Soft Publish High · Google RL-QEC on Willow

Google RL-QEC on Willow (Nature)

Development / What changed

In-situ reinforcement learning steers >1,000 control parameters from QEC detection events during compute — calibration unified with runtime (not stop-the-world retune).

Simple

The chip retunes from error beeps while still running.

Technical

Policy-gradient RL on a detector-rate surrogate; d=5/7 surface codes + d=5 colour code on Willow.

What we know (numbers)

• ~20% LER cut vs expert calibration
• Up to 3.5× LER stability = RL + decoder steering under injected drift (control-alone 2.4×)
• Record LERs 7.72e-4 (surface d=7) / 8.19e-3 (colour d=5) = multi-ingredient

What we don’t / REJECT

• “AI solves QEC”
• Injected drift ≠ all natural drift
• New code claim = false

Why it matters

FTQC calibration-downtime bottleneck — orthogonal to inventing better codes.

Sources

Nature 655, 879–884 (2026-07-08) s41586-026-10759-2 · Zenodo 10.5281/zenodo.17566521

Soft Publish High vs PDF · Quantinuum C₄-Helix

Quantinuum C₄-Helix on Helios

Development / What changed

Early-FT architecture demo (not memory-only): [[20,2,6]] memory + 2-logical Clifford under correction + heterogeneous GHZ to d=5 surface — beats physical baselines without postselection (headline comparisons).

Simple

Corrected logical operations that beat the physical qubit baselines — not a memory-only stunt.

Technical

ε≈4.6e-5/LQ/cycle; Clifford ε_C≈2.8e-4 vs physical ~1.2e-3; GHZ lower bound ~99.925% vs physical ~99.54%. Helios 98-qubit machine (Nature machine paper Jun 17, 2026 confirmed).

What we don’t / REJECT

• No T/magic demonstrated
• Company-only preprint; thin independent media
• Early-FT 10⁻⁶–10⁻⁸ targets = simulations
• Naked “breakthrough” / “universal FT computer”

Why it matters

Architecture progress toward early fault tolerance — still not a finished FT stack.

Sources

arXiv:2609.03194 · Helios Nature 10.1038/s41586-026-10676-4