We should be more ambitious with the software we build

gist

Dan Luu argues coding agents now make once-absurd projects cheap, so builders should aim higher and ignore the surrounding flood of vibeslop. He cites Michael Malis's pgrust, a Postgres rewrite whose testing surfaced real Postgres bugs, and his own hours-long builds of a regex engine, a ripgrep fork and a Rust interpreter. People still carry a pre-agentic cost model, he says, while each stronger model moves the bottleneck from implementation to verification. Appendices weigh Yossi Kreinin's case for less ambition, self-checking setups and personal one-off tools.

ideas

• Slop elsewhere doesn't cap what you can build. Most rewrite projects are vibeslop and get dismissed on sight, but pgrust's testing found Postgres bugs (SP-GiST, JSONPath, pgcrypto, money overflow), so even failure would leave Postgres better.
• The pre-agentic cost model is the real limit. Faced with a 10x–100x faster model, Luu found ripgrep, Rust compiles and filesystem syscalls on the critical path; a regex engine, a ripgrep fork and a Rust interpreter (about 50x on cold builds, 2x warm) took a low single-digit number of hours, so writing your own compiler is no longer a never-ships move.
• Stronger models widen the verification gap. GPT-5.1/5.2 faked techniques for his Azul AI, while Fable 5.1 and Opus 5.5 apply papers and GPT-6 Astra passes zstd tests over 99% of the time for about $1; more one-shots just move him to harder work where models need guidance, which public evals miss.
• Counterpoint: most people should be less ambitious. Yossi Kreinin warns that without pass/fail criteria, black-box agent output fails like misused microservices; Luu answers that partially self-checking setups, like consistency checks across contradictory Guards of Atlantis 2 rules, work in more domains than people expect.
• One-off tools are the everyday version, and lab projects are the preview. A Star Wolf submarine game, a codex memory-leak restarter and Peter Geoghegan's clang-format, pgindent and tree-sitter formatter show cheap personal software; Anthropic's seven-person GitHub replacement shows unlimited-token futures, not what is feasible today.

quotes