GPT-5.6 Sol 通过 Codex 在 MIT 执行量子芯片校准

OpenAI says its GPT-5.6 Sol model, harnessed through Codex, helped run routine quantum computing experiments in MIT's Engineering Quantum Systems Group. In the demonstration, graduate student Beatriz Yankelevich used an agent to coordinate measurements on a superconducting six-qubit chip, shifting a substantial part of a repetitive calibration workflow from manual execution to AI-assisted operation. The significance is not that an AI system independently replaced a physicist. It is that the system was applied to a defined, hardware-connected scientific process: selecting measurement parameters, operating laboratory equipment, analysing results and determining the next step. As described in OpenAI's case study on Codex and quantum computing experiments, the result gave the researcher more time for experiment design, data analysis and planning. For businesses watching AI move beyond chat and content generation, the MIT work is a useful example of a more practical direction. AI agents can become valuable when they are connected to real tools and given bounded, repeatable workflows with clear outputs, while people retain responsibility for ambiguous cases and higher-level decisions. From routine measurements to an agent-led workflow Quantum-chip calibration involves interdependent measurements that need to be performed repeatedly. OpenAI says the Codex-enabled agent handled an end-to-end set of clearly defined tasks, including identifying qubit transition frequencies, calibrating control and readout pulses, and estimating coherence. Those steps matter because calibration is not a single isolated command. Measurements affect later choices, and the workflow requires the system to interpret data before proceeding. In this deployment, GPT-5.6 Sol selected parameters, ran the relevant measurements through the lab hardware, assessed the results and chose subsequent actions with minimal human intervention. The work took place in MIT's Engineering Quantum Systems Group, or EQuS, where Yankelevich is a graduate student. The case study therefore describes a real laboratory setting rather than a purely simulated benchmark. It also narrows the claim appropriately: the agent automated a substantial portion of a routine workflow, not the full scientific process of developing quantum computing research. Workflow element Routine manual approach MIT demonstration using GPT-5.6 Sol through Codex Measurement parameters Researcher performs repeated calibration choices Agent selected measurement parameters Laboratory execution Researcher performs routine measurements Agent operated lab hardware for the workflow Results and next actions Researcher analyses results and plans follow-up steps Agent analysed results and decided subsequent steps Ambiguous data Researcher judgment is required Human guidance may still be required What the demonstration does and does not show The strongest conclusion from the case study is that an agent can coordinate a structured experimental sequence, rather than merely answer questions about one. That is a meaningful advance for laboratory automation because the work combines tool use, data interpretation and repeated decision-making inside a defined process. It does not establish that AI can reliably manage every experimental condition without oversight. OpenAI explicitly notes that noisy or ambiguous data may still need human guidance. That limitation is important in quantum research, where measurements can be difficult to interpret and calibration decisions can have consequences for later experiments. The practical model is therefore automation with an escalation path. The agent can take on routine, well-specified work. The researcher can intervene when data quality, exceptions or scientific judgment require it. This division of labour is also more realistic for companies than a promise of fully autonomous operations. Why this matters outside a quantum lab Most businesses do not operate a superconducting quantum chip, but many have workflows with the same basic structure: gather data, apply a known procedure, inspect the result and determine the next action. The relevant lesson is not to deploy an AI agent everywhere. It is to identify work that is repetitive, measurable and connected to approved systems. Potential candidates include: recurring simulations or test runs with established parameters quality-control checks that produce structured measurements data-processing workflows that need routine follow-up actions operational tasks that require moving between software tools and interpreting defined outputs The MIT example also highlights the prerequisites. An agent needs access to the relevant tools, an understandable workflow and boundaries for when it should stop or ask for help. In the quantum demonstration, the tasks were concrete: locate transition frequencies, calibrate pulses and estimate coherence. Businesses should seek a similarly specific starting point rather than beginning with an open-ended request to automate an entire department. That approach can help teams concentrate scarce expert time on exceptions, planning and higher-value analysis. It can also expose process weaknesses. If a workflow cannot be described clearly enough for an agent to execute safely, it may need better documentation, cleaner data or more consistent operating procedures before automation will deliver dependable results. For companies exploring agent-led operations, Scalevise can help turn a repeatable process into a controlled implementation that connects AI to the tools your team already uses. Our AI workflow automation service focuses on reducing manual handoffs, defining practical human review points and building workflows around measurable business outcomes. Start by identifying one high-volume task with clear inputs and decisions, then discuss an AI automation project with Scalevise. Frequently Asked Questions What did GPT-5.6 Sol do in MIT's quantum computing experiment? OpenAI says the Codex-enabled agent selected measurement parameters, operated lab hardware, analysed results and chose subsequent steps for routine measurements on a superconducting six-qubit chip. Did GPT-5.6 Sol fully automate quantum computing research at MIT? No. The demonstration automated a substantial portion of a defined calibration workflow. OpenAI notes that noisy or ambiguous data may still require human guidance. Which quantum computing tasks did the agent handle? OpenAI says the workflow included identifying qubit transition frequencies, calibrating control and readout pulses, and estimating coherence. What is the business lesson from this MIT demonstration? The case suggests that AI agents are most useful for bounded, repeatable workflows that combine tool access, structured data and clear points for human intervention. Conclusion The MIT demonstration shows a concrete application of GPT-5.6 Sol through Codex: coordinating routine quantum-chip calibration tasks in a working research environment. Its wider relevance lies in the workflow design. AI agents can take on meaningful operational sequences when the tasks, systems and escalation points are clearly defined, while expert people remain essential for uncertainty and scientific judgment.