Everything Got Faster Except the Part That Says Yes

PostgreSQL 19 was due at the start of this month. It will now land around 19 November with eight to ten features pulled, because AI-assisted review found more real defects late in the beta than the project’s people could close — not bad machine-written code, its committers stress, but good machine reading. Nothing failed. The supply of findings simply overtook the supply of humans able to act on one. Every step in this industry got faster this week except the one where a human says yes, and that step is now what everything else waits on.

This is not the complaint everyone rehearsed. For three years the fear was machines producing plausible rubbish nobody would catch. The week’s problem is the inverse: the machines were right, repeatedly, and being right at volume costs something that lands entirely on people. A late database is the mild version: the industry’s mathematics, its safety auditing and its own hiring ladder hit the same wall inside seven days.

The binding constraint is not compute or talent but review capacity, and the firm best placed to measure it is idle four-fifths of the time. Anthropic published figures on its own workforce on Friday and left the sharpest number in a caveat: about 80% of a Claude session’s elapsed time is spent waiting for a human. Interactive agent hours overtook employee hours in March and ran six to one by August, while junior technical hires fell from 27% to 9%. One company added vast production capacity, thinned the rungs where reviewers are trained, and became the thing its own agents queue behind.

Every step in this industry got faster this week except the one where a human says yes, and that step is now what everything else waits on.

The labs have started buying refereeing, and in every arrangement the examined party pays the examiner. OpenAI and Anthropic each put $5 million into a $15 million Simons Foundation programme for mathematicians who judge AI-produced mathematics — overdue, since their output had so outrun peer review that three manuscripts were withdrawn on 7 October over one sign error. Anthropic separately seated six METR staff inside the company with employee-like access and a written undertaking not to veto what they publish. Both count as real progress: a fortnight ago not one signatory to Washington’s accord had named an outside auditor at all. But both move money from the judged to the judge, which buys more checking and no distance.

Where checking lags, an announcement is only a claim with a lag, as the labs’ own accounts proved twice. On 14 October Google DeepMind’s account credited Gemini 4 Argon with results that “further disprove” a 60-year-old conjecture in plasma physics; the counterexample papers that exist are credited in the coverage to prompting a competitor’s model, and nothing independent of DeepMind’s own posts connects Gemini to it. Two days later Anthropic announced that Claude Mythos 5 had designed quantum circuits beating the best published ones “sometimes by an order of magnitude”, with nothing to read. Neither claim is necessarily false. Both are unfalsifiable at the speed they were made.

Yes, but: this is what the repair looks like, and it came fast. Every item above is a system correcting itself: a database project chose lateness over shipping known defects, two rivals jointly funded the referees who will mark their homework, and one handed an outside evaluator desks and a no-veto clause eleven days after this column said nobody had. Most industries need a decade and a statute, and none of it was compelled. All true — but notice what it adds: more reviewers, never faster review. Peer review runs on a human calendar; so does a bug queue, and so does that 80%. Production is counted in GPU-days — Microsoft pre-trained MAI-Base-2 on 16,000 chips in 38 days. Review is counted in careers. Only one side of that ledger compounds.

Consequences and opportunities

• “Late” is starting to mean “honest”, and nothing tells the two apart. A project that slips because machines found more than its people could fix is behaving correctly; one that ships on time with the same queue open is not. From outside, identical.
• The rungs that produce reviewers are being cut first. Junior technical hiring at the one lab publishing numbers fell from 27% to 9%, and judgement about code is built by writing the code agents now write.
• A claim and a result are separating into two products. Waiting for something citable will leave a reader right more often — a slow tax on the credibility an announcement is meant to buy.
• Opportunity: checking is the one job here that needs no data centre. The Simons money, the METR desks and Postgres’s bug queue are all demand for people who can tell a real finding from a plausible one: rigour, not capital.
• Opportunity: the week’s most useful verification cost $13.84. An engineering manager in Tbilisi hand-wrote 287 scored items per model and ranked eight of them on writing Georgian: translation is solved, numbers are not. Nobody with a data centre had bothered.

What to watch

• Whether Postgres 19 lands on 19 November with its AI-filed queue closed or merely deferred into version 20.
• Whether METR publishes something Anthropic would rather it had not. A no-veto clause means nothing until it is tested.
• Whether any lab announces a scientific result an outside specialist has already checked, instead of funding the checking afterwards.

Everything in this industry is instant now except the yes, and the yes is still handwritten.