A learner steers a quantum code without stopping it
A Nature paper from a Google Quantum AI team describes a reinforcement-learning controller that treats error-detection events inside a quantum error-correcting code as a training signal, so control settings can shift while correction continues instead of stopping for a full recalibration. On a Willow superconducting processor, steering the controls made an injected drift about 2.4 times more stable in logical error rate, and about 3.5 times more stable when the decoder was steered too. After ordinary calibration, further learning reduced logical error by about a fifth. The reported distance-7 surface code averaged roughly 7.7 logical errors per 10,000 cycles. The experiment is a quantum memory in the laboratory; uninterrupted steering of one long algorithm is explored in simulation.
The record ends here. Everything below this line is invented.
What if a quantum computer learned from its errors and never paused?
Part one
The Unpaused Code
Lina's shift had no handover. The code on the cold stage had been correcting itself since the week she was hired, and the only control she was allowed to touch was the log of small nudges the learner made to the pulses.
She liked the nudges. They were thinner than a calibration sheet, and they never asked the machine to forget the state it was protecting. A drift in one coupler showed up as extra detection clicks. By the time her coffee was cool, the learner had walked the pulse back.
Tonight the clicks did not walk back. They gathered on one measure site and stayed, a fine rain on a single tile. The learner's spread, which usually tightened when it was sure, opened again. The card on the wall called that exploration. It was how the machine kept up with a slow drift. It was also how a living computation could be poked until it failed.
Lina set her palm on the abort that would dump the logical state and start the old calibration from nothing. The clicks stepped sideways, onto a second site, as if the trouble had moved rather than grown.
She lifted her hand. On the trace, the observable that had been fading held its line.
What’s real
- Error detections can train controls during correction
- Injected drift was stabilized without a full stop
- Further learning cut logical error after calibration
- The demonstration was a laboratory quantum memory
What’s invented
- A code that runs for years without a pause
- Lina and the abort under her palm
- A drift that steps to a second site
- An observable that holds because she waits
Part two: The Spread She Would Not Close
Part two isn’t on sale yet. Check back soon.
Filed under physics, quantum error correction, reinforcement learning, superconducting qubits.