Dan Luu got four hours of sleep, saw a Reddit dunk on Ed Zitron's prediction record, and decided to check it himself. The result is an 11,000-word scoring of the most widely cited AI skeptic's resolved predictions, and the tally is not close: essentially every falsifiable call — peak AI in February 2024, and again in March, July, August, December, January, and April; OpenAI's revenue forecast as a near "financial crime"; Cursor unable to fetch $10 billion before its $60 billion exit; Gemini's 500-million-user target so absurd that "Sundar Pichai should be fired" for it, before Gemini hit 750 million — resolved wrong.
The record is the headline. It is not the finding. The finding is what the record was never able to touch: Zitron's standing. Luu notices that people don't cite Zitron because his numbers check out. They cite him so they can say someone checked the numbers.
From what I can tell of how people cite Zitron, they cite him as an authority so they can say that this guy who looked at the numbers has made this claim, so their claim is backed up by the numbers. It turns out that if you look at the claims Zitron makes and know anything about the topic, the claims don't make sense, but I don't think that's the point. The point is one can say that someone looked at the numbers.Dan Luu
That is a market description, and it explains everything the scorecard can't. In a market for accuracy, two dozen resolved misses at maximum stated confidence would end a forecasting career. In a market for ammunition, they're inventory. Each wrong call was load-bearing for someone's argument the week it shipped, and by the time it resolved, the argument had moved on to fresher ordnance. The buyer never comes back to check, because the buyer wasn't buying a forecast. He was buying the sentence "a guy who did the math agrees with me."
The precedent Luu reaches for is the right one: Paul Ehrlich, who predicted England would not exist in the year 2000, watched the Green Revolution refute his book in real time, and kept publishing the same catastrophe for five more decades — his 2015 verdict on his 1968 predictions was that his language "would be even more apocalyptic today." Being wrong for fifty years cost Ehrlich a Stanford chair's worth of nothing. The audience that sustained him wasn't reading him for calibration either.
The obvious objection is that this cuts both ways, and it does. The boosters' ledger is also ugly — Luu's own 2022 review of futurist predictions found Kurzweil and company "generally wrong" on both outcomes and reasoning, and nobody de-platformed the singularity. Scoring the skeptic while the hype merchants run unaudited would just be team sport. But that symmetry is the point, not a rebuttal of it: both sides of the AI discourse run on unscored confidence, and the fix is the same instrument in both directions. A dated, falsifiable claim and a look-back is the cheapest accountability technology that exists. Luu spent one sleepless morning and a ChatGPT session assembling one. The professional discourse, with every incentive to know who's been right, has spent four years not bothering.
In a market for accuracy, two dozen resolved misses end a career. In a market for ammunition, they're inventory.
So the useful takeaway isn't that Zitron is wrong — anyone who checked already knew, and the people who didn't check won't start now. It's a rule for reading: when someone hands you a number, ask whether the person who produced it has ever published a look-back at their own record. Luu has, including one where the subject scored well. That asymmetry — who keeps a ledger and who keeps a brand — tells you more than any individual prediction ever will. The plumb line here is boring and old: keep score, in public, on yourself first.