OpenAI says GPT-5.6 Sol autonomously calibrated routine measurements on a six-qubit MIT chip
- MIT EQuS connected GPT-5.6 Sol in Codex to laboratory control software and tested it on an uncalibrated six-qubit superconducting chip used to benchmark the group's fabrication process.
- Using measurement-specific skills and chip design targets, the agent selected parameters, ran hardware measurements, analyzed data, then refined a measurement or saved its result for the next step.
- When signals were clear, Codex completed standard calibration with little human intervention, identifying transition frequencies, calibrating control and readout pulses, and measuring quantum-information retention time.
- Weak or noisy signals took the agent longer to tune and sometimes required an experienced researcher, leaving ambiguous physical results outside reliable autonomous handling.
- EQuS says it now regularly uses agents for routine measurements on standard chips that can take a researcher several days to characterize, shifting researcher time toward data analysis and experiment design.
Hacker News opinions
I wonder why OpenAI is publishing posts so close together. Is it reacting to controversy around the Navier-Stokes post?
AI has sped up work enough that frequent announcements may be normal now. At my company, two months of internal releases feel like more than people can absorb, and we no longer announce every small improvement.
I think OpenAI is rushing toward an IPO before the bubble bursts.
This reads as a blog post, not a major product announcement. The FLT proof was much closer to the latter.
I had qubit bring-up and calibration automated with Python in 2011, including spectrum scans, lifetime tests, Rabi and Ramsey measurements, gate calibration, and process tomography. AI would save routine programming time, but chip quality and qubit lifetimes were still my real limits.
Even if I used AI here, I would want it writing Python scripts rather than following an analysis.md file.
The linked factorization-record paper is about contrived large numbers with known easy factorizations, so it is unrelated. This post is about an LLM calibrating a chip, though a Python expert system may still do this job better today.
I see LLMs as a very powerful but dumb boulder that people have pushed downhill without knowing where it will land. That power, rather than superhuman intelligence, worries me.
I have seen good programmers become much faster when an LLM handles grunt work, while they still control the final code quality. It can also produce unmaintainable slop, like any badly used tool.
The business model may be to buy enough compute to spend unused capacity on patentable scientific work and models competitors cannot run yet. Capital has long been used to hire labor and produce patents, so this is not entirely new.
I dislike the framing that AI frees people for other work. People should prepare for AI doing much more of the work itself, rather than taking comfort from that wording.
OpenAI seems to run two messages at once: models are helpful assistants, but AI also poses frightening risks.