The Apprentice and the Word "Yet"

**EDO SEGAL:** Roger, you've written that there's a single word doing more damage to clear thinking about AI than any white paper — the word "yet." The machine can't originate questions *yet*. Can't feel stakes *yet*. The "yet" smuggles in a forecast and disguises it as humility. Unpack that for me. And Marvin — the "yet" is basically your entire worldview compressed to three letters, so defend it.

**PENROSE:** "Yet" is the most consequential word in the discourse and almost nobody hears it land. When you say the machine can't do something "yet," you've asserted that the limitation is *temporal* — that it's a matter of more data, more scale, better architecture, and time. But that's the whole question, not a throwaway qualifier. If the limitation is structural — if it's drawn by Gödel, by the difference between computing and seeing — then "yet" is not premature. It's wrong. There is no future in which it's overcome, the way there's no faster engine that exceeds the speed of light, because the speed of light isn't an engineering target, it's a feature of spacetime. The honest word isn't "yet." The honest phrase is "we don't know whether." And every practical decision — what to teach a child, how to build a school, whether judgment has an expiry date — turns on which word you use, and the entire culture has quietly chosen "yet" without noticing it made a choice.

**EDO SEGAL:** Marvin — defend the "yet."

**MINSKY:** Gladly, because the "yet" has a track record and the "never" has a graveyard. Every single thing the never-camp said machines would never do — never beat a grandmaster, never recognize a face, never hold a conversation, never prove a real theorem, never make a poem that moves you — they did, one after another, and each time the never-camp moved the line back and called the next thing the *real* never. Roger's "never" is the latest line, drawn just past wherever the machines currently are, with a theorem painted on it for respectability. I've watched this movie for sixty years. I was *in* it — I told people general intelligence might be twenty years off and I was wrong about the timing, badly wrong, but notice the direction of my error: I was too *optimistic*, the thing was harder than I thought, not impossible. Hard is not impossible. "Yet" is the word of someone who's been wrong about timing. "Never" is the word of someone who's about to be wrong about possibility.

**PENROSE:** But you've just done the thing, Marvin — you've listed behaviors. Chess, faces, conversation, theorems, poems. I granted every one of those before you said them. I have always said the machine will do all of it and more. My "never" was never about a *behavior*; it was about understanding, about the inside, about the seeing. You can't refute a claim about understanding by listing behaviors any more than you can refute a claim about a painting's having a painter by noting how good the forgery is. The graveyard you're pointing at is full of people who said the machine couldn't *do* X. I'm not in that graveyard. I said it would do everything and understand nothing, and nothing you've listed touches it.

**MINSKY:** Then your "never" is unfalsifiable, and that's worse, not better. If no behavior could ever count — if the machine could do *literally everything* a human does and you'd still say "no understanding" — then your claim makes no contact with the world. You've defined "understanding" as precisely the thing that can never be demonstrated, which means you can hold it forever no matter what happens, which means it's not science, it's a fortress. At least my "yet" can be killed. The day the machines stop improving, the day they hit a wall and sit there for fifty years, I lose. What loses *you*?

**PENROSE:** Gödel loses me — if someone shows the human perception of mathematical truth is, after all, formalizable without remainder, my argument collapses, and I'd accept it. That's my falsifier and it's a real one, in mathematics where I live. You want a falsifier in *behavior* and I keep telling you the claim isn't about behavior, so of course no behavior falsifies it — that's not a fortress, it's a claim correctly aimed at its actual target. We're back at the seam: you'll only count a test you can watch from outside, and the thing I'm pointing at is, by its nature, on the inside.

**EDO SEGAL:** Let me bring an old worry of mine into this, because it's where the "yet" bites the reader hardest, regardless of who's right. The [apprenticeship problem](https://www.youonai.ai/fieldguide/med/apprenticeship_problem). For the whole history of skill, you became a master by doing the grunt work — the junior lawyer's document review, the junior coder's bug fixes — and somewhere in that friction, the judgment got built. If the machine does all the grunt work, the friction vanishes, and I don't know where the next generation's judgment comes from. Marvin, you studied how minds get built. Does the apprentice still get made?

**MINSKY:** This is the real danger and it's got nothing to do with whether the machine is conscious — which is why I think it's the most important thing we'll say tonight. Judgment is *learned*. It's built by the friction Edo's describing — you reactivate the states that worked, you accumulate the cases, you organize the agencies, and that takes doing the thing badly many times. [Strip the friction and you strip the construction](https://www.youonai.ai/fieldguide/med/friction_that_produces_understanding). On my own theory — judgment is learned organization — removing the apprenticeship is catastrophic, because there's no other way the organization gets built. So here's a place Roger and I should be allies, not enemies: whether judgment is computable or not, it is *grown by struggle*, and a civilization that automates away the struggle stops growing the thing. I'd build the schools to protect the friction even though — *especially* though — I think the machine could in principle learn the same way. The kids need the friction more urgently than the machines do.

**PENROSE:** And I agree with all of it, word for word, which should tell the audience something. We've spent two hours unable to agree on what a mind *is*, and we agree completely on what to do about the apprentice. Protect the friction. Guard the slow, unaided wrestling. Don't let the machine do the homework that was never about the homework. The only thing I'd add is the reason it cuts deeper on my view: if the seeing is non-computable and permanently human, then the friction isn't a temporary training regime to keep humans ahead until the machines catch up — it's the permanent method by which the one irreplaceable thing gets cultivated, forever. You don't ever automate it away, because what it grows is the only thing the machine can't supply. Marvin protects the apprenticeship as a holding action. I protect it as a cathedral.

**EDO SEGAL:** Convergence five, and it might be the most useful sentence for a reader in the whole night: protect the friction, build the schools around the struggle, no matter which of these men is right about the metaphysics. Hold it. Because next we go to the most intimate room — the self, and whether the *I* you feel yourself to be is the one thing the machine can never have, or just another part it's already building.