Case Study: Forecasting Honestly — What Would Change Your Mind?
Executive Summary
Quantum computing forecasts range from "five years to commercial advantage" to "never useful," and both extremes are usually asserted rather than argued. A forecast that cannot state what evidence would falsify it is not a forecast; it is a position.
This case study builds a disciplined forecast: decompose the question into measurable components, assign explicit uncertainties, state falsification conditions, and identify which indicators actually carry information. The method matters more than the numbers — the numbers will be wrong, and a well-structured forecast tells you how it was wrong.
Skills applied
- Decomposing a compound forecast into independent components.
- Assigning and combining uncertainty.
- Specifying falsification conditions.
- Distinguishing high-information from low-information indicators.
Phase 1: Specify the question
"When will quantum computers be useful?" is unanswerable — it bundles several distinct questions. Pick one and make it checkable:
Q: By what year will a quantum computer solve a commercially valuable problem faster or cheaper than the best classical alternative, verified by an independent third party?
Every clause does work: commercially valuable excludes sampling milestones; best classical alternative excludes strawman baselines; independent third party excludes vendor self-reports.
Phase 2: Decompose
The event requires several things jointly:
| Component | Requirement | Status |
|---|---|---|
| A. Logical qubits | ~100+ at algorithmic error rates | ~50–100 at low distance (QuEra 2026) |
| B. Logical gate fidelity | $10^{-8}$ or better per logical operation | Not demonstrated |
| C. Real-time decoding | At scale, microsecond latency | Small-scale only |
| D. An algorithm with advantage | Proven or strongly evidenced on a valuable problem | Simulation is the candidate |
| E. Classical baseline holds | Classical methods do not close the gap | Uncertain — they keep improving |
All five must hold. This is why the forecast is not simply an extrapolation of qubit counts: four of the five components are not qubit counts.
Phase 3: Estimate each
Being explicit about reasoning, and about it being judgement rather than measurement:
A. Logical qubits (100+ at useful distance). Current trajectory doubles verified logical qubits roughly annually, but at low distance. Reaching 100 at distance ~15 needs both count and quality. Estimate: 2030–2036, median ~2033.
B. Logical gate fidelity. Requires below-threshold operation with margin plus demonstrated logical two-qubit gates. $\Lambda \approx 2.14$ today needs to reach ~10. Estimate: 2031–2038, median ~2034.
C. Real-time decoding at scale. Currently the least-discussed component. FPGA decoders work at small scale; terabit-rate hierarchical decoding is unbuilt. Estimate: 2030–2037, median ~2033. Low confidence — this is the component I would most expect to be surprised by, in either direction.
D. An algorithm with advantage on a valuable problem. Quantum simulation of strongly correlated materials is the leading candidate. The obstacle is state preparation (Chapter 16) as much as circuits. Estimate: 2032–2042, median ~2036.
E. Classical baseline holds. DMRG, tensor networks, and neural-network wavefunctions keep improving. Probability the target problem remains classically hard when quantum hardware arrives: ~70%.
Combining (not independent — B and C correlate with A, D is largely separate):
$$\text{Median estimate: } \mathbf{2036\text{–}2038}, \text{ with } \sim70\% \text{ probability by } 2045$$
Phase 4: State the falsification conditions
This is the part that distinguishes a forecast from an opinion.
I would move the estimate substantially earlier if: - A logical two-qubit gate is demonstrated at error below $10^{-6}$. - $\Lambda$ exceeds 10 on a device with 100+ physical qubits per patch. - A qLDPC code operates below threshold on hardware, cutting overhead ~20×. - A peer-reviewed result shows advantage on a problem with commercial value, independently reproduced. - Real-time decoding is demonstrated at 100+ logical qubits.
I would move it substantially later if: - Below-threshold operation fails to improve beyond $\Lambda \approx 3$ for five years. - A classical algorithm dequantizes the leading simulation advantage. - Decoder latency proves an unexpected hard barrier. - Funding contracts sharply, slowing all components at once.
I would abandon the forecast entirely if: a fundamental obstacle to scalable fault tolerance were identified. None is known, and the threshold theorem argues against one existing — but this is the condition that would matter most.
Phase 5: Which indicators carry information
Not all news is evidence.
| Indicator | Information value | Why |
|---|---|---|
| $\Lambda$ (error suppression factor) | Very high | Directly sets overhead |
| Verified logical qubit count | Very high | The actual currency |
| Logical gate error rate | Very high | Component B directly |
| Published resource estimates | High | Moved the RSA number 50× (Ch. 15) |
| Code-theory advances (qLDPC) | High | 24× overhead reductions |
| Decoder latency demonstrations | High, under-watched | Component C |
| Physical qubit count | Low | Says nothing about quality |
| Funding rounds | Very low | Tracks sentiment |
| Partnership announcements | Very low | Usually cloud credits |
| Stock price | None | Uncorrelated with capability |
The three numbers to track: $\Lambda$, verified logical qubit count, and logical gate error. Everything else is commentary, and the bottom four rows are noise that dominates coverage volume.
Phase 6: The discipline
Three habits that separate a useful forecast from an assertion:
- Decompose before estimating. "When will quantum be useful" invites intuition. "When will logical gate error reach $10^{-8}$" invites evidence.
- Write the falsification conditions first. If you cannot say what would change your mind, you are not forecasting.
- Track the high-information indicators only. Following press coverage produces a forecast that tracks press coverage.
Applied honestly, this method produces a wide interval — 2036 to 2045, with real probability outside it. That width is not a failure of the method; it is the correct representation of a state of knowledge in which four of five components remain undemonstrated. A narrow forecast in this domain is a red flag, whichever direction it points.
Discussion Questions
- Four of five components are not qubit counts. Why does public discussion focus on the one that is?
- Component E — the classical baseline holding — was given 70%. Argue for a substantially different number.
- Real-time decoding was flagged as the most likely surprise. Is that right, or is another component more uncertain?
- Build your own forecast for the same question, stating your falsification conditions.
Your Turn: Extensions
- Forecast a narrower question — "when will a logical two-qubit gate reach $10^{-6}$ error" — and specify falsifiers.
- Track the three high-information indicators for six months and note whether your estimate moves.
- Find a published quantum forecast and identify whether it states falsification conditions.
- Construct the strongest case for "never commercially useful" and identify its weakest premise.
Key Takeaways
- Decompose compound questions into measurable components; four of the five here are not qubit counts.
- State falsification conditions explicitly — a forecast that cannot be wrong is a position, not a prediction.
- Track $\Lambda$, verified logical qubit count, and logical gate error; funding, partnerships, and physical qubit counts carry little or no information.
- Honest decomposition yields a wide interval (roughly 2036–2045 here), and that width is the correct representation of the uncertainty.
- Narrow confident forecasts in either direction are a warning sign, since no forecaster has access to evidence that would justify one.