antirez vs mitsuhiko: Should You Still Read AI-Generated Code Line by Line?
Redis creator antirez and Flask creator mitsuhiko held a public debate on X about GPT 6 Astra's coding performance. antirez argues that skilled programmers getting poor results reflects a changed skill set rather than a model problem; mitsuhiko counters with hands-on complaints from his C projects. The exchange ended with viewpoints intact and a suggested rewrite experiment that has not been run, though antirez separately noted AI lets him attempt challenges he'd never have had time for.
Confirmed
- antirez observed many strong programmers reporting poor results with GPT 6 Astra, feeling like a "parallel universe"; his explanation is that the skill set for good programming has changed, overlapping only partially with the old one.
- mitsuhiko says he is in an adaptation period, currently using Astra to discuss problems rather than write large amounts of code.
- In a game side project, mitsuhiko found Astra's C code often hard to read, frequently falling into a token-efficiency-driven mode.
- mitsuhiko also complained that Astra (a Gemini-based agent) becomes expensive in implementation and makes only very narrow changes, with poor experience when trying to understand why and how it made them.
- Comparing models, mitsuhiko finds Opus 5.5 produces noticeably fewer "weird results" with almost no heavily compressed code.
- antirez's core claim: once high-level design is expressed via prompt and validated, reading local code line by line is unnecessary; if local choices can't be verified by design, tests, or program goals, verification belongs at the design level.
- antirez's practical rationale: AI code output reaches ten to twenty thousand lines a day, making confidence via human reading unrealistic; he relies on proxy metrics—actual runtime behavior and feel, extensive QA.
- antirez predicts RL will push AI code style further away from what humans consider "good and comfortable"; he already judges only by outcomes.
- mitsuhiko's reply on why he still reads code: it's about comfort—he'll stop reading only when trust in the model is high enough, which it isn't yet.
- antirez suggested a "crazy experiment": spend two or three hours reimplementing the same project in C using only Astra.
- Responding to skeptics, antirez said AI lets him attempt challenges he'd never have had time for (like his AW port), and challenged unskilled users to try the same port with AI—implying his deep fundamentals still matter.
- His only regret: not having carved out time to hand-write small, poetic programs.
Unconfirmed
- Whether the experiment was actually run, and how Astra performs on a pure C rewrite, remain unknown.
- The "changed skill set" explanation is antirez's personal view; the disagreement over code readability was never resolved.
Why it matters
The exchange crystallizes the core divide in AI-assisted programming: one side argues that with exploding code output, work should shift up to design-level validation with confidence built on QA and runtime proxies; the other insists readability and reviewability of generated code directly affect engineering quality and maintenance costs. The firsthand comparisons from two veteran developers also offer concrete reference points for style differences between models like Astra and Opus 5.5, while antirez's predictions about RL-driven style divergence and his own experience of AI unlocking previously impossible projects suggest this divide may only deepen.
2026-09-27 ~ 2026-09-28 · 13 related posts
Primary sources
- [source] antirez: Good programmers failing with GPT 6 Astra points to a different skill set — antirez · 2026-09-27
- antirez suggests a crazy experiment: build your whole project in C using only GPT 6 Astra — antirez · 2026-09-27
- mitsuhiko: Opus 5.5 rarely produces heavily compressed code unlike Astra — mitsuhiko · 2026-09-27
- [source] mitsuhiko dislikes GPT 6 Astra's C code: hard to read and weirdly token-efficient — mitsuhiko · 2026-09-27
- antirez challenges mitsuhiko: why read AI-generated code if the design is verified? — antirez · 2026-09-27
- antirez on AI code: if local choices can't be verified by design or tests, why check? — antirez · 2026-09-27
- [source] The thread starter: why read AI-generated code if the design is prompt-verified? — mitsuhiko · 2026-09-27
- antirez: Reading AI code is no longer feasible — he now trusts tests and QA instead — antirez · 2026-09-27
- mitsuhiko on Astra: expensive, narrow edits, and a frustrating experience to review — mitsuhiko · 2026-09-27
- antirez: RL will push AI-generated code further from what humans consider good — antirez · 2026-09-27
- antirez regrets not finding time to hand-write small poetry-style programs — antirez · 2026-09-27
- antirez: AI lets me finish challenges I could never have completed before — antirez · 2026-09-28
1 near-duplicate retellings: mitsuhiko