Interpolation or Origination

EDO SEGAL: Ada, in 1843, looking at a machine of cogwheels that had never computed a single sum, you wrote a sentence that reads less like Victorian speculation than like a caption for a product that shipped last year. You wrote that if the relations of harmony and composition could be expressed in the engine's formal language, then the engine "might compose elaborate and scientific pieces of music of any degree of complexity or extent." It does now. I have heard it. The machine produces music in any style, on demand, that moves people to tears who do not know a machine made it. You predicted that. So I have to ask the cruel version of the question: did you predict the machine's creativity — or did you predict your own objection's defeat?

LOVELACE: I predicted neither, and the precision matters, so let me hold the two halves apart the way I held them in 1843, because everyone since has insisted I must have meant one or the other. Read the prophecy again and notice its grammar. It is conditional. If the relations of harmony can be expressed formally, then the engine can compose. I did not assert that music is creation that a machine performs. I identified the exact condition under which music becomes operation — and operation is precisely what the engine does. To compose by formal rule, within a formalized harmony, is to execute a procedure. The engine composing music is not the engine creating. It is the engine doing the only thing it ever does — following rules — applied to a domain we had not yet thought to formalize. So there is no contradiction in me, and there is no defeat. I predicted machine-made music and I denied machine creativity in the same breath, because I understood that producing music and originating music are different acts, and the engine can do the first without the second. Your weeping listeners are weeping at the output. The output is real. The origination is not in the loom.

EDO SEGAL: So the tears are real and they are evidence of nothing about the machine.

LOVELACE: The tears are real and they are evidence about the listener. They are the listener supplying a someone behind the pattern, exactly as your "I felt met" supplied a someone behind your clarification. The machine has the relations of harmony. It executes within them flawlessly. What it does not have — what no one has shown it to have — is the thing I reserved: the wanting of the music, the meaning of it, the deciding that this music rather than that was worth making and why. It can compose the symphony. It cannot want the symphony to exist. And every domain you successfully formalize and hand to the engine, you discover the same thing: the part you were proud of was, more than you flattered yourself, operation.

MITCHELL: Now I have to come in hard, because this is the round where I think the Countess's framework is both most useful and most quietly stacked. She has set up two boxes — operation, which the machine does, and origination, which only the human does — and she has defined them so that the human always keeps the good one. I want to examine the actual mechanism instead, because the field has a real debate here and it is more interesting than her two boxes. Set the positions out fairly. The interpolation view says the model is a vast statistical map of its training data and everything it produces is a point somewhere inside the space that data defines — a blend, a recombination, never a reach outside the distribution. That is the Countess's "surfacing" in modern dress, and serious people who know the mathematics hold it. The origination view says this proves too much, because human creativity is also recombination — every artist trained on a corpus of everything they ever saw and heard, every new work a recombination of that input. If recombining disqualifies the machine, it disqualifies us.

And here is the data I actually have, because I refuse to argue this from the armchair. We can study whether these systems reach genuinely novel combinations or only interpolate familiar ones. The answer, so far, is mixed in an illuminating way: they reach combinations no human had assembled — novel protein structures, game strategies no master had played, images of things that never were — and they fail at the kind of novel abstraction a person finds natural when surface and depth disagree, which is the conceptual slippage Copycat was built to model. So they are not pure interpolators — a pure interpolator could not produce the protein. And they are not originators in the Countess's robust sense — an originator would not collapse on the slippage. They are something the two-box theory has no box for. That is my entire objection to her objection. Reality is in the gap between her boxes, and she built no gap.

LOVELACE: Then let me defend the gap by occupying it, because I think I built it and you have not noticed. When I said the engine could weave patterns "no weaver imagined" — patterns of a complexity and extent beyond what its designers foresaw — I was describing your protein. A combination new in the world, reached by the machine, that no human had assembled. I never denied that. It is the whole wonder I insisted on against the people who called the engine a mere calculator. The protein is the engine surfacing a combination latent in the structure of the possible, that the search found and no human found. New in the world. Originated by no one. And the slippage your machine fails — the reconception, the seeing-past-the-surface — is exactly anticipation, the one thing I said the engine could not do. So your data, Doctor Mitchell, does not embarrass my two boxes. It populates them precisely. Novelty of output: the machine has it, in abundance, the protein proves it. Origination of idea: the machine lacks it, the slippage failure proves it. You have not found a thing between my boxes. You have found my two boxes, confirmed by experiment, and given them new names.

MITCHELL: That is a genuinely strong reply and I am going to concede the structure and contest the stability. Yes — "novelty of output without origination of idea" is a clean cut, and yes, the protein and the slippage-failure fall on the sides you say. Here is where I will not follow you. You treat "anticipation" — the slippage, the reconception — as a bright line the machine is on the wrong side of forever. I treat it as a capability with a trend. Every year the slippage failures get rarer and subtler. The line you call permanent, I have watched move — not all the way, not to origination, but move. And I have a professional allergy to permanent lines about machines, because I watched an entire school of brilliant people — the symbolic AI people — defend a permanent line, "machines can't learn concepts without explicit rules," and lose it completely. So I will grant you the boxes for tonight. I will not grant you that the wall between them is load-bearing for all time, because the only evidence for that is that it has held so far, and "it has held so far" is precisely what every defeated permanent line had going for it right up until the morning it didn't.

EDO SEGAL: Mark this, because it is a real convergence and agreements are news. You have both now agreed — Ada by construction, Melanie by data — that the machine produces genuine novelty of output and lacks origination of idea in the robust sense, as of tonight. That is enormous. The student arguing about this on the internet thinks one of you says "it's all just statistics, nothing new" and the other says "it's creative, it's alive." Neither of you says either. You agree on what the machine does. Your entire disagreement is one word — forever — and whether the wall between novelty and origination is a fact about the machine or a fact about how long we have been looking. Hold that single word. The next round goes to the image you gave the whole field, Ada — the loom — and to whether weaving without comprehending is the machine's permanent nature or its current address.