The Doom Predictions, Graded
A decade of AI executives warning the world about their own product, checked one by one against what actually happened. Most of it is unfalsifiable marketing. Three predictions were specific enough to test — and the newest grade just came in, from Anthropic's own numbers.
Am I The Idiot — the strongest case against our own take.
The piece's strongest move, catching Amodei's bioweapons prediction "checking out," is also its weakest evidentiary link. The only evidence offered is Anthropic's own threat-intelligence report: the company that made the 2023 prediction is also the sole source grading it three years later. A report describing attempted misuse is not proof of successful uplift, and it doubles as exactly the kind of safety marketing this outlet is elsewhere skeptical of; a security team publishing "we caught five attempts" is also a security team advertising that its detection works. If Anthropic has a commercial incentive to be seen as both dangerous enough to matter and safe enough to trust, this report serves both goals at once, and the piece doesn't fully reckon with that beyond naming the source.
The jobs-prediction grade leans on "early labor-market data" that isn't cited. If that data reflects a broader tech hiring slowdown, rate environment, post-ZIRP correction, an unrelated sector contraction, rather than AI substitution specifically, the "trending toward confirmed" call is simply wrong, and there's no named dataset in the piece for a reader to check that against.
There's also a harder challenge to the whole framing: calling Musk's, Altman's, and CAIS's statements "unfalsifiable marketing" assumes they were designed to be untestable rather than that existential risk is, by its nature, a claim that can only be evaluated probabilistically and in hindsight. A genuine, good-faith worry about a low-probability, high-severity outcome will always look like a mood before the fact. Treating vagueness as evidence of insincerity is a real interpretive choice, not a neutral fact, and a reader could reasonably reject it.