Tyler Cowen makes a series of excellent points in a recent post where he muses over the value of large model hallucinations. He notes that he would not like for these systems to stop hallucinating, since the hallucinations have value in representing something. What this something is, he suggests, could be our statistical average view of the world. Where a hallucination is factually wrong, it reveals a deeper pattern of ‘wrongness’ in our overall understanding of the world. Maybe, then, hallucinations can allow us to map areas of our understanding that are candidates for research, improvement and factual discovery?
‘Hallucination’ is not merely a large models phenomenon. People make things up too — a lot of the time, actually. You have surely done so as well at some point, when you did not know the answer to something – and you made, say, ‘an educated guess’. The educated guess is simply a statistical prediction given the knowledge you have trained on, and so exploring hallucinations will be a lot like exploring the nature of guessing.
A real question here is how guessing incorporates more than linguistic knowledge – and if our guessing is better or worse than that of language models. Can we guess better if we are embodied? Guessing might be a much more important mode of cognition than we have allowed for – and may provide a really helpful mental model for thinking about artificial intelligence; not as an omniscient being, but a different guesser.
Our default position has been to assume that what these models do is that they predict the next word or set of words in any situation. This opens the possibility of thinking more carefully about the relationship between guessing and predicting — and where we do one and where we do another. In a closed setting, such as a game, we predict – and in an open setting, or open world, we guess – or something such. There is a need for some conceptual clarity here.
What are some ideas we could explore? Maybe the below.
- A guess is a prediction with a certain probability range.
- A guess is a prediction but without a probability assignment at all – and where no such assignment makes sense. I.e. guesses are about uncertainty and predictions about probability.
- A guess is very different from a prediction in that it serves as a basis for our actions as it is made – we do not evaluate guesses as we evaluate predictions. (Somewhat unclear – what is the key we are getting at here? That guesses are reflected in actions?)
- A guess is never about a single fact, but about a complex system.
- A guess is about the unfolding of a narrative, a prediction is about an isolated proposition.
- A guess is revealed prediction (as in when we say, thoughtfully, “I guess I believe he did not really mean it…”) through introspection.
There is more to be done here — linking up the statements we make about the future in ways that help create a grammar of the future and how we relate to it in different ways. This grammar now changes with the advent of more powerful artificial intelligence, perhaps?