The answer key

A numerical bug that can make valid SP-GiST operations fail. A format-cache bug that can return wrong JSONPath results. An overflow in the money type. Those are Postgres bugs, and they were found by pgrust, Michael Malis's rewrite of Postgres. Dan Luu leads with it in We should be more ambitious with the software we build, the essay at the top of Hacker News this weekend.

Luu's argument is that programmers still carry a pre-agentic cost model. Projects that would have taken him a year now take a few hours: a regex engine, a ripgrep fork tuned to his own workloads, a Rust interpreter that he says made cold builds 50 times faster. He's right about the cost model. But his examples share something he mentions only in passing. Every one of them was built against an answer key that someone else had already written.

What AI has made cheap isn't software. It's software that has to match something that already exists.

pgrust has Postgres. Every query already has an expected answer, and when the two disagree, someone files a bug report, sometimes against Postgres. The regex engine and the ripgrep fork have ripgrep. The interpreter has rustc. Luu's zstd eval has a registered test suite, which GPT-6 Astra passes more than 99% of the time for about $1. His Azul bot had a win column. Peter Geoghegan's custom Neovim formatter, another of Luu's examples, exists because pgindent fails his CI when a line is off, and the same CI job tells him whether the formatter works.

Luu names the underlying pattern himself, describing why each new model leaves him further from "it just works":

Implementation is racing ahead of verification. Where verification already exists, as an old implementation, an RFC with tests or a scoreboard, the race doesn't matter, and the ambition is close to free.

Yossi Kreinin raises the obvious objection in Luu's appendix. Most problems don't come with "clear pass/fail criteria," he writes, so most people should be less ambitious, not more. Luu's answer is that you can often build a self-checking setup anyway, and his example is honest about the price. The rules of Guards of Atlantis 2 contradict each other so badly that the designer has repeatedly changed their mind about what they mean. Luu set up consistency checks across every conflicting explanation of the rules, spent a month or two of a $200-a-month subscription on it, and finally got a working rule set by comparing his against someone else's implementation. That works. It also shows where the cost moved. Writing the game was cheap. Writing the answer key was the project.

Most product work looks like the board game, not the database. The spec is a ticket, a Slack thread and whatever the customer meant. There is no Postgres to diff against. So the ambition that pays off first is the second version of something: a port, a rewrite, a replacement. When seven people at Anthropic replaced GitHub internally in five months, GitHub was the answer key. Whether you're scoping your own project or reading someone else's pitch, the useful question about AI-built software isn't how big it is. It's what it was checked against.

So be ambitious. Just expect that ambition to land first on software that already exists. pgrust may never replace Postgres. It has already corrected it.