We should be more ambitious with the software we build
B · Useful, limited A widely read engineering essayist, Dan Luu, makes a clear argument that builders still use a pre-agentic cost model, but by his own note the post was written fast and without rigorous benchmarks, so the evidence is personal anecdote.
Summary based on the full text·Original: about 25 minutes to read
In short
An undated essay by Dan Luu (discussed on Hacker News on 24 October 2026), built around a remark by Michael Malis of the pgrust project: "we should be more ambitious with the software we build." Luu argues that many programmers still carry a "pre-agentic cost model" and do not attempt projects, such as a faster regex engine, a Rust interpreter or a database rewrite, that coding agents now make feasible. He pairs this with a warning: such work is only worth anything with serious testing and benchmarking, and each more capable model widens the gap between what it can build and what it can verify.
Key numbers
- about 5 hourshis time to build a superhuman Azul board-game AI with GPT-5.1/5.2 and no prior AI knowledge
- more than 99%share of runs in which GPT-6 Astra passed all registered zstd tests, across all conditions
- about $1cost for GPT-6 Astra to implement zstd
- 50xcold-compile speed-up from an agent-built Rust interpreter (2x on warm builds), unbenchmarked
As reported by the source.
English·Web page·opens danluu.com
What it reports
- pgrust, a Rust rewrite of Postgres, is tested thoroughly enough to have uncovered real Postgres bugs (SP-GiST, JSONPath format cache, pgcrypto, money overflow, dictionary-loader memory safety).
- Luu built a tunable regex engine, a ripgrep fork and a Rust interpreter in "a low single-digit number of hours", work he estimates would have taken him a year or more to reach "works for me" quality.
- Azul AI: about 5 hours to superhuman, about 20 hours to what appeared to be the strongest AI at the time, with GPT-5.1/5.2; a later builder beat it in less time using Fable 5.1 and Opus 5.5, which could implement techniques from papers.
- Toy eval: GPT-6 Astra passed all registered zstd tests more than 99% of the time across all conditions, for about $1, so he moved to harder targets (IMAP RFC, Pandoc).
- Newer models do not make things "just work"; they move him to harder problems where verification is the constraint.
- Counterpoint from Yossi Kreinin, quoted: most people should be less ambitious, because many problems lack clear pass/fail criteria and AI-sped code can be overfit and unmaintainable.
- Other examples: an Anthropic team of seven built an internal GitHub replacement and code review tool in five months; 3 of the top 10 Azul players on Board Game Arena last season were reported to be cheaters.
How they got it
Personal projects and conversations, written up quickly. Luu states he did "a quick comparison that satisfied my curiosity" rather than his usual comprehensive benchmarks.
Limits
- No systematic measurement; speed-ups and hour counts are the author's own estimates.
- Examples skew toward expert programmers working on tools for themselves; the essay itself concedes production quality takes much more work.
- Page is undated; model names and costs reflect a fast-moving period.
Why it matters for companies
For engineering leaders, the useful question is which internal tools, migrations or rewrites were ruled out as too expensive and should be re-costed now. The essay's own guardrail is the part to adopt: fund the testing, property-based checks and benchmarks before the build, since the cost of producing code has dropped faster than the cost of knowing it is right. Kreinin's counterpoint is a fair warning against managers ordering teams to "make it fast with AI" without that discipline.
Summary based on the full text. Read the original for the full argument and data.Read the original on danluu.com
- Tags
- Coding agents