Quantum Error Correction
Google shows in late 2024 that more qubits reduce errors instead of adding them; 2026 brings records with dozens of logical qubits.
Open in the interactive tree →Qubits are so error-prone that many physical qubits are bundled into one “logical” qubit that continuously detects and corrects errors. This only works if individual errors are below a threshold; then the total error shrinks with every enlargement. Showing this “below threshold” had been open since Peter Shor’s proposal in 1995.
As of October 2026
Google’s Willow chip (105 qubits, 9 December 2024) halved the error with each enlargement of the surface code from 3x3 to 5x5 to 7x7 (about 0.14% error per cycle). Since then Quantinuum’s Helios (98 trapped-ion qubits; the company advertises up to 50 logical qubits) and Harvard/QuEra (algorithms with up to 96 logical qubits on neutral atoms) have raised logical-qubit counts, and in September 2026 Quantinuum reported its Helix code at 4.6e-5 error per logical qubit per cycle on two logical qubits. On 7 October 2026 DARPA advanced IBM, Atom Computing, Diraq and IonQ to Stage C of its Quantum Benchmarking Initiative, which tests whether a useful quantum computer is possible by 2033.
Open steps
- Real-time decoders on control hardware High AI leverageDecoders must read error syndromes and answer within about a microsecond on chips next to the qubits, for ever larger codes and new code families.
- Cheap non-Clifford gates Medium AI leverageUseful algorithms need T gates from magic states; making them with few qubits and low error is a leading cost of fault tolerance.
- High-rate codes with real wiring Medium AI leverageqLDPC codes need far fewer qubits than the surface code but long-range couplings that most chips lack, and no full demonstration exists yet.
- Automatic calibration at scale High AI leverageThousands of control parameters drift; reaching 100+ logical qubits needs calibration that keeps error rates stable during runs without human tuning.
Where AI could help
Medium AI leverage. Neural decoders and learned calibration already help at the next scale step, but the error rates of the physical qubits are a hardware matter.
- Neural decoders that keep up with microsecond cycles at larger code distances
- Reinforcement-learning calibration of control parameters during operation
- Search for lower-overhead codes and fewer-gate circuits
- Large-scale noise simulation to plan next chips
Shown so far
- In December 2025 (preprint, revised March 2026) AlphaQubit 2 decoded surface codes up to distance 11 in under 1 microsecond per cycle on commercial accelerators, and color codes up to distance 9. source
- In November 2024 Google's AlphaQubit neural decoder made 6% fewer errors than tensor-network decoding and 30% fewer than correlated matching on Sycamore data, but was too slow for real time. source
- In July 2026 Google Quantum AI reported that reinforcement learning kept logical error rates about 20% lower and 3.5 times more stable than periodic recalibration on its Willow chip, steering over 1,000 control parameters (secondary report). source
Prerequisites
- Information Theory1948Quantum error correction extends Shannon's classical error-correcting codes
- Error-Correcting Codes1950Quantum codes were modelled on classical error-correcting codes
- Quantum Algorithms (Shor)1994
- Cold Atoms: BEC & Laser Cooling1995Trapped-ion and neutral-atom logical qubits start from laser-cooled atoms
- First Quantum Processors2019