Looking Back at AI Scaling with Anthropic Co-Founder Jared Kaplan

I mean I came to work on this mostly just because I guess at a very fundamental level maybe two of the most interesting questions you can ask are what is the universe and what is intelligence, what is the mind, that sort of thing. I think another perspective on this also is that what we can actually do with intelligence is almost in itself kind of a feature of our universe. It's not a law of physics, but the fact that certain kinds of algorithms are possible to execute in a reasonable amount of time and others aren't, I think is actually kind of an interesting feature. There's a way that any of us could beat Shakespeare. We could write every possible book that could ever exist and then we could just read all of them and go through and pick out the ones that were as good or better than Shakespeare. And if we had good judgment, all we would need is good judgment and we could do that. But that's obviously a ridiculous idea because all of the books that could ever be written are completely computationally infeasible to construct or write. And so that's sort of a silly example, but points to this idea that what kinds of things we can and can't do in terms of processing information really fundamentally constrains what's possible. So those are some super abstract philosophical reasons to care. Mostly I care because I think there's been really dramatic progress in AI using neural networks in the last 10 years. And I expect that that's going to continue. And I think a perspective that I as a physicist have on this is that there are actually scaling laws in AI. And this is something that I've worked on directly and it helped inspire GPT-3. So if you look at how smart a program that's writing is, and you look at how smart it is as a function of how much data it reads, we talked about 200 billion words, but it could have been 20 billion words or two billion words etc. If you look at how smart it gets as a function of how much data it has and how big the neural network is, how big its brain is, then it turns out there are actually very precise scaling laws. They're actually very reminiscent of things like thermodynamics. In the sense that these networks are so big and the data sets are so big that there's a kind of law of large numbers that when you try to optimize their performance you get some consistent predictable result and so I think that's a perspective that is very much the way that a physicist would think about a system, just sort of look at what happens when you change certain parameters and then try to formulate a precise kind of law of nature like a mathematical formula that tells you what changes as you change these inputs like the amount of data or how big the neural network is.

Exactly, exactly so.

Yeah, that's exactly right. So these are artificial neural networks that are being run on big clusters of computers and it's exactly what you said. As you either make these neural networks bigger or you increase the amount of data and there are various different ways you can combine these things, there's a predictable sense in which the intelligence of this system improves and I mean you can quantify that very straightforwardly. So for example in our GPT-3 paper you can look at the SATs, these are the test that American high school students take to get into college and they're analogy there are various word problems in the SATs there are analogies and you can ask how well does this neural net do at answering SAT questions? Could it get into college in the US or not? And the answer is the most recent model on some of these questions, even though it wasn't taught to do the SATs, it just read it's an autodidact. It read everything on the internet in some sense. It learned the largest model was actually better than the average American high school student at these multiple choice questions. And you can also see that there's steady improvement. So if you take models that have less data or that have smaller networks they don't do as well but there's kind of a steady predictable scaling of how intelligent these networks are as you increase these inputs and I think this is one of the many reasons why I don't think that this whole neural network AI deep learning field is overhyped. I think there are these predictable trends which indicate that we're really going to be able to make even more intelligent systems in the relatively near future.

Yeah, that's all correct. It's what people often call unsupervised learning because we just give the model all this data, all these words, all these web pages and essays and books downloaded from the internet and it just reads them and learns and we just sit back and watch. So that's absolutely true. Whether or not what kind of learning this is, whether or not it's achieving some notion of true intelligence or true understanding, is fairly controversial. There are a lot of people with different opinions. I tend to be fairly optimistic on the side that it really is learning a lot of subtle important features of the world and how to think about it. And I think combining this with other learning techniques that maybe are more specific, more aimed at actually doing a specific task, I think we will have models that really do understand and really can think in all external respects. I mean, you see a duck and you call it a duck because it looks like a duck. And similarly here, if you have something that's capable of doing all of these different things at once that we associate with human intelligence, then I think I would give it the benefit of the doubt. And imagine that it actually does have some model of the world that means it understands the world and is making predictions and doing tasks based on that model.

Yeah, I actually do think it's possible. I don't think it's going to happen very soon, but yeah, I think that humans seem like we're very dominant as a species on the planet because we're intelligent and because we are able to work together and preserve knowledge within our civilizations and pass it on to our children. But there's a whole spectrum of intelligence. There's paramecium, there are mice, there are dogs, there are dolphins and us. And I don't think there's any sense in which that continuum of intelligence, of course I'm oversimplifying a lot by calling it a continuum. I mean all these different species have different skills, but anyway, there's no reason to imagine that you can't have a creature or an AI system that is much more intelligent than we are. And therefore would be much better at understanding the laws of physics than we are. And I do think that is achievable. One of the benefits of being a physicist is that the field of artificial intelligence research is very traumatized. There was this idea in the sort of 60s and 70s that they were going to be able to build systems that were smarter than people in just six months or a year. And I'm not even really sure that very many people believed that or claimed it widely. But anyway, it has this reputation that the field of AI overclaimed. They said they were going to make tremendous progress and then they failed and there was this AI winter. And so people in the AI field I think are very reticent to make any kind of ambitious claim because they're afraid of being accused of crying wolf. And as a physicist, I don't really have any of that kind of collective trauma. I'm just sort of looking at it with fresh eyes. And it really just seems like there's going to be a trend of significant improved performance in intelligence in the foreseeable future.

Sure, sure. I would be happy to come back. Yeah, it was a lot of fun and thanks for all the questions and for having me.