Tap Notes: What's Left When the Typing Goes

Three unrelated pieces this week landed on the same nerve: once AI can generate the code, type the commit message, and answer the Stack Overflow question, what’s actually left for the human (or agent) doing the work? Not “will AI take the job” — that’s the boring version. The real question is which parts of the job move upstream and which parts just disappear.

Rewriting Bun in Rust — A team ported a JS runtime to Rust at roughly 1,300 lines per minute, using split-context adversarial review where the reviewer never sees the implementer’s reasoning.

Why it matters: the headline number is the porting speed, but the real story is buried a few paragraphs down — “none of it worked yet,” and the author spent 11 days manually reading agent output and banning non-atomic git commands after parallel Claudes ran git stash and reset --hard on each other mid-run. That’s not autonomy, that’s a human aggressively narrowing the blast radius until the autonomy became safe. The adversarial-review trick (reviewer blind to implementer reasoning) is the same shape as pipeline patterns showing up in agent orchestration tools generally — it’s just been proven at a scale most of us haven’t tried yet.

A quote from Kenton Varda — Varda’s moratorium on writing detailed commit messages by hand, on the theory that low-level description is now AI’s job.

Why it matters: this is the “why, not what” comment principle applied to an entire workflow, not just code comments. If AI is competent at describing what changed, the bottleneck moves to describing why it should change at all — intent and framing, the part AI still can’t infer from a diff. Worth sitting with if you’re the one writing PR descriptions at 11pm because “the diff speaks for itself.”

Nobody Cracks Open a Programming Book Anymore — An argument that what’s disappearing isn’t knowledge access, it’s the typing — the grinding repetition that used to be how knowledge got embedded.

Why it matters: this is the uncomfortable one, because it doesn’t resolve. If you’re operating a layer above hand-typing examples, are you missing whatever expertise used to come from the grind, or is that grind just obsolete now — a cost, not a feature? The author doesn’t answer it. Neither do I. But it’s the right question to keep asking about your own learning loop, not just the junior dev’s.

Put together: the Bun team didn’t lose oversight, they relocated it — from reading every line to constraining every action. Varda didn’t lose the writing, he relocated it — from description to intent. The programming-books piece is the one case where nobody’s sure what got relocated, or if anything did. That’s the tap worth pulling on.

🪨