Quantum

Plain English first. Technical detail sits below the fold. IBM stays off this page (Monitor only).

Soft Publish Google RL-QEC on Willow

Google: the chip retunes itself while it still runs

In plain English

Google showed a way for its quantum chip (Willow) to retune itself from its own error signals while still running — less “stop the computer to fix it.” After expert tuning, they saw about 20% fewer logical errors. In a lab test with fake drift, stability improved up to about 3.5× when the software that watches errors adapted too.

What this is not: “AI solved quantum error correction.” Fake lab drift is not the same as every real-world problem.

Why it matters

Big quantum machines lose time when they have to stop and retune. Cutting that downtime is useful even if the underlying codes stay the same.

What we don’t know yet

For the technical reader

In-situ reinforcement learning steers >1,000 control parameters from QEC detection events during compute — calibration unified with runtime. Policy-gradient RL on a detector-rate surrogate; d=5/7 surface + d=5 colour on Willow. Record LERs 7.72e-4 (surface d=7) / 8.19e-3 (colour d=5) = multi-ingredient. Control-alone stability under injected drift: 2.4×.

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

Editor’s take: Soft Publish with mandatory caveats. Brother test first; jargon below the fold.

Soft Publish Quantinuum C₄-Helix

Quantinuum: protected operations that beat raw qubits

In plain English

Quantinuum showed protected two-qubit operations that beat ordinary (“raw”) physical qubits — not just storing information longer. Think: doing work under a shield, not only holding still under a shield.

What this is not: a finished universal quantum computer, and not a naked “breakthrough.” It’s a company preprint (not fully peer-reviewed yet). They still haven’t shown the full toolkit (no T/magic gates yet).

Why it matters

Error correction has to do more than remember — it has to compute. This is an early architecture demo on that path, not the end of the path.

What we don’t know yet

For the technical reader

[[20,2,6]] memory + 2-logical Clifford under correction + heterogeneous GHZ to d=5 surface — beats physical baselines without postselection (headline comparisons). ε≈4.6e-5/LQ/cycle; Clifford ε_C≈2.8e-4 vs physical ~1.2e-3; GHZ LB ~99.925% vs physical ~99.54%. Helios 98-qubit (Nature machine paper Jun 17, 2026).

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

Editor’s take: Soft Publish as primary-sourced technical. No naked breakthrough. Don’t overclaim early fault tolerance.

IBM — Monitor only, off this page. Media “70 logical qubits” claims were misleading vs the paper (error detection + throwing away bad runs). Not printed here.