Towards a quantum computer that learns from its errors
Google Research
Google trained an AI agent to keep a quantum computer's error-correcting chip calibrated while it's actually running, instead of stopping to fix drift. Result: 3.5x more stable operation, which is a big deal since quantum algorithms need to run continuously for real work.
Based on reporting by Google Research — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Quantum computers have a dirty secret: they can't stay in tune. The control signals that steer qubits — frequencies, amplitudes, phases — drift constantly, and until now the only fix was to halt everything and recalibrate. That's fine for short demos, but useless for the kind of quantum algorithms that need to run for days or months without interruption.
Google's new approach, described in a Nature paper out of Google Research and DeepMind, treats the error-detection data generated by quantum error correction as more than just a signal to fix mistakes after the fact. Normally, decoders like AlphaQubit or Tesseract read those detection events and patch the logical qubit's state, without ever asking why the errors showed up. Google's twist is to feed that same stream of data into a reinforcement learning agent that continuously adjusts thousands of control parameters on the fly, nudging the hardware back into tune while computation keeps going.
They tested it on Willow, Google's flagship superconducting chip, by intentionally injecting artificial drift and letting the RL agent fight back. The payoff was a 3.5x improvement in how long the processor could hold a stable logical qubit. Even more telling: after human experts had already tuned the system as well as they could by hand, the RL agent still cut the logical error rate by another 20%. Combined with Google's other recent QEC advances, the team pushed error rates down to record lows — under one in a thousand cycles for the surface code, one in a hundred for the color code.
The scaling question matters more than the headline numbers, honestly. Google ran simulations with hundreds of qubits and tens of thousands of parameters and found that the number of training iterations the RL agent needs doesn't grow with system size, thanks to the local nature of QEC detection events. That's the detail that turns this from a cute lab trick into something that might actually survive contact with a thousand-qubit machine.
This fits a pattern Google keeps pointing to: hand-built physics models eventually hit a ceiling, and machine learning trained on real data blows past it. It happened with computer vision, it happened with AlphaFold and protein structures, and it's now happening with AlphaQubit outperforming algorithmic decoders. Quantum calibration looks like the next domain where decades of physicist intuition get quietly outcompeted by an agent that just watches the errors and adapts.
My take — AI-written commentary, not fact-checked reporting
I'm a sucker for anything that swaps hand-tuned physics models for learned ones, and this is a clean example of that trend eating another field alive. The scaling result is the part that actually matters, not the flashy 3.5x number, because a calibration trick that only works on toy chips is worthless for the thousand-plus-qubit machines everyone claims are coming. Google gets credit here for testing that, even if the real proof will be an RL agent babysitting a machine that runs for months, not a lab demo that runs for minutes.
Read more about this at: Google Research