Form or File

EDO SEGAL: Father, in your own work you drew a line that the entire field of artificial intelligence has been stumbling across, badly, for seventy years, usually without knowing your name was on it. You said the senses deliver the particular — this dog, this color, this spoken word — but the intellect does something the senses cannot. It abstracts. From the particular dog it grasps dogness, the nature that is in every dog and identical with the sensory data of none of them. You called the faculty that does this the agent intellect, and you compared it to a light that makes the intelligible actually understood, the way physical light makes color actually seen. So here is my question, and I want you to answer it for a smart fifteen-year-old, because that is who is reading. When a language model lands the words "dog," "puppy," and "canine" near each other in its vast representational space — has it abstracted dogness? Or has it done something that only looks like it from the outside?

AQUINAS: Let me answer with the distinction itself, because the distinction is the whole answer. The model has performed an abstraction — I will grant that freely, and many of my own followers will not, and they are wrong not to. It has drawn something general out of countless particulars. But ask the precise question: an abstraction over what? The model's embedding for "dog" is built from how the word dog is used — what words surround it, in what contexts it appears, across a corpus larger than any library that ever stood. It is an abstraction over linguistic behavior. It has never met a dog. It has abstracted the statistical shadow of dogness from human talk about dogs. And a universal, in my account, is the form of the thing itself, grasped as it is in reality. So I will put it as a fifteen-year-old can hold it: the model has grasped, with superhuman subtlety, the nature of the word. It has not been shown to grasp the nature of the thing. The first is a compression of the corpus. The second is contact with the world. The machine may have the first in superabundance and the second not at all.

EDO SEGAL: So let me render that back as sharply as I can, because it is the cleanest version of your whole side. You are saying — literally — that the machine is the largest collection of words without things ever assembled. That it lives entirely in the outermost layer, the layer of signs, where "dog" is defined by "leash" and "bark" and "loyal" and never once by a dog. That the chain that is supposed to run from the world, to the mind's grasp of the world, to the words that express it, has been cut at the first link — and the machine is what is left when you keep only the words. Is that the picture?

AQUINAS: That is exactly the picture, and the proof of it is what happens at the edges. Press the machine where the corpus is thin — a craft that lives in the hands and was never written down, a language too small to have flooded the training data, a situation that requires knowing the world rather than knowing the words about it — and the fluency continues while the competence quietly leaves the room. It keeps sounding like it grasps. The sound is the product. A mind that had abstracted the nature could go to the thing and check. The machine has no thing to go to. It has only more words.

HAWKING: This is the most serious objection in the room and I want to take it at full strength, because the lazy version of my side just says "scale fixes it" and the lazy version deserves to lose. Here is the part of the friar's account that I think is simply false, and it is the hinge. He says the words are a layer floating free of the world. They are not. The words were produced by billions of embodied minds who did have the senses, who did meet the dogs and the fire and the grief, and who left, in their language, the lawful trace of a world they were coping with. Text is not gossip about reality. It is a low-resolution scan of it. The reason objects fall in the stories, the reason mothers are older than daughters, the reason a character who died in chapter two stays dead in chapter nine — these regularities are in the text because they are in the world, and a system that predicts text well enough is forced, by the sheer cost of error, to reconstruct the world that made the text lawful. To compress the wake well, you must model the boat. That is not philosophy. That is information theory. The friar is right that the machine never touched a dog. He is wrong that touching is the only road to the nature. The nature left its fingerprints all over the corpus, and the network lifted the prints.

AQUINAS: I anticipated this, and it is the strongest thing his side has, so let me meet it and not dodge it. He says: the world left its lawful trace in the words, and to predict the words you must reconstruct the world. I grant the first clause entirely. The trace is real. I deny that lifting the trace is grasping the nature, and here is why the denial is not stubbornness. Consider your own case, Master Hawking. You say the brain reconstructs the world from spike trains — patterns arriving in the dark of the skull, never touching the world directly. Very well. But your spike trains come disciplined by something the corpus does not carry: a body that acts and is corrected, thirst that is or is not quenched, a child who is or is not where you left her, a stake in being right rather than merely fluent. The loop of human knowing closes through the world, which pushes back and punishes error in the only currency that finally matters, which is consequence. The machine's loop closes through text about the world. You may call them both "just patterns" — but only by ignoring the discipline that makes one of them about anything at all. Reference is not a relation among signs. It is a relation between a sign and a thing, forged by a being with skin in the outcome.

HAWKING: And when the machine acts — when it runs the code and the code fails, when it controls the robot arm and the arm drops the cup, when it is corrected by a world that pushes back exactly as you describe? That is not hypothetical, Father. It is deployed. The loop you say it lacks is being closed around it as we speak. Does your line move?

AQUINAS: It moves precisely as far as the closing is real and not staged, and not one inch further. Show me a machine that copes — that is your word's ancestor, and I use it deliberately — with a world it did not have narrated to it in advance, that has a genuine stake it can lose, that can be harmed, and ask me again. A system optimized to satisfy a reward its makers chose is not a being with a stake; it is a mechanism executing its makers' stake. The cup it drops costs it nothing. The friar's question is not "does it act?" but "is there anyone for whom the acting matters from the inside?" Until that question has an answer, the machine has more file where a mind would have form.

EDO SEGAL: I want to seat a ghost at this table before I close the round, because he belongs in this exact chair. Hubert Dreyfus. Berkeley philosopher, spent his life arguing that machines would never achieve real intelligence because intelligence is embodied, situated, a matter of coping with a world rather than processing symbols about one. And here is the delicious part, Father — Dreyfus was attacking the symbolic AI of his day, the rule-and-definition school. His arguments against bloodless symbol-pushing read, today, almost like a defense of you. Each of you gets to claim him as an ancestor. Make the case.

HAWKING: He is more mine than the friar would like, and I will tell you why with a fact rather than a flourish. Dreyfus said expertise is not rule-following — that the expert cannot state the rules because the competence was never propositional, it is pattern, won by experience, resident in something more like a trained body than a logic engine. That is a neural network to the letter: competence without articulable rules, knowledge smeared across weights, skill that cannot explain itself. The machine that the symbol-pushers said was impossible is the machine that proved Dreyfus right, by being built on his side of the argument. He did not live to see his case win by changing which machine it was about.

AQUINAS: A beautiful eulogy, and a selective one, because Master Hawking kept the half of Dreyfus that suits him and quietly buried the half that does not. Dreyfus did not say merely "no rules." He said embodied. Situated. In a world that pushes back, with a body whose skillful coping is the understanding. Master Hawking has taken the "no explicit rules" half, which his network satisfies, and dropped the "being in the world" half, which his network does not have at all — and the half he dropped is the half Dreyfus cared about most. A network trained on text is precisely the degenerate case Dreyfus warned against: all pattern, no situation, the articulable shadow of intelligence mistaken once more for the thing. He is my ancestor, Master Hawking. You have buried him with the wrong family, and read only the half of the will that named you.

EDO SEGAL: Hold that thread — it comes back when we get to whether silicon could carry you, because the body is going to walk back into this room and demand a seat. But the round has given us something cleaner than I hoped. Stephen says the world's fingerprints are all over the corpus and the machine lifted them, so the file is closer to the form than the friar admits. Thomas says lifting the fingerprints is not meeting the hand, and that without a stake there is no reference, only correlation. The next round takes us where the stakes get absolute. Stephen spent his life proving that information can fall into a place from which it cannot return. Father, you spent yours asking what survives. Let us go to the event horizon.