GPT‑5.6 Sol calibrated four quantum bits (qubits) on a new MIT chip; researchers intervened in 4 of 40 measurements
GPT‑5.6 Sol calibrated four quantum bits (qubits) on a new MIT chip; researchers intervened in 4 of 40 measurements
On September 8, OpenAI described work by Massachusetts Institute of Technology graduate student Beatriz Yankelevich. Through Codex, the agent controlled measurements on a new chip with six superconducting qubits: it read results, selected parameters, and ran the next experiment.
Calibration prepares the chip for experiments: the result of one measurement determines a setting for the next. Each qubit is read through a coupled resonator, an electrical circuit whose response changes with the qubit’s state. The lab first identifies the resonator frequency and readout signal power, then finds the qubit frequency, adjusts the control pulse, and measures how long the qubit retains its quantum state. These parameters determine the next step.
In the technical case study, software developed by MIT’s Engineering Quantum Systems (EQuS) research group sent pulses from room-temperature instruments to the chip, collected the response, plotted the results, and saved the resulting parameters. GPT‑5.6 Sol had access to measurement programs, current settings, plots, raw data, logs, and the measurement database. For each task, the researchers prepared instructions containing code, a list of completed calibrations, criteria for good and poor results, and possible physical causes of failure. At each step, the agent selected parameters, ran a measurement, analyzed the signal, and used the result either to refine the search or to save the value for the next measurement.
On the new chip, the agent found all six resonators and established an initial readout signal power. It then completed a standard sequence of interdependent measurements for four fixed-frequency qubits. Researchers intervened to improve the results in only four of the 40 target measurements.
The two qubits whose frequencies could be tuned with magnetic flux produced less distinct signals. The agent determined how their frequencies varied with magnetic flux, but Yankelevich suggested widening the range and changing the parameters. The model incorrectly judged one early scan to be acceptable: according to Yankelevich, the spectrum in the preceding scan probably extended beyond the scan range. After taking measurements over a wider range, the agent refined the result.
“I can assign agents hours of measurements overnight or while I’m working in the cleanroom,” Yankelevich says in OpenAI’s account.
The agent ran the standard measurement cycle for many hours, saving settings after each step. Yankelevich used the time this freed up to analyze results, design subsequent experiments, and examine cases where a plot supported several possible physical explanations.