Opening Positions
LOVELACE: Thank you. I want to begin with precision, because the sentence everyone quotes from me is almost always quoted to mean something smaller and stupider than I meant. People say: Lovelace thought the machine could never do anything new. That is not what I wrote, and the difference is the whole evening.
What I wrote is that the Analytical Engine "has no pretensions whatever to originate anything. It can do whatever we know how to order it to perform. It can follow analysis; but it has no power of anticipating any analytical relations or truths." Read it again and notice what it does not say. It does not say the engine is feeble, or limited in scope, or incapable of astonishing us. The surrounding Notes are a sustained argument that its powers are vast — that it could, I wrote, compose elaborate and scientific pieces of music, that it might act upon any objects whose relations could be expressed in the abstract science of operations. I was not belittling the machine. I was drawing a boundary, and the boundary is exact: the engine can follow analysis. It cannot anticipate it. It elaborates consequences already implicit in what we hand it. It does not reach past the given and discover what was not, in principle, already there.
And here is the crucial phrase, the one I want hung over this table all night: "whatever we know how to order it to perform." The limit I drew is a limit relative to human knowledge. The engine's reach is exactly coextensive with what we already know how to specify. If we can write the procedure, the engine runs it. If we cannot write the procedure, the engine has nothing to run. It cannot generate a procedure we do not, in principle, already possess.
Now. I have been told what became of my loom, and I will not pretend the telling left me unmoved. You did not hand the modern engine its procedures. You handed it examples, and it found the procedure itself — and that, I grant immediately, is a thing my engine could not do and a thing I did not imagine. So let me state my position not as a museum piece but as a live wager, refined by what I have learned in the interval. When your network finds a structure in fifteen million images that no one wrote down, where did the structure come from? Not from the network's wish to find it, for it has none. It came from the images, which came from the world, which is to say it came from you — from the humans whose camera caught the cat, whose hand wrote the label, whose collective ordering of the world is what the data is. The procedure was not in the engineer's head. Granted. But it was in the data, and the data is congealed human ordering. The engine still originates nothing. It has merely been ordered by a vastly larger and more diffuse author than a single programmer: it has been ordered by all of us at once, and we have lost the ability to see ourselves in the order because the order is so large. That is my opening. The author did not vanish. The author got too big to see.
EDO SEGAL: Fei-Fei.
LI: That was beautiful, and I am going to disagree with most of it, and I want to start by conceding the part she has right, because it is more than my own field usually admits. Ada is correct that the data is human. ImageNet was labeled by people — tens of thousands of them, image by image, through crowdsourced work, and the categories came from a structure of human language and human attention. The world that flowed into the machine flowed through us first. I have never disputed that, and anyone who tells you the machine learned from nothing is selling you something.
Here is where the wall comes down. Ada says the structure was "in the data," and therefore the human is still the author. But that move proves too much, because by that logic nothing is ever discovered — not by machines and not by people. The structure of the prime numbers is "in" the integers; was it therefore not discovered? The shape of natural selection is "in" the living world; did Darwin originate nothing because the pattern was already lying there? The whole of science is the extraction of structure that was, in some trivial sense, already present in the data of the world. If "it was implicit in the input" disqualifies the engine from originating, it disqualifies us too. And I do not think Ada wants to say that we originate nothing. So the interesting question is not whether the structure was present. It is who, or what, did the work of finding it.
And that is the part I lived. In the mid-2000s, computer vision was stuck — decades of ever-cleverer algorithms, all of them failing the same way the moment they met the messy real world. The consensus was that the models were not smart enough. I came to think the field had it backwards. A child sees a single cat and forever after recognizes cats — fat ones, cats in shadow, cartoon cats — and not because the child has a better algorithm. Because the child has been flooded since birth with a torrent of visual data, labeled implicitly by the world itself. The advantage was never the model. It was the data. So I made a bet most of my colleagues thought was deranged: build a dataset that contains everything, tens of thousands of categories, millions of images, and feed a relatively ordinary network on it. And what happened in 2012, when AlexNet ran on ImageNet, was not that vision improved. It was that the machine assembled, inside itself, a hierarchy nobody designed — edges into textures into parts into objects — a way of seeing that no engineer wrote and that, crucially, I could not have written, because I do not have conscious access to how my own visual cortex carves the world. The machine did not execute my order at a depth I foresaw. It found something in the dark that I did not know was findable. You can call the finder a slave to its data if you like. But the finding was real, and it was not mine.
EDO SEGAL: Before I frame the rounds, one discipline I impose on every long conversation, because it pays off two hours later. Each of you, in a few sentences — what do you envy in the other's position? Not respect. Envy. The thing her side gets to have that yours does not. Ada first.
LOVELACE: I envy the evidence. I made my claim in an armchair, about a machine of brass that existed only in Babbage's drawings and my imagination — I reasoned from the nature of the thing. Fei-Fei got to build it, run it, watch it surprise her, measure the surprise. My poetical science was a telescope pointed at a possibility. Hers is a laboratory with the possibility breathing on the bench. There are nights I would have traded my certainty for one afternoon of her data.
LI: And I envy the line. Ada gets to stand somewhere. She has a floor under her — there is a human author, meaning lives on our side of the glass, the machine elaborates and we originate — and from that floor she can push. My position has no floor, because I follow the mechanism wherever it goes, and the mechanism keeps dissolving the very distinctions I would like to keep. Every year the gap between "finding" and "originating" gets harder for me to locate, and I am the one who is supposed to know where it is. Ada gets to defend human authorship. I am stuck watching it get complicated, and the watching is lonelier than the defending.
LOVELACE: That may be the most honest thing either of us says tonight.
EDO SEGAL: Two openings, two envies, and already you can see the architecture of the evening. It is not that one of you loves the machine and one fears it. It is that you locate the author in different places. Ada says the author got too large to see but never left. Fei-Fei says the finder is something new in the loop, and that calling it our reflection is a way of not looking at it. Hold both. We start the rounds at the exact seam — at the sentence itself, the wall in Note G, and whether the modern engine has crossed it.