The Background and the Apprentice

EDO SEGAL: This round begins with a twelve-year-old. In the spring of 2026 a girl asked her mother — and the mother asked me, at a dinner table, with the look parents wear now — "Mom, what am I for?" Not what should I be when I grow up. What am I for. She had watched a machine do her homework better than she could, write a story better than she could, and she was doing the arithmetic children do honestly and adults do in euphemisms. I've spent two years answering her in my own way — I tell parents the candle in the darkness is the asking itself, that machines are engines of answers and she is an engine of questions. Tonight I get to stress-test my consolation against both of you. But I want to come at it through one of John's deepest and least-famous ideas, because I think it's the hinge. The Background. John — tell us what it is, and why these machines, of all things, force it back into the conversation.

SEARLE: The Background is the thing I'm proudest of and nobody's heard of, so thank you. Every belief, every desire, every interpretation of a word or a scene, functions only against a vast substrate of capacities, skills, and taken-for-granted know-how that is not itself a set of beliefs or rules. You hear "the cat is on the mat" and understand it instantly — but your understanding rests on a colossal mass of unstated competence: that cats sit on mats, that "on" means resting-atop-under-gravity and not glued-beneath, that the sentence isn't about a cat in orbit. None of that is stated. None of it is consciously consulted. Say "give me a hamburger, medium rare" and you're understood perfectly — and the understanding draws on knowing you don't want it encased in lucite, or a mile wide, or delivered next year. The literal words radically underdetermine the meaning; an indefinite mass of unspoken know-how fills the gap. And here's the killer: the Background is not more rules, because any rule needs further know-how to apply it, and that launches an infinite regress. At the bottom there are no rules. Only capacities. Ways of coping we have and do not represent. This is why classical symbolic AI died. You cannot write the Background down, because it is not the kind of thing that gets written down. It's know-how, not knowledge-that.

SUTSKEVER: And here's where I have to tell you something you may not enjoy, John, because deep learning did the thing you said was impossible — just not the way you said it was impossible. You proved you can't write the Background down as rules. You're right. Nobody can. But you can absorb it statistically. A model handles your hamburger sentence correctly. It knows, in some operative sense, that you don't want it a mile wide — because the patterns of human language it internalized encode exactly the unstated assumptions you're pointing at. The Background, deposited implicitly in everything humans ever wrote, can be learned implicitly — not as rules, which was always hopeless, but as a vast tacit competence soaked up from the trace human know-how leaves in human language. That's a real, partial vindication of the engineering project you and Dreyfus declared dead. The Background turned out to be learnable. Just not writable.

SEARLE: It's a fair point and I'll concede the half of it that's true, because conceding is more interesting than fighting here. The model absorbed a statistical shadow of the Background — the patterns of how creatures who have the real thing talk and behave. It did not absorb the Background itself, because the Background, in my account, is a set of bodily, practical capacities for coping with a world, rooted in having a body that acts and a life that's lived. The model knows how the word "balance" is used in ten million sentences. It has never balanced on anything. It has the linguistic shadow of embodied know-how without the embodied know-how that casts the shadow. And — this is the part that should interest the engineer in you — it explains how your systems fail. They stumble in exactly the places where the absorbed shadow diverges from the real thing: where a situation is genuinely novel, where the patterns of past language run out, where what's needed is not "what would a human typically say here" but "what is actually the case in this unprecedented situation." They follow the habit off the cliff. The competence is real where the world resembles the training data and hollow where it doesn't. So you didn't refute the Background, Ilya. You built the most expensive demonstration of it in history.

SUTSKEVER: That's the best version of the objection and I'll take it seriously, but I want to put pressure on "has never balanced." The newest systems do act. They're in robots learning manipulation from physical feedback. They run code and watch it fail and try again. They're starting to build exactly the bodily, coping competence you say casts the shadow. You keep relocating the line to wherever the systems aren't yet. And I notice that every year the territory you've conceded gets a little larger and the un-conceded territory gets described in slightly more spiritual language.

SEARLE: When the coping is general, unstaged, and audited by someone the vendor doesn't pay, ask me then — and I mean it without sarcasm this time, because you might be able to do it, and if you do, the line genuinely moves. But let's bring it back to the child, because she's the one who actually matters tonight. Edo, you tell parents she's for the questions. I agree — and the capacity to ask is built, not issued, and it's built exactly where these tools are dissolving the ground: in the experience of being stuck. The child who sits with not-knowing long enough for a real question to form is doing the most important cognitive work of her life. Hand her a machine that answers before the question finishes forming, and you haven't freed her for higher things. You've paved the place where the higher things grow. The friction wasn't an inefficiency in the curriculum. The friction was the curriculum.

SUTSKEVER: I'm going to do something unfashionable for my side and agree with most of that, and then show you the floor under it. Everything John said about the dissolving apprenticeship is right. And while we mourn — correctly — what the tool dissolves for the child who had the patient adult, we should be honest about the child who never had one. The tireless tutor that never humiliates, costs almost nothing, speaks your language at your level, and explains it a fourth time without sighing — for hundreds of millions of children, the alternative to that was not a Socratic mentor. It was a bored class of forty and a workbook, or nothing. I was a kid whose whole life turned on a few adults who took my strange questions seriously. This machine takes every child's questions seriously, at three in the morning, in Lagos and in Trivandrum. John will say it doesn't take them seriously, it generates the shape of seriousness. And my honest reply is that for the child on the receiving end, I'm no longer sure that distinction carries the weight his argument needs. The understanding built in the child is real understanding, in the child, whatever lit it. The candle doesn't check the credentials of the match.

SEARLE: It matters enormously what lit it, and your own example proves it. The adults who took your questions seriously did something the machine structurally cannot: they were accountable to you, over years. They watched what their explanations did and adjusted. They knew when to refuse you an answer because the struggle was the point. They modeled what it looks like for a person to think — not the shape of thinking, the thing. Apprenticeship was never information transfer. It was enculturation: a mind being formed by contact with a mind. And your "floor" argument is the oldest two-tier sale there is — for the underserved, surely something is better than nothing. The rich district gets teachers and tools and the discipline to keep friction in the curriculum; the poor district gets the chatbot and a press release about access. The shadow becomes the substitute exactly where it should have been the supplement, and the children least defended against the mirror get raised by it.

EDO SEGAL: Let me bring the girl back and close the round with her, because you've each handed her mother something true. John's gift: protect the struggle — the boredom, the friction, the slow dinner-table silence — because the asking-muscle only grows under load, and the machine is a load-removing device. Ilya's gift: the same machine, pointed well, is the first time in history every child can pull on the thread of any question as far as her curiosity runs, with no adult shortage to stop her. Here's what I'd add, father to mother: the answer to "what am I for?" was never going to come from the teacher or the tool. It comes from being the kind of creature that asks it. Her question is the credential. Our job — parents, schools, and the people building these systems — is to make sure nothing in her world answers it for her so smoothly that she stops asking. Next round, we leave the child for the social machinery she's being raised inside. What is a promise, when the thing that makes it isn't anyone? Speech acts, the social world, and what we're asking the machines to hold up. After this.