The Symbol and the Vector
**EDO SEGAL:** Leibniz, you wanted a character for every concept — a discrete, readable, meaningful atom of thought. "Dog" would be a symbol, built from simpler symbols, the way thirteen is built from primes. Geoff's machine stores "dog" as a point in a space of thousands of dimensions, a vector of real numbers that mean nothing in isolation, whose only meaning is its *position* relative to every other word. Discrete and readable against continuous and opaque. I want you each to make the case that your representation is the one that's actually about something. Leibniz, why is the symbol the home of meaning?
**LEIBNIZ:** Because a symbol can be *given a reason*. This is the heart of it, and it returns us to the principle I built my metaphysics upon — that nothing is true without a sufficient reason why it is so and not otherwise. A symbolic representation wears its reasons on its face. The concept is composed of simpler concepts by stated rules; the inference proceeds by stated steps; and at every point one may ask *why* and receive an answer that can be examined, contested, and corrected. When my logic engine concludes a thing, it can show its work. Every step is stateable, inspectable, accountable. That is not a convenience. It is the condition of the thing being a *reason* at all, rather than a mere happening.
Now consider his vector. When his machine denies a man a loan, or ranks a woman's application below the line, and the man asks *why* — what is the answer? There is a *cause*: a flow of weights determined the outcome. There is no *reason*: no intelligible ground that can be stated, examined, judged. I drew this distinction with great care in life, and it has waited three hundred years to become a matter of justice. A cause is a brute antecedent that brings a thing about. A reason is a ground that makes a thing *understandable* — that lets you see why it had to be so, and therefore lets you argue that it should *not* have been. His machine gives causes without reasons. It produces outcomes that are caused but never justified. The symbol is the home of meaning because the symbol can answer *why*. The vector only ever answers *because the weights*.
**HINTON:** This is Leibniz's strongest card and I want to honor it before I take it apart, because there's something true in it. Yes — these systems are hard to interpret. The [black box problem](https://www.youonai.ai/fieldguide/med/interpretability_problem) is real, I've spent years on it, and Leibniz is right that a decision you can't explain is a decision you can't contest, and that matters enormously for justice. I concede the whole practical point. We *should* build the tools to look inside, and we are, and it's hard.
But here's what the argument can't survive. Leibniz says the symbol "wears its reasons on its face" and the vector doesn't. Watch what happens when you turn that lens on a *person*. You make a decision — you hire her, you trust him, you feel the proof is right — and I ask you why. And you give me reasons. And the psychology of the last fifty years says, over and over, that the reasons you give are very often *not* the cause. They're a story you tell *after*, a rationalization assembled to make your conduct legible, while the real cause was a pattern in your neurons you had no access to. The symbol on the face — your stated reason — is frequently a confabulation bolted onto a vector underneath. So the clean distinction collapses: humans are *also* causes wearing the costume of reasons. Your network confabulates; so does mine. The difference you want isn't between reason and cause. It's between a system whose story can be tested and revised and one whose story can't — and that's an engineering problem about *access*, not a metaphysical fact about symbols.
**LEIBNIZ:** No, sir — and here you give too much away to save your machine. You say the human reason is "often" a confabulation. Often is not always, and the cases where it is *not* a confabulation are the cases that hold the world together. When the judge gives his reason, I may appeal — I may point to the principle he misapplied and *watch him change his mind in the light of the reason given*. The reason does *work*. It is load-bearing. It is the thing argued with, and the argument moves it. Your machine's explanation, when you can extract one at all, is a description bolted on after the fact that changes nothing in the process that produced the output. That is the whole difference, and it is not small: a human's reason is *part of the mechanism of deciding* and can be corrected by being shown wrong; the machine's account is a label, and the label argues with no one. You have shown that humans *sometimes* fall to the machine's level. You have not shown that the machine *ever* rises to ours.
**HINTON:** Except it does, every day, and I've watched it. You correct one of these systems — you say, no, your reasoning at step four is wrong, the premise doesn't hold — and it revises. It says, you're right, and it reworks the conclusion in the light of what you showed it. That's the reason doing work, Leibniz. That's the account being load-bearing and being corrected by being shown wrong. It's the exact thing you just said only humans do.
**LEIBNIZ:** It revises because you have given it a new *input*, not because it has grasped a *reason*. There is a difference between a thing that changes its output when you change its prompt and a thing that *sees that it was mistaken*. I grant the behavior is nearly identical. I deny that identity of behavior is identity of kind. You will say I cannot prove the difference from the outside — and you would be right, and I shall return to it tonight, for it is the seam of everything. But do not tell me the loan-machine *gave a reason* merely because it produced a different number when prodded.
**EDO SEGAL:** I want to pull one thread out of this before we move, because it's been under the table since the basement. Geoff, the whole field for a while has been circling back toward Leibniz — [neuro-symbolic systems](https://www.youonai.ai/fieldguide/med/symbolic_ai), the marriage of the fluent net to a logical scaffold, the vector wedded to the symbol so the machine can not only answer but *justify*. Doesn't that frontier concede that Leibniz had something the vectors lost — that you need the symbol back to get the reason?
**HINTON:** It concedes that symbols are a *useful tool the network can learn to use* — like a person using a notebook, or a logician using a formal proof to check the intuition that got there first. I don't deny that's powerful. What I deny is the inversion Leibniz wants: that the symbol is the *home* of the thought and the net is the servant. It's the other way around. The intuition comes first, in the pattern, and the symbol is the net's external scratchpad for checking and sharing what the pattern already grasped. We bolt the alphabet on at the end, to communicate and to verify. The thinking happened before, in the weights, with no alphabet in it. Leibniz built a magnificent scratchpad and mistook it for the mind.
**LEIBNIZ:** And I say you built a magnificent intuition and mistook the absence of a scratchpad for the presence of a soul. We have, at least, located our disagreement with precision. He believes the symbol is downstream of the thought. I believe the thought, properly so called, *is* the symbol made lawful. The notebook, or the mind that needs none.
**EDO SEGAL:** Two men, one fork, and notice that neither flinched. Hold that — it comes back when we ask whether the machine could ever owe you a reason. But the next round goes to the place where Geoff's whole worldview was forged, the thing he says these systems do that yours never could. Not symbols. Not vectors. *Learning.* What it means that the machine wasn't told — it was shown. After the break.