The Baby and the Bathwater
Let us be neither too rapturous, nor too fearful, of LLM’s.
Since the advent of ChatGPT, we’ve been whipsawed between rhapsodic infatuation, and a kind of weary lamentation of its lack of true understanding.
There is no doubt that these Large Language Models (LLMs) are created from the NWITS paradigm. Predicting the next word in the sentence certainly works well for texting: as you begin to compose a word, the texting app predicts the word you might be in the midst of typing, just as gmail tries to fill in the rest of the sentence you’re writing. In other words, predicting the next word in the sentence works very well for predicting the next word in the sentence.
What is surprising, however, is the extent to which this technique seems to be able to function in abstractions. If you ask it to create a scene outline from a given movie, it will do that. It ‘understands’ what an outline is, and then proceeds to offer up the structure of the movie. Let’s not throw out the baby with the bathwater.
There are two possibilities for what is happening here. One is that it’s an illusion: it understands nothing, but is able to answer these questions in a convincing way. The other is that it is, in fact, able to synthesize ideas.
Which is it?
Ilya Sutskever of OpenAI recently said, “One answer is that information about the world, even the visual information, slowly leaks in through text .” How it does this isn’t exactly clear, but it sure does seem that there are abstractions that are emergent.
This argument harkens back to The Chinese Room argument, as put forth by the philosopher John Searle:
Searle imagines himself alone in a room following a computer program for responding to Chinese characters slipped under the door. He understands nothing of Chinese, and yet, by following the program for manipulating symbols and numerals just as a computer does, he sends appropriate strings of Chinese characters back out under the door, and this leads those outside to mistakenly suppose there is a Chinese speaker in the room.
To my mind, it doesn’t really matter. What is undeniable is that it does have the ability to act as if it understands. And for a human interacting with it, that is all that matters. (I’m skirting the question of whether or not someone falls in love with the AI, or is offended by it, etc. These are important questions, but are not germane to the one at hand.)
The point is that it is, to a surprising degree, doing what we ask. How much of this is due to Reinforcement Learning with Human Feedback, is unclear. The more humans get involved with the application, and the more they send feedback, the better the model becomes. And because there are over a million users now of ChatGPT, they’re getting a lot of feedback. First-mover status is, as always, extremely valuable to a company, and the rest will now play catch-up.
It looks to me like ChatGPT is, by whatever means, creating emergent properties which look a lot like understanding.