We Should Be More Ambitious With the Software We Build

danluu.com, October 2026, by Dan Luu. (The post’s own title is in quotation marks — it borrows Michael Malis’s phrasing.)

One-line: Coding agents have made things that used to be too expensive cheap enough to just do — a year of work to “works for me” quality now takes a few hours — and most productive programmers haven’t updated their cost model.

Key claims

• The pgrust case. Michael Malis’s pgrust, a Postgres rewrite, is easy to dismiss as AI slop, but Luu trusts Malis’s work, and its testing has turned up real Postgres bugs: a numerical SP-GiST bug, a JSONPath format-cache bug, a pgcrypto bug, memory-safety issues in the dictionary loader, a money overflow. Even if pgrust fails, the attempt improves Postgres. Luu still says not to use it as a primary production database.
• Skepticism is rational, but beside the point. Most rewrite projects are low quality. The real question is whether you can use these tools for your own work, whatever others publish. Luu would have written “can” instead of “should.”
• His own tooling, rebuilt in hours. Looking at where his time went once models got 10–100x faster, the bottlenecks were ripgrep, Rust compilation and filesystem/syscall overhead. In a few hours each he built a regex engine tunable to his workloads, a ripgrep fork using it, a Rust toolchain change that sped up his builds — and switched from Mac to Linux for filesystem speed.
• The pre-agentic cost model. Writing your own compiler or game engine used to be what people who never ship did. Now it is reasonable, and it “doesn’t even occur to people.”
• Board-game AI as a benchmark. His Azul AI took ~5 hours to become superhuman and ~20 to be the strongest he knew of, with GPT-5.1/5.2 needing heavy steering. Gabriel Cangussu later built a stronger one in less time on newer models — evidence the models understand the domain better.
• Harder evals. Having models implement zstd, newer models passed registered tests over 99% of the time even under handicaps, cheaply — so he moved to the IMAP RFC and Pandoc. Newer models one-shot more, so he uses them on harder projects, where they still struggle; the gap between what a model can verify unaided and what it can implement is widening.
• The counterpoint (Yossi Kreinin): most people should be less ambitious — agent loops need crisp pass/fail criteria, black boxes are hard to understand, and AI speedups produce overfit, unmaintainable code. Luu agrees on the risks; the answer is self-checking systems, not timidity. Example: consistency checks for the contradictory rules of Guards of Atlantis 2, for a month or two of a $200/month subscription — cheaper than writing a clean spec by hand. Fuzzing and property-based testing work in domains people assume they don’t.
• Other examples: personal games (a submarine-combat game after Star Wolf, a Scout-like), Peter Geoghegan’s Neovim formatter combining clang-format and pgindent, and Anthropic’s internal source-control replacement built by seven people in five months. Luu sees a game renaissance driven by coding agents.
• Caveats. Don’t overbuild — Steve Yegge’s Gas Town and Ralph loops are cited as tooling that didn’t clearly help — and personal tools can be obsoleted by what the labs ship. The post was written quickly, with little benchmarking.

Notable quotes

it doesn’t even occur to people to do these things that are now easy

things that were previously too expensive to do

I would not recommend using it as your primary production database today.

there’s a game renaissance going on right now due to coding agents.

How it connects

• Ambition and How to Be More Ambitious — A Tutorial for Humans With Models (Felix Rieseberg) — the same month’s other argument that models should raise the ceiling on what you attempt.
• Developer Productivity and Agentic AI.
• We’re not a PE fund… (tweet) — the enterprise version: “So many people are still writing code by hand!”, with the caveat that legacy codebases need unsexy work before agents can run.
• Dan Luu — author; his earlier “Willingness to look stupid” is already in the wiki via Willingness to Look Stupid… (tweet).