelevich therefore limits agents to narrower assignments while keeping responsibility for interpreting results and choosing the research direction.
The reported result supports automating defined, repeatable laboratory tasks rather than handing an entire investigation to an AI system. Connecting a model to lab software and formalized measurement procedures may extend how long equipment can operate without continuous attention. The stated limitations remain consequential: ambiguous physical behavior still needs scientific judgment, so the experiment does not establish general autonomy in research.
Sources
Originally published by AI Feed: OpenAI Tests GPT-5.6 Sol on Six-Qubit Chip Calibration.
Top comments (0)