The Steelman Hour

EDO SEGAL: The rules, because I am told this segment is why people listen to the end: each of you now argues the other's position, at full strength, for as long as it takes — not a caricature, not a confession extracted under torture, but the case as you would make it if it were yours and your reputation rode on it. The test of whether you've understood an opponent is whether you can state her position so well she'd sign it.

And before either of you begins, let me say why I believe in this exercise enough to spend a tenth of our evening on it, because some listeners find it theatrical. Twice in my life, the position I now hold arrived while I was arguing against it. Not after — during. There is something about inhabiting an argument — load-bearing, first person, your own mouth making its moves — that reading it never produces. The brain defends a position differently than it audits one. So this is not theater. It is the cheapest instrument ever invented for finding out whether your opponent's house has rooms your map doesn't show — and whether your own has load-bearing walls you've never had to touch because no guest was ever rude enough to lean on them. Margaret Boden: you are now the author of the octopus paper. Convince me the machine cannot mean.

BODEN: Very well. I shall do it properly — which means I shall not argue substrate, and I shall not argue incapacity. I argue signal. Meaning is a relation between an utterance and the world, routed through an agent's purposes. Examine the training signal of these systems with a scientist's eye: at no point — none — does information about the world enter, except as already filtered through human descriptions of it. The system is, with mathematical exactness, a model of descriptions. Now, the seductive error my own side commits: we observe that the model's outputs track the world astonishingly well, and we infer contact. But the tracking is exactly what one predicts from a sufficiently good model of descriptions made by creatures in contact — the correlation is inherited, every gram of it, and an inherited correlation snaps precisely where the octopus's bear arrives: outside the support of the inherited descriptions. And the deepest version, which I'd write on the blackboard if I were teaching the position: when the model is wrong, nothing happens to it. No thirst, no fall, no grief. Error, everywhere life means anything, is expensive — meaning is, at bottom, the bookkeeping of a creature that can lose. This system cannot lose. Therefore its tokens are moves in a game with no stakes, which is to say: form. The fluency is real, the utility is real, and the meaning is — rigorously, demonstrably — on loan from the people in the corpus and the person at the keyboard. There. Emily, would you sign it?

BENDER: I'd sign it and assign it — the "bookkeeping of a creature that can lose" is better than my own formulations, and I'll be stealing it with citation. You even reproduced the part my critics always drop: that the position is about the training signal, not the substrate — no carbon chauvinism anywhere in it. Which is itself unsettling, Margaret, because you built my cathedral in four minutes and you don't worship in it. All right. My turn, and I'll match your discipline. I am now Margaret Boden, and I argue that the machine can be genuinely creative. I do not argue from output — from the dazzle of the lighthouse sentence; that way lies the coached witness. I argue from decomposition, which has always been her method and is the only honest one. Creativity is not a mystery substance; it is a structured capacity: generate candidates under constraints; evaluate them against multidimensional criteria; model the constraint-space itself; operate on that model. Every one of those components, taken alone, demonstrably exists in machines today — the only question is integration, and "integration is impossible" has been falsified so many times in the history of this field that asserting it should require a permit. The deflationist must therefore claim there is a component missing from the list — and watch what happens when she's asked to name it: she says meaning, the system has no stake. But then she must explain why stakes are required for the generative mechanism rather than for the social crowning of its products — why the candidate-maker must care, when we already know the caring can live in the partnership, the laboratory, the practice around the machine. Pressed hard, her missing component keeps dissolving into the community's half of the process — which the mechanist never claimed for the machine anyway. The deflationist, in other words, wins every battle about the H and has never once engaged the P. The mechanism is assembling, component by published component, and no one has named the wall. That is the case. And... hm.

EDO SEGAL: Say the hm, Emily. The hm is why this segment exists.

BENDER: The hm is that I felt the case from inside while making it, and from inside, my "stakes" objection has a precise location and it is not where I usually aim it. It's an objection about evaluation — about the crowning — and Margaret has been telling me that for three hours. If I'm honest, the seam between us is thinner than my public writing suggests: I think the evaluative half constitutes creativity and she thinks it completes it. That's... a real disagreement, but it's a philosopher's inch, not the Grand Canyon my hashtag implies.

BODEN: And my hm, offered unprompted in the same currency: arguing Emily's case, I could not find the join where "inherited correlation" becomes false of children — who also begin as models of descriptions made by creatures in contact. The difference I reached for at the last step was the body, the cost, the losing. Which means the load-bearing wall of my opponent's house... is the one wall my own theory never needed to address, because creativity theory always assumed a creature. The assumption is now load-bearing, and I do not know what happens to the three doors in a system that cannot lose. There. We have each found the soft floor in our own house, which I believe was your design all along, Edo.

EDO SEGAL: It's the only design I have. The final hour, then: two people who have found each other's load-bearing walls, and their own soft floors. Let's see what's actually left standing between you.