OpenAI and MIT Announce AI Successfully Takes Control of Quantum Computer

Just now, OpenAI announced in collaboration with MIT: AI has successfully taken control of a quantum computer!

In a joint report titled “Case Study: Agent Calibration for Superconducting Qubits,” researchers from MIT’s Quantum Engineering Systems Group (EQuS) confirmed that AI agents, powered by GPT-5.6 Sol and Codex, can now autonomously run quantum computing experiments end-to-end.

In other words, GPT-5.6 Sol can now conduct physical experiments autonomously.

During this process, the AI achieved autonomous operation of the quantum computing experiment, result analysis, and qubit calibration.

From releasing microwave pulses by controlling laboratory hardware, to reading and processing massive volumes of complex data, and dynamically adjusting the next experimental plan based on results, AI operates around the clock.

Although physicists won’t be losing their jobs anytime soon, the rules of scientific research have been completely rewritten at this moment.

AI for Science has advanced from "data-assisted" to "automated research."

Large models act as a research brain, directly coordinating experimental hardware, thereby completely breaking through the efficiency bottleneck in fundamental physics.

In the future, the core barrier in cutting-edge science will be "super AI + autonomous labs," and AI will finally become a reality in hardcore scientific fields!

This marks the real-world application of AI in hardcore scientific fields.

In reality, building and controlling a quantum computer is an extremely difficult task.

At MIT’s Engineering Quantum Systems group, doctoral student Beatriz Yankelevich did something bold: she integrated GPT-5.6 Sol with Codex’s autonomous agent framework to directly connect the AI’s API to the lab’s hardware control systems.

After months of iteration, researchers equipped the AI with "experimental context"—feeding GPT-5.6 all the details of experimental setups, chip design schematics, and specific measurement skills, including template code, common physical failure causes, and examples of successful and unsuccessful charts.

An unprecedented AI closed-loop automation experiment has officially begun.

To test the limits of AI, the research team gave GPT-5.6 a challenging task: a newly created chip containing six superconducting qubits that had never been measured before.

Four of them are fixed-frequency qubits, and two are tunable-frequency qubits. The chip’s parameters are completely unknown, and the AI must explore and determine the actual physical parameters based on the design objectives.

From hardware control software to microwave signal sources, and on to data acquisition cards—all are opened to AI.

Researchers equipped Codex with specialized "skill packs" tailored to specific measurement scenarios, instructing it on the physical logic and evaluation criteria for each quantum test.

Subsequently, humans let go with their hands.

On screen, GPT-5.6 Sol's chain of thought flows like water.

It executed a series of hair-raising physics-based闭环 operations without real-time human intervention.

1. Autonomous parameter inference: Based on the chip’s design objectives, autonomously determine the initial microwave pulse frequency, power, and width for measurement;

2. Underlying hardware scheduling: Directly invoke control software to manipulate hardware and emit microwave pulse sequences toward the chip deep inside the dilution refrigerator;

3. Data Reiteration and Denoising: Receive the weak quantum-state reflected signals, then digitize, clean, fit, and apply Fourier transforms to them;

4. Adaptive Decision Iteration: If the signal is optimal, automatically extract the qubit's resonant frequency and store it in the database as a reference for the next measurement; if the data is anomalous, adjust the measurement boundaries in real time and initiate a new round of trial-and-error optimization.

From identifying qubit transition frequencies and calibrating microwave pulses for control and readout, to ultimately measuring the coherence time—how long quantum information can be retained—the entire standard physics measurement workflow is autonomously executed by AI.

In 40 complex target measurements on fixed-frequency qubits, AI autonomously handled the vast majority, with human researchers intervening only four times!

This is AI conducting a textbook-perfect "autonomous hands-on" operation at the forefront of physics.

To understand the significance of this breakthrough by GPT-5.6 Sol, we must first lift the cold yet fascinating veil over quantum computing’s foundations: why does calibrating quantum chips drive the world’s top physicists to despair?

In classical computers, the state of a transistor is strictly binary: either 0 or 1, completely stable.

In the world of superconducting quantum computing, qubits are called "artificial atoms."

They use superconducting Josephson junctions to create nonlinear discrete energy levels.

To make these "artificial atoms" behave as desired, scientists must use microwave pulses precise to the nanosecond or even picosecond level to flip quantum states between different energy levels, thereby performing quantum logic gate operations.

But "artificial atoms" are too fragile and too finicky.

First, coherence is extremely fragile.

Extremely weak electromagnetic stray signals in the environment, or even temperature drifts of a few millikelvins (mK) in a dilution refrigerator, can cause qubits to undergo "decoherence," causing the carefully prepared quantum information to instantly collapse into noise.

And "parameter drift" is also a challenging issue.

The physical properties of qubits are never constant. Affected by microscopic material defects and two-level system fluctuations, the resonance frequency measured on a chip yesterday may have drifted today.

Moreover, multiple factors in the study are interdependent and intricately linked.

To perform a precise calculation, you must first determine the resonance frequency of each qubit; once the frequency is found, calibrate Rabi oscillations to determine the pulse amplitude; then measure Ramsey fringes to precisely correct the phase; next, measure the energy relaxation time $$T_1$$ and the phase decoherence time $$T_2$$…

The data from the previous step is an absolute prerequisite for the next measurement. Even a minor systematic error in the prior step renders all subsequent measurements invalid.

The EQuS team at MIT faces an extreme experimental environment daily: quantum chips are placed inside a large cylinder called a dilution refrigerator, cooled to an astonishing millikelvin temperature (close to absolute zero, colder than outer space).

Once the chip package has cooled, humans can no longer touch it directly; all interactions must be conducted through software.

Researchers program microwave pulse sequences on a computer at room temperature, send these pulses through coaxial cables deep into a refrigerator to interact with the quantum circuits on the chip, then amplify, digitize, and bring back the faint reflected signals to the real world for analysis.

The daily routine at MIT’s EQuS lab involves frequently fabricating a variety of standard quantum chips for benchmarking.

In reality, performing a complete characterization of each chip often requires a trained physics graduate student to monitor the equipment continuously for days, performing hundreds or even thousands of interconnected manual measurements and adjustments.

It’s an open secret in the physics community: contemporary quantum physicists often spend more than 70% of their youth on repetitive experimental labor known as “tuning screws.”

Now, GPT-5.6 Sol has broken this deadlock.

According to details disclosed by OpenAI and MIT, this test was not a pre-scripted "automated macro," but rather a highly intelligent dynamic博弈.

Faced with this uncalibrated 6-qubit chip, GPT-5.6 Sol demonstrated highly human-like "physical intuition" and "adaptive decision-making":

When quantum signals are in a high signal-to-noise ratio with distinct characteristics, Codex performs flawlessly.

It first emits a wideband sweep signal, keenly detecting the resonance dip in the cavity reflection caused by quantum level transitions against a noise background, then automatically narrows the scan window to precisely lock onto the qubit's transition frequency.

Next, it seamlessly transitions to time-domain control, precisely fitting the relationship between pulse amplitude and flip angle, calibrating the readout and control pulses, and ultimately measuring the time window during which the qubit maintains its coherent state.

The entire process flows seamlessly, requiring no human intervention whatsoever.

In the leaked technical documentation from OpenAI, GPT-5.6 Sol’s chain of thought reveals its reasoning process. Faced with a damped sine curve exhibiting fluctuations, it internally mutters:

The current fitting residual is unusually large, possibly due to drift in the reference baseline for demodulated phase, causing I/Q component mixing... I need to first perform phase unwrapping; if the residual remains above the threshold, I should increase the microwave detuning by 5 MHz and rescan...

This kind of error correction and branching decision-making, based on an understanding of physical meaning, is precisely the barrier that traditional script programs could not overcome.

Of course, the official report also reveals the very real technological boundaries:

When signals become weak or are overwhelmed by physical background noise, GPT-5.6 Sol begins to appear "clumsy."

In low signal-to-noise ratio edge regions, AI models must repeatedly test parameters, causing time consumption to rise exponentially; sometimes, when physical chips exhibit unpredictable anomalous physical behaviors, the AI may misinterpret the results and fail to accurately attribute the cause, ultimately requiring experienced researchers to intervene manually.

OpenAI stated frankly in its report: current AI agents can execute well-defined standard workflows with exceptional stability, but the intuition of top scientists remains irreplaceable when interpreting highly ambiguous, complex results filled with physical anomalies.

Even so, it has accomplished the most difficult leap in human history: fully liberating scientists from tedious, routine tasks.

However, after the introduction of GPT-5.6 Sol and Codex, the laboratory's relations of production were completely restructured:

Role elevation: Humans focus on high-level design (solutions, mechanisms, overall planning), decoupling from low-level parameter tuning.

Multi-agent collaboration: Establish three AI roles—theory, control and monitoring, and simulation—operating in parallel.

Real-time self-evolution: Codex handles tactical execution—once code is generated, it is directly deployed onto real quantum chips, enabling real-time reverse debugging based on physical feedback to form an ultra-fast closed loop.

This super闭环 of "theoretical deduction - simulation design - physical hardware control - data fitting - iterative optimization" has reduced the scientific research iteration cycle, which used to take weeks, to hours or even minutes.

This is an elevated assault on the entire paradigm of experimental physics.

Today, OpenAI is making headlines worldwide—they claim to have achieved a major mathematical breakthrough on the "Millennium Problem," proving the Navier–Stokes equations.

The OpenAI proof indeed shows that there exist certain fluids which, while initially perfectly normal, reach infinite velocity in finite time under the Navier-Stokes equations.

In other words, under certain conditions, this equation may not reliably reflect physical reality.

This is the first time AI has truly solved a mathematical problem on the level of the Riemann Hypothesis. This problem is not only famous for its extreme difficulty but also for its wide-ranging applications.

The Navier-Stokes equations can be used to simulate weather, ocean currents, water flow in pipes, the motion of stars in galaxies, and airflow around airfoils, as well as to design aircraft and vehicles, study blood circulation, design power plants, and analyze pollution effects.

But another mathematician, Tristan Buckmaster, raised doubts, claiming that after hearing about the research by him and Anthropic researcher Levent Alpöge, OpenAI rushed to release what it called a "solution."

Tristan Buckmaster also claimed that OpenAI attempted to exclude Anthropic researcher Levent Alpöge from authorship on the paper.

OpenAI claims they only solved the singularity problem for the Euler equations, while they solved the Navier-Stokes equations.

Buckmaster said they were one step away from solving the Navier-Stokes equations and, to avoid being scooped, hastily published their findings before OpenAI.

According to reports, OpenAI deployed 10,000 agents, ran the process for 88 hours, and consumed 130 billion tokens.

Altman also said the two approaches appear different, but admitted that Anthropic’s actions were indeed a catalyst: “We did explore this direction because last week there were rumors online that Anthropic’s model had solved a Millennium Problem, and we were curious whether our model could do the same.”

Some netizens joked:

I heard Anthropic discovered a room-temperature superconductor—hopefully no competitors will waste 88 hours and 10,000 post-Astra agents trying to rush out a press release.

Ultraman replied: "Let’s try" (stay tuned).

Everyone exclaimed, "Math is doomed," and the real survival crisis for mathematics has erupted.

But whenever it comes to the physical world, large models often seem powerless.

But this time, OpenAI's announced case has completely shattered the barrier between the virtual and physical worlds.

AI has truly entered the human physical world for experimentation.

Reference materials:

https://cdn.openai.com/pdf/case-study-agentic-calibration-of-superconducting-qubits.pdf

https://openai.com/index/codex-quantum-computing-experiments/

https://x.com/anabology/status/2097450250067689558

https://x.com/LuminaBench/status/2097421659552518226?s=20

This article is from the WeChat public account "New Intelligence Yuan," authored by ASI Revelation; edited by Aeneas David.