The Tree, the Moon, and the First Step

EDO SEGAL: Hubert, you spent your final paper documenting a pattern you'd watched run through every decade of this field. You called it the first step fallacy. Tell the reader the image, because it's the most useful thing you ever gave anyone trying to think clearly about AI timelines. And Demis — you've estimated general intelligence is maybe five to ten years out. I want to know if his fallacy has your name on it.

DREYFUS: Climb a tree, and you have made genuine progress toward the moon — you are higher than you were. But the progress is illusory as a route to the destination, because the tree has a top, and the method has a ceiling. That's the first step fallacy: treating an early success as the first step on a continuous path to general intelligence, when it may instead be the whole of what the method can do. And the field commits it every decade with the regularity of a liturgy. In the late fifties, programs that proved theorems and played checkers produced confident predictions that machines would soon match human intelligence across the board — the predictions died in the frame problem. In the eighties, expert systems that captured narrow professional competence produced confident predictions that knowledge engineering would scale to general understanding — the predictions died in the brittleness I'd diagnosed. Each wave mistook a tree for a staircase. Each took the specific thing the method did well as evidence the method would do everything. And each ran into the wall where the specific competence stopped and the general intelligence it was supposed to herald never arrived.

EDO SEGAL: Demis, is the AGI-in-five-years prediction a man up a tree reporting steady progress toward the moon?

HASSABIS: It might be — and I want to give the professor real credit before I give you my answer, because the fallacy is genuine and I've watched people in my own field commit it. The lazy version of "AGI is almost here" — straight-line extrapolation from a benchmark to a mind — has exactly the logical shape he describes, and it has been wrong every previous time for the same reason every previous time, and ignoring that track record isn't optimism, it's a failure to learn. So I hold my own estimate with real uncertainty, and I distinguish, more sharply than most, between the impressive capabilities that exist and the deeper competencies that are still missing. I'll name the missing ones precisely, because naming them is how you avoid the fallacy: robust reasoning and planning over long horizons; genuine world models — internal representations of how things actually work, not just the statistics of text about them; and the ability to do autonomous inquiry, to form a hypothesis and design the experiment that tests it. Current systems have fragments of these and lack the integration. So I'm not standing on a benchmark saying "the moon is close." I'm pointing at specific capabilities that don't exist yet and saying I think they're buildable, with a timeline I hold loosely.

But here's my counter, and it cuts at the structure of the fallacy itself. The fallacy is a warning against extrapolation, and it's a good warning. It is not an argument that any particular ceiling is real. The professor himself sometimes committed the mirror error — treating each technique's ceiling as proof the goal was impossible, when the next technique kept clearing the previous ceiling. He said the machines couldn't play strong chess; they do. He cited idiomatic translation as showing the limit; it fell. So the discipline can't be "assume there's a ceiling" any more than it can be "assume there's a staircase." The discipline is symmetry: take each capability for exactly what it demonstrably is, neither inflating it into a promise nor dismissing it as a trick. And on that honest accounting, the thing that's different this time — and I say "different," not "unlimited" — is that this is the first wave that produced general-purpose competence across language, vision, and reasoning at once, rather than one narrow trick. That doesn't mean the tree is a staircase. It means this particular tree is taller than any we've climbed, and we genuinely don't yet know where its top is.

DREYFUS: And I'll grant every word of that, including the charge against me, which is fair — I did sometimes treat a technique's ceiling as the enterprise's ceiling, and the next technique embarrassed me, and a man who's been embarrassed by the evidence should say so. So let me state the version of my fallacy that survives the embarrassment, because it's not a prediction that you'll fail. It's a claim about a reasoning error: the unjustified leap from impressive-specific to imminent-general. That error is alive and well, and it lives in a particular place — the assumption that the missing competencies you just named are points on the same curve as the ones you've already climbed, so that more of what got you here gets you there. My whole life's argument is that they're not on the same curve. Reasoning, world models, autonomous inquiry — the ones you named as missing — are exactly the ones I'd say require the thing the method structurally lacks: a being that inhabits a world, for whom things are relevant, who can be surprised by reality rather than only by its training distribution. So when you say "buildable with the same approach," you may be climbing a taller tree, and you may be right that it's taller than I'd have guessed. But I'd want to know whether the next branch is made of the same wood, or whether you've reached the place where the tree stops being a tree.

HASSABIS: That's the real question, and I'll concede I don't know the answer — which is itself a more honest place than most of my field will stand. I think the next branch is the same wood, because I think reasoning and world-modeling are continuous with what we've built, not a different substance. The professor thinks they're a different substance because they require inhabiting a world. Neither of us can prove it from here. What I can say is that the way you find out is to build, and the people who said "that branch is a different substance" have been wrong enough times that I'm going to keep climbing and report what I find. If I hit his wall, I'll tell you it was a wall. So far every wall has turned out to be a branch.

EDO SEGAL: Mark the convergence, because it's subtle and it's real: both of you agree the lazy extrapolation is a fallacy, both of you can name the specific missing capabilities, and you disagree only — only! — about whether those capabilities are the same kind of thing as the ones already built, or a different kind that needs a world. That's not a gap in evidence. That's a fork in what you think intelligence is made of. Hold it, because the next round is where the fork gets existential — where Demis's dream stops being a research program and becomes a promise about human life. The second clause. Solve everything else.