What makes Ilya Sutskever believe that superhuman AI is a natural extension of Large Language Models?

Ilya Sutskever, the chief scientist of OpenAI has appeared in a podcast, of which a very telling 2-minute snippet has been published on YouTube:

I have created a decent transcript (that is, taken the transcript and fixed the transcript errors) so I can quote it here and add a few comments. I’m not going to argue in full here, it is just that I think it is very useful that people understand the conviction Ilya and friends hold.

This is a good moment to point you to The “ChatGPT and Friends” Collection. Watch the clarifying presentation or read my guest blog on Erik. J. Larson’s substack for some key parts on LLMs so that you understand what a ‘token’ is, because Ilya’s argument is: ‘token prediction’ is enough for superhuman intelligence.

Here we go:

Ilya:
I challenge the claim that next token prediction cannot surpass human performance

It looks like on the surface it cannot. It looks on the surface if you just learn to imitate, to predict, what people do, it means that you can only copy people. But here is a counter argument for why might not be quite so: If your neural net.. is if your [base?] neural net is smart enough..., you just ask it — like — "what would what would a person with like great insight and wisdom and capability do?". Maybe such a person doesn't exist, but there's a pretty good chance that the neural net will be able to extrapolate how such a person should behave. Do you see what I mean?

The key argument here is that LLMs can take data created with ordinary human performance, but a neural net should be able to use that data and come up with (extrapolate to) better answers when it is tasked to ‘think like a person that is smarter than the people who created the training data’ (i.e. us). The presenter spots the issue:

Presenter:
Yes, although where would it get the sort of insight about what that person would do if not from...

Ilya:
From the data of regular people. Because, like, if you think about it: what does it mean to predict the next token well enough? What does it mean actually? It's actually ... it's a much ... it's a deeper question than it seems. Predicting the next token well
means that you understand the underlying reality that led to the creation of that token.

It's not statistics. Like, it is statistics, but what is statistics?

In order to to understand those statistics, to compress them, you need to understand what is it about the world that creates those statistics. And so then you say: okay, well I have all those people, what is it about people that creates their behaviors? Well they have — you know — they they have thoughts and they have feelings and they have ideas and they do things in certain ways, all of those could be deduced from next token prediction. 

The first key element here are “predicting the next token well means that you understand the underlying reality that led to the creation of that token” (see this for how that works out). The second key element is: that understanding, especially understanding people’s thoughts, feelings, ideas, etc. can be deduced backwards from next token prediction. Kind of: I hear what you say, so I know what your thoughts are.

And the combination of these two leads to the conclusion that from token prediction (based ultimately on what humans have produced — regardless of the quality and truthfulness of that material by the way) comes understanding, and because we can ask “answer like you’re a four-year old” we can also ask “answer like you’re a superhuman intelligence” and the token prediction system will be able to do that, eventually:

And I'd argue that this should make it possible — not indefinitely, but to a to a pretty decent degree — to say: well, can you guess what you do if you took a person with like this characteristic and that characteristic? Like, such a person doesn't exist but
because you're so good at predicting the next token, you should still be able to guess what that person would do this hypothetical imaginary person with far greater mental ability than the rest of us.

The short summary of this is: According to Ilya, we can get Large Language Models to become a superhuman intelligence by simply ‘asking’ (prompting) the model to act like a superhuman intelligence. Easy-peasy!

Here is the video:

I had a go at asking the current GPT4 model to be more intelligent than what had gone in, but we’re clearly not there yet and there was no sign of anything beyond what is in the training data. But that doesn’t change the conviction Ilya and Friends are working under.

This article is part of the Understanding ChatGPT and Friends Collection.