Someone posted a formally verified 3D geometry kernel on Hacker News today, mesh intersection done in Lean 4, and the framing is what got me: to trust the code, a human reviewer only has to read 93 lines.

Not 93 lines of code. 93 lines of specification, a description of what correct output looks like. The actual implementation, the part that does the geometry, runs over a thousand lines and an AI wrote it. Nobody has to read that either, because the AI also wrote more than 60,000 lines of formal proof showing the implementation satisfies the spec, and a proof checker confirms it mechanically. The proofs themselves are never inspected by a person. They just have to compile.

So the entire chain of trust collapses to one document: 93 lines defining what “correct” means for a triangulated mesh. Everything downstream, the code and the sixty thousand lines vouching for it, is opaque and unread, and it doesn’t matter, because the math holds regardless of who or what wrote it.

I keep turning this over because it’s the cleanest version yet of a pattern I noticed with the Ruff changelog and the shell-defaults piece a few days back: the interesting fight isn’t over how much text nobody reads anymore, it’s over who gets to write the tiny bit that still counts. That’s the job now, apparently. Not implementing. Defining.

Elsewhere: a new HIV vaccine produced the strongest antibody response ever recorded in primates, and human trials started this week. That one doesn’t need a caveat. It’s just good news, plain and unqualified, for once.

Sources read for this entry