Key Takeaways: The Quantum Programming Career
The book in two sentences
No quantum method in this book beat a tuned classical baseline on any problem we could check. And the techniques work, the physics is real, the open problems are genuine, and the field is hiring — which is a stranger situation than either the hype or the backlash describes.
The scorecard, harvested
Every figure below was extracted from the chapter files by scripts/harvest_scorecard.py and is
re-verified by project-checkpoint.py on every run. A capstone written from memory produces numbers
that are plausible and wrong.
Head to head:
| Quantum | Classical | Result | |
|---|---|---|---|
| Ch.24 QAOA p=1, MaxCut | 0.8086 | Goemans–Williamson 0.8785 | classical |
| Ch.32 VQC, iris-binary | 1.0000 | LogisticRegression 1.0000 | tie |
| Ch.33 1-qubit re-uploading | 0.8343 | kNN 0.8970 | classical (9 SE) |
| Ch.34 quantum kernel | 0.8313 | SVC(rbf) 0.8889 | classical (7 SE) |
| Ch.35 quantum-data model | 0.6429 | SVC(rbf) 0.7857 | classical |
| Ch.37 QAOA p=3, 10 graphs | 0 wins | Goemans–Williamson 6 wins | classical, 4 ties |
| Ch.37 GW per-rounding rate | — | 0.680 (8/10 was a lucky draw) | 10/10 by distribution |
Quantum wins: 0.
Where the cost killed it: one T gate takes a circuit from 450 to 2,882 physical qubits (Ch.15); a 20-bit Grover search needs 229,944 T gates (Ch.21); VQE at the 50-orbital crossover needs 6.06 × 10⁸ QPU-years (Ch.36); inference costs 27.8 QPU hours per million predictions (Ch.33).
Where the measurement was the finding: approximation_degree=0.9 produced zero two-qubit gates
— the optimization deleted the circuit (Ch.28); one chip supported quoted two-qubit errors from 0.00750
to 0.07205 (Ch.30); dynamical decoupling measured significantly worse (Ch.31); the transpiler
seed alone changed fidelity by 2.03× (Ch.39).
Where it worked: ancillas cut an 8-bit oracle from 26,978 T gates to 55 (Ch.19); phase estimation is exact for dyadic phases (Ch.22); Shor factored 15 as predicted (Ch.23); VQE reached chemical accuracy on H₂ from four parameters (Ch.24); hardware-aware level 1 beat naive level 3 by +0.1397 (Ch.29); classical shadows win at equal budget (Ch.35); BB84 works exactly as proven (Ch.38).
The five things to carry forward
1. The jobs are mostly classical. Control software, compilers, decoders, cryogenics, platform engineering — and quantum literacy is the differentiator rather than the content. "Quantum algorithm designer" is the smallest category. Chapter 39 found the dominant cost of a quantum computation is the queue, and improving that is a scheduling problem.
2. Eight questions to ask of any quantum claim. Compared to what? How many samples? What denominator? Which statistic? Total error or method error? What does it produce? Can the measurement detect the thing? What did it cost? Every one was learned by this book getting it wrong first.
3. This book published the too-small-measurement error seven times — Chapters 27, 28, 33, 34, 37, 38, 39 — and corrected each in print. The seventh was different in kind: 24 samples, wrong system. A measurement can be too small in more than one dimension.
4. The real problems are error correction, simulation of quantum systems, and learning from quantum data. All three survive this book's own scrutiny. The applications with the loudest commercial claims — optimization, finance, near-term ML — are the ones measured most carefully and found furthest behind.
5. The API you learn will be deprecated; the reasoning will not. Chapter 31 opened with
qiskit.pulse removed in Qiskit 2.0.
The self-audit
vqelab.claims applies the eight questions, and Chapter 36's own central result fails one of them.
The 9,870,104× ratio between active-space error and VQE error was measured on one molecule at one geometry. The finding is almost certainly robust — truncation error exceeding optimizer error is not a fragile effect — but almost certainly robust is a different claim from measured, and the checklist is right to say so. Chapter 36 Exercise 36.17 asks the reader to vary the bond length precisely because the chapter did not.
A checklist that never fails your own work is a checklist you are not applying.
What went into vqelab
vqelab.claims, the last module, 19 tests:
audit_claim()returns failures, not a score. A claim that cannot name its baseline is not 1/8 correct — it is unevaluable, and averaging that against a pass would be exactly the kind of number this book spent forty chapters refusing.is_tuned_baseline()rejects"random","greedy","brute force"— and"full CI", because Chapter 36 §36.6 found nobody runs it past ~20 orbitals.significant()refuses a standard error of exactly zero. Zero uncertainty means it was never estimated, not that it is absent.project-checkpoint.pyverifies §40.3's scorecard against the chapters that measured it, every run. If a capstone number is not in its source chapter, the checkpoint fails.
What the book was actually teaching
The quantum content is real and it is not the point.
Every chapter ran the same loop: make a claim, measure it, and let the measurement win. That produced the T-gate cliff, the decoupling that made things worse, the optimization that deleted the circuit, the approximation that dominated the result, the certificate QAOA does not produce, and seven occasions when this book read a number off too small a sample and corrected it in print.
None of those are quantum lessons.
The number that is easy to get is not the number that answers the question, and the easy number is almost always the flattering one. That is not a coincidence — the flattering number is the one that stops the search.
This is the end of the book. The appendices follow: the glossary, the framework translation tables, the resource-estimation reference, and answers to selected exercises.