Continual learning remains a key unsolved capability

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 1:57

   Is that just a limitation of the large language model paradigm? Well, look, I think there is, and I'm definitely a subscriber to the idea that maybe we need one or two more big breakthroughs before we'll get to AGI. And I think they're along the lines of things like continual learning, better memory, longer context windows, or perhaps more efficient context windows would be the right way to say it.

2. 2:19

   So don't store everything, just store the important things. That would be a lot more efficient. That's what the brain does. And better long-term reasoning and planning. Now, it remains to be seen whether just sort of scaling up existing ideas and technologies will be enough to do that, or we need one or two more really big, insightful innovations. I'm probably, if you were to push me, I would be in the latter camp.

3. 2:41

   But I think no matter what camp you're in, we're going to need large foundation models as the key component of the final AGI systems. Of that, I'm sure. So I'm not a subscriber to someone like Jan LeCun who thinks that there's sort of some kind of dead end. I think the only debate in my mind is, are they a key component or the only component? So I think it's between those two options. And for me, this is one advantage we have of having such a deep and rich research bench. We can go after both of those things at maximum with maximum force, both scaling up the current paradigms and ideas. And when I say scaling up, that also involves innovation, by the way, pre-training especially, I think we're very strong on.

4. 3:27

   And then really new blue sky ideas for new architectures and things, you know, the kinds of things we've invented over the last 10 years as Google and DeepMind, you know, of course, including Transformers. Can something with a lot of hard-coded stuff ever be considered AGI? No, I think, well, depends what you mean by a lot. I think that I'm very interested in hybrid systems is what I would call them or neurosymbolics.

5. 3:52

   Sometimes people call them, you know, AlphaFold, AlphaGo are examples of that. to some of our most important work. combines neural networks and deep learning with things like Monte Carlo Tree Search. So I think that could be possible, and there's some very interesting work we're doing using the LLMs with things like evolutionary methods, AlphaEvolve, to actually go and discover new knowledge. You may need something beyond what the existing methods do. But I think learning is a critical part of AGI. It's actually almost a defining feature. When we say general we mean general learning can it can it learn uh new knowledge and can it learn across any domain that's the general part so for me learning is synonymous with intelligence and always has been okay so if learning is synonymous with intelligence and these models still don't

6. 4:45

   have the ability to continually learn no uh like i said earlier it has goldfish brain yeah it can search the internet and it can be like i figured this out yeah but it doesn't change the model it's just, we'll forget it after the session. Do you have a theory as to how the continual learning problem can be solved? And do you want to share it with us all day? I can give you some clues. We are working very hard on it. We've done some work on, you know, I think the best work on this in the past with things like AlphaZero, you know, that learn from scratch, versions of AlphaGo, AlphaGo Zero also learned on top of the knowledge it already had. So we've done it in much narrower domains. Games are obviously a lot easier than the messy real world. So it remains to be seen if those kinds of techniques will really

7. 5:30

   scale and generalize to the real world and actual real world problems. But at least the methods we know can do some pretty impressive things. And so now the question is, can we blend that, at least in my mind, with these big foundation models? And so of course, the foundation models are learning during training, but we would love them to learn, you know, out in the wild and including things like personalization. I think that's going to happen. And I feel like that's a critical part of building a great assistant is that it understands you and it works for you as technology that works for you. And we've released our first versions of that just last week. Personal intelligence is the sort of first baby steps towards that. But I think to have it, you want to do it more than just having your data in the context window. That's you want to have something

8. 6:18

   a bit deeper than that, which is, as you say, actually changes the model over time. That's what ideally you would have. And that technique has not been cracked yet.