GPT-5.6 Sol runs a quantum experiment. AI is no longer just writing code; it is choosing the next measurement

An MIT experiment shows a shift from AI as a programming assistant to AI as an active component of a physical laboratory. Beatriz Yankelevich of the Engineering Quantum Systems Group connected Codex and GPT-5.6 Sol with software controlling measurement equipment. The model can prepare code, start a measurement, load the results, evaluate them and choose the next experimental step. This is not merely proposing a procedure in chat, but a closed loop between the model and laboratory instruments. Yankelevich describes the practical advantage succinctly: she can let agents run measurements for many hours overnight. According to OpenAI, the system was able to operate with little intervention when signals were stable and easy to read. When noisy or ambiguous results appeared, however, human judgment remained necessary. This is an important correction to ideas of an autonomous laboratory. The model can dramatically accelerate routine iterations, but weak signals and unexpected phenomena remain areas where human experience matters. Economically, the utilization of expensive laboratory infrastructure changes. Measurement equipment often sits idle outside working hours or waits for a person to prepare the next run. An agent can extend the usable time of the apparatus and remove delays between experiments. Return on investment does not arise automatically, however. If the system performs unnecessary measurements, it consumes capacity, material and energy. The laboratory therefore needs limits on the number of steps, the range of permitted parameters and rules for when an experiment must stop. For instrument and laboratory-software manufacturers, this is a signal that interfaces for safe automation may become a new competitive advantage. An instrument that provides precisely defined commands, status information and machine-readable error messages will be easier for agents to operate than a closed system. It will also be important to separate simulation from real action: the model should first prepare a plan and only after review receive permission to change sensitive device parameters. The regulatory impact will be greatest in fields where experiments affect safety, healthcare or manufacturing. There it is not enough to prove that the model can perform the task. The organization must document who approved the action, what data the model saw and how the system behaves when the result is abnormal. The MIT case is therefore not proof that a laboratory can be handed over to AI, but a very concrete guide to where to start: repeatable experiments with clear boundaries and measurable outcomes. In practice, it will also be necessary to separate the right to propose an experiment from the right to actually run it. A model may have broad access to documentation and historical measurements but narrower permissions for the equipment itself. This graduated access can use AI's speed without allowing one incorrect interpretation to damage a sample, an instrument or an entire experimental run. KEY TERMS Qubit: The basic unit of quantum computation, using properties of quantum mechanics. Calibration: The process of setting and verifying device parameters before the actual measurement. Closed loop: A procedure in which the result of one measurement automatically determines the next step.

GPT-5.6 Sol runs a quantum experiment. AI is no longer just writing code; it is choosing the next measurement

The full article is available to users registered at Hard Skills