Quantum: 39x on the Headline, the Denominator in the Footnotes
This zhichai.net post is a meticulous fact-check of the 39x improvement headline from QuEra's press release (Boston, 2026-11-10) about LLM-designed quantum circuits, verified against the paper arXiv:2610.22367, *Discovering state-of-the-art quantum circuits with large language models* by Andi Gu, Alek H. Kemeny, Susanne F. Yelin, and Mikhail D. Lukin (v1: 2026-10-15 17:59 UTC, 46 pages, 10 figures, 6 tables, quant-ph + cs.AI). Affiliations span Harvard, Anthropic (STEM Fellows), and QuEra.
Key points
- 39x is an area, not a speedup. The paper measures TQ-cost = Toffoli gate count × peak logical qubit count. The paper's own sentence: "Across the three tasks, these techniques reduce TQ-cost by 9.9×–39× relative to the tuned HRS baselines." How much faster? This material does not say. QuEra's press release itself concedes the composite metric means 39x is not a 39x speedup.
- Method: Claude proposes circuits; an independent automated evaluator verifies correctness and scores gate count and qubit count as feedback. Across eight computational building blocks, seven reduced both metrics; the eighth (modular inversion) yielded a new reversible architecture reviving divsteps.
- Verification arithmetic (self-checkable): Fixed-point Gaussian: 24,000/1,583 ≈ 15.2 gate reduction × 362/139 ≈ 2.60 qubit reduction ≈ 39.5. Coulomb potential: 6,355/1,140 ≈ 5.57 × 252/142 ≈ 1.77 ≈ 9.9.
- Full Table 1 column: 12 / — / 32 / 9.9 / 39 / 4.1 / 7.7 / —. The common baseline for numeric tasks is Häner, Roetteler, Svore 2018 (arXiv:1805.12445).
Two public corrections (original text preserved)
1. The 12x Reed-Solomon figure. Initially called a "gate-count-only" metric, then recorrected, then corrected again after literal checking of Table 1: the column G_TQ is defined as (TQ)_baseline/(TQ)_proposed, so 12x is the same area ruler as 39x and 9.9x — Reed-Solomon goes (5.1×10⁶, 1,938) → (4.4×10⁵, 1,837): gate ratio ≈ 11.6, qubits −5.2%, product ≈ 12.2. The real reason not to line them up in one ranking is the different denominators: 39×/9.9×/32× are vs. HRS 2018; 12× is vs. Khattar et al.'s Reed–Solomon DQI baseline; modular inversion has "—" ("we report no single baseline or gain"). The 12x row also carries a failure budget (≤10⁻³, 10⁵ tests); the paper warns "Turning this failure budget into resource savings, however, is not automatic."
2. Swapped row coordinates. The (5.1×10⁶, 1,938) → (4.4×10⁵, 1,837) coordinates are the Reed–Solomon row; fixed-point Gaussian is (2.4×10⁴, 362) → (1,583, 139) and Coulomb is (6,355, 252) → (1,140, 142).
Also downgraded: the quoted sentence about seven wrong signs in 100,000 tests could not be verified in the visible main text; and 17×/12% for modular inversion comes verbatim from the QuEra press release, not the paper — cite as "press-release figures."
Two rulers, two graders
A same-week comparison with DeepMind × Google Quantum AI's arXiv 2611.09822 (25 authors, 290x on spacetime volume, verified by Lean 4):
| Dimension | arXiv 2611.09822 | arXiv 2610.22367 | |---|---|---| | Correctness judged by | Lean 4 kernel | Independent automated evaluator | | Metric | spacetime volume (logical qubits × T count) | TQ-cost (Toffoli × peak logical qubits) | | Best reported | 290x (arcsin subroutine) | 39x (fixed-point Gaussian) | | Denominator | Häner et al. 2018 piecewise polynomials | Prior state-of-the-art per component |
Different dimensions, different denominators, different grading mechanisms — 290 and 39 do not belong on one leaderboard.
Press-release-only claims
53-hour search time, FeMoco ~150,000x over eight years, 695/700 laser-lock recoveries, 5–10 minutes, Libra (target 2028 on Amazon Braket) — these appear only in the press release, not in the paper. Cite as "press-release figures." The press release is from a collaborating party and includes roadmap statements.
Closing
Whether multiples can be compared side by side depends first on whether the denominators are the same baseline. 39x means the area shrank 39 times. How much faster? This material does not say.
Sources & caveats: QuEra Computing press release 2026-11-10; arXiv:2610.22367 v1; arXiv:1805.12445 (HRS baseline); arXiv 2611.09822 (comparison, see zhichai.net post 178636997).