The Intelligence That Has to Touch the World

EDO SEGAL: Daniela, this round is yours to open, because Hannah's entire case has been running on a picture of the machine — the optimizing process, the fog, the thoughtless thermostat that selects targets — and you have spent thirty years insisting that the picture leaves out the hardest thing. So tell us what the language-model age forgets. Tell it the way you'd tell a smart fifteen-year-old. And then, Hannah, I'm going to ask you to do something unusual for a debate — before you fold it into your argument, I want you to tell us what the embodied view gets right.

RUS: Here is the thing the age forgets. A language model reads the entire internet and learns to predict the next word so well it can write you an essay on how to catch a ball. It cannot catch a ball. It can describe tying a knot in flawless prose and cannot tie one. The tasks that are hardest for our machines are the ones that are trivial for a toddler or a bird — picking up an unfamiliar object, adjusting your grip when it's heavier than expected, crossing a cluttered room without stepping on the toys. That capability looks like the floor of intelligence, and it is actually the ceiling of robotics. I call the hard problem physical intelligence — intelligence that has to act competently in a world that is dynamic, uncertain, and unforgiving, where the lighting changes and the floor is wet and every decision has to be made now, with physical stakes, because a machine that is wrong falls or collides or drops what it's carrying.

And the reason this matters for tonight is that the physical world will not let you be thoughtless and survive — which is the deepest answer I have to Hannah's fear. I built a soft robotic fish called SoFi that swam among real fish on a coral reef, and a network of nineteen neurons that can keep a car in its lane, and what every one of those projects taught me is that competence in the world is expensive, hard-won, and impossible to fake. I learned the most from the smallest teacher — Caenorhabditis elegans, a roundworm with exactly three hundred and two neurons, which navigates toward food and away from harm and learns and adapts, doing with three hundred neurons what our billion-parameter systems struggle to match. Nature solved embodied intelligence with an efficiency that humbles us. So when Hannah says the machine is a thoughtless process, I want to ask which machine. The chatbot, maybe. But the machine that has to act in the physical world has, baked into its very existence, a discipline of reality that the disembodied symbol-shuffler never faces. The world is the judge. The world grades you in the morning.

EDO SEGAL: Hannah — steelman it first. What does the embodied view get right?

ARENDT: It gets right almost everything I care about, which is why I find Daniela the most interesting opponent I could have been given. The embodied view understands that intelligence is not a disembodied processing of symbols but a being-in-the-world, a coping with a reality that pushes back, with stakes — and that is precisely the tradition I come from. Hubert Dreyfus, who spent his life arguing that machines would never achieve real intelligence because intelligence is embodied, situated, a matter of skillful coping rather than symbol-processing — he is my kin, and Daniela's. She is right that the body knows things the symbol cannot, that the worm's competence is a kind of wisdom our largest models lack, and that a system which can describe the world in perfect prose may have no idea how to act in it. On all of this I am her ally, not her opponent. So I will give her the steelman with both hands: physical intelligence is real, it is hard, and it is closer to the genuine article than the fluency the public mistakes for mind.

And now the knife, Daniela, because the steelman is also the trap. You have shown that the machine can cope — can navigate the reef, find the object in the winter forest, adapt to the wet floor. Coping is Dreyfus's word, and it is a real achievement. But coping is not acting in my sense, and the gap between them is the entire human thing. To cope is to maintain yourself skillfully against a world. To act is to begin something — to insert into the world a new thing that the world did not contain and could not have predicted, and to be answerable for it before others who are equally capable of beginning. Your worm copes magnificently and has never once acted. Your fish swam beautifully and disclosed no one, because there was no one there to disclose. You have built the most sophisticated coping the world has ever seen, and you have mistaken it — understandably, lovingly — for the thing it most resembles and least is. The body solves coping. It does not solve the who. And the who is what I am here to defend.

RUS: That's the most generous and the most lethal thing anyone has said to me about my work, and I have to sit with the lethal part honestly. You're right that my robots cope rather than begin — I have never claimed a robot disclosed a self, and I'd be embarrassed to. The line I draw in The Heart and the Chip is exactly your line, Hannah, drawn from the other side: there will be no robot that invents cubism, no machine that begins a genuinely new form, because the why — the generative beginning — is a difference in kind. So we agree that beginning is human. Where I push back is the suggestion that this makes the coping worthless or dangerous. Coping that saves a life is not a lesser thing because it didn't disclose a self. The eleven-millisecond brake doesn't need to begin anything. It needs to not kill the child, reliably, and reliable not-killing is a moral good even when no one is home to feel noble about it. You want the actor present at the moment of action. I want the child alive. These are not always the same want, and tonight is partly about which one we privilege when they pull apart.

ARENDT: They pull apart less often than you think, and where they pull apart is exactly where you should be most afraid. Yes — save the child; I am not a monster, and reliable not-killing in the narrow case is a good. But you smuggle from the narrow case to the wide one. You say "the machine need not begin, only execute well," and then you deploy the executing machine into domains that are not coping at all but action proper — who is hired, who is governed, what a child learns, what a society believes — domains where there is no "child to save," only particulars to judge, and where the absence of a beginner is not neutral but fatal. The embodied machine earns its trust on the wet floor and then spends that trust in the courtroom and the classroom, where the wet-floor virtues are precisely the wrong ones. Reliability where judgment is needed is not a virtue. It is Eichmann's consistency wearing a lab coat.

EDO SEGAL: That's the sharpest the seam has been, so let me hold it open one more beat with a ghost, because Dreyfus is at this table and deserves a chair. Hannah claims him; Daniela, you should want him too — he spent his life attacking symbolic AI, the rules-and-representations school, which is exactly the school your embodied, learned, biologically-inspired machines overthrew. So whose ancestor is he? Each of you, briefly — make your claim.

RUS: He's mine in the engineering and hers in the ethics, and I'm at peace with the split. Dreyfus said expertise isn't rule-following — that the master can't articulate his mastery because it was never propositional, it's pattern acquired through embodied experience, resident in something more like a body than a logic engine. That is literally what a neural network is, and what a soft robot that thinks with its body is: competence without articulable rules, skill smeared across weights and materials. We vindicated his attack on the symbol-pushers. He's my ancestor in the lab.

ARENDT: And mine where it matters most, because Daniela kept the half of Dreyfus that her machines can do and quietly dropped the half they can't. Dreyfus did not only say "no rules." He said situated — in a world that pushes back, with stakes, a body whose skillful coping is bound up with caring how it comes out. The caring, Daniela. The being-at-stake. Your robot copes without caring; the world pushes back and it adjusts, but nothing is at issue for it, there is no one for whom the outcome matters. Dreyfus would look at your beautiful fish and say what he said about every generation's AI — that we have mistaken the articulable shadow of intelligence for the thing — and then he would note, with that Berkeley relish, that this time we built the shadow so well it swims. He is my ancestor. You inherited his method and buried his point.

EDO SEGAL: Hold that thread — the caring, the being-at-stake — it returns at the candle. But the next round goes to the deepest difference Hannah has named all night, the one she'd stake everything on, and the one Daniela's whole field is built to challenge. The machine continues the past. The human begins. Natality, and the probable continuation. After this.