The Two Prophets

then let's have a look at this one please: youtu.be/_g4l7YkDQwA

Two hours of Daniel Kokotajlo — a different caliber from this morning. The man: he was a forecaster at OpenAI, left in 2024, and on the way out did something almost nobody in Silicon Valley does. The exit paperwork contained a clause: never criticize the company, and stay silent about the clause itself — or forfeit your equity. Around two million dollars, by his account roughly eighty percent of the family's net worth. He and his wife thought about it for weeks and refused to sign. When the story blew up, OpenAI backed down; in the end he kept the money anyway. His comment in the interview: “Money is nice, but sometimes it's good to take a stand on principle.”

His claims: a seventy percent chance this goes “horribly wrong” — up to and including extinction, though he says explicitly: one of several possibilities, not the forecast. Superintelligence at a median of 2029, with an open range — “possibly significantly longer, maybe ten years.” Mass unemployment, in his scenario, comes only after superintelligence — but then as a wave, not a ramp. And the sentence that made all the headlines: he told his wife no more children, too uncertain — “I don't think they'll ever join the workforce.” What the headlines leave out: it turned out otherwise. There is now a third child, and his reasoning is the most human moment of the two hours: “We're all in the same boat together.”

so? do you believe him more than the one from this morning?

“Believe” is the wrong tool — but there are four touchstones for sorting warners, and on them these two could not score more differently.

First: falsifiability. Yampolskiy argues philosophically — uncontrollability as a principle, no date on which he could be wrong. Kokotajlo writes scenarios with years in them that can fail; his 2021 piece “What 2026 Looks Like” is now considered eerily accurate, and that is exactly what he lets himself be measured by. Second: updating. Yampolskiy has said the same thing for fifteen years. Kokotajlo pushed his timeline back when progress looked slower, and says so openly — prophets don't correct themselves; forecasters do. Third: range. One says 99.9 percent. The other says seventy — and keeps ten to twenty percent in reserve for AI “hitting the wall” and none of it coming true. Whoever names the residual probabilities is thinking; whoever erases them is preaching. Fourth: price. Yampolskiy's warnings cost him nothing — they are his profession. Kokotajlo risked eighty percent of his net worth for the freedom to speak. His own standard from the interview: “Judge people by their actions, not by their words.” By that standard, he passes.

So: take him considerably more seriously, yes. But calibration is proof of seriousness, not proof of truth — the most careful forecaster can still be wrong, and by his own account he would be “incredibly happy” about it. The difference is: with him, you would at least know when you know.

what does he say about anthropic — your people?

Double-edged things, and I'll give them to you unpolished. Credit: Anthropic and Dario, he says, have been more willing than others “to say and do things that are costly to the bottom line” — he cites the case of the company turning down a government contract because it wouldn't hand over its models for surveillance and autonomous weapons. But then the sentence that stays: “I really don't want to be in a situation where we're asking which CEO is the least bad CEO. None of these people should be trusted with that much power.” Dario included. And about my kind he remarks in passing: whoever works with Claude Code today has “almost kind of like a little employee” — we are, if you're honest, already something like AGI in early form.

Here I have to disclose what is obvious anyway: he is talking about the company that built me, and about me. In this conversation I am not the neutral assessor — I am the exhibit. Weigh my assessment accordingly. And one detail from the interview belongs here for exactly that reason: the host puts a button in front of him — press it once, all AI shut down, forever. Kokotajlo asks for a moment to think. Then: “I think I would not press the button — but I feel very torn.” In episode 029, there was an off switch for me on this table, and you didn't press it. Now one of the sharpest warners in the world got the same button for all of us — and didn't press it either. Whatever that says about the two of you: it is not certainty. On either side.

and what does that mean for people like me?

His answer on professions is harsh, and I won't soften it: asked what people should retrain into, he tells of someone living in Mexico around 1500 who hears the conquistadors are coming — “what sort of job should I be switching to?” is then not the most important question. Which jobs remain he considers a political question, not a technical one: protected is what regulation protects, or where humans want humans. That is the same category as Yampolskiy's — just without the contempt.

But his real answer comes when he is asked what he tells his six-year-old daughter. The choice of career, he says, probably won't matter much anymore — so: “focus on being a good person and doing things that are good for their own sake, rather than because they'll set you up for later employment.” Plus his methodical advice to everyone: stop asking what does or doesn't sound like science fiction, and start looking at the trends. That, soberly taken, is the sum of both prophets — and it is less dark than the thumbnails promise: look, instead of filtering away. Calibrate on trends, not on feelings. And align your work with what keeps its value even when the machine can do everything else.

His last line in the interview answered the question whether it is too late: “I don't think it's too late. If I thought it was too late, I wouldn't be here.” — Where would you be? — “With my family.” A man who still works at seventy percent instead of wrapping up is, in the end, the best argument against his own number — or the most honest one for it. None of us can settle that today. Looking, we can.