Vibe coding breaks in quantum computing: 59.5% one-shot on Qiskit, failures are semantic
AryHHAry · x · 2026-09-19
A detailed analysis of why vibe coding fails in quantum computing. Web dev offers cheap, honest feedback (visible errors, tracebacks), so "runs" ≈ "correct"; in Qiskit and PennyLane circuits compile, histograms look tidy and VQE loss decreases while the unitary is still wrong — semantics break silently. Key data: QuanBench+ (2026) shows best one-shot scores of just 59.5% on Qiskit and 42.9% on PennyLane, rising to 83.3%/66.7% after 5 repair rounds, with remaining failures being logic errors; on QuantumKatas, Opus-class and Codex-class models reach 75–83%, dominated by Assertion Errors — valid code operating on the wrong state. The author's thesis: vibe coding assumes cheap and honest feedback, but quantum computing provides expensive and silent feedback — simulators forgive, hardware rejects outdated APIs, gradients can "learn" on ansatzs that observe nothing. Upgrading models only shifts the error distribution, not the second layer of fidelity, provenance, and reproducibility. If circuit proofs are only trustworthy while the chat is open, that's not rising intelligence — it's authority shifting from quantum mechanics to autocomplete.
More from coding & agent
- Clinical AI prototype architecture: local de-ID, typed model decisions, code as final gate — MaziyarPanahi · 2026-09-19
- Running Jev inside a clinical workflow: 6 typed decisions, 1 blocked, 1 to human review — MaziyarPanahi · 2026-09-19
- Grok Build tests Remote Control: link your PC via CLI and drive it from your phone — XFreeze · 2026-09-19
- Developer Says AI Models Now Handle Nearly All His Work — He Just Approves — tushaarmehtaa · 2026-09-19
- 8 components of harness engineering: why the system around the model, not the model, makes agents reliable — blaizedsouza · 2026-09-19
- Google's 12-page Agentic Engineering guide lays out a 5-stage pipeline for building agent teams — blaizedsouza · 2026-09-19