The Exchanges

Q Yeah, and the sort of, that was actually going to be my next question, like, three years after the publication of the book, like, who would be the best candidate, like, the closest thing to the master algorithm? Is that reinforcement learning? Is that, like, you know, whichever other one?

A When I was, you know, thinking of writing the book, one of the things that I did was I went and talked with my machine learning colleagues. First of all, asking them for stories, you know, because, you know, the book needs stories. Some of them had very good ones. But also asking them what they thought of the idea of a master algorithm, right? And some people were very gung-ho about, you know, believing that there's, some people don't believe that there's a master algorithm, right? That it's like, you know, the philosopher's stone or the perpetual motion machine. That may be. Some people strongly believed in it. Two of the biggest believers were Rich Sutton and Jeff Hinton. Of course, they had different notions of what the master algorithm was. For Rich Sutton, the master algorithm is enforcement learning, and, and, you know, for Jeff Hinton, the master algorithm is, you know, reverse engineering the brain and figuring out how it learns. And, you know, I think the, um, and we'll, but we'll see, right? I think, um, if you look at what's happened in the last few years, there is no doubt that backprop, right, has gone from strength to strength. And if you look at, you know, the gamut of real world applications of deep learning, they all use Backprop. It's amazing how that one single algorithm with the right architecture, et cetera, et cetera, right, can do so much, right? And the …

AI assessment note: “I actually do not believe that that's the case. I think Backprop will not get us there.”

Q finance, um, alternative data is a, is, is a big trend. So this idea that there's all sorts of companies out there that can create a digital exhaust, um, of, uh, data that, uh, when analyzed can, can, can provide signal. Is that, is that something that you're excited about? Is that, uh, is there, is there truly signal? Is that noise? Is that very exploratory? We don't know yet.

A I think alternative data is a perfect example of what we were just talking about, right? So there's data from the markets themselves, which again, back in the nineties, this is really what people were using. But part of what you can do today is that there are these rivers of data, right? And now, of course, the question is like, some of that data is irrelevant. Some of that is relevant, but in ways that are not obvious. A lot of the value maybe isn't combining data in different ways. There's, you know, there's the, all of the issues that we deal with in machine learning, I think, are present there, except on a, on a, on a much larger Scale. Not just in terms of volume, right? But just in terms of all the issues that arise. But I think at the end of the day, you know, using, I mean, if you think about like, you want to value an asset, you want to trade, you want to make markets efficient, you, you know, thinking about it from a Bayesian point of view, right? You're supposed to use all the data that is relevant. Not some, but all the data, right? And until you're using all the, and you may not be able to use it because the cost is too high, right? But up to what the cost allows,

AI assessment note: “thinking about it from a Bayesian point of view, right? You're supposed to use all the data”