OpenAI Partners with MIT: AI Autonomously Completes Full-Process Qubit Calibration
nashnova research
OpenAI and MIT have confirmed that an AI agent powered by GPT-5.6 Sol and Codex can run quantum-computing experiments end-to-end — humans stepped in only 4 times across 40 calibration tasks, compressing weeks-long cycles to hours, and freeing researchers from repetitive tuning.
How "unknown" was this chip, and what did the AI have to do?
The test subject was a brand-new superconducting chip with 6 qubits — 4 fixed-frequency, 2 tunable — all physical parameters completely unknown.
This means → the AI was not running a known script; it had to characterize every parameter from scratch, much like a first-year physics student encountering a new device.
The workflow: the AI called lab-control software directly, fired microwave pulses at the chip sitting inside a millikelvin dilution refrigerator — a device that cools to near absolute zero — then digitized and fitted the return signals.
Only 4 human interventions in 40 tasks — how?
Across 40 routine measurement tasks on fixed-frequency qubits, the AI handled nearly all of them; human researchers stepped in just 4 times.
The pipeline has three stages: broadband sweep to catch the resonator dip (the qubit's "signal gateway") → lock the qubit transition frequency → switch to time-domain control for pulse calibration, ultimately measuring the qubit coherence time — how long a qubit holds its quantum state; longer is better.
In plain terms = a well-trained physics PhD student used to spend days camped by the instruments, running hundreds of manual measurements. The AI runs 24/7, compressing week-scale iteration to hours or even minutes.
What lets the AI make "physics judgments" instead of just following a script?
OpenAI's published chain-of-thought logs show GPT-5.6 Sol identifying why a fit residual — the gap between measured data and theoretical model — was too large, then deciding to run phase unwrapping or re-scan with an adjusted microwave detuning.
This means → the AI is making branching decisions grounded in physics — "this deviation likely comes from a phase jump; let me try a different approach" — rather than "script error, wait for a human."
Researchers also injected "skill packs" into Codex: the physics logic, evaluation criteria, template code, and example plots (successes and failures) for each quantum test — essentially giving the AI an experiment manual with built-in judgment.
Where does the AI hit its limits — can scientists be replaced?
Low signal-to-noise edge cases are the current bottleneck: the AI must repeatedly probe parameters, and time cost rises exponentially.
When the chip exhibits unpredictable anomalous physics, the AI falls into misjudgments and still needs experienced researchers to intervene manually.
In plain terms = the AI can reliably handle "well-defined standard workflows," but when results are deeply ambiguous and full of physical anomalies, a top scientist's intuition remains irreplaceable.
What does this mean for quantum computing?
Qubit calibration has long been one of the most time-consuming steps in quantum-computing research — material defects and environmental noise cause chip parameters to drift constantly, trapping large amounts of research labor in repetitive tuning.
This means → once the AI takes over standard calibration, researchers can redirect toward higher-level experiment design and mechanism research.
This reflects a shift in AI's role in scientific experiments — from "tool" to "junior researcher." Whether it can keep expanding its autonomous boundary in more complex, more ambiguous physical scenarios is the key test for whether this paradigm can truly scale.
市场有风险，内容仅供研究参考，不构成投资建议。