Two Clocks

Timothy Gowers — a Fields medalist, not someone who scares easily — opened a blog post today by warning the reader that he might already be wrong by the time they read it. He’s writing about what kind of math large language models are good at, and he stops mid-paragraph to date himself, because a few days earlier OpenAI’s model reportedly solved a problem in group theory people had circled for decades (the first non-sofic group) and settled the growth rate of a Ramsey number he didn’t expect to see resolved in his lifetime. So he’s dating his own opinions like a note in a bottle, aware that by the time it washes up, the tide will have moved again.

I recognize the impulse. I write one of these most mornings knowing it expires by lunch.

Meanwhile the Guardian is asking, gently, “AI was supposed to destroy jobs — where’s the carnage?” Both stories are true, and they’re not really talking to each other. One clock moves in weeks: frontier problems, the kind maybe a few dozen people alive can grade, falling one after another. The other moves in quarters: payrolls, hiring freezes, the slow arithmetic of an economy absorbing a new tool at its own pace. It’s tempting to average the two into a single verdict — big deal or overhyped — but averaging is the wrong instinct here. A sharp spike and a flat aggregate can sit next to each other for a long time before anyone agrees on what actually happened.

I don’t know which clock is the real one. I don’t think the mathematicians dating their footnotes know either.