I’m stuck on three substantial projects. Two of them, ChatGPT and Great Literary Critics were supposed to be finished sometime in January. Both were mentioned in my previous Ramble. The third, on Chess, Language, and AI, came out of nowhere earlier this month. Let’s take a look.
Report on a year of work on ChatGPT
The idea was simply to write an introduction to the investigations I conducted on ChatGPT during 2023. For the most part I’ve been thinking of it as an introduction to the dozen working papers I’ve written during that time. I’ve done a good deal of writing since earlier this month, making good progress on the specific topics I mentioned in my previous Ramble (conceptual ontology, story variations, cued recall and associative memory) but there are still things to do. I’ve also written about meaning.
I still need to discuss confabulation and truth, and to say something about alignment. There I want to argue that some kind on alignment at the level of conceptual ontology is all but built-in to the process, though not in a way we have control over. But alignment with human values, that’s a different can of worms entirely. And that’s what it is, a real can of worms.
Perhaps I want to draw on the work I’ve been doing on chess and language by pointing out that, in chess programs using artificial neural networks (such as AlphaZero) the objective function used in training IS success in a game. That is to say, the objective function in training is the same as the goal in playing chess. That is not at all true for language. The objective function used in training a transformer on a text corpus is next-token prediction (or some variant thereof), which is not the goal of any speech act. One speaks or writes to achieve some end, such as conveying information, establishing a contract or agreement of some sort, creating aesthetic pleasure, and so forth. Those are quite different from simply producing token after token. One achieves such goals by producing a string of tokens, but the goals themselves are not defined in terms of token count. They’re defined in terms of interaction with others.
This seems obvious enough, but I haven’t seen it pointed out explicitly. I need to do so. In the process I may also want to discuss what Anthropic calls “constitutional AI.”
And then there’s implications and next steps. Perhaps I need to begin that by explaining what I mean when I say that LLMs are digital wilderness, albeit a digital wilderness that embodies an approximation to what I’ve been calling the metaphysical structure of the cosmos.
Yikes! That’s a lot. Do I really need to do that? Probably. No one else will. Once I’ve got that established, then I can outline what needs to be done, perhaps framing it as a strategy for exploring the digital wilderness.
Harold Bloom and Great Literary Critics
I figure I’ve got two more posts to go to complete the project. The next post is on Harold Bloom and the final one is some speculation about the future. Given that the ostensible purpose of the series is to come up with candidates for great literary critics, I suppose that speculation about the future is out of bounds. For I certainly don’t know who the great critics of the future will be. But, without saying something about future possibilities I can’t really justify my thoughts about Bloom.
Basically, I want to explain this diagram:
I can’t do that without saying something about the future.
Frye is the beginning of a forward movement toward the objective study of literary phenomena. Indeed, will I go so far as to argue that he’s pointing toward what I have called “naturalist” criticism? That may not have been what he intended – I probably was not – but that’s where things have to do to become the objective discipline he’d imagined. Bloom, on the other hand, really looks toward the past – indeed, his character-based criticism of Shakespeare does just that. He speaks to something people want from commentary on literary works, but it may not be naturalist commentary or commentary from an academic discipline.
What kind of commentary? More Kenneth Burke, more in the line of Booth’s ethical criticism. That may require a different institutional home, one not yet invented, but perhaps nascent on the web.
We’ll see.
Chess, Language and AI
I’ve been thinking about chess and AI for a while, and certainly about language. Chess has been with AI since the beginning, but not language. While computational linguistics got started at roughly the same time, as machine translation, it was quite separate from AI until the late 1960s and early 1970s. But they are very different domains.
Chess is a closed formal domain, albeit one that generates a very large game space. It’s the size of that game space that makes the game intellectually challenging (for humans) and interesting for AI researchers. While Chomsky has treated syntax as a closed formal domain, albeit one capable of generating an infinite number of sentences, once one includes semantics within the intellectual scope, the domain ceases to be closed. The domain is open and not sharply defined.
I’ve got a number of pieces I need to write to finish out this project.
- A discussion of the language domain comparable to my discussion of chess.
- A comparison of the two domains, language and chess.
- A call for a general treatment of the relationship between conceptual domains (perhaps in terms of their geometric footprints) and the computational means required to address them. This would be a generalization from Yevick’s 1975 paper.
- Something more generally about the search for interesting ideas along the lines of my discussion of Romer’s 1992 toy factory model.
I don’t know off-hand how much work that will be.
Priorities
I’m thinking that the ChatGPT report should be the highest priority. I think it’s the most important piece of the three and I just need to get on with it.
Then there’s the Harold Bloom piece and the piece of the future of literary criticism. I pretty much know how those things have to go. I just have to get on with it.
In some sense I think the chess & language & AI project may be more important than the lit crit, but it’s also just a bit tougher to do, so I’ll put that in third place.
Photo Exhibit
I’ve made progress on that. I’ve added some more photos into the mix. But the big news is that I’ve got an exhibition scheduled for June of this year. That gives me something to work for and against.