I’ve been busy. Time to sit back for a second and take a quick look.
Photos
It looks like I put my camera(s) down from August through the end of November. That’s four months that I’ve been hibernating in a funk when damned near everything shut down, not just camera. Just look at the number of blog posts over there at the right. You can see my mood in the numbers, something I wrote about back in 2018, A Mind Over the Long Haul: My Posting Patterns.
Here's my first shot upon coming back:
Chatting with ChatGPT
It’s not photography that brought me back, though, it’s ChatGPT, the newest and baddest chitter-chatter from our friends in Silicon Valley. I’m exhausted from all the playing, thinking and working I’ve been doing with ChatGPT. My document of Chat interactions has almost 50K words over 147 pages.
There are things you learn that you add to your current repertoire, little things, big things. There’s a bit of that here. But there’s something much deeper. Some things lead you to reorganize some, a lot, of what you already know. That’s harder work, burns more energy, touches more cells. That’s what I’ve been doing.
There’s Benzon’s second law:
If you want to know what it’s like to drive a car, you have to sit in the driver’s seat. Going along for the ride isn’t the same.
Thus, while the effect of working with ChatGPT has been deep, it hasn’t much affected my fundamental ideas about computing & AI. It doesn’t work that way. We’re not going to get from here to there – whatever wherever there is – simply by scaling up our current technology. We do need to go back and recoup some stuff from the classical era of symbolic AI and computation linguistics. And we need a more thoroughly embodied intelligence that moves about the world easily and fluently.
Intelligence without consciousness or intention?
Moreover, you know, the classical arguments against AI, the arguments by Dreyfus and Searle (Chinese Room), remain valid. But the context has changed radically. Back then AI technology couldn’t do anything really impressive. Oh sure, it was impressive to those who built it because they know how difficult it is. And there is that Eliza Effect, which involves humans attributing intentionality where it really isn’t there. But what the tech could actually do, wasn’t impressive. The ARPA Speech Understanding project of the mid-1970s was a massive and impressive project spread across three or four institutions and a half dozen sites for three years. But the end result was a system that couldn’t do much, certainly much less than the smart phones since forever.
Where I’m going is that those classical arguments, the arguments from intention, got most of their rhetorical force, as opposed to the cognitive force, from the impoverished abilities of that technology. The current technology does not diminish the cognitive force of those intentional arguments, but the capacities of current tech ultimately weakens the rhetorical force of those arguments. We’re on the threshold of something that might as well be called intelligence without intention and consciousness. That’s a very strange beast.
Where does that put us? It’s been clear to me for some time that this technology, based on deep learning, is unlike anything we’ve seen before. We’re in “here be dragons” territory. ChatGPT is NOT a conscious intentional (in the philosophical sense) actor. It is a very complex inanimate contraption of some sort. But what sort?
I am provisionally thinking of them as devices that exhibit intelligence (whatever that is) without having either consciousness or intentionality – though perhaps one cannot have intentionality without consciousness, certainly not consciousness without intentionality. If that seems weird, well then, it is. And I’m comfortable with it, provisionally.
Back to the Future
All this GPT chatting has gotten me thinking about the future once again, about the Fourth Arena: first there was inanimate matter, then life, after that we have human culture, which circles the glob though the brains of very clever hairless apes. We’re now stepping on to the Fourth Arena, just what that will turn out to be like, that remains to be seen. I’m thinking of it as a community in which these new creatures will be interacting with us, through desktop, online, and handheld systems, as they currently, but also through robots. They’re also be interacting among themselves.
Of course, current systems are already doing this kind of thing. But we’re headed to something new, something strange, something exciting, but also dangerous. Once more I’m thinking about Neal Stephenson’s The Diamond Age: Or, A Young Lady's Illustrated Primer as a loose and suggestive framework in which to situate my thinking about the future.
A Methodology for describing and analyzing interactions with ChatGPT
When I started playing with ChatGPT at the beginning of December I was doing just that. But as I got deeper and deeper into it, I began a more systematic style of prompting. I was looking for things. I began to see patterns in what ChatGPT was doing.
The upshot is that I now see the possibility of constructing a competence grammar (to borrow from Chomsky) of ChatGPT’s discourse. I don’t know what this grammar will look like, but Joseph Becker’s phrasal lexicon (from the ancient days, the mid-1970s) is very suggestive. From the abstract of his 1975 paper:
This paper will focus on the contrary aspect of language, wherein utterances are formed by repetition, modification, and concatenation of previously-known phrases consisting of more than one word. I suspect that we speak mostly by stitching together swatches of text that we have heard before; productive processes have the secondary role of adapting the old phrases to the new situation. (For more, see my posts, The phrasal lexicon [System 1 vs System 2 for language] and GPT-3, the phrasal lexicon, Parry/Lord, and the Homeric epics.)
That’s the stuff. Whereas he was looking at relative short strings, I want to open it up to include longer strings, 10, 20, 30 words, maybe more. That’s what ChatGPT is working with at the discourse level.
My next step is to get people in the industry to recognize the usefulness of this work: 1) it will point the way to understanding what those (otherwise) opaque artificial neural networks are doing, and 2) help in creating more satisfactory benchmarks to evaluate the capabilities of these systems. To that end I’ve begun thinking about an explicit methodology and setting up a workshop to develop and teach that methodology.
No comments:
Post a Comment