The Joe Walker Podcast - Nassim Taleb — Meditations on Extremistan

Episode Date: September 19, 2024

Nassim Taleb is trader, researcher and essayist. He is the author of the Incerto, a multi-volume philosophical and practical meditation on uncertainty. Full transcript available at: https://josephnoelwalker.com/nassim-taleb-158/See omnystudio.com/listener for privacy information.

Today I'm speaking with Nassim Nicholas Taleb. He has influenced me perhaps more than any other thinker. I discovered his work when I was quite young, at the end of 2016. I read his books out of order. I finished with Fooled by Randomness, and I started with The Black Swan. That's the correct order. The Black Swan was the book that got me hooked. For me, that book was not so much about black swans as about what Nassim calls the platonic fold. And this year I've had the pleasure of meeting him in person.

He has a certain magnanimity. He's been very kind to me. So it's an honor to have him on the podcast. Welcome Nassim. Thank you for inviting me. So naturally I have many questions and I guess the theme of my questions is probably best summed up by the title of your technical book, The Statistical Consequences of Fat Tales. But I'd like to start a little bit abstract and then get more and more real. So, first question, it only takes one black swan to know that you're an extremistan. But if you're in a particular domain, which has yet to experience a black swan, how do you know whether or not you're an extremistan? Okay, so let's not use the word black swan and use extreme deviation. Black swan is something that carries large consequences. It tends to happen more easily in an environment that produces large deviation, so what I call extremistan. more easily in an environment that produces tail large deviation. So what I call extremist time.

Mhm. So let's ignore the terminology black swan here because it may be confusing. And let's say that the following asymmetry, let's present, discuss the following asymmetry. If I am using a thin-tail probability distribution, let me see, I can be always surprised by an outlier with respect to my distribution, a large deviation. That would destroy my assumption of using that distribution. If on the other hand, I'm using a large deviation model, or a model that's an extreme stand model, the reverse cannot be true. Nothing can surprise you.

A quiet period is entirely within statistical properties, so is a large deviation, which is why you have to assume that you're in the second class of models unless you have real reasons to a real robust representation of the world to rule it out. Right. For example, we know that with height that you're, you're, you're from Australia. Even, I mean, in Australia, you may run into someone who's two meters, 40 centimeters tall.

Uh, but have you, I mean, even in Australia, they don't have people five kilometers tall or 500 kilometers tall. Why? There are biological limitation. The person needs to have a mother. When we use a maximum entropy representation, the Gaussian is the maximum entropy distribution with known mean and variance. So you're bounding the variance. You see?

If you're bound to variance, it's the equivalent of bounding the energy. So you see what I'm leading at. You can't have unlimited energy. So you know that a lot of mechanisms have these physical limitations. Right. You see? So you can rule out based on knowledge of the process,

biological understanding, physical understanding. But if you don't know anything about the process, or the process is in concern multiplicative phenomena, such as contagions, pandemics, or simply processes that don't have a limit to their movement. Like, for example, a price, you and I can sell, buy from one another this at a billion dollars. There's no limitations. There's no physical limitation to a price. Therefore, you could be an extremist and you cannot rule out a thick-tailed distribution. Right.

So you mentioned height as an example of a Gaussian process. Yeah, or actually pseudo-Gaussian, more like log-normal, but with low variance, yes. Sure. Because it's bounded on the left. Yeah, okay. So what are some heuristics you use to judge whether you have a compelling reason to believe that something

has a gaussian process it's not a i mean uh you see it when you you know it when you see it okay if we're talking about uh weight height such phenomena then you can rule out extremely large deviation yeah not completely but but those deviations that occur are still going to be acceptable. In other words, you may have a five meter tall human with some kind of deregulation, hormonal deregulation or something like that, but you're not going to get a 500 kilometer tall human.

In finance, you can't rule out the equivalent of 500 kilometer tall or 5 billion kilometer tall person. Yeah. And finance, you can't rule out the equivalent of 500 kilometers to a whole or 5 billion kilometers to a whole person. Yeah. Okay. So basically you need to couple

the absence of extreme events with some kind of very compelling explanation as to why the data is

Explanation that rules out these deviations based on either energy or more knowledge of the physical process.

Yeah. The generator is physical after all. So it's interesting that not only do power laws appear everywhere in the natural and social world, but perhaps certain tail exponents appear to be intrinsic. So last week I was chatting with your friend and collaborator,

Raphael Douardy, and he mentioned that he has this view that the tail exponent for financial markets seems to be three. He's wrong, but that's Rafael. There was a theory of why it was called the semi-cubic theory that he is following, and someone figured out that the tail exponent for a company size, the size of companies was 1.5. So therefore their orders are going to impact the market. Hence, by using a, I mean, by using a square root model

of impact, in other words, where the quantity impacts the price following some

kind of square root effect, okay, then you end up with markets having a,

what they call the cubic, going from half cubic to cubic. It is a nice theory, but I think the tail exponent in financial markets is lower than that from experience. And I don't like these cute theories because the distribution of concentration is not 1.5 half cubic. And with technology, it's much, much higher.

But you also see it in other domains, like a lot of people have commented on the fact that city size seems to have... I would not get married to these. Okay. Is that because there's always going to be, there's always the possibility of an even more extreme event to kind of screw up the exponent? Or less extreme event. I mean, coming up with an observation that's very noisy and generalizing to a theory of cubic or half cubic,

or there used to be a square law and a lot of things. I mean, it's a very noisy representation. Okay, so I have a couple of questions about finance. How long before 1987 did you realize that volatility shouldn't be flat across strike prices

for options? And how did you realize? I mean, I saw deviations and I realized and I had you know an unintentional reward from having a tail exposure so I realized

you don't have to be a genius to figure out if the payoff can be so large as to swap the frequency. So I think that I was pretty convinced by September 1985 after the Plaza Accord, we had a 10-sigma move. At the time, we didn't have access to data like today.

But we saw prices, and I noticed that effectively you had a higher frequency of these very, very large deviations across stocks. I mean, you had mergers, you had stuff like that. So it was obvious. And then therefore the black trolls or the equivalent black trolls, they call it black trolls, but black trolls didn't really invent that formula. They just justified it.

The formula is from Bachelier and others and a collection of others who rediscovered it or repackaged it differently that you need to have a higher price for tail options. So I got in the business of collecting tail options. But one has to be

pretty blind not to see that you have winner-take-all effects in finesse, which is not compatible with a Gaussian representation. Right. Yeah, it's pretty crazy how blind so many people have remained to that observation. So your books have become very famous. Universa has done very well. Mark Spitznagel has also written books which have sold well. Why hasn't the tail hedging strategy now been fully priced in by markets because of uh

thank because of mba uh lecturing modern portfolio theory because people get blinded by theories And also because you, if you're trading your own money, you're going to be pretty rational about it. If you're dealing with the institutional framework, you need to make money frequently. And the trap of needing to make money frequently will lead you to eventually sell volatility.

So there's no incentive to buy volatility for someone who's employed for a finite period of time in a firm. No incentive. Right. Yeah. Are there any markets that do price in convexity? They all do in a way, but they don't know how to price it. Right.

Interesting. So I have a question about venture capital, but it perhaps has broader applications. There's a kind of inconsistency I noticed. So on the one hand, as a consequence of the power law distribution of returns, one recommendation to say public market investors is they may want to pursue a barbell strategy, which you've written about. So say you have like 90% of your portfolio and very safe things like bonds. And then with 10%, you can take lots of little speculative bets to maximize your optionality. The same logic could also be pursued by say book publishers where you might want to take because the success of books is power distributed you might want to take

lots of little bets to maximize your chance of publishing the next Harry Potter. On the other hand I've heard venture capitalists say that a reason from the exact same premises, the power law distribution of startup success, but come to an opposite conclusion, which is that they want to concentrate their bets really heavily in a handful of companies. Because,

okay, the way you need to look at venture capital is that it's largely a compensation scheme. Largely like hedge funds, compensation scheme. Okay. Compensation. The two and 20. No, no, the mechanism.

So they don't make their money, venture capitalists, they don't make money by waiting for the company to really become successful. They make their money by hyping up an idea. Okay. Getting new investors and they're cashing in as they're bringing in new investors. Which I mean, look at how many extremely wealthy technology entrepreneurs are floating around while not having ever made a penny in net income. So the income for venture capital comes from a greater fool approach.

Okay. So a Ponzi kind of dynamic. Not necessarily Ponzi, but you're selling hope. You package an idea. It looks good. So you sell it to someone and then they have a second round or third round. They keep, keep it around so you can progressively cash in.

Got it. It's not based on your real sales. Okay. Or your real, your real cashflow, your real, particularly in an environment with low interest rates where there

was no penalty for playing that game. Do you think there's any skill in Vasey?

I mean, they have skills, but most of their skills are in packaging, not in...

Not for the things people think. Exactly. Packaging, because they're trying to sell it to another

person. It's a beauty contest. You know, the Keynesian one. The Keynesian beauty contest. So they package a company and look at the compensation of the venture capitalists. You can see it. I mean, either you have financing rounds where someone cashes in at high price, or you have an initial public offering.

So I come from old finance, old school finance, where you haven't really succeeded until the company gets a strong cash flow base. All right, so I have some questions about behavioral economics and empirical psychology. Behavioral economics. Yeah, I thought that was the center. Well, I'm not a behavioral economics podcast, but I do have a lot of questions about this. So, first question.

If I take the inserto chronologically, you seem much more sympathetic to empirical psychology and the biases and heuristics research program in full by randomness and at least by the time you get to… Okay, so let me tell you the secret to full by randomness. Okay. I wrote full by randomness and it became very successful at the first edition and it had no references and it had no behavior, just behavior aside from how humans don't understand probability. Minimal of that.

Then I met Danny Kahneman in 2002. In Italy. In Italy. And then okay, I spoke to him. He said, you don't have a lot of references for stuff like that, a lot of comments. So I said, no problem. And I got about 100 books in psychology.

I read them over a period of, say, six months. Okay. Went through the corpus, everything figured out. You know, they think that their math is complex and math is trivial and wrong. And then I cited and I remodeled the prospect theory because prospect theory itself, because it is convex, concave, it tells you itself that you should take, if you're going to lose money, you take a big lump. make money slowly because people like to make a million dollars a day for a year rather than 250 million and then nothing. Right. But this is a reverse for losses.

And there's a lot of things in it that's correct. So I like that aspect. So anyway, and I start putting references on sentences I've written before not knowing anything about it which was not the most honest thing but it was

to link my ideas to that discipline right it is it's not like I got the idea

from these books I got the ideas and then found confirmation in these books then I met Danny from these books. I got the ideas and then found confirmation in these books. Then I met Danny for the very first time. I told him,

your ideas don't work in the real world because they underestimate people in the real world. They underestimate the tail event whereas in world, they overestimate it. But there's a difference. In the real world, you don't know the odds.

And you don't know the payoff function very well. In your world, you know the odds and the payoff function. So he liked the fact that I gave him a break in that sense and still used his prospect theory. Because the idea that losses are in the loss domain is convex, I liked the idea. But by then, I knew enough about the psychology literature and about all these decision-making theories. So by then, I built myself a knowledge of that. I revised it full by randomness. I put a section in the back connecting my ideas to that literature.

And then they started liking it in the world. Brother Chiller didn't like it. He said, you had a great book. It was genuine. Now you have an academic tome. That was Schiller. Right. But the other people liked it.

So that was my first encounter was on prospect theory, which I believe is correct for that function, but not necessarily for the underestimation, overestimation of probabilities and decision making for reasons that

I show here. Because you never have a lump loss except with lotteries. Typically, it's a variable, and there's no such thing as a typical large deviation. Right. You see?

It is technical, but maybe your viewers will get it better if it was an explanation. We'll get there next. Yeah. And then I started looking at stuff viewers will get it better with an explanation. We'll get there next. Yeah. And then I started looking at stuff done in behavioral economics, such as the Benartian-Taylor.

The Benartian-Taylor assumed that, so I thought it was a mistake, Benartian-T Nancy and Taylor assumed Gaussian distribution and then explained why people prefer bonds to stocks. That was the idea at the time. And I said, maybe it's right. And then therefore it was irrational. They went from the standpoint as irrational to not have more stocks, given the performance.

But I tell them that the risk is not the one you see. See, you have tail risks that don't show in your analysis. I told Taylor, Taylor said, well, assuming it is a Gaussian, then my theory works. I'm not assuming the world were coconuts. A lot of things would work. So the world is not a Gaussian. But you're recommending that for 401k and stuff like that.

So then I noticed that's the first mistake in Thaler. There are other mistakes in that discipline, like this idea of rationality. And to me, rationality is in survival, not in other things. And I discovered, and then I spoke to smart people like Ken Binmore. When you speak to smart people,

you realize these people are not making the claims that are common in that, I'd call it industry in that field. For example, there are things that are deemed irrational such as, let me take a simple example. People use as a metric and it was not contest tested, the transitivity of preferences. That I prefer apples to pies, pies to, say, bananas, all right? Okay. But then bananas to apples, all right?

So you're violating transitivity of preferences. But I said, no, maybe that's not the way the world works. If I always prefer apples to pie and I'm presented that choice, nature wants to make me eat other things and also wants to reduce the stress on the environment of people always eating the same thing. So it's a good way for nature to make me vary my preferences, either to protect nature

and to protect myself. Right. So it's not necessary, you know, the transitivity of preferences is not a necessary criterion for rationality. It's a way nature makes you randomize your choices,

for example, for a broader. So that's one thing. So if now if I were to structure this conversation about the defects of behavioral and cognitive sciences as linked to economics and decision theory, we have things linked to misunderstanding and decision theory. We have things linked to misunderstanding of probability structure and things linked to misunderstanding of the dynamic aspect of decision making, what we call erudicity. So let's use these categories.

So we have the equity premium bias comes from equity premium, the fact that people don't invest. Their explanations come from poor understanding of probability structure. The aspect of prospect theory that is wrong comes from misunderstanding of probability structure. That if you have an open-ended distribution with fat tails, then you don't have the same result. The idea of... What's the other idea? The fact that people, if you give them 10 choices, the one over N, okay? One over N is optimal under fat tails.

Right. But, so this is, again, I think Thaler has one over N papers saying that, you know, you should reduce people's choices because they spread them too much. But that's an optimal strategy. There's another one about probability matching where you think that probability matching is irrational. Probability matching means that if something comes up 40% of the time and something comes up 60% of the time that you should invest 100% of the time in the higher frequency you want. But in nature and in animals,-style modeling, if I have 10 horses and I've got to allocate among the 10 horses, if I want to maximize the expected return, how do I allocate? The proportion is probably of when.

So these are the errors linked to probabilistic structure. There's another one also. There's intertemporal choices. Like if I tell you, do you want a massage today or two massages tomorrow, you're likely to say, okay, two massages tomorrow. Or let's

assume that you, when facing this choice, you take the two massages tomorrow, not one today. But if I tell you in 364 days, you have a choice of one versus two, you see, you would reverse. Actually, let's say that you have it the other way, that you take the one today rather than two tomorrow, but you reverse. That's not if you use a different probability distribution or different preference structure.

Right. Plus there is another one that, I mean, how do you know the person offering you that bet will satisfy tomorrow? You see? As I say, the bird in the hand is better than

Right. than some attract one in the future on some tree. So if you say the person is full of bologna, maybe full of bologna, I'd rather have one today, let me take it today, or maybe bankrupt. But if he is 364 or 365 days, the effect is not that big. So it depends on what kind of preference structure you have

or what kind of errors you have in your model. So this is the first class, misunderstanding of probability. It can go on forever. The second one is more severe, misunderstanding of dynamics. Like, we had a Twitter fight with Taylor while running a Ruri where he couldn't understand why you can refuse a bet of 55% win

versus 45% probability of losing, that someone can refuse such a bet and be rational. Okay? Number one is to realize that of course you can

refuse such a bet because you got to look at things dynamically. Yeah, if you keep taking those bets you'll eventually blow up. Of course you can refuse such a bet because you got to look at things dynamically. Yeah. If you keep taking those bets, you'll eventually blow up. I take risk in life of that nature all the time. Yeah. You see? And it would bring you closer to an uncle point.

Yeah. I could probably do it for a dollar, but maybe not $10 or not $100. That's certainly not a million dollars. Yeah. See? So he couldn't understand the ergodic thing. And that's the Kelly criterion, it shows it clearly. But Kelly criterion is just one example of getting that result for that optimized for growth. My whole idea is surviving. It's simple like saying, hey, you know what? The trade-off of smoking, one cigarette, look at how much pleasure you derive versus how much risk you're taking.

So it's irrational. So yes, but do you know people who smoke once? You got to look at the activity, not an episode. And he couldn't get it. That's one example. There are other similar examples. Oh, let's talk about mental accounting. Mm-hmm. There are other similar examples. Oh, let's talk about mental accounting.

I think he started with mental accounting, say there. And he finds it irrational that, say, a husband and wife have a joint checking account. The husband visits the department store, sees a tie, doesn't buy it, it's too expensive. Goes home and then sees this gift and gets all excited that he got it from

his wife for his birthday. So, you know, that mental accounting is irrational. So, yeah, but how many birthdays

do you have a year? Okay? Yeah. So, it's not frequent. So, you know, so this is where, you know, you got to put some structure

around the mental accounting. Another mistake he makes, there's mistake, the mistake is that it's irrational when you go to a casino to increase your betting when you win money from a casino. That's mental accounting, that money won from a casino should be treated from an accounting standpoint the same way as money that you had as an initial endowment. Yeah, but think about it. If you don't play that game, you're going to go bankrupt. This is what we call it because playing with the house money. So it's not irrational.

So practices that have been around for a long time are being judged by that industry. I call it an industry because it became an industry just producing papers. And they don't have a good understanding of the real world and not a good understanding of probability theory. So this is why we shouldn't be talking about it. And actually I hardly ever talk about them anymore. I mean I discussed them initially when I went in and found effectively, yeah, we are fooled by randomness, but not in the way they think.

And they are more fooled by randomness in other areas. So let me pull out of this. I pulled out of this, but in my writing, I hardly ever discuss them. Yeah. At least, I guess, like distinguishing empirical psychology from behavioral economics. My quick take on empirical psychology is that a lot of the heuristics that, say, Danny and amos found are actually

descriptively pretty good approximations of how humans think but the problem was the additional step they took of then labeling those as you know the use of many of those heuristics as irrational against their normative actually their normative benchmark use the word irrational they but yeah they were careful they were careful with that. But they still indirectly use it, only because they had a war with some… Giga-Ranta. No, after the…

I think, was it the Lisa paper? The one with the bank teller? Linda. The Linda problem, yeah. They had a lot of problem with philosophers who

and then they avoided use the term in the whole industry the term rationality right but effectively they find it this is not

something yeah rational but they don't use the word rational yeah yeah the okay but but you forget something rational, but they don't use the word rational. Yeah, yeah. Okay, but you forget a few things

that one, I had a lot of people in the advertising industry knew these tricks. And then also even in the psychology literature, a lot of things

have been done. But their mark is to show how decision making by human is messed up. It's like what Tversky said, I don't specialize in artificial intelligence, I'm into natural stupidity. But effectively, they are the ones who are stupid. I mean, people in that industry. Not humans. I mean, we've survived doing these things. And also, there's the school of Gehrenser who finds that these heuristics are... Ecologically rational. Are rational, but you don't have to go out of the way

to show that these things are rational. I just don't have to go out of the way to show that these things are rational. I just don't want, my problem is that I don't want the practitioners of that field who understand, barely understand probability to get anywhere near the White House. And we dangerously came closer, but we were during COVID. I mean, first, remember that we had Cass Sunstein, who to me is about as dangerous as you can get. Okay, what I call, actually, I wrote IYI, the Intellectual Idiot,

based on him and Taylor, right? Because I knew Taylor well. Sunstein I met once, but it's sort of kind of thing that sort of like instant revelation, oh, he is it, right? The way they reason, okay?

And so we had these people advising initially against reacting to COVID. Again, misunderstanding of probability. Why? They say, well, this is the empirical risk, and the risk of Ebola is very low compared to the risk of falling from a ladder. They were on it. I remember that article. That was before and when COVID started, Sunstein was advocating ignoring COVID because he said, look how the risk is going to lower. He mixed a multiplicative process with an additive one. And by the way, now, if you'd ask me to figure out the difference, is for me, you get fat tails via multiplicative processes. Not all fat tails come from multiplicative processes,

but you need the, but multiplicative always generates some kind of either log normal or fat tail, but log normal is very fat tailed by the way. Yeah. And at high variance, it acts like a power law. Right. Whereas at low variance, it acts more thin tail. It looks at low variance like a Gaussian. Yeah.. Whereas at low variance, it acts more thin-tailed. It looks at low variance like a Gaussian.

Yeah. It's strange, isn't it? That's log normal. There was an Australian, there was an Australian person, I think his name was Haydee, who spent all his life on a log normal.

Oh, really? Yeah. Are there examples in the real world of log normal distributions? Yeah, of course. There was a big dispute between Mandelbrot and anti-Mandelbrot saying that from

Gibra, you look at wealth, yeah, but what happens with the thing, when you start multiplying, you see, you get a lognormal. Right. Naturally.

And the way it's technical, sorry. No, technical is good. Yeah, so if I take a Gaussian distribution and take the exponential of the variable, you see, because you know that the log is additive, right? Okay. Okay, so when you multiply. So you take the exponential, you get a log-normal distribution. Okay, log-normal distribution. And the mu and sigma of log-normal distribution are preserved. Right.

They're not the mean and variance of the log normal. They're mean and variance of the log of the log normal. Okay. It's misnamed. It should be the exponential. But there was another name called exponential for another distribution. Okay. So Gaussian, you exponentiate, you get log normal.

Now, there's a distribution that's thin-tailed but slightly further tailed than a Gaussian. Barely, right? The exponential,

the gamma, you know that class. Okay. You exponentiate, what do you get? A power law. You see? So you're very,

so which one you're exponentiating, your base distribution needs to be Gaussian for it to end with a log normal. Right. Or

fatter tail than a Gaussian. Okay. And the next class is a gamma over, you know, the exponential.

And you get a Pareto. Right. Yeah. And then, of course, there's an exponential get a Pareto. Right. Yeah. And then of course there's an exponential of a Pareto it's called log Pareto. Okay.

And here as I say you're no longer in Kansas you're not in Kansas anymore. This is a little above my pay grade but it seems to make sense.

So just a couple of final questions on behavioral economics then I want to move on to some other stuff which results in behavioral economics do you think are robust we've spoken about the

loss do they call the asymmetric loss function and prospect theory is there anything else no no nothing else let me think about it No. Nothing else? Let me think about it.

I think that, I mean, we know a bunch of things that are part of that school, but they're not central. For example, how people react, framing, how people react based on how we present things to them. A lot of these things work. But whenever they make a general theory or a recommendation that connects to the real world, they get it backwards. I mean, I took Thaler. I told you, Thaler, all his papers, you interpret them backwards.

If he says, okay, you should uh a concentration okay an optimal concentration of stock you go over one or n you saw my podcast with danny kahneman last year i did not see it i just read the segment the segment where he said that that he accepted that. Uh, I mean, he said it publicly, but he had told me privately. Yeah, I agree. He says it doesn't work in, in, uh, under fat tails. It turned out to be one of his last podcast interviews. What did you make of his answer? I mean, obviously you already knew the answer, but.

He made it public. He made it public. He made it public. Yeah. He said in Taleb's world. I mean, I'm talking about the real world. I don't own the world. I'm not a…

In the world you live in. It's also the world the rest of us live in. But it showed great integrity. It shows integrity. It shows also… No, it shows realism, and it shows also he didn't want to upset me because he was always scared of me going against him. Oh, okay.

You see? Right. Even though he's not on Twitter. He definitely, I mean, one thing about him, I'm certain that he knows everything that was said about not on Twitter. I mean, one thing about him, I'm certain that he knows everything that was said about him on Twitter. Okay. I mean, I'm saying he doesn't believe

he should be up there. He's normal. He himself would tell you, I'm normal. Yeah. I told him, why did you write a book if you know that you have a loss aversion? In other words, one bad comment hurts you a lot more than a lot of praise. He's going to say, I shouldn't have written a book.

That's funny. Yeah, I don't have the same loss aversion, right? I don't mind. I have the opposite function.version, right? I don't mind. I have the opposite function. Oh, really? Yeah. A little bit of praise from people, right, for me is offsets pages of hate.

Oh, interesting. But you definitely have, I assume you have loss aversion in other aspects. No, no, of course, of hate. Oh, interesting. Yeah. But you definitely have, I assume you have loss aversion in other aspects. No, no, of course, of course. But it's not the same kind of loss aversion reputationally. Got it.

Yeah. You see, that's my idea of anti-fragile. Right. Because I didn't start as an academic. I started in the real world. Yes.

I mean, look at it now. I mean, I started, when Gaza started, I felt honorable to go in and defend the Palestinians when nobody was defending them. It took a while for a lot of people to jump on the train. And in the beginning, I had probably 15 people attacking me

for every one person supporting me. And now, of course, that's switched because maybe they found it less effective to attack me. People tend to attack those who can be intimidated. So there's this sense of honor that sometimes makes you feel rewards from saying something unpopular or risky. Right. Worry about integrity, not reputation.

Yeah. I mean, as you advance in age, you go back. If you're doing things right, you go back to your childhood values. You see, your childhood values were about honor and taking a stand when needed.

And then you continue. And every time I take a stand, I feel it's existential. I feel I've done something. Right. What I'm saying is that Danny doesn't have the same representation, and someone complained about him amongst his circle of friends, jokingly. He said, for me, happiness has different value.

For Danny is eating mozzarella and Tuscany. That's his idea of hedonic. Therefore, he analyzed everything in terms of hedonic treadmill. But I'm sure deep down, Danny was not like that. He realized that that was

not what life was about. Yeah. It was more about goals and aspirations and values. But he was an atheist. You know that. And the first time I met him, he ate prosciutto. I told him, prosciutto. He said, there's not a single religious bone in my body.

So then I realized that it's a different customer. Right. And when you're not religious, there are a lot of good things, but there could be bad things. You're too materialistic about your view of the world, and you're coming here to maximize mozzarella and prosciutto. It's very different.

Yeah, it starts to taste a bit boring after a while. So if your sympathy towards biases and behavioral economics was something you changed your mind about, are there any other big things in the inserto that you think you got wrong or you've changed your mind about? No, no, I didn't change my mind.

If you don't, go read the Fool changed your mind about? No, no. I didn't change my mind. Okay. Sorry. If you don't, go read the Fool by Randomness. Yeah. Read it. I mean, that… And you'll see that there's nothing. It's just I changed my mind about one sentence, about praising that industry. Yeah.

Okay. I changed my mind about the industry, but what I wrote about, I didn't change my mind. Okay. Okay. Because I used them for some of the ideas I had when there was no scientific literature on them. But that didn't change my mind. Okay, my whole point, as I started, by humans were idiots when it comes to fat tails, particularly under modern structure because of the way we present probability

to them. And Kahneman liked that. But the idea that humans should… I never had the idea that humans should avoid one over n, should avoid mental accounting, should avoid… Oh yeah, I don't think you'd believe that. Exactly, exactly. I never changed my belief. I never believed in the equity premium puzzle. Sure, sure. So, but I found initially in the industry

things to back up. Although, in the industry, they believe and people who hate tail options keep signing the industry. Right.

Because in that very paper that I like for the convexity of the function, uh, Conaman, uh, you know, shows that people overestimate the odds. You see, so I praised that pad and never changed my mind on the paper. You see, I never, they, she and I, I said, it's completely wrong. It's only because you're clipping the tail that, that, that it shows that the missing tail shows in the, uh, in the probability jumping up. Well, let me ask generally then, are there any big things in the insert

O that you've changed your mind about? Important, important things, nothing beyond the sentence. Okay. Okay. So, so far, I've corrected a lot of sentences here and there. Like, I've removed that sentence. I said something about Tetlock

that I qualified. I said, okay, when he said his study saying that forecasting, people can't forecast. The industry was okay but not the consequences

that everybody drove it to weird conclusions. So I kept taking from the industry that people can't forecast very well. We can't forecast very well. We can't forecast very well. But they never want the next step is that you build a world where you're forecasting.

Doesn't matter. And or you have payoff functions that are convex, where forecasting errors actually fuel expectation. In other words, the payoff improves from that. All right, well, let's talk about forecasting. So I've got some questions about forecasting and then about the precautionary principle, then war, then pandemics. Okay.

So if you had to boil it down, how would you describe the substantive disagreement between you and the broad intellectual project of super forecasting? I, I, I… Is it just about binary versus continuous payoffs? Yeah, there's one thing, it was, first of all, it's the quality of the project, aside from the, and the discussions, they didn't understand our, because've got a bunch of people involved with me and the replies are insults.

So the first one is binary versus continuous. And I knew that as an option trader that the naive person would come in and think an out of the money binary option would go up in value when you fatten the tail. In fact, they go down in value when you fatten the tail because the binary is a probability. So, I mean, just to give you the intuition, if I take the Gaussian curve, plus or minus one sigma is about 68%. If I fatten the tail,

exiting, in other words, the probabilities of being above or below, actually they drop. You see? Why? Because you have the

variance is more explained by rare events. The body distribution goes up. Yeah, the variance is more explained by rare events. The body distribution goes up. Yeah, the shoulder is narrow. Exactly. You have more ordinary because you have higher inequality. And the deviations that occur are much more pronounced. Right. Okay. So, and that we know

so in other words, you're making the wrong bet using binary options or using anything that clips your upside. That we know as option traders and rookies usually or people who are not option traders, sometimes PhD in economics or something,

they always express their bet using these, right? And we sell it to them because it's a net of two options. So, and there's a difference between making a bet where you get paid $1 and making a bet where you get paid a lot. And in Fool by Randoms, I explained that difference by saying that I was bullish. All right. The market, but I was short.

How? Well, I was bullish in a sense. What do you mean by bullish? I think the market had higher probability of going up, but it's the expectation of being I was short. Well, I was bullish in a sense. What do you mean by bullish? I think the market had a higher probability of going up, but the expectation of being short is bigger. So these things don't translate well

outside option trading. And of course, these guys don't get it in forecasting. The other one is the subselect events these guys don't get it, okay, in forecasting. The other one is they subselect events you can forecast because, but they're inconsequential, you see? They're very small, restricted questions. They're inconsequential, so, and also, they're events.

There's no such thing as an event like for example will there be a war yes or no I mean there can be a war it could kill two people it could be a war it could kill 600,000 people

yeah so in extremist that's the one thing one sentence Mandelbrot kept repeating to me there's no such thing

as a standard deviation in extreme extent. Yeah. You see, so you can't judge the event by saying, oh, there's a pandemic or no pandemic. Right. Because the size is a random variable.

Let me give you an example. Right. If there were, if you have scale, that's the idea of having scale-free distribution versus on scale. The ratio of people with 10 million over people with 5 million is the same as the ratio approximately 20 million over 10 million. This is a Pareto. Sorry, that's a Pareto. It's almost how you define it. But look at the consequences

of that. The consequence of that is that it tells you that there's no standard event. Right. There's no typical event. No typical event. You cannot say the typical a typical event. No large deviation. So to give you an idea, if I take a Gaussian, the expected deviation above three sigma is a little more than three sigma. And if you take five sigma, it's a little more than five sigma. It gets smaller. It's above zero sigma, it's about 0.8 of a sigma. As you go higher, it shrinks.

It's like saying, what's your life expectancy at zero? It's 80 years old. But at 100, it's two years, it's 80, two additional years. So as you increase

a random variable, or as an extremist stand, the scales stay the same. Yeah. So the expected life, if we were distributed like company size, the expected company, as I said, what's the expected company? Higher than 10 million in sales? 15 million. 100 million in sales, 150 million. The average.

2 billion in sales, 3 billion. So it's the same as saying, oh, he's 100 years old, he has another 50 to go. How many? He's 1,000 years old, another 500 to go. You can't apply the same reasoning with humans. We know what an old person is. Because as you raise that number, things shrink. For extremist standards, you raise that number, things don't shrink. As a matter of fact, proportionally they stay the same, but in

absolute they explode. So this is why that explosion tells you that there's no standard large deviation. And that was Mandelbrot's sentence. And just looking at the world from that standpoint, that there's no characteristic scale changed my work better than a crash of 87. Because now I had a framework that is very simple to refer to and they are probability basins. So this is why I learned a lot working with Mandelbrot. And people weren't conscious of that stark difference, like operationally. Hence, I wrote this book, Statistical Consequences of Fat Tales.

And this is why I dedicated The Black Swan to Mandelbrot, based on that idea, that characteristic scale that I explained in the black swan, if you use that, then you have a problem with forecasting. You see? Because it is sterile in the sense that what comes above has a meaning. See, it's higher than 10 million, higher than 100 million, it has a meaning. So this is

where. I've written another thing about forecasting, a paper, and I think we insulted Tetlock only

because it's good to insult people who do such work, and also only insulted them because

he spent like five years, you know, that's why I call him the rat. Someone's timing going to go back. So we explained that, and I called it, what did I call it, on the, about a single forecast, a single forecast, point forecast, right,

on why one should never do point forecast for a fat tail variable. What was the title of the paper? Single point forecast for fat tail variables. Yeah, but I forgot what was the beginning, on the inadequacy or something. And in it, I wrote it with Cherry Lowe and Yannir Mariano, who were then active on COVID. We did the data, we published the Nature Physics paper on distribution of

people killed in pandemic. And guess what the tail exponent is? It's like less than one, isn't it? It's half. Yeah, it's less than one. It's like the levy. Infinite men. Yeah. Actually, it's clipped, not infinite men. Some transform becomes infinite men, but that is the same with wars. Yeah, because you can't kill more than a billion people. Exactly, you can't kill more than a population. It attracts for a large part of it.

And if you do a log transform, then it's very robust. Anyway, so we were then involved in pandemics and all the people said, oh, he's super forecasting that how many people will be killed in the pandemic. And I said, no, it's foolish to forecast and it's even more foolish to critique someone's forecast. He misforecast because 95% of the observation will be below the mean. Yeah, it's crazy. So if you have a... It's exactly like my trading.

If 98% of the time you lose money, you can't say, well, he was forecasting he's going to lose money this year. You get the idea. It's meaningless. Actually, on that, it's funny to think that Winston Churchill probably would have had a terrible briar score. He was wrong on all these questions like the gold standard, Winston Churchill, the gold standard, India, Gallipoli.

That's one that's very close to home for Australians. He was wrong on all these calls. But he was right on the big question of Hitler's intentions. So he was right in payoff space, like when it mattered. Yeah, in payoff space it mattered. Yeah, he was wrong in the small. It's like you lose a battle and win the war. It's like versus Napoleon.

Yeah. Napoleon was only good at winning battles. Yeah. And he won, I don't know, if you're numerical, look at how many battles he won. I don't know if you numerically look at how many battles he won. He did pretty well. He did well except for Waterloo.

The reverse Churchill. Yeah, the reverse Churchill. And he's hyped up because look how many battles he won. They were significant maybe compared to the rest. And after a while actually he stopped winning them. It became harder because people learned from him.

So there's one thing about frequency space is a problem because in the real world, you're not paid in frequency. Right. You're paid in dollars and cents. Yeah. It reminds me of that anecdote in Filled by Randomness, the trader who I assume is you, is simultaneously bullish on the market going up

over the next week. Yeah, that was the one I was explaining. But also short the market. Yeah, that was the one I was explaining. In frequency space, I'm bullish. In payoff space, I'm bearish. Yeah. But do these binary forecasts

have some I agree that the value is limited but don't they have some value like I feel like if someone

I haven't seen many functions because it assumes that you get a lump sum I mean for elections the binary

and there's another bias that I wrote a paper on about how to value election to integrate the variance and the price. But you don't have a good understanding of how to translate binaries into real world. And then we discovered another thing also with the binary. In the fat-tailed variable, if you want to get the exact probability,

you see, it doesn't match the payoff. To give you an example, let's say that I have a... I'm good at forecasting the rate of, the rate of growth of COVID. Okay? You cannot translate that

into how many people will be killed. Because the rate of growth is the rate of growth. You see, if you have to translate it the rate of growth is the rate of growth. You see, if you have to translate it in the number of people, you take the exponential rate of growth. You say Wt equals W0ERt.

Okay. And a small error in R can be entailed. But if it's exponential, Wt will be Pareto. You see? So you can have an infinite expectation on W

with a finite expectation on R. This is a problem. We tried to explain it to that paper. It didn't go through. So now what we discovered also later on, and this also applies to something

what I call the VAR-CVAR dilemma, is that people thought we were good at value at risk and not good at CVAR. Value at risk is saying, okay, you have within 95% confidence you won't lose more than a million. And I thought it was flawed because that's not the right way because conditional on losing more than a million,

you may lose 200. Right. Okay. So that remaining 5% is where the action was. But someone pointed out in a discussion group or discussing the answer to Titlok, and then mentioned that my application of the exponential transformation also applies for value at risk. Because he said if you want to get the probability, you don't know the probability is

distributed in centales. Because it's bounded between zero and one. Exactly, it is centaled. It's a frequency. It is entailed. Right. Okay. It's a frequency. It's like a bet. It's a random variable. This is why they have Reier's score, all that thing. Yeah.

But then, the transformation of that probability, okay, outside the Gaussian, okay, you have what you have. You have the inverse,

you see? You want to go from a probability to X, rather than, if you got an F for probability, you see you want to go from a probability to x rather than, if you get an f for probability you see, that transformation of course is a concave convex function

so it is explosive you see so I understand for comparing your approach I guess extreme value theory to extreme value theory okay sorry okay comparing how you think about forecasting or the impossibility of forecasting to the super forecasting approach how important is it as evidence the fact that you have made a lot of money, and as far as I can see, there are no fabulously rich super forecasters?

Yeah, I always say something, that people good at forecasting, like in banks, they're never rich. I mean, you can make them talk to customers, and then customers remember, oh, yeah, he forecast this. But there's another thing I want to add here about the function. If you take a convex function and you're betting against it, and we saw that we were doing Ruri in the same week we had a fight

with Richard Taylor. So I showed something that I showed you in Ruri. Yeah. That you could, if you have a function, let's say that you're predicting volatility, right? And you're an option trader and you're, that was a fixed thing. And the volatility comes steadily, all right, you're going to break even. So in other words, let's assume the level of volatility breaks even.

Now, if volatility comes unsteadily, use your shirt. You can move up the expectation by showing that, hey, you're predicting steadily and you make $1. But the volatility comes in lumps because the way you can express a bet against volatility is going to be nonlinear. It comes in lumps, comes the other way. So I said, okay, I'm 30% overestimating volatility, and I'm making money. All right? He is buying volatility, 30%.

He's selling volatility with a big discount and losing money. So this is where I take that function, and the function is you break even at one. So you have five one and two zeros you make money but if you have uh six zeros and one uh five you lose your shirt in squares so so you realize that that's my my thing about there's no uh i've never seen a rich forecaster so if it came to light in a few decades time that super forecasters had been doing really well not blowing up would that update you in favor of super forecasting we're saying ifs okay let me see i mean i i don't like

these uh uh conditionals right so when So when you see super forecasters find a way to make money outside of being paid, you know, to forecast, but like the function makes money, then it would be very interesting. Okay. But I think that in the real world,

we view the thing differently. You can't isolate the forecasting from the payoff function. Right. So this is what my central problem is. And we tried to explain it to Tetlock. I even brought my friend Bruno Dupierre. Somehow Kahneman invited us to lunch. Actually, I ended up inviting him. Said, let's have lunch. It was Tetlock. He wants to discuss his super forecasting thing. I brought Bruno De Pierre, who's a friend of mine, and his guy has one paper, the most influential paper I think in all of the history, one paper, nothing else.

And it was published in a magazine called Risk Magazine. The guy is, you know, to talk about he figured out, of course, quickly the difference between binary and vanilla and stuff like that. So we had lunch. We realized

I mean, Danny doesn't make claims, but Tetlock didn't even know what we're talking about, right? So, but there's something, how do I know if someone understands probability? They understand probability if they know that probability is not a product, it's a kernel. It's something that adds up to one, right? So whatever is inside, okay, cannot be isolated. It's a kernel. Got it.

You see, it is a thing that adds up to one. It's like saying the densities are not probabilities, but they work well within a kernel. We even had, at at some point people using negative probabilities, just like in quantum mechanics they use negative probabilities. And smart people understand that yeah, you can use negative probabilities because it's a kernel. The constraints are not on the inside, the constraints are on the summation, on the raw summation. Right.

So when you say what is a kernel, therefore it has its properties. Okay? Completely different. So you should look at what you're doing with probability. It by itself doesn't come alone. So you're multiplying within an integral, P of X, with some function G of X. Yeah. Okay?

P of X by itself has no meaning yeah all right g of x all right has some meaning now if you're doing a binary g of x is an indicator function if x is above 100 0 or 1 whatever well however you want to phrase it or it could be continuous could be convex could be concave could have a lot of other shapes. And then we can talk. But talking about probability itself, you can't. Yeah. You can't separate P of X and talk about that by itself. Exactly.

You can't talk about that by itself. Yeah. That's the whole point of a probability density function. Yeah. Density, not probability. Yeah. For the mass function, it may resemble the probability because of the frequency to be there, but it's just like something that has one attribute that is a derivative of something that's never decreasing, and a function that is never decreasing and goes up between zero and one.

So it's a derivative of a function, right? Because you reintegrate to use it so that's the way you got to look at it yeah so and we our option traders don't talk about probabilities we talk about value of the option and the other option is like that part of distribution is valuable because you get a continuous payoff there. Yeah. I've got some questions about the precautionary principle. So I want to stress test it with you or explore its application in practice. So I want to get your take on this critique of the precautionary principle. So the critique would be something like it's possible to tell a story that all sorts of risks

might be multiplicative systemic risks. And ultimately, policymakers need to prioritize between those risks because contingency planning can be… I believe in survival. So if you don't take it seriously, society doesn't survive. I just want a structure where those who don't survive don't bring down the whole thing. Because I think that there are two things.

The precautionary principle as understood, and there's what we call the non-naive precautionary principle that has restrictions on what you got to have precaution about. Because a lot of people couldn't get it that why are we so much against technology? We're not against technology. We're against some classes of engineering that have a reversal effect. And it was a huge standard error. And when I discussed on the podcast or the probability book, whatever you want to call this one with Scott Patterson discuss the mouse story well it was caused a great family famine was

trying to get rid of sparrows sparrows yeah okay and then they killed all the sparrows or they tried to kill as much further they could that sparrows eat insects right so they had an environmental problem with insects proliferating on the strain, and they didn't see it coming. Now you say, okay, this is a case, this clear cut of disrupting nature at a large scale, and something we don't quite understand.

This is exactly what our precaution is about, except that we added multiplicative effects. Like, we don't exercise precaution on nuclear. This is why we're trying to, the way I wanted our precautionary principle to work is to tell you what is not precautionary. And for us, nuclear was not precautionary. Why? Because you can have small little reactors and that one explodes in california doesn't impact one in bogota the harm is localized

exactly it's localized so unlike pandemics yeah hey everyone this is joe i want to give a quick plug to my weekend newsletter before we return to the conversation. Every weekend I send out an email with a bunch of links to things I've been reading, watching or listening to. Some of the links relate to research for upcoming podcast episodes, but more often they're just random interesting things I've discovered during the week, papers, articles, videos, etc. Now, according to the platform I use to send these emails, each weekend about 20% of my mailing list clicks at least one of these links, which if you know anything about email marketing is an extremely good click-through rate. But there are only a few thousand subscribers on this mailing

list, whereas my podcast audience is an order of magnitude bigger. And that tells me I need to do a better job of telling you all about this newsletter you're missing out on. As I said, it's very high signal. I only share stuff I've actually consumed during the week and that I think is worth sharing. I don't force myself into a template where I give you the same number or type of links each weekend. Sometimes you'll get three links, sometimes you'll get 11. It's purely stuff I've actually been reading and usually you'll learn something. To sign up, go to my website, jnwpod.com. That's jnwpod.com and click newsletter. Okay, back to the conversation. So to focus on technology, my understanding is that you wouldn't seek to apply the precautionary principle to the development

of a technology that could pose systemic irreversible risks just to its deployment because otherwise you would be going back and and like setting fire to mendel's p plants because that knowledge could ultimately lead to GMOs. So there's obviously got to be a line… No, we're against implementation of GMOs in nature. We're not against research about whether you can modify something. You can't stop research. People can do research.

Yeah, got it. people can do research. Yeah. Got it. So applying that to artificial intelligence, obviously as the technology currently stands, it doesn't warrant application of the precautionary principle because it doesn't impose systemic harms. If we got to the cusp of the technology being able to recursively self-improve, which the most plausible way that would happen is that we could use AI to automate AI research itself.

I have problems with discussing AI in terms of precaution because I don't immediately see anything about AI, why you should stop AI, that it will self-reproduce given a robot cannot climb stairs. So you're afraid of robots, scale of robots, multiplying and becoming a robot colony that will take over the world. I mean, these things are a stretch of imagination. We have bigger problems to worry about. I don't think most people who think about AI risk view robotics

as a constraint. So what is it? Because... Technology would... The whole thing would become risky if technology becomes autonomous. Right. So in other words, that's my understanding that that's what they're worried about.

And it becomes autonomous. It has to, first of all, you can shut down your computer, right? And it no longer impacts our life here. You can't hit the water because it's down the computer. The other one, for it to be systemic and taking over the whole planet the information systems i mean this is very strange that people could not understand the geomote threat are now obsessing over ai because it's tend to surprise them when they uh when they ask you the question it tends to be if you're surprised by AI, you have a problem.

It means maybe that's, for me, an intelligence test to figure out what AI can do or cannot do. Okay. There's a lot of things it can do that helps. Okay. Okay? But for it to become a, how can I, autonomous, in other words, a colony of just like humans, like biologically equivalent to humans, you have so many steps to make.

Yes, but all that needs to happen is the first major step is it needs to automate AI research itself. And then as soon as it can make itself smarter through recursive self-improvement, all the other problems

like robotics become much easier to solve. Okay, let's see if it can do that. Okay, but if it could, let's worry about it. Then you put constraints. You can't put constraints ahead of time on a research. You've got to worry about an event happening. Okay. I mean, you've got to see or talk in speculative.

Okay, one quick final side note on AI. A lot of people have remarked on the fact that LLMs haven't produced any original scientific insights. Yeah. And that may be because they're fundamentally Gaussian. Have you thought about? No, no, it's not. That's not the reason. It's because they are, they may actually produce insight because of the randomizing stuff and may make a mistake one day. Right. But so long as they don't make mistakes,

it's just representing what's out there. Yeah. It's probably a weighted thing. Okay. As a matter of fact, it's the reverse of scientific research because how does LLM work work it works at reflecting what makes sense all right probabilistically so i try to trick it by asking it uh you saw on twitter in the beginning

say okay how i'm gonna trick it because that's if you know how it functions and and again thanks to uh my uh genius friend wolfram i got how is it i got this blog post he sent me i read it and i got the book said okay now i know it works all right it works by probability matching by the way all right it doesn't give you the same answer all the time and it's not going to do all the homework so it doesn't have to connect the pieces directly. So use probabilistic methods. So that's what reflects the consensus. So I asked it during the Congress of Berlin. There was a war between the Ottoman Empire on one hand, and then you had Greece on the

other hand, among other allies. And there was a fellow, Kara Theodori, who was the father of the mathematician Kara Theodori, who was representing someone there. Who did he represent? They said, oh, he's a foreign affairs minister of Greece. You see? It's not like a search engine giving you facts. It is using probabilistically how things occur together. He has a Greek name. Therefore, in fact, he was representing the other side, the Ottoman Empire. As a matter

of fact, it was, I think, the Victorian days, that he said, oh, meeting with a representative of the Mohammedan world. It was an article in the Times, and his name was, he had a Greek name. I think, is it Constantine Caratheodorus? Or his son is Constantine, whatever. So I asked Shantipati, he made that mistake. So how do you make money in life? How do you really improve? How do you write a book? How do you, okay.

Think people didn't think about. Because if you're going to start a business that makes sense, guess what? Someone has thought about it. Okay? I think something. And strategy PT is designed to tell you what makes sense based on current information.

Right. Not look for an exception. There may be a possible modification. I don't know. To make strategy PT only tell you what makes no sense. And that would hit one day. It's like our usual adage at Universa is if you have a reason to buy an option, don't buy it.

Because other people will also have the same reason. So it's the same thing with starting a business. You're not going to make money on a business that makes sense. Because a lot of people have tried it. Maybe some pockets here and there. People have tried it. So the idea of strategy PT coming up with genuine insights is exactly the reverse of the way it was modeled.

And like everyone, it was vague for me until I saw the book by Wolfram a couple of years ago, two summers ago, or last summer. It was... The guy is very clear. He thinks like...

He's very systematic and extremely intelligent. I never met anybody more intelligent than him. Yeah. I did a four and a half hour podcast with him last year yeah in connecticut and it was one of the more surreal experiences i've had really the guy is you write down the formula he gets it right away he understands things uh like effortlessly yeah and his his intellect isn't domain dependent he can apply it across all aspects of his life

yeah I mean I don't know I don't wanna but like he thinks about business really well he has a

he has a business yeah but he's regimented in the way he operates and collects data on himself sorry the way he collects data on himself.

Yeah, no. But anyway, so he's, I mean, I enjoy hiking with him once a year. And I, anyway, so thanks to him, now we have an idea how these things work. Okay. It was clear. I mean, maybe there's some other text but but if when when i if i need the text i'd rather read his treatment yeah because of the way uh i got used to thinking and also

because i don't haven't seen the quality elsewhere yeah it's a great book is uh primer on llms so i have some questions about war some questions questions about COVID, and then we're finished. Yeah. So one of the deepest things I've picked up from you in recent times is the concept of the shadow mean. And I guess the intuition here is that we have some historical data for some phenomenon, whether that's market drawdowns or deaths from war

or deaths from pandemics. And those data can appear to follow a thin-tailed distribution, but it's naive to assume that the process that's generating them is actually thin-tailed because in the background and behind the curtains of reality, it could actually be a fat-tailed process that's generating the data. It's just that it takes a really long time for extreme events to show up. So fat-tailed distributions can masquerade as thin-tailed ones.

And bringing this to statistical moments, the mean of the data we've observed is better thought of as the sample mean. And you have this approach where you work out what you call the shadow mean, which I guess is equivalent to the population mean. That is the mean of the process that's actually generating the data. And you've done this for warfare and I want to talk about that specifically but just first generally for others who may want to explore this approach can you outline the basic steps in your process is it number one estimate the alpha

number to plug in estimation let? No, no, no. Let's explain to the viewers or listeners what do I mean by shadow mean. Let's take a one-tail distribution. You have visibly in a sample of 30 automation, you're not going to get events that happen less than 1% of the time. You agree? Yes.

So for a Gaussian, it's not a big deal because these that happen less than 1% of the time have less impact on them. The probability gets increasingly smaller, so it doesn't matter much. So with a small sample, you don't have a big shadow mean effect. Actually, with a Gaussian, it has to be a one-tailed gaussian so so a low variance like normal right like height okay so you observe a bunch of people and you have an idea what what the average height in town is okay

now when we talk about things that are open-ended and fat-tailed, visibly, most observations will be below the mean. So when you compute the mean, it's going to be biased down from what they call empirical observation. So the empirical distribution is not empirical. And that's what is central for us. So I take the S&P 500, and you can figure out that

if you want to stress test it over the next X days, taking the past 10 years low, the worst deviation past 10 years low, the worst deviation the past 10 years is not represented because of insufficient sample as you go further in the tail. You take industries

like biotech, for example. It is a heavy-tailed industry. So what you observe is less than I think I wrote it in a black swan, the observed mean underestimates the true mean. Whereas for insurance, it overestimates the true mean. Right.

For banking. Because one is to the right, one is to the left. So I looked at what has a positive shadow mean and what has a negative shadow mean. If you're selling volatility, you have a

shadow mean that's going to be way lower than your observed mean. But if you're talking for wars, even without survivorship bias, which is another story, we have a process that's vastly nastier than what we observed. About three times nastier. Okay, three times nastier okay three times last year yes so in other words um the the historical process underestimates the true process and and we published in uh we published about the shadow mean in in in various venues we have a paper in in physica a on wars but we applied it in

quantitative finance to operation loss. I published a journal called Quantitative Finance and we applied it to other domains. But that's an idea that I wrote about in The Black Swan. But only where is the invisible?

Because visibly by definition, the 100-year flood is not going to be present in five-year data. Okay? So you have a shadow mean if you limit it to five years. Yeah. So the other big innovation of the work that you did on war was this concept of inter-arrival time.

And if I remember correctly, the waiting time for wars with deaths above a threshold of 10 million people is a bit over a hundred years. Yeah. And that means that because we haven't, just because we haven't observed any, like the last, the last conflict with deaths of more than 10 million was World War II,

nearly 80 years ago now but we can't infer from that that violence is declining the client plus another thing that we discovered that's very robust is inter-arrival time is has an exponential distribution. Like a poisson, you know? The inter-arrival time of poisson, it means it's memoryless. Right. In other words, if it arrives on average every, say, 100 years, and then we haven't had one in 100 years, you don't say, oh, it's coming.

It's memoryless. So you wait another 100 years the expectations stay the same yeah yeah so what structural explanations do you think are generating the fat tautness of war is it just the development of increasingly destructive technologies and then maybe also some globalization and the fact that violence can spread memetically i don't i mean i i looked at the data i reflected the data violence did not decline i did not put my concerns and my concerns that in the past to do what's being done in gaza now required much more so we have a lot more destructive the ability, I mean,

to kill is greater. In the past, it would take a long time to kill so many people. You have to do it manually. And now we industrialize the process, which is very sad.

And then I have started branching out into foreign policy, realizing that effectively there's some things in that SGD, Society of Judgmental Decision Making, when they analyze the Vietnam War, and there are a lot of good things

in that industry. And all the biases. You realize that we have the United States, the most dynamic country, very vital, was completely incompetent State Department. So you realize the decision for war. I mean, think of Afghanistan, how naive it is not to figure out

what's going on. So, they're going to make mistakes, of course, more mistakes, of course. And these alliances, like you back up

not understanding consequences. So, it's sort of like Mao's sparrows. You back up bin Laden not realizing that you helped bin Laden, you built a machine that will turn against you.

Right. It's like the Hydra. Like? The Hydra. It cut off. Yeah, yeah. No, no, but they created it. So if an interventionist foreign policy on the part of the United States

and then it involves spreading democracy and stuff like that, it's actually more dangerous than just isolationism. So the culture is very different today. Right. Which is why, you know, outside of our statistical work, I have to say that there's this incompetence, rest, and sophistication that makes the world more dangerous. So then if we move back through the historical data,

the wars become less fat-tailed as you move into the past? No, the fat-tailness is the same, what we call the scale. The alpha doesn't change, the scale changes. So I think one of the things that you and Professor Pasquale Cirillo found was that in the past death counts were exaggerated both because conquerors and victims had incentives to exaggerate. Obviously the conquerors want to appear more intimidating. No, no, no, no.

I made this comment later on after looking at the data because when we analyze past wars, we try to figure out a robust way to look at the structure of the random variable by taking for every war different accounts, and then randomizing between the high and the low. Say Algeria's war, the French had 280,000. For example, the Algerians had 1 million. Since then, everything has been revised.

So we took both numbers and randomized. So we created 150,000 histories between all the numbers that we had with permutation from within, the high and the low estimate. And we figured out that, boom, they all give the same alpha. Right. So we were, we, but the motivation was that people lie about numbers.

And do that. Is that true? And ours is to remove the effect of different estimates. Yeah. Them or their enemies, you see. Okay. So aside from that and the non probabilistic way I myself observed that a lot of people like to exaggerate

their killings yeah like Genghis Khan because it was optimal mm-hmm you know you don't have it if people think that that you're gonna kill a lot of people they won't oppose you so which is why you do a lot of stuff for show. Yes. A lot of devastation for show. Yes. That makes sense. Victims exaggerating their suffering was less intuitive to me,

but then I remembered Tom Holland's work or Rene Girard's work or even your treatment of Christianity in Skin in the Game. I realized what makes Christianity unique is the valorization of the victim. work or renee gerard's work or even your treatment of christianity and skin in the game i realized what makes christianity unique is the valorization of the victim christianity and shiite islam right only the two religion yeah that uh that that that have this uh glorification of victimhood yes which is is christianity and shiite islam yes shiite islam when they have a martyr you know like and there's still been but after the murder of uh hasan and hussein you know 1300 years of mourning or stuff like that glorification

basically for for just being killed yes so i I was wondering if the glorification of victimhood, if the spread of Christianity is maybe what was driving the exaggeration of death counts on the victim side? I don't know. We don't have good records of what happened in the period right before Christianity dominates, simply because we had a big transition, and history is written by the winners, of course,

by the Christians. So we don't have a clean record of what happened before, but we know that there are some purely invented, fabricated series of events of martyrdom in what's now North Africa and Southern Mediterranean and Roman Southern Mediterranean. Yeah.

So we know a lot of them existed and we know a lot of them didn't exist or exist the same story in 17 different places. Right. Or 7 different places. So we know that it either existed too much or did not exist. Yeah, yeah.

So one of the implications of your work on war with Pasquale is that because of these inter-arrival times, we really should wait about 300 years without seeing a conflict of the scale of World War II. Yeah, if you had to wait 300 years, then you'd say, oh, the distribution has changed. Yes, then we could say... But we have had no information statistically

from the past 80 years. Yeah. And that was the thing about Pinker thinks that the world has changed and he couldn't understand our insults. Just like Tetlock, he couldn't understand the statistical claim against that.

Yeah. So you think that, I mean, it's possible that the data generating processes could change. It's just that we haven't seen anything that would overturn the null hypothesis. That's exactly the point. That's one way to look at it. I don't like the null hypothesis story

because that's mostly for applied statistician working in the medical lab or psychology department. But the gist of it is there. That's the intuition. Yes. And so we have no statistical grounds on which to say violence has declined. None. Yeah. And we don't even go to the second step.

I've seen it has increased, which is what I saw, but I don't want to make that point statistically. Yeah. Well, it's super interesting and important work. I want to talk about COVID. So, oh, actually, sorry. work i want to talk about covid so oh actually sorry one one maybe can i just ask you one technical question on the the war stuff before we move on so i'm not sure if this is an interesting question but let me test it on you so generally how much does it change the conclusion of analyses like yours with Pasquale on war if you impose soft ceilings

like the eight billion deaths? Zero. Okay. Because you stress tested it for war. No, no, no. That soft ceiling, you mean it's only an artifact to show that in log space, it is a power law. But you have to go very up to 5 billion

doesn't make a difference whether it's ceiling or no ceiling. Okay. For both. Yeah. It doesn't make a difference because the ceiling is continuous. It's like a log function that turns the maximum to infinity. Okay. Okay, but it only happens asymptotically.

Okay. Okay. All right. Yes, I want to talk about COVID. So in late January 2020, you wrote a memo with Yanir, a mutual friend. Yeah, it started, yeah.

I mean, Yanir and I were friend. Yeah, it started, yeah, and, I mean, Yanir and I were concerned about Ebola before that. Yes, back in 2014?

Yeah, we were obsessing over pandemic because I wrote on the Black Swan. Yeah. And it was picked up by a bunch of people

in Singapore. So, we were like all concerned about, you know, the big pandemic because it would travel faster than the Great Plague. So this is why we were very concerned when it started

and we wanted to invite people to kill it in the egg. And you wrote this memo which was then shared with a friend in the White House. Can you tell me the story of that? Is there anything you can share that you haven't shared publicly before? No, no. The paper by itself is meaningless because we would have written one in your and I separately.

But there was no particular novelty to that idea. Sure. But when we started seeing what's happening in China, we realized that there was a problem and then I start looking at ways to how do you mitigate something that's fat-tailed

you lower the scale how do you lower the scale? by cutting off the distribution to parts reduce connectivity reduce connectivity and it's very strange that the uh trump administration

did not they I mean they spent all this money all right I'm giving money handing out money all of that and then hit him that that you're most effective by having uh controls at the border or you test people I mean, in the past, we used to have very effective lazarettos where people were

confined or quarantined. And now, we can do it more effectively with testing. Do you think your memo with Yanir is what convinced the White House

to close the borders to China? I don't care less about the White House. There's something that disgusts me about the Trump administration. I don't want to.

You just do your duty and you move on. Do you sense that governments and policymakers say in the US have gotten any better at thinking about how to deal with tail risk? No. I think if I'm saying their effort to deal with risk, increase tail risk, because you end up with people like Cass Sunstein and these pathologizers, I call them. They make you stupid for worrying about things

because their textbook tells you they shouldn't worry about it. And they don't understand fat tails. Once you understand fat tails, things become very easy. You start thinking differently about AI, differently about other things. You see?

I tell you, yeah, once AI stops multiplying, let me know. And stuff like that. This is my department. You see? I'll tell you, once AI starts multiplying, let me know. All right? And stuff like that. This is my department. Fat tails and precaution requires fat tails.

Yeah. I mean, you can have precaution at different levels, but the one we're concerned with at a higher micro level

requires fat tails. Do we need any new social institutions to better deal with FATELs? I have no idea. Okay. At this point, I'm too disgusted with these bureaucrats

and the way they handle both sides. Via negativa. The way they, exactly. I mean, you want a simpler world. Yeah. It creates a complex world,

institutions that make it more complex. Sort of like you ask foreign policy. You go to Afghanistan, then you have to handle the government of Afghanistan. So it's like you get involved into a series of feedback loops you never thought you'd get into. Yeah.

So, Nassim, I'm finished with my main questions. I had a few random questions. Let's continue, yeah. It's just a random sampling of different things I don't do the podcast and interviews so well I very much appreciate you speaking with me so okay what's the biggest thing most people in social science get wrong about correlation that's an important question they don't know what it means. I mean, there are SGD people who really think that experts have a problem, and there are good results there.

And they ask the people to do the regression. What does it mean? And they can't explain their own results. They the equation they couldn't explain the graph how much this represents that so uh there are a lot of incompetence in in social science and they use metrics they don't understand and like people a lot of people thought correlation was was an objective thing it's a measure that depends on some sample and then has a very limited meaning and also they don't realize that when or visually, that correlation of 50 is not halfway between zero and one.

It's much closer to zero. You have this saying, so people are familiar with the phrase, correlation isn't causation. You have this phrase, correlation isn't correlation. Yeah, exactly. I had a lot of Twitter fights with people, and that was fun because I didn't know that people didn't think of correlation that way. That's another thing. If you look in finance, naively,

you see that the effect of the mean of correlation, it appears to be like, say, x and y are correlated. Your expectation of delta x is going to be rho sigma x over sigma y based on delta y

you're linking you observe the effect of x based on observation of y but for betting and decision making it's not that it's more like a factor that's

something like rho square or 1 minus rho square or like similar to the minus log 1 minus rho square so in other words very non-linear in other words low correlations are noise or like similar to the minus log of one minus row square. So in other words, very nonlinear. In other words, low correlations are noise. And again, 50 is not halfway between zero and one.

Right. And one is infinity. That's for decision making. And you put that into your either Kelly criterion or any form of decision making. And then you realize how much more I should bet on X knowing Y or on something knowing some side information

and simplify it. When I made a graph showing how it has this how visually you can see it mutual information which is an entropy measure is vastly more It has this, how visually you can see it. Mutual information, which is an entropy measure, is vastly more informative.

That's in a linear world. And now as you go nonlinear visibly, if you have a V curve, zero correlation and infinite mutual information. So that mistake was correlation. But there are other mistakes in correlation not well explored. I didn't go into it because I'm into cycling. I'm too lazy to go into it. But I showed

that basically it's not it's sub-additive. To give you an example, if I take a correlation of a row, it's not going to be rho in the positive. If you sum up the quadrants,

positive, you know, x positive, y positive, x negative, if you sum up the quadrants, you don't get rho. Because visibly the mean shifts according to every quadrant. So it's going to be sub-additive in absolute terms.

Which is a problem. It tells you that sub-sampling taking a correlation of sub-sample that will give you a correlation of the whole and that's not well well known and i wrote a paper i don't feel like publishing it because the problem with referees is it's hard to get good referees so on the last paper we had a guy says tell me i'm substituting correlation with mutual information and say do you have evidence that correlation is a metric you don't say you have evidence scientific evidence that correlation works it is a metric right by definition so you can use it for evidence. So I said, okay, you've got to give up on publishing too much because of contact with referees who are not sophisticated unless you find the journals that have the good referees.

So maybe I'll publish these results because the practical implication is monstrous. And maybe I'll put it here. I'm on the second edition, third edition. I add correlation. Smart people get it. Smart people. But you have to know math to know that correlation is not what it means.

Right. And then your regression. The regression was an R square of 0. Right. And then your regression, they do regression with an R-square of 0.05. And they think anything above 0.5 is kind of celebrated in social science. I see,

but the problem is if you include model error, okay, it dilutes to 0.5 big time. Right. It's crazy. I mean,

there's just so much of social science is built on correlation. Exactly, and it is. It's so huge. Plus, the other thing is how to translate a result. Let's say that you see papers. You see a huge cloud, okay, and it tells you, oh, look, IQ and education,

or IQ and wealth, all right? Okay, very good. Or IQ and income. First of all, it's wrong. Income is fat-tailed. IQ is, by design, sent-tailed. So you can't regress them. Yeah.

But let's say we did that. They got a big noise. In other words, if you hire someone based on IQ, you get such a low probability in your favor for a sample of one. You need large numbers. They don't get it. So, you know, oh, you should hire or no, because

with such a weak correlation, the law of large numbers doesn't start acting until you hire a whole town or something. You see? You get the idea. You're getting noise. So you're getting noise.

So that metric is noise unless you have wholesale. Yeah. Because of visual variations. Yeah. So the way the law of large number works, I explore it here, even for thin tails, it's misunderstood. What I call the reverse law of large numbers

if you take a property say how much hypertension is going to be lowered by this medication and reverse it and look at what are the odds of it working on your patient you get a completely different answer from the one they think because on average

it works say 4, four points. But some people, it's going to be a lot higher and so forth. So this is where the interpretation of the statistical claims that they're making, it can be messed up. I mean, I saw it in the IQ. First of all, they don't know how to compute a lot of things, and they don't know how to read correlation, but also how they interpret it. We're going to tell them, okay, put a graph with a noise. And you see a graph and you realize, at the best of their claims, in these papers that show the effectiveness of using IQ, even with the circularity.

In fact, if you're going to take an exam, you're going to have a high IQ, but you're also going to get a good college degree, and that helps your income in the beginning. We're not talking about wealth or stuff, so it's for employees. So even taking all of these, you look at the cloud and say, well, you know what? You can't use it for any individual hire. You need a batch.

And then they start. There's a lot of other things in IQ that tells me that either these people, I used to think that they're mean. Like, in other words, like a lot of race science comes from people being, you know, having some kind of problem, all right? Sociopathic problem. So I thought that, but I think, no, that's just plain dumb. And you can see it in the real world. Think about it. If these people know anything,

they'd go make money and then go continue doing psychology, but they can't. It's very true. Okay, next random question. Yeah. Maybe you know the answer to this,

maybe not, but historically, culturally, how do you explain the perspicacity of the Russian school of probability? What was in the water in Russia? No, they, I mean, schools emerged when you start having norms and groups of smart

people together. And there's a depth in Russian approach to mathematics. But during the Soviet, they had to make themselves useful. You know, science had to contribute to society. So they can be remarkably practical while at the same time there's that constraint. And I mean, a lot of it is French as well. I mean, when you of it is French as well. I mean, when you look at the big results,

you always have a combination. But I think the Russians have contributed the most to probability. Followed by, of course, the French. And, of course, the English school of probability is just like Galton. And all these regression, all these things that are bad come from this English school of probability. And usually, they have an agenda. Like Galton wanted to prove that Irish were stupid by measuring the Ukrainian.

Right. And the linear regression, the hypothesis testing, the Fisher thing, all these are completely different. Yeah. But one is probability. The other one is what we call standardized statistics. But you cannot go at non-standard statistics without knowing probability. So we have a class of people who can only use Gaussian. And I have this theory that every single problem

needs a new class of estimators adapted to the problem. That seems like a pretty good heuristic. Yeah. So if you don't know how to redo an estimator, how to redo the theory. Yeah. You see? The only thing in common is a lot of large numbers. That's it. Right.

And you want to know what it applies to. So when you ask me something about the alpha, the law of large numbers sometimes works a lot better for the alpha than that's what I mean. Yeah, because the, the, um, the tail exponents follow a thin tail distribution, right? It follows an inverse gamma distribution. Okay. And you get it.

It's a process that's clean. Which is a specific type of thin-tailed. Yeah, yeah, yeah. And if you get it, if the process is clean, Okay, you have a... It's remarkable how quickly you get the alpha. Yeah, that's cool. I showed you at Ruri,

reversed, try to get the means all over the map. Yeah. You get the alpha always within like... Yeah, it's really neat. It's really neat. Yeah. Standard error on the alpha is low. Yeah. You got the alpha always within like... Yeah, it's really neat. It's really neat. Yeah.

Standard error on the alpha is low. Yeah. Standard error on the mean is huge. Yeah. Yeah. So you think Hayek's knowledge argument can't support prediction markets. And obviously Hayek argued that knowledge was consolidated through prices and arbitrages, trading products, services, financial securities. Yeah. argued that knowledge was consolidated through prices and arbitrages,

trading products, services, financial securities. Is the principal difference there just that these things that Hayek was considering were continuous and that logic can't be extended to aggregating binary forecasts? Or what's the difference? Hayek's idea is that no, it's more explicit versus implicit. That for him, knowledge is not explicit, that it's implicit. The difference between knowledge that can be taught

and formalized and knowledge that is embedded in society. And that one expresses itself through the manufacturing and then the end price. And why a systematized economy, you're systematizing something that is not explicitly, led itself to explicit phrasing, is what harmed the Soviet. So I would not I would not I would say that this applies to probability the wrong way for you, which is that

using a probabilistic model is trying to be systematic about things that are too rich for you to express them systematically. So in other words, his knowledge is what's embedded in society, not what is formalized. Otherwise, the Soviets would have taken the formula and applied it. Okay, maybe

I'm too slow today, but so how does that preclude extending the knowledge argument to prediction markets? Because we're not just talking about prediction. We're talking about function predictions. Okay.

They're all embedded. You can have what appears to you a bad predictor in frequency space, but the function turns out to be better. Got it. See? And you don't know the functions. It's still systematizing something that should be you know,

not, I mean, you should look at the result of the process, not the exact replication of that process in the lab environment. Yeah. Okay, I'll ask my last random question. So, I know that generally

you prefer mean absolute deviation to standard deviation. Why has standard deviation become such a traditional measure? Like historically, how did that happen? Okay, because I think I discovered here a paper claimed by Fisher, I think, who found that in the Gaussian case, it's more efficient than mean absolute deviation.

Because, again, to tell the viewers, a lot of people mistake one for the other. Standard deviation is the square root of the average sum squares. It doesn't have a physical intuition.

What a standard deviation is what is the average so for example if you have the process right with all the observation

at zero and and one observation at a million for an average of a million the standard deviation be 500 times mean deviation

right and the Gaussianussian world is about 25 percent higher like square uh square root of uh you know uh the usual square root 2 over pi is mean deviation of standard deviation got it the so this is the i would i would i would say that it's another basic thing, that a lot of people, we wrote a paper. People don't know what we're talking about when we talk about volatility because they would use, we're talking about people who are practitioners and people who are students, PhD students in mathematics,

of finance and then we asked them to try to interpret some kind of financial data where you're showing standard

deviation volatility and then they would give you mean deviation interpretation so yeah

yeah it's more intuitive than standard deviation yes yeah no so there's a wedge

both of them the fat tails, the way I'm interested in the measure, not because of, you know, to pick on practitioners who make mistakes, but because the ratio

of standard deviation and mean deviation is the best indicator of fat tailness. Yeah. See? Yeah. And for Gaussian,

it's, I said, 25% higher yeah see yeah and for gaussian it's i said 25 higher for for for for koshi is infinite yeah not infinite i mean for something that has uh not koshi uh anything with with an alpha below two it's gonna be infinite because one is infinite, the other is finite. Final, final question. Is there anything you can tell me about your next book, The Lydian Stone? I have no idea what my next book,

what shape it will take. For the last three books, last two books, Skin and Game and this one, I had no conversation with them. I've just finished the book yeah and i don't like this so you know you gotta write a plan people get excited yeah yeah yeah all that i'm i'm working now on really uh the difficult work so next book has to lose time with time scale and uh and and probability okay there's a lot of entropy stuff in it but but i'm at a point where i'm writing for myself now yeah what what makes the most fun that's great there's nothing more fun than this because you know an hour two hours

day of math you feel rested after that. Yeah. You see? So I'm doing more math. Great. Well, I wish you much more math and much more enjoyment. Yeah, but I don't want to be identified, and I don't want my grades to say I'm a mathematician. I'm just enjoying using it for problems that are non-mathematical in nature. So it's not like I'm trying to improve the math, I'm using it. But math is fun and relaxing. So this is why I like it.

Yeah. Well, Nassim, you've been so generous with your time. Thank you so much. It's been a real honor. Thanks, thanks. Thanks for inviting me. And hopefully next time we do a podcast you reverse. You start with random questions and then you go to structure. Okay. Sounds good. That's more Hayekian. Thanks. Bye everyone. Thanks Nassim. Thanks so much for listening

to my conversation with Nassim Taleb. Two quick things before you go. First you can find the episode video as well as transcript which is full of hyperlinks relating to the concepts we discussed on my website, jnwpod.com. That's jnwpod.com. Second, as you can tell, researching for and producing this conversation took a lot of work. If you'd like to help me out, the best thing you can do is share the podcast, whether that's messaging your friends or sharing a link to it on Twitter.

The main way my podcast grows is through my audience. So I'd really appreciate your help. Thanks again. And until next time, ciao.