Friday, April 19, 2024

Will Team LLM ever catch up to Team Atlas? [+ escape from a maze as a facilitating analogy]

I don't know when I first saw a video of Atlas. But whenever it was, I'm sure I was astounded. As astounded as I was with ChatGPT? I don't remember. And of course, I couldn't play around with Atlas. In any case, that would have been much more difficult than playing around with ChatGPT.

I don't know when I first saw a video of Atlas. But whenever it was, I'm sure I was astounded. As astounded as I was with ChatGPT? I don't remember. And of course, I couldn't play around with Atlas. In any case, that would have been much more difficult than playing around with ChatGPT.

The thing is, the researchers who built Atlas had to develop a profound understanding of the dynamics of humanoid motion. In contrast, the reseachers who work on LLMs don't need to know much of anything about language. That's worth thinking about.

Consider, for example, these remarks that a pioneering computational linguist, Martin Kay, made awhile back:

Symbolic language processing is highly nondeterministic and often delivers large numbers of alternative results because it has no means of resolving the ambiguities that characterize ordinary language. This is for the clear and obvious reason that the resolution of ambiguities is not a linguistic matter. After a responsible job has been done of linguistic analysis, what remain are questions about the world. They are questions of what would be a reasonable thing to say under the given circumstances, what it would be reasonable to believe, suspect, fear, or desire in the given situation. [...] What we are doing is to allow statistics over words that occur very close to one another in a string to stand in for the world construed widely, so as to include myths, and beliefs, and cultures, and truths and lies and so forth. As a stop-gap for the time being, this may be as good as we can do, but we should clearly have only the most limited expectations of it because, for the purpose it is intended to serve, it is clearly pathetically inadequate. The statistics are standing in for a vast number of things for which we have no computer model. They are therefore what I call an “ignorance model.”

LLMs did not exist in 2005, when Kay made those remarks. As he died in 2021, before the release of ChatGPT, I don't know how it would have reacted to it. I see little reason to believe that he would alter those remarks in a fundamental way. Perhaps he would remove the word “clearly,” and maybe “pathetically” as well. 

LLMs, however, are still inadequate models of human linguistic behavior. The industry’s current infatuation with them is perhaps an ironic testament to the cliché that ignorance is bliss. Martin Kay also remarked that, in resting content with a statistical view of language, “one turns one’s back on the scientific achievements of the ages and foreswears the opportunity that computers offer to carry that enterprise forward.” I agree, though perhaps not in the way Kay meant those words. For I believe that LLMs have an important role in developing a detailed understanding how language works. The intellectual monoculture that has grown up around LLMs seems unable or unwilling to appreciate that – a profound failure of the imagination.

Language is grounded in the operations of the human brain. Our ability to probe and maniuplate the brain is quite limited. That is not the case with LLMs. Here’s a facilitating analogy I am working on for a report I am preparing about my work with ChatGPT over the last year. I’m talking about using ChatGPT to generate stories:

The model is structured such that, when it starts generating a text from a certain location in its activation space, it will have created a coherent text – a story in this case, word-by-word, by the time it exits that region of the space.

As a crude analogy, consider what is called a simply connected maze, one without any loops. If you are lost somewhere in such a maze, no matter how large and convoluted it may be, there is a simple procedure you can follow that will take you out of the maze. You don’t need to have a map of the maze; that is, you don’t need to know its structure. Simply place either your left or your right hand in contact with a wall and then start walking. As long as you maintain contact with the wall, you will find an exit. The structure of the maze is such that that local rule will take you out.

“Produce the next word” is certainly a local rule. The structure of LLMs is such that, given the appropriate context – a prompt asking for a story, following that rule will produce a coherent a story. Given a different context, that is to say, a different prompt, that simple rule will produce a different kind of text.

Now, let’s push the analogy to the breaking point: We may not know the structure of LLMs, but we do know a lot about the structure of texts, from phrases and sentences to extended texts of various kinds. In particular, the structure of stories has been investigated by students of several disciplines, including folklore, anthropology, literary criticism, linguistics, and symbolic artificial intelligence. Think of the structures proposed by those disciplines as something like a map of the maze in our analogy.

Unfortunately, students of those various disciplines have not reached a consensus on how to characterize those structures. Linguists are entertaining a variety of proposals about the nature of sentence-level syntax and students of those other disciplines haven’t converged on a way to describe story structure. Still, we have a starting point for constructing our story maps, even if it is somewhat confused and ambiguous.

If we are to exploit LLMs in ways that analogy suggests, then we are going to have to use symbolic models to do it. First we propose a symbolic model for some aspect of the structure we know how to probe or manipulate. Then see whether an appropriately prepared LLM behaves in the way our model predicts that it should. If it doesn’t, then we revise and repeat.



and again,

and again.... 

* * * * *

Addendum:  The NYTimes has published an article about Atlas, noting that it will retire to the a museum of decomissioned robots in the Boston Dynamics lobby.

No comments:

Post a Comment