In September 2026 Scott Alexander challenged Steven Pinker to a public debate on whether AI could end us. Pinker declined. A debate, he wrote, is “a spectator sport, not an epistemic tool,” and he suggested bets on verifiable predictions instead. Alexander answered with an open letter, itemized and checkable.
Read both letters and the two largely agree. Both say the dangers are real and want oversight. The difference is degree and timing. Alexander thinks the danger is close enough to act on now. Pinker doubts the extinction scenarios and does not share that urgency. On urgency I side with Alexander.
On the manner of continuing, I side with Pinker, though Alexander’s written reply is a step in the right direction. Urgency tempts escalation: the challenge, the open letter, the list of grievances. That is exactly when the slow manner matters most. Writing beats a stage, because a reader can check each claim and nobody wins on a zinger. A bet beats both. It forces each side to name what would change its mind, which neither letter did, and it adds a deadline and a score.
Tetlock showed that forecasters who keep score get better, and a bet keeps score. You cannot collect on extinction. The bets belong on the steps toward it: an AI that copies itself, resists shutdown, or runs a long task on its own. An adversarial collaboration with resolvable predictions would settle more than a third open letter. It would also force the Dennett test, since you cannot bet against a position you have not stated well.
The same split runs through the optimists and doomsayers on AI, and through every quarrel our slow republic was built to cool.