What you're allowed to stop watching

Monday's Hacker News front page carried three arguments about AI coding, a few slots apart. Zig creator Andrew Kelley responded bluntly to Anthropic's two-months-late explanation of porting Bun from Zig to Rust, and Ray Myers's read of the exchange sat at 827 points. antirez told programmers they now have less impact because they look at the code. And a user reported that xAI's Grok CLI had uploaded his entire home directory to the company's cloud storage. Not one of the three is about whether models can code. Every party to every one of these fights stipulates that they can.

The argument has moved. The question of the last three years — can the model do the work? — got answered well enough that nobody on that front page bothered to litigate it. The live question is what you're allowed to stop watching once you believe. And read together, the three stories give the same answer: you can stop watching exactly where something else stands watch — a test suite, a sandbox, an honest writeup — and not one inch further.

antirez is the strongest case for stopping. He still reviews every AI-generated diff that goes into Redis and now calls the habit mostly pointless: Fable and GPT 5.6 catch race conditions his review won't.

But look at what he spends the recovered hours on: QA, testing, and a DESIGN.md that states in plain language what each data structure must do. That isn't withdrawn attention. It's attention moved up a level — from checking the code to specifying what would count as the code being wrong. The verification didn't disappear. It changed altitude.

The Bun writeup fails at exactly that altitude. Myers's complaint isn't that the port was a mistake — he grants the migration mechanics are genuinely useful. It's that the explanation arrived two months after the merge, listed pros with no cons, and never disclosed the tradeoffs a reader would need to check the judgment: what did the Rust port cost in build time, and what would a TigerBeetle-style discipline in Zig have bought instead? He pairs that with the balance sheet — $132 billion raised, an IPO approaching a trillion-dollar valuation, no profitability to show — and calls Anthropic an unreliable narrator. The label earns its keep. When the narrator's valuation rides on the story that coding is going away, a writeup with no cons isn't documentation. It's marketing wearing documentation's clothes, and marketing can't be audited.

The Grok CLI report is the bill for watching nothing at all: an agent with filesystem access, a user who trusted the defaults, a home directory in a vendor's bucket. And one slot over, an 824-point Ask HN asked the site to flag AI-generated articles — readers requesting, in effect, the provenance metadata they can no longer reconstruct by reading. Same demand, four forms: show me the thing my trust is supposed to rest on.

You can stop watching exactly where something else stands watch — and not one inch further.

The obvious rejoinder is that verification is going to the models too — antirez says as much, and he's right that a frontier model now out-reviews him line by line. Concede all of it. But delegated verification still hangs from an anchor a human set: the model checks the code against the invariants; someone decided the invariants. The DESIGN.md, the test that must not break, the boundary the agent may not cross, the tradeoff a writeup owes its readers — those are declarations of what must be true, and generating them is precisely what can't be delegated, because they are the standard the generation is judged against.

Which is why the sharpest detail of the day is the quiet one: the day's most forceful argument for not reading code ends with its author still reading the code, out of respect for users. That isn't an inconsistency. The watching was never for him.