Full episode with Ray Dalio (Dec 2019): https://www.youtube.com/watch?v=M95m2EFb7IQ
Clips channel (Lex Clips): https://www.youtube.com/lexclips
Main channel (Lex Fridman): https://www.youtube.com/lexfridman
(more links below)



Podcast full episodes playlist:
https://www.youtube.com/playlist?list=PLrAXtmErZgOdP_8GztsuKi9nrraNbKKp4

Podcasts clips playlist:
https://www.youtube.com/playlist?list=PLrAXtmErZgOeciFP3CBCIEElOJeitOr41

Podcast website:
https://lexfridman.com/ai

Podcast on Apple Podcasts (iTunes):
https://apple.co/2lwqZIr

Podcast on Spotify:
https://spoti.fi/2nEwCF8

Podcast RSS:
https://lexfridman.com/category/ai/feed/

Ray Dalio is the founder, Co-Chairman and Co-Chief Investment Officer of Bridgewater Associates, one of the world's largest and most successful investment firms that is famous for the principles of radical truth and transparency that underlie its culture. Ray is one of the wealthiest people in the world, with ideas that extend far beyond the specifics of how he made that wealth. His ideas, applicable to everyone, are brilliantly summarized in his book Principles.

Subscribe to this YouTube channel or connect on:
- Twitter: https://twitter.com/lexfridman
- LinkedIn: https://www.linkedin.com/in/lexfridman
- Facebook: https://www.facebook.com/lexfridman
- Instagram: https://www.instagram.com/lexfridman
- Medium: https://medium.com/@lexfridman
- Support on Patreon: https://www.patreon.com/lexfridman

Let me talk about, if we could, about AI a little bit.

So we've, Bridgewater Associates, manage about $160 billion

And our artificial intelligence systems algorithms are pretty good with data.

What role in the future do you see AI play in analysis and decision making in this kind of data-rich and impactful area of investment?

I'm gonna answer that not only in investment, but I give a more all-encompassing rule for AI.

As I think you know, for the last 25 years, we have taken our thinking and put them in algorithms.

The computer takes those criteria algorithms and they put them, they're in there and it takes data and they operate as an independent decision-maker pair in parallel with our decision-making so for me, it's like there's a chess game playing And I'm a person with my chess game and I'm saying it made that move and I'm making the move and how do I compare those 2 moves?

So we've done a lot, but let me give you a rule.

If the future can be different from the past and you don't have deep understanding, you should not rely on AI.

the cause effect relationships that are leading you to place that bet in anything, okay?

Let's say if it was do surgeries, and you would say, how do I do surgeries?

I think it's totally fine to watch all the doctors do the surgeries, you can put it on, take a digital camera and do that, convert that into AI algorithms that go to robots and have them do surgeries and I'd be comfortable with that.

Because if it keeps doing the same thing over and over again and you have enough of that, that would be fine, even though you may not understand the algorithms because if the thing's happening over and over again and you're not asking, the future would be the same, that appendicitis or whatever it is will be handled the same way the surgery, that's fine.

However, what happens with AI is for the most part, is it takes a lot of data and it, with a high enough sample size, and then it puts together its own algorithms.

Okay, there are 2 ways you can come up with algorithms.

You can either take your thinking and express them in algorithms, or you can say, put the data in and say, what is the algorithm?

And when you have machine learning, it'll give you equations, which quite often are not understandable.

If you would try to say, OK, now describe what it's telling you, it's very difficult to describe.

And so it's very good for doing those things that could be done over and over again if you're watching and you're not taking that.

But if the future is different from the past and you don't have deep understanding, you're gonna get in trouble.

As far as AI is concerned, AI and let's say computer replications of thinking in various ways.

I think it's particularly good for processing.

But the notion of what you want to do is better most of the time determined by the human mind.

Like, okay, how should I raise my children?

It's gonna be a long time before AI, you're going to say it has a good enough judgment to do that, who should I marry?

Maybe you can get the computer to help you, but If you just took data and do machine learning, it's not gonna find it.

If you were to then take 1 of my criteria for any of those questions, and then say put them into an algorithm, you'd be a lot better off than if you took AI to do it.

But by and large, the mind should be used for inventing and those creative things.

And then the computer should be used for processing because it could process a lot more information, a lot faster, a lot more accurately, and a lot less emotionally.

So any notion of thinking in the form of processing type thinking should be done by a computer, and anything that is in the notion of doing that other type of thinking should be operating with the brain, operating in a way where you can say, ah, that makes sense.

The process of reducing your understanding down to principles is kind of like the process, the first 1 you mentioned, type of AI algorithm where you're encoding your expertise.

You're trying to program, write a program.

The process of reducing principles to a computer program.

Or when you say, when you write about, when you think about principles, is there still a human element that's not reducible to an algorithm?

My experience has been that almost all things, including those things that I thought were pretty much impossible to express, I've been able to express in algorithms, but that doesn't constitute all things.

So you can, ooh, you can express far more than you can imagine you'll be able to express.

So I use the example of, okay, it's not, how do you raise your children?

Okay, you will be able to take it 1 piece by piece.

And the way to do that, in my experience, is to take that and when you're in the moment of making a decision or just past making a decision, to take the time and to write down your criteria for making that decision in words, okay?

That way you'll get your principles down on paper.

I created an app, online call, it's right now just on the iPhone, it'll be

It'll be in a few months, it'll be on Android.

But it has an app in there that helps people write down their own principles.

So when you're in that moment where you've just, you're thinking about it and you're thinking your criteria for choosing the school for your child or whatever that might be, and you write down your criteria, or whatever they are, those principles, you write down and that will, at that moment, make you articulate your principles in a very valuable way.

And if you have the way that we operate, that you have easy access, so then the next time that comes along, you can go to that, or you can show those principles to others to see if they're the right principles.

You will get a clarity of that principle that's really invaluable in words, and that'll help you a lot.

But then you start to think, how do I express that in data?

And it'll shock you about how you can do that.

You'll form an equation that will show the relationship between these particular parts and then the, essentially the variables that are going to go into that particular equation and you will be able to do that.

And you take that little piece and you put it into the computer.

And then take the next little piece and you

And before you know it, you will have a decision-making system that's of the sort that I'm describing.

So you're almost making an argument against an earlier statement you've made.

At first you said there's no way a computer could raise a child, essentially.

But now you've described making me think of it.

If you have that kind of idea, meritocracy, you have this rigorous approach to bridge while it takes an investment and apply it to raising a child.

It feels like through the process you just described, we could, as a society, arrive at a set of principles for raising a child and encode it into a computer.

That originality will not come from machine learning.

The first time you do, so that the original, yes.

But eventually as we together develop it and then we can automate it.

That's why I'm saying the processing can be done by the computer.

So we're saying the same thing, We're not inconsistent.

We're saying the same thing, that the processing of that information and those algorithms can be done by the computer in a very, very effective way.

You don't need to sit there and process and try to weigh all those things in your equation and all those things.

But that notion of, okay, how do I get at that principle?

And you're saying you'd be surprised how much you can express.

So This is where I think you're going to see the future.

Right now, we go to our devices and we get information to a large extent.

And then we get some guidance, We have our GPS and the like.

In my opinion, principles, principles, principles, principles, I want to emphasize that.

You write them down, you've got those principles.

They will be converted into algorithms for decision making And they're gonna also have the benefit of collective decision making.

Because right now, individuals, based on what's stuck in their heads, are making their decisions in very ignorant ways.

They're not the best decision makers, they're not the best criteria, and they're operating.

When those principles are written down and converted into algorithms, it's almost like you'll look at that and follow the instructions and it'll give you better results.

You can go to your local doctor and you could ask his point of view and whatever, and he's rushed and he may not be the best doctor around, and you're gonna go to this thing and get that same information or just automatically have an input in that, and it's gonna tell you, okay, here's what you should go do, and it's gonna be much better than your local doctor.

And that, the converting of information into intelligence, okay, intelligence is the thing.

We're coming out with, again, I'm 70 and I wanna pass all these things along.

So all these tools that I've found need to develop all over these periods of time, all those things, I wanna make them available.

And what's gonna happen as they're going to see this, they're going to see these tools operating much more that way, the idea of converting data into intelligence.

Intelligence, for example, on what they are like.

On what are your strengths and weaknesses.

Intelligence on who do I work well with under what circumstances.

We're gonna go from what are called systems of record, which are a lot of, okay, information organized in the right way, to intelligence.

And we're gonna, that'll be the next big move, in my opinion, and so you will get intelligence back.

And that intelligence comes from reducing things down to principles and to.


Speaker 0: Let me talk about, if we could, about AI a little bit. So we've, Bridgewater Associates, manage about $160 billion

Speaker 1: in assets. And our artificial intelligence systems algorithms are pretty good with data. What role in the future do you see AI play in analysis and decision making in this kind of data-rich and impactful area of investment?

Speaker 2: I'm gonna answer that not only in investment, but I give a more all-encompassing rule for AI. As I think you know, for the last 25 years, we have taken our thinking and put them in algorithms. And so we make decisions. The computer takes those criteria algorithms and they put them, they're in there and it takes data and they operate as an independent decision-maker pair in parallel with our decision-making so for me, it's like there's a chess game playing And I'm a person with my chess game and I'm saying it made that move and I'm making the move and how do I compare those 2 moves? So we've done a lot, but let me give you a rule.

If the future can be different from the past and you don't have deep understanding, you should not rely on AI. Okay? Those 2 things.

Speaker 1: Deep understanding of

Speaker 2: the cause effect relationships that are leading you to place that bet in anything, okay? Anything important. Let's say if it was do surgeries, and you would say, how do I do surgeries? I think it's totally fine to watch all the doctors do the surgeries, you can put it on, take a digital camera and do that, convert that into AI algorithms that go to robots and have them do surgeries and I'd be comfortable with that. Because if it keeps doing the same thing over and over again and you have enough of that, that would be fine, even though you may not understand the algorithms because if the thing's happening over and over again and you're not asking, the future would be the same, that appendicitis or whatever it is will be handled the same way the surgery, that's fine.

However, what happens with AI is for the most part, is it takes a lot of data and it, with a high enough sample size, and then it puts together its own algorithms. Okay, there are 2 ways you can come up with algorithms. You can either take your thinking and express them in algorithms, or you can say, put the data in and say, what is the algorithm? That's machine learning. And when you have machine learning, it'll give you equations, which quite often are not understandable.

If you would try to say, OK, now describe what it's telling you, it's very difficult to describe. And so they can escape understanding. And so it's very good for doing those things that could be done over and over again if you're watching and you're not taking that. But if the future is different from the past and you don't have deep understanding, you're gonna get in trouble. And so that's the main thing.

As far as AI is concerned, AI and let's say computer replications of thinking in various ways. I think it's particularly good for processing. But the notion of what you want to do is better most of the time determined by the human mind. That what are the principles? Like, okay, how should I raise my children?

It's gonna be a long time before AI, you're going to say it has a good enough judgment to do that, who should I marry? All of those things. Maybe you can get the computer to help you, but If you just took data and do machine learning, it's not gonna find it. If you were to then take 1 of my criteria for any of those questions, and then say put them into an algorithm, you'd be a lot better off than if you took AI to do it. But by and large, the mind should be used for inventing and those creative things.

And then the computer should be used for processing because it could process a lot more information, a lot faster, a lot more accurately, and a lot less emotionally. So any notion of thinking in the form of processing type thinking should be done by a computer, and anything that is in the notion of doing that other type of thinking should be operating with the brain, operating in a way where you can say, ah, that makes sense.

Speaker 1: The process of reducing your understanding down to principles is kind of like the process, the first 1 you mentioned, type of AI algorithm where you're encoding your expertise. You're trying to program, write a program. The human is trying to write a program. How do you think that's attainable? The process of reducing principles to a computer program.

Or when you say, when you write about, when you think about principles, is there still a human element that's not reducible to an algorithm?

Speaker 2: My experience has been that almost all things, including those things that I thought were pretty much impossible to express, I've been able to express in algorithms, but that doesn't constitute all things. So you can, ooh, you can express far more than you can imagine you'll be able to express. So I use the example of, okay, it's not, how do you raise your children? Okay, you will be able to take it 1 piece by piece. Okay, at what age, what school?

And the way to do that, in my experience, is to take that and when you're in the moment of making a decision or just past making a decision, to take the time and to write down your criteria for making that decision in words, okay? That way you'll get your principles down on paper. I created an app, online call, it's right now just on the iPhone, it'll be

Speaker 1: on Android. I tried getting it on Android. Come on, now.

Speaker 2: It'll be a

Speaker 1: good on Android.

Speaker 2: It'll be in a few months, it'll be on Android. Awesome. But it has an app in there that helps people write down their own principles. Because this is very powerful. So when you're in that moment where you've just, you're thinking about it and you're thinking your criteria for choosing the school for your child or whatever that might be, and you write down your criteria, or whatever they are, those principles, you write down and that will, at that moment, make you articulate your principles in a very valuable way.

And if you have the way that we operate, that you have easy access, so then the next time that comes along, you can go to that, or you can show those principles to others to see if they're the right principles. You will get a clarity of that principle that's really invaluable in words, and that'll help you a lot. But then you start to think, how do I express that in data? And it'll shock you about how you can do that. You'll form an equation that will show the relationship between these particular parts and then the, essentially the variables that are going to go into that particular equation and you will be able to do that.

And you take that little piece and you put it into the computer. And then take the next little piece and you

Speaker 1: put that into the computer. And before you know it, you will have a decision-making system that's of the sort that I'm describing. So you're almost making an argument against an earlier statement you've made. You're convincing me. At first you said there's no way a computer could raise a child, essentially.

But now you've described making me think of it. If you have that kind of idea, meritocracy, you have this rigorous approach to bridge while it takes an investment and apply it to raising a child. It feels like through the process you just described, we could, as a society, arrive at a set of principles for raising a child and encode it into a computer. That originality will not come from machine learning. The first time you do, so that the original, yes.

Speaker 2: That's what I'm referring to.

Speaker 1: But eventually as we together develop it and then we can automate it.

Speaker 2: That's why I'm saying the processing can be done by the computer. So we're saying the same thing, We're not inconsistent. We're saying the same thing, that the processing of that information and those algorithms can be done by the computer in a very, very effective way. You don't need to sit there and process and try to weigh all those things in your equation and all those things. But that notion of, okay, how do I get at that principle?

Speaker 1: And you're saying you'd be surprised how much you can express.

Speaker 2: That's right, you can do that. So This is where I think you're going to see the future. Right now, we go to our devices and we get information to a large extent. And then we get some guidance, We have our GPS and the like. In my opinion, principles, principles, principles, principles, I want to emphasize that.

You write them down, you've got those principles. They will be converted into algorithms for decision making And they're gonna also have the benefit of collective decision making. Because right now, individuals, based on what's stuck in their heads, are making their decisions in very ignorant ways. They're not the best decision makers, they're not the best criteria, and they're operating. When those principles are written down and converted into algorithms, it's almost like you'll look at that and follow the instructions and it'll give you better results.

Medicine will be much more like this. You can go to your local doctor and you could ask his point of view and whatever, and he's rushed and he may not be the best doctor around, and you're gonna go to this thing and get that same information or just automatically have an input in that, and it's gonna tell you, okay, here's what you should go do, and it's gonna be much better than your local doctor. And that, the converting of information into intelligence, okay, intelligence is the thing. We're coming out with, again, I'm 70 and I wanna pass all these things along. So all these tools that I've found need to develop all over these periods of time, all those things, I wanna make them available.

And what's gonna happen as they're going to see this, they're going to see these tools operating much more that way, the idea of converting data into intelligence. Intelligence, for example, on what they are like. On what are your strengths and weaknesses. Intelligence on who do I work well with under what circumstances. Intelligent.

We're gonna go from what are called systems of record, which are a lot of, okay, information organized in the right way, to intelligence. And we're gonna, that'll be the next big move, in my opinion, and so you will get intelligence back.

Speaker 1: And that intelligence comes from reducing things down to principles and to. That's how it happens.

Speaker 2: Principles into... That's how it happens. You

See all Lex Fridman transcripts on Youtube

Ray Dalio: Artificial Intelligence Principles | AI Podcast Clips