Two recent lines of work have gotten me thinking, my reconsideration of John Horgan’s The End of Science and my dialog with GPT-3 about a Jerry Seinfeld bit. Let’s start with the end of science and work our way to Seinfeld.
Horgan set himself a difficult task in arguing for the end of science, for he is, in effect, predicting the future. And that’s difficult. When I wrote about his argument 25 years ago I took him to be making a philosophical claim, though I didn’t think about it in those terms, and countered with a claim of my own. Twenty-five years have passed and Horgan is still claiming that science is in trouble and I’m still agreeing with him, but also arguing that there’s a difference between being in trouble and being at an end.
I’ve done a lot of work in the interim and that’s what I’ve been thinking about. There’s my book on music, Beethoven’s Anvil, my return to neuroscience through Walter Freeman’s neurodynamics, a great deal of descriptive work in literature and film, along with my methodological and quasi-theoretical piece on literary form (Literary Morphology: Nine Propositions in a Naturalist Theory of Form), my excursions into computational criticism, and my encounter with Latour, culminating with Living with Abundance in a Pluralist Cosmos: Some Metaphysical Sketches, which sure looks like an outline of a philosophical system.
And that’s what I’ve been thinking about: Am I sitting on the scattered pieces of a philosophical system? If so, what would it take to pull it together? In a way it begins with the little piece David Hays an I wrote on complexity and natural selection, A Note on Why Natural Selection Leads to Complexity (1990), for that provides a justification for why, over the long course, living systems become more complex, cultural systems to. So that links our earlier paper on the brain (Principles and Development of Natural Intelligence, 1988), in which we located functional principles at major junctures in vertebrate phylogeny, with our more recent work on cultural evolution (e.g. The Evolution of Cognition, 1990). What would it take to spell out those connections and then to link the whole contraption to my more work on pluralism?
I note that, while Hays and I talked about abundance, under the rubric of fecundity, in private conversation, we hadn’t published about it at the time he’d died. But I introduced it in my original treatment of Horgan’s book, Pursued by Knowledge in a Fecund Universe (1997), where I also introduced the idea of implementation, taken from computing. The general idea is that the routines of a higher realm (of biological behavior, of cultural practices) are implemented in the mechanisms of the pre-existing and lower realms. The idea is that evolution in complexity proceeds by elaborating variations and extensions of existing mechanism until, somehow, a phase shift happens so as to reorganize the system in a new way. In that reorganized system new mechanisms emerge to govern older systems, using them as vehicles to implement its new goals and behaviors. My immediate objective is not, however, to explain how such phase changes occur – I’ll happily leave that to others, others with technical skills I lack – but only to (provisionally) recast our work on the principles of natural intelligence and those of cognitive evolution in culture in the same terms. And then to take that and see how it resonates with the Realms of Abundance notion from the pluralism paper.
And that brings me to my recent interaction with GPT-3 over a Seinfeld bit. I don’t myself have access to GPT-3, but I know someone who does, Phil Mohun. And so one day last week we spent 15 or 20 minutes chatting back and forth on Slack while Phil quizzed GPT-3. It was a lot of fun. When you read the post you’ll see that it took awhile to get a sense of how GPT-3 ‘understood’ the joke’s punch line.
We need to free these systems from our narcissistic investment in them.
THAT’s what’s interesting. I don’t really care whether GPT-3 understands language in the way you and I do – it doesn’t – nor whether or not it’s one Big Giant Step toward AGI – I don’t find the idea of AGI (artificial general intelligence) very interesting, or even coherent. But it does seem to me the GPT-3 is ¬one Big Giant Step toward SOMETHING. But just what that is, I don’t know, nor do I think anyone else does either.
At the moment I don’t even believe we’ve got the terms and concepts we need to think about that. If GPT-3 doesn’t understand Seinfeld’s bit in some way, if it isn’t thinking about its knowledge when responding to the questions Phil and I put to it, then what IS it doing? What words do we use? “Process” seems too generic. It’s a cop-out, useless. If you will, we need a phenomenology of machine intelligence that's different from our phenomenology of ourselves.
As long as we insist on anthropomorphizing these systems we're deceiving ourselves about them. They really are radically new and not merely halting attempts to re-create ourselves in machine form. We need to free these systems from our narcissistic investment in them.
I keep thinking about a comment Yann LeCun made about end-to-end systems (GPT-3 is such a system) in an article by Kenneth Church and Mark Liberman: “If you can train the entire thing end to end—that means the system learns its own features. You don’t have to engineer the features anymore, you know, they just emerge from the learning process.” I believe that’s analogous to the argument that Sydney Lamb has made about meaning in networks, that the meaning of a node is a function of its position in the network (I discuss this in, for example, a recent post, Minds are built from the inside [evolution, development]).
I’m tempted to say that the meaning of a node in a relational network is a function of its relationship to both the input and the output side of the network. What does the mean in the case of GPT-3? Does it make any sense to say that its internal structures ‘gain’ meaning in the process of interacting with humans? It seems to me that such systems do have something in common with the human brain, something deep and fundamental, that they ‘construct’ themselves. That’s something we must understand, and we’re not going to create that understanding as long as we’re worried about whether or not it’s really got a human-like mind. It doesn’t. But it has something.
In my working paper, GPT-3: Waterloo or Rubicon? Here be Dragons, Working Paper, Version 2, I talk about why GPT-3 is always going to have trouble with so-called common-sense reasoning. Someone with mathematical competence needs to take that argument and develop it. Common-sense knowledge is ‘close’ to our physical interaction with the world. Ultimately it depends on that knowledge. But GPT-3, and similar systems, are trained on text only, so they simply don’t have access to the sensorimotor foundation on which common-sense reasoning is based. That places severe limitations on what they can do, making flexible and robust common sense reasoning always just beyond reach.
How do you make up that deficit? Sure, you can train a system on images and sounds, touches, and smells, but it also has to move about in the world. You can see where this is going, can’t you? We have to create an android that starts out the capabilities of human infant and then grows, matures, and learns. How do we create an android that grows like a human? At the moment I’m thinking that if we want to create an artificial creature with all the capabilities of a human being, we’re going to have to create an artificial human being. We haven’t got the foggiest idea of how to do that. But what’s the point?