The Age of Prediction | Igor Tulchinsky and Chris Mason

[00:00:06] I think the biggest job of the next generation, like this is what I would do if I was figuring out what to study and train in, is predictor.

[00:00:16] Like we've seen examples of this. We saw the movie Moneyball where they used data about baseball players to predict which baseball teams would do the best. And I used to write algorithms for the stock market to use data to pry to predict how stocks would act.

[00:00:33] And we're going to talk about that more today. And people use data also now to analyze the genome to predict what diseases someone might have and on and on.

[00:00:42] But the technology has gotten so good, so fast. I'm happy to talk with the two authors of the book, The Age of Prediction, Christopher Mason and Igor Tulchinsky.

[00:00:57] Now Igor is interesting from the financial perspective. He has a $7 billion hedge fund which analyzes millions of pieces of data around the world to predict stocks, to predict simply what's going to happen with stocks tomorrow or an hour from now or 10 seconds from now.

[00:01:15] And Christopher Mason uses data from the human genome to study what diseases we can start curing, what sicknesses someone might have or what traits someone might grow up with.

[00:01:28] So I want to know what is the state of this industry? Like how much can we really predict? How can we get better at it? What are the limitations?

[00:01:35] And then we just had a fun time while I pitched different ideas. So here's Igor and Chris, authors of The Age of Prediction. This isn't your average business podcast and he's not your average host. This is The James Altiger Show.

[00:02:08] You know, Chris, I'm amazed at all the things you were able to discover about people by analyzing their genetics. Yeah, basically it's a predictive algorithm that's in every cell. So you have bits of DNA, RNA, proteins and you leave these everywhere you go.

[00:02:26] So I think of the forensics chapter you're probably thinking about maybe or even just the cancer diagnostics we can do with DNA. It's extraordinary. So the book is called The Age of Prediction, Algorithms, AI and the Shifting Shadows of Risk is the subtitle.

[00:02:39] If I can try to summarize at a 20,000 foot level, it seems like there's three types of prediction. One is where you use statistics slash AI to model things that can either be very predictable or somewhat predictable,

[00:02:54] like things ranging from insurance risk to cancer risk using genes to stock market predictions. Then there's the kind of prediction where it's just someone's opinion like, you know, Rifkin analyzing what the economy is going to be like five years from now.

[00:03:12] And then there's the kind of prediction where it's like pretty definite. So the solar system is going to eventually collapse. I can predict that and I know 100% chance I'm correct. Would you say that's roughly three categories of prediction?

[00:03:26] Yeah, based on facts based on optimism and based on reality. Yeah, I like that. What about based on pessimism? It's the same as optimism, right? It's just a flip the side.

[00:03:39] It just goes negative direction. But you can still use pessimism and look at the same facts and view it as, you know, it could almost be depressing to you if you think, oh, we're doomed because the sun will engulf the earth in a few billion years.

[00:03:53] But you could think, well, no, that means we know when we got to get moving by. It could be exciting and get you moving.

[00:03:59] Exactly. So you point out the exponential growth in data. Maybe you could describe that a little bit like how much more data we are generating now than even 10 years ago? And what that means for prediction?

[00:04:12] Well, I'll jump in. Like, as first is I think the amount of data certainly in genomics and just the ability in biomedicine to generate data is what's often been called at least in genetics is genomic amounts of data.

[00:04:24] Most people think of astronomical amounts of data as being really big and involving exabytes or yodobites of data, which is trillions of terabytes. But it's actually genetic data and genomic data are now eclipsing the amount of data made by telescopes and astronomy.

[00:04:41] So there's a paper that just described is called, is it genomical data or astronomical data? Which one's bigger and concluded that there's actually more genomic and biomedical imaging data than there is astronomical data.

[00:04:51] So when you think of really large, you have to think of things in trillions of terabytes, not quite today, but in the near future. And that really basically means every day that you wake up there's more data than any other day in human history.

[00:05:04] Let's do a little thought experiment. So let's say I wake up and now I can sample my DNA in seconds and do some diagnostics like right away or some AI or some statistics can do some diagnostic right away.

[00:05:18] What theoretically could I learn about myself this morning from my DNA? If you grab your DNA, so there's a lot of DNA in your body. About half of it is actually microbial DNA that moves and changes and evolves sometimes every 20 minutes the bacteria are dividing.

[00:05:35] So you'll learn about any changes in your microbiome, which are the small creatures in and on and around you that have moved. So maybe you could pick some up from you can pick up obviously like a pathogen like flu or COVID.

[00:05:47] The SRCV2 which causes COVID you could of course, you get sick but a lot of them are actually things that are the anchor for a full ecosystem that itself is a little pharmacy.

[00:05:57] So you can see if there's any changes in your gut microbiome. If you have problems with your gut for example, you can look at your epigenetic changes or it's not your DNA but also how it's packaged and regulated.

[00:06:07] And of course you can get new mutations every day you get mutations. Most of them are harmless but some of them could be the beginnings of a cancer that you could see if you saw it that very first day.

[00:06:17] What about from the DNA itself? And this is not something you would find in the morning when you sequence my genome or whatever you would get all the kind of hard coded things about my DNA. What can you see from that?

[00:06:31] Well, from there you can tell a lot about you like your ancestry. So are you for example, do you have any Jewish ancestry? We had a fun part of the book we talked about Igor as a Jewish but then at the end of the chapter he got even more Jewish because the database has got updated.

[00:06:45] So we can see the databases, this is a good lesson of prediction is that your predictions are only as good as your training data and your databases. And when you change the databases you'll get slightly usually improved predictions but sometimes they can go the other way.

[00:06:58] But so we can look at ancestry or risk taking likelihood how fast you process caffeine or other drugs. So you can really be predictive about how and what way drugs and molecules will be processed in the body of any person.

[00:07:11] So some of this seems predictive. Some of this you know for a fact. So for instance with the DNA there are some genes where and I'm going to simplify it with my language but if they're on you have like taste acts disease for instance and if they're off you don't.

[00:07:26] And so it seems like some things they had enough data that they were able to figure out which single mutation genes cause which diseases and some data though is more predictive like oh is this person more likely to be happy or sad or Jewish or not Jewish.

[00:07:44] And you do that by matching tens of thousands of humans who have sequenced their genome. You know what they this person was Jewish and happy and this person was something else and also not happy maybe and then you can start to build together probabilities based on a new genome.

[00:08:04] Yeah in a nutshell that's in the ballpark right. And basically you look for differences in the phenotype or what people express as trait and you compare that to the genome and it could be everything from height for example which is there's no one gene for height there's not even two there's probably several hundred genes that really

[00:08:22] can immediate how tall you are but it is very heritable if you look at tall parents of tall kids and short the inverse we so it is a very heritable but very complex so it's a what's called polygenic meaning just more than one gene that influences that trait in a highly heritable way.

[00:08:39] And so as you as we found more of these genes they'd be built into the models we can predict to within about an inch or so how tall you'll be so if you take a baby at birth sequence the DNA we can get down to probably about and within an inch of how tall they'll likely be.

[00:08:52] So we're going to be from the technology to manipulate a gene at birth or a sequence of genes at birth to change someone's height.

[00:09:01] We're actually doing it not for height but we're doing it for diseases and modifying DNA as we speak so you can actually modify it embryos have had disease genes removed for example for hypertrophic cardiomyopathy or heart disease gene or even in an adult there's been treatments to get rid of beta thalassemia by doing gene editing in the person's body as an adult for

[00:09:22] cell and beta thalassemia these blood disorders.

[00:09:25] Now these singles single gene diseases correct so with multiple gene diseases the this is where the data is like immense like if the permutate the possible permutations of which genes could be like let's say height is caused by 100 different genes out of what like 32,000 genes or some outrageous number.

[00:09:45] But 60,000 total yeah yeah yes.

[00:09:47] So the permutations are in the I don't know quintillions quadrillions so it's impossible to use statistics or a computer for that is this something like AI could start to figure out when there's multiple genes involved like feeding it through neural networks the way they did with chat APT.

[00:10:03] You could basically have feed the data from you know basically millions and millions of patients and they're in a clinical metadata and their traits and essentially learn what are the new signatures that are driving some of these changes but but it wouldn't just have to be genetic data could also be what was in your diet or what was other factors in the environment.

[00:10:20] What else mediates that last few percent you can build into some of these models as well. But how far are we from really the technology to really understand things like height or intelligence or cancer you know all of these factors that involve hundreds maybe thousands of genes.

[00:10:39] We're in some cases were very close I'd say a height for example is pretty well teased out even autism risk even though autism is complicated and several hundred genes many of them are now consistently being identified so we can explain a lot more autism than we could certainly 10 years ago and almost.

[00:10:54] Can very barely do it all 30 years ago so I think even complex diseases or complex traits we can now explain pretty well and you can edit them you can edit.

[00:11:03] Dozens of genes at one time with what's called multiplex editing in pigs for example they've done up to 60 genes at a time is new trials they can give you up to 100 edits all at the same time all over the genome.

[00:11:14] The worry though is it's not perfect so you can it's like sending some of the bunch of erasers to your book and if it's correct you would very precisely do changes in the text of life but if it's messy then you'd of course be hard to read the book because you've made to many mutations so that's what we're.

[00:11:30] Because the more genes involved the more you could have side effects and editing them like what if some set of the genes for height are also related to genes for I don't know some disease or with our cancer.

[00:11:40] As I was the risk there are new methods for CRISPR ones called prime editing where instead of breaking both strands of DNA and swapping out a chunk and then having that occur sometimes off target effects meaning not where you want them to be.

[00:11:52] Prime editing breaks only one strand and actually is much more precise so you can use some of those methods or there's actually a quest by many companies right now and a lot of money being invested to try and find newer and more precise methods for editing but the technology to hear today and they're only going to get better.

[00:12:07] So Igor this strikes me as so you've been involved in like for instance predictive algorithms for stock market predictions.

[00:12:15] This strikes me as a little different than that because the difference between the genome and its relationship to diseases and traits in the body that's probably accurate like once they figure it out they know once we know these genes affect height we know forever those genes affect height but with the stock market.

[00:12:35] The more people know something the less likely it is to work next year. So for instance if you're going to predict new additions to next year's Russell 2000 so you start buying the stocks now well every once everybody starts predicting that it's too late to use this algorithm.

[00:12:50] That's right that's right in genomics what you figure out does not affect the subject but in finance the fact that you figure it out is going to change the way the system behaves eventually.

[00:13:04] I mean I take some blame on that of some algorithms that I used to use stopped working after I wrote about them so because I have the misfortune of loving trading but also writing and for what you know.

[00:13:20] So I had an algorithm that it might have been you know it seems statistically significant to me but the past 80 times roughly the cues gapped up between 0.4 and 0.6% you could short at the open and they would be flat at some point within the next hour or so.

[00:13:40] And then I and it was like an ATM machine for me like every time it happened I made money until I wrote about it and then it was actually random after that. Well you you made the market more efficient on the other hand. I'm a hero. Very hero.

[00:13:59] So there's this spectrum between OK given data we can know some facts like for instance on the genomic side or given data we can predict but we don't want to tell people our predictions if we're making use of it. So there's like a spectrum there.

[00:14:22] So like Moneyball the book by Michael Lewis predicting baseball outcomes that's more like a stock market style prediction because if you know that everybody if you could draft people who who are good at walking then everyone's going to draft it and the arbitrage goes away for teams.

[00:14:39] That sounds right. That sounds right. What are other categories like that maybe even traffic right if Google algorithm is routing everyone in one direction.

[00:14:50] So it will not be empty and where there is no traffic there will be traffic and possibly most algorithms that make predictions in systems in which the user is participating will be like that.

[00:15:05] It's only if you're predicting something away from your ability to influence it that it should not change too much.

[00:15:13] Did you ever get nervous in your hedge fund career that you just were going to run out of algorithms that eventually they are all the arbitrage because now there's like 30,000 PhDs and every hedge fund trying to find these discrepancies in the data or in the outputs.

[00:15:30] Did you ever worried you would all the arbitrage will be gone the market would be smoothed out and that's it.

[00:15:36] In the beginning I used to worry about it and you know have these interviews and everybody was worried about it that the market would become efficient but it never has and actually simply logically thinking somebody has to make it efficient right.

[00:15:56] So somebody is going to be there making it efficient no matter what it may just not be you that's a problem.

[00:16:05] But can't my predictions also be mean reverting so for prediction worked regularly for a while because presumably it modeled some mob psychology and then if it stops working for a while won't it mean revert and eventually work again.

[00:16:19] Yeah and when it does you know it'll come back to life it'll get turned back on. Okay yeah that makes sense.

[00:16:29] And you know also obviously you talk in the book a lot about insurance and insurance risk in the entire insurance industry is built on predictive modeling but that's been the case for let's say hundreds of years what's new in insurance in predicting human behavior now that has helped the insurance industry.

[00:16:47] It's the immediacy in real time data you can use real time information coming let's say from people's driving to adjust their rates you can use body sensors to change outlook on the health of an individual and so on.

[00:17:05] I imagine it used to be that you fill out a form once a year and that's kind of where it's so now there's more and more data coming in so the insurers are getting a clear picture which theoretically is good for everyone.

[00:17:23] Our car insurance for example you know they put a little device that goes in your glove box basically it's okay tracks your speeds and GPS coordinates and looks to see if you're speeding basically but when we got our recent insurance my wife didn't want it in the car so I don't want that thing in our car so we purposely took the higher rate of insurance just because she didn't want that in the car.

[00:17:43] So you know they're in the insurance companies now are saying well you get a discount if we put this device in your car so they can build better models of you but you could always just not take it I guess but then they make you pay more for it.

[00:17:54] Yeah have you ever read the book 2041 by Kai Foo Lee.

[00:17:58] So Kai Foo Lee is a big AI technologist from way back essentially the father of speech recognition and he wrote this book basically a bunch of scenarios about how predictive AI might work in 2041 he uses in the first story he uses insurance as an example where this family gives over all their emails in exchange for a discount but then the insurance company could model them better.

[00:18:22] Their insurance rates spiked because I guess one member of the family from her emails that could be determined that she was in a let's say not pleasant relationship and so her risks of accidents they knew would go up according to their data you know so there was kind of pros and cons to the enormous amount of data we can now capture to use for prediction.

[00:18:43] But my question here is this actually is even is closer to the genomics kind of model where just because you know something it's still light up prevented from happening as opposed to the stock market predictions.

[00:18:56] Yeah you mean like if you know it's necessarily faded in the sense that you know it's a most genetic risk or probabilities but some things are really hard to avoid like taste act disease or cystic fibrosis where you almost certainly get some or Huntington's disease there are some diseases where it's going to be really hard to avoid.

[00:19:13] But even that was right in the book because of the CRISPR and different genome modification systems you're no longer subjugated to the shuffle of genetic lottery you got as an embryo.

[00:19:23] You can in theory modify it or tweak it or think about what you hand down to the next generation for the first time really in ever we have the ability to kind of tweak what is that risk I feel like it's still like yes for this again for the single gene mutations that are causing diseases you can turn it on and off switch and get rid of the disease.

[00:19:43] Using technologies like CRISPR but most things are more complicated and I'm just wondering what are the first complicated things that we're going to be able to solve.

[00:19:53] I think you know like heart disease yeah or even like heart disease would be one in there you know there's mutations that can drive this and there's even known genes for hyperchlorostemia that we could target.

[00:20:04] Some of those have already been targeted actually in terms of again it's usually one or two genes that have been targeted and there's more than one.

[00:20:10] So in those cases you know or for example if you look at certain kinds of cancer they're driven by like a beer say one and two for example a lot of breast and ovarian cancer driven by a handful of genes probably the top 30 genes along would explain 90% of the cases or so

[00:20:26] 90% of the cases even for ovarian cancer so we know what those genes are you can constantly be scanning in the blood for anyone to see do you see a spike of any mutation so I see it.

[00:20:37] And we're going to kind of like whack them all take and get rid of that mutation and if another comes mutation comes in a different gene see it and then go after that target so I think it would end up being you rarely would need to go after say 15 genes at once.

[00:20:49] You'd probably do it over time for more complex diseases that like cancer but for some things like height if you really wanted for example to be if you if you think you're going to be for example really you know let's say four foot two and you wanted your kid to be taller so again this sounds hypothetical but something like this will probably happen where someone's okay I'm going to have a safe way to do to make sure your kid is tall and do it in IVF clinic.

[00:21:12] And it doesn't have been yet but the close thing is a company called orchid which is doing this for embryos they sequence the GM of each embryo and then you pick the one that you want based on that selection.

[00:21:22] Because you could predict okay this this embryo is going to be a female tall athletic high IQ and this embryo and let's say with it with you know odds on each one but pretty good odds and this embryo there's odds that oh low IQ not athletic.

[00:21:42] A male and so I'm just going to not you I'm going to abort that embryo and give birth to this one.

[00:21:48] Yeah we're just sort of happening today with orchid at least there's one company there's other ones that are also coming on into the market that are trying to guide IVF basically but but it's embryo selection and that's let's just so I see so embryo selection before embryo modification.

[00:22:05] It's a little easier to do that because modification we don't know yet we were just making guesses on the odds like we don't know 100% chance this person is going to be six foot three but always like a 60% chance which is better than these other embryos so then we don't have to take our risks with modification.

[00:22:23] What do you think in China they're doing what they're doing there is a lot more of the somatic methods I've seen really doing things I mean in your body as an adult modifying yourselves.

[00:22:32] They're been looking a bit more at embryos and also actually being much more aggressive with which new modified cell therapies basically genetically modified cell cells re infuse them into the patient.

[00:22:45] And then you know have a new targeted therapy that happens but it's targeting something that from our own publications we wouldn't recommend because it's if you're targeting what you think is only on a cancer cell but turns out it's also on 20% of all the other cells in your body that probably very painful.

[00:23:00] Because you suddenly unleash all these angry immune cells that are attacking what is thinks is out of cancer but is also on regular cell.

[00:23:08] Yeah like it seems like you know this could be really powerful for anti aging if you're you know depending on which model of anti aging you believe if you attack telomeres and change the genes in them so that they don't shrink over time or you are able to inject new ones and then they attach to I don't know how it all works that they attached to yourselves or whatever.

[00:23:29] It seems like this could be really great. Like these things called Yamanaka factors that they're researching in Asia I guess but there's overlap with cancer. Apparently the more you get this kind of treatment the more likely you are for cancer so there's all these risks.

[00:23:44] Yeah absolutely and so I think the you have to balance what is going to be the likely benefit from it from any possible side effects or you know do no harm as the basis for most medicine.

[00:23:54] And in some cases we might not know though do it with think we know or we looked in a mouse but we don't know yet any human so that's why clinical trials always start small.

[00:24:04] I start with 10 people maybe eight people start small for people that really need it and then you slowly expand.

[00:24:10] What about the kind of stuff that like Palantir does so given a set of bank transactions and a bank customer they could make a guess or prediction as to whether this customer this bank customer is a terrorist or not for instance or involved in some kind of a

[00:24:24] financial fraud. And again this is more related to kind of the stock market prediction stuff because as you know more how they're predicting you could modify your behavior but how close are we to really modeling human behavior like that.

[00:24:40] I mean credit card companies do that today I think like they'll look at any spending patterns look at any changes in behavior try and guess whether one just to guess whether it's you did someone take your credit card and someone else or did someone grab your phone.

[00:24:56] And then the other thing that they'll do is just look to see you know are you a risk and some other capacity and so I think I mean other thing it does is for example is iTunes or Spotify.

[00:25:06] Try and make a playlist based on what you've listened to before which is great but then when my daughter got my phone and started listening to all of her songs it totally screwed up my algorithm so now it's not it's no longer my ideal playlist.

[00:25:19] So given that that kind of prediction like the Netflix or Amazon style like since you bought this you might like this that's been around for a while.

[00:25:25] Yes I'm sure it's improved but what's really cutting edge with that like I get it that with genomics this kind of mapping the all the permutations of possible data to real diseases and human traits that's important.

[00:25:40] Obviously the stock market is an immense problem that can never be fully solved what things are are are blowing away on the kind of social modeling you know given the sheer amount of data we now have as compared to 10 years ago.

[00:25:53] So the movies that get recommended to me I actually like you know fair percent of the time so life is getting easier less less thinking to do.

[00:26:06] Yeah and also I guess I used to get called all the time by credit card companies saying oh this is suspicious behavior but it was just me being me now I don't get as many calls it seems like they are better at modeling if a credit card is you or not.

[00:26:22] You might still be a suspicious person just by your weird habits yes but at least it knows that now. Yeah but OK what do you do in situations like March 2020 when the advent of covid happens and you talk about this in your in your book.

[00:26:36] The market falls it's it's it's the equivalent of what it seemed to level call blacks one event like the market falls eight standard deviations more than normal in a short amount of time something that should never happen one trillion times and yet it happened.

[00:26:50] What do you do when there's really no model and again the stock markets considered like a fat tailed kind of curve as opposed to a bell curve. How do you take into account situations like that where you can suffer significant loss treating it like a bell curve.

[00:27:05] You cannot divide things into ripples and the waves so the waves are things like you described that these are gigantic events that are specific and you can see them identify them and the repulsor are you know you know the stock of a toothpaste company moves 0.1% when something else happens so we trade the ripples but we stay neutral to the waves so when they don't like.

[00:27:35] 2020 happens.

[00:27:38] It shakes us somewhat but we're more or less neutral so we we stand through it but where there are waves there are the ripples to and the ripples start going in all kinds of different directions I see that that totally makes sense so like for instance if something big is happening.

[00:27:54] It's affecting the entire market over a long period of time whether it's a day or weeks or months you kind of that's that's not your business but let's say for 20 seconds.

[00:28:05] The Canada markets deviate from the US markets by a wider spread than usual you can say okay within the next 20 seconds they usually snap back and you could play things like that whether regardless of the larger wave that's happening in the markets.

[00:28:18] Yes it's like that but it's not only the time the ripple can have a long duration but it's just very weak so that nobody else is really trading it but you.

[00:28:44] I see so but that could be so you know there's an AI of course you know there's this difference between unsupervised learning and supervised learning the much like how chat GPT was built initially you could use unsupervised learning to find where they are.

[00:28:59] The AI finds that the context of some patterns of language is related to some movements of stock we don't know what the connection is but there seems to be a connection and then from there you could just start figuring it out.

[00:29:13] You just need the statistical relationship you don't need to understand why it works it may not be possible to understand why it works and by the time you understand why it works it won't work so.

[00:29:24] Yeah I mean no things you do with the data I'm trying to think of other layers of that data so for example where do we see higher rates of different diseases or cancer or infections could be related to what is your other talkings nearby do we see you know other.

[00:29:38] Direct like a geospatially informed view of health care could help you stratify risk and even find causes or you know essentially why some people would be getting sick.

[00:29:47] It is something you can maybe use all the Google imaging data for example from Google Earth and look for trends there a simple thing is in number of trees or night lights in a road can influence health in some ways and so you could use that if you could make a lot of money on that but you could at least stratify risk and help people to try to better outcomes.

[00:30:04] I wonder if there are things that we could look at that we're not looking at that could help us decrease let's say a spike in accidents and some geography or you know things that we just haven't even thought of because we don't know the connection we don't really comprehend why there would be a connection but there is one.

[00:30:21] Yeah that's a good thing about like a lot of the tools for stratification of these AI tools will find the patterns and then can build that into a model you can essentially leverage that to make a better prediction.

[00:30:34] We actually do this for example when you sequence a potential pathogen from a sample from your own sample for example we look at all the facets of the data not just what species is there but you know statistics on the fragments of DNA that came out or the pH of the urine or other factors that could better diagnose the ETI and everything goes into the models and essentially we don't even need to know why the model gets better but if it can predict better what pathogen is present then we can use it and actually some of these are under review now by the FDA.

[00:31:04] I think that's a good thing about the AI algorithm is that they can embrace some of the AI algorithms because they work really well and they'll lead you to a better way to do diagnostics and care.

[00:31:13] What about just like pundit predictions so somebody goes on CNBC and says well I think gold is going to go up because of geopolitical stress blah blah blah do you think humans have gotten better given you know more understanding of history more data about recent events more opportunities.

[00:31:34] To predict and see how those predictions turn out do you think humans have gotten better at being essentially pundits. The answer is paradoxically no because the more the ability to predict things improves.

[00:31:50] The more people lean on those predictions the more they're used and what the what remains is a more and more unpredictable world that gets harder and harder to predict.

[00:32:02] So by the time you know somebody is saying something about gold on CNBC everything he knows has already been figured out and other you know more notorious pundits and there's probably nothing left to pundit above.

[00:32:19] What about a macro trader like someone like George Soros who famously predicted the collapse of the pound in the early 90s I think it was 1991 or 1992 and you know was there any was it just luck that some macro trader succeeded and others didn't or was there something else that they have some special insight that maybe has kind of disappeared from the markets now.

[00:32:44] I think they had insight and understanding and there was not too much competition for a high level of insight but these days there is.

[00:32:56] Yeah it seems like that's the most and the trading arms of all the banks seem to be very quant focused because they do all the high frequency trading and so on.

[00:33:05] So again like where's definitely in medical there's opportunity you have a lot of data so now we can start figuring out which multiple genes you know relate to what characteristics and traits although there it seems like a math problem like you have to deal with these exponential size math problems that computing can't do just is quantum computing a solution for figuring some of these things out if and whenever there is quantum computing I'm not sure there ever will be.

[00:33:33] I'm not sure I understand it but is that a solution to the exponential problem.

[00:33:38] If it does what it's supposed to do it could help and give us just that much more compute capacity as on the planet and so that would certainly help but I think a lot of it also be some of the testing will be done on the ground you know or you could do something with model systems but I mean I would be the first one to jump in line if we had solid kind of quantum computing up and running it'd be great.

[00:33:59] There are classes of problems of quantum computing can crack and the classes of problems that still remain unattainable. Really what's the type of problem that quantum computer can crack.

[00:34:12] You know these days there they're they're selling encryption algorithms which are quantum proof so that's that's one type of algorithm which is a good example because a classic encryption right now let's say the way big coins encrypted.

[00:34:27] That can be solved by a quite it cannot be solved factoring a hundred digit prime number cannot be solved by a thousand super computers linked together but a quantum computer can do it in a second.

[00:34:40] So that's a classic example and now you're saying there's algorithms that could make cryptography quantum proof. There are there's no computer that can solve all problems.

[00:34:50] So because what I what I worry about is are we hitting a point of let's call it peak data where we have the maximum amount of data for certain categories that we can basically handle because the computers are not fast enough and not fast enough not because the chips are slow but because mathematically it's too exponential a problem.

[00:35:12] Yeah that can happen the data the rate of the growth in the data may simply exceed the computing power's ability to digest the extra data. And I don't think we're there yet but it certainly could be because generating so much data.

[00:35:32] Yeah baby but peak data indicates that the data will have less utility or that will have peak past the ability to use the data.

[00:35:40] So I don't know if I call it peak data just unwieldy we've reached a pass to being able to wield and manipulate data efficiently to only having various degrees of efficiency because I think the data assuming it's clean data should still get more useful as you get more of it and over time.

[00:35:55] Right so you'd have to sort of come up with more categories of that you're studying in the data but like again genomic data seems like it's there.

[00:36:04] Like for 15 years we could predict Tay-Sachs and other single gene mutations like what single genes cause which diseases but I feel like for something like possibility later in life of stroke or predicting IQ something like that which requires hundreds of genes were just we're never really going to be able to.

[00:36:24] The data is there and it's possible the algorithms are there but we just our computers are not never going to be fast enough to solve them.

[00:36:32] I mean maybe today but not I wouldn't say never because I never say never because I think they still could be you know in 50 years there could be something else even beyond quantum or that's some other variation of a computing or even just efficiencies of the algorithms could be improved in ways we can't imagine now so so.

[00:36:51] Maybe in the short term yeah but I think long term.

[00:36:55] Yeah I would say now I mean if you go back 200 years ago it would have been inconceivable that people would routinely fly through the air and airplanes it was you know what no one would have believed you right and no one believed even the right brothers for a while so I think you know a century is a long time with current humanities.

[00:37:12] Well it's really interesting you say that about you know the algorithms might improve that's an area I haven't thought so obviously over the years over the centuries statistics has improved so instead of just trying to like match something against a normal curve.

[00:37:26] Now there's all these very sophisticated algorithms for speech recognition vision recognition and and so on how much more do you think the math can improve because that also seems to be very we might be a peak math in terms of like how much more do you think the math can improve because that also seems to be very.

[00:37:38] We might be a peak math in terms of like how much math is actually useful that's coming out of the academia.

[00:37:45] Yeah like math departments there's no new yeah math I mean there are some there's a lot of work on you know since you're reconciling quantum mechanics with Newtonian physics and that that that is still something that should be solved at some point or but there's no new kinds of numbers being discovered or entirely new kinds of calculus right most of the time.

[00:38:08] That's been done since the 16th century and so I think there are some I mean Newton probably would have argued he was peak math I don't know I'm not having a calculus but.

[00:38:18] But even statistics has with with again the rise of the need for pattern recognition I feel statistics has evolved in the past 30 or 40 years like these hidden Markov processes and other techniques to really do sophisticated pattern matching and I just wonder can that be improved.

[00:38:35] Depends depends what you call an algorithm right maybe the basic algorithms can't be improved and don't need to be but if you look at something like Alpha zero or charge you P.T. is an algorithm then yes from time to time a groundbreaking algorithm does appear.

[00:38:54] Yeah but I wonder how much of that was OK more advances in you know adverse neural networks versus the speed of computers finally look the speed of computers were there to solve more or less computer vision 20 years ago.

[00:39:10] But I feel only in the past five 10 years it was fast enough to handle large language models like chat GPT and that was purely a speed thing how much was related an algorithm thing. I think there were elements of both.

[00:39:23] Yeah that makes sense and also just the amount of data like that you can build large language models until you had lots of data to look at I mean so something all those things together coming to the data the algorithms compute then you can do it but you really couldn't do it 15 years ago.

[00:39:39] Yeah actually to your point we didn't even have the data we didn't have all written text up until last year you know stored anywhere in one in one easy to use place.

[00:39:48] So but even that with the speed it took a bunch of supercomputers a year and a half to crunch the large language model that's now chat GPT and then another year and a half of supervised learning. So it'll be interesting to see how that speeds up.

[00:40:02] So Chris I know you're interested mostly in predicting breakthroughs in medical technology and Igor and financial stock market predictions. What other things would you like to predict.

[00:40:14] I think the methods you know you can use them anywhere right what we're doing now is we're actually predicting data itself data has this property that you mentioned before that when you predicted the data.

[00:40:28] The data doesn't change from the fact that you predicted so we're predicting data and you can model most things as data.

[00:40:38] So you then you just start going into different industries and how much can each industry be improved through prediction and algorithms and through just getting rid of that 90% of the work that's kind of mechanical in nature.

[00:40:55] If you consider the chart GPT is a mechanical thing in the end.

[00:41:00] Do you think I can predict what a hit song will be so let's say I take every hit song of the past 20 years and feed it into my statistics slash AI machine and then I use AI to create a video with a beautiful woman or guy and boom do I have a hit song you think that's possible.

[00:41:20] Good beat good voice.

[00:41:23] I mean there are there's something that are catching us to the there's definitely signatures of songs that are poppy if you will and so and they're popular so I think you could you could you could build a model and maybe you could get 99% screw but I don't know if you could ever say 100% like any model.

[00:41:39] You can you may have to model influencers and feed the song to the right influencers together popularize.

[00:41:45] Yeah that's a good point but maybe I can create an influencer is using modeling so okay here's here's everything I said on Instagram over the past year that got you know a million likes or more and break down those and then here's what every influencer look like.

[00:42:04] So now I'm going to come up with like the average like super influencer and then boom now I'm going to feed that influencer song and make a record label and use that. To vertical integration.

[00:42:16] Yeah people wonder if this rise in predictability is going to bring down creativity but I sort of think it's going to bring up creativity because now it's going to free up your resources in some ways to come up with even more creative ideas well what do you guys think.

[00:42:32] It certainly frees the up to ask questions and get very quick answers which frees up creativity doesn't save time. I think it's like any new technology can be a tool or a weapon right in this case could be a great tool for creativity.

[00:42:46] Of course people are afraid of AI because it could be in theory weaponized but on the creativity side I think it's going to time saver is going to be inspiring you can you can craft then with the stable disused and all the other tools.

[00:42:57] You can make a land you can just describe the landscape you're imagining and it creates it for you and you can build from there very quickly.

[00:43:05] These amazing portraits I think I think it's phenomenal lets you do what I always wish I could do as a kid with like describe something a scene an idea of say you know a rabbit with 12 different tentacles that was playing.

[00:43:19] You know the harpsichord and also juggling pool cues like I could never really draw that but I can get an album to make that in five seconds right so it's it's phenomenal.

[00:43:28] But maybe there's a problem there where you know one of the reasons why people always say oh the book was so much better than the movie is because when you're reading the book you're kind of constructing the movie in your head instead of it being you know and then suddenly you see the movie and it's like you're so disappointed because it wasn't good as that movie you built in your head.

[00:43:45] In your head but now we're going to be able to see basically the movies that we build in our head much faster.

[00:43:51] You would choose yeah I think which is still I mean it might be it could be good and bad but mostly good because you can get in the movies out like you'll see what's present faster but then also you could have 55 variations of you can change the seed kernel for most of the art for example.

[00:44:06] So you make 50 versions of the thing you were just thinking and you actually might then imagine things that you weren't quite originally thinking so you don't just get one imagination to get 50 of them.

[00:44:16] Or you can do some pruning maybe a print down to 25 that you like but I think that's you know just imagine if you could have 50 brains instead of one you kind of can have that today which is pretty amazing.

[00:44:27] So what do you think so that we're in the age of prediction obviously obviously like I should tell all my kids to be data analysts because that's going to be just this huge profession for the next 20 I'm making that prediction that's going to be a huge profession for the next 20 or 30 years.

[00:44:41] 20 30 years from now what do you predict will be seeing in our predictive abilities and how we use it in society that will just blow our minds curing cancer by the way won't blow my mind because I expect that.

[00:44:53] That's supposed to happen. Yes. But it could be you know every toilet I'll be monitoring like every morning you get a little update report on every molecule in your body. You would get information about the environment around you around your home or your apartment. I think you would.

[00:45:10] I think we'll see prediction coming even from other planets like for example the Perseverance rover landed on Mars. It took too long for a signal to go from Mars. So it had to use image recognition software during the landing to get there.

[00:45:23] So we'll start to see prediction algorithms and tools even send data back from other planets like Mars. You can get to be able to get news from the future because it will be mostly predicted on a micro scale.

[00:45:36] I can say oh this person crossing the street moves like someone who's going to rob a car in the next day. So you know kind of minority report style predictions. You know the movie with Tom Cruise. So what types of news events do you think will be predictable?

[00:45:51] Maybe elections. Elections might mean we saw this in Cambridge Analytica that might be tweaked before they even happen. So you might know it. You might know the future because you've made the future. It is possible.

[00:46:03] I think if you examine news headlines and just examine the news you will find patterns in it already that so much news follows other kinds of news. It is predictable and so on and so forth. But nobody is really putting that into a news service.

[00:46:26] That's fascinating to see these days. So what you're doing there is you're disconnecting news a little bit from reality which is what newspapers probably do anyway. And you're saying tomorrow's headline is more based on today's headline than in the actual events that happened today. Yes.

[00:46:46] So you may be able to print tomorrow's newspaper better than the actual paper that's going to be printed tomorrow because tomorrow's paper will have some noise in it from news sources, et cetera, et cetera. But your prediction is based on mathematics.

[00:47:04] Well the age of prediction such a fascinating book I really do think the job people should be preparing for is data analysts because that's going to be used in every single industry more than prompt engineers or AI coders because AI is going to write its own code.

[00:47:22] Being able to understand what data to look at and why and how to make use of it whether it's the medical industry or sports or stocks or insurance or art.

[00:47:31] This is going to be such a valuable skill to have and it's just beginning field like the creativity there is going to be amazing. But the age of prediction is like a guidebook to what's happened and what's going to be happening

[00:47:43] and all the ways people use prediction technology and such a great book. How did you guys team up to write it? Like why did you write it together? How do you know each other?

[00:47:53] We really met actually at lunch at Cornell's campus and then just started brainstorming about data and then started walking through the lab chatting about the use of data for medicine and overlapping with finance and it just became exciting to think about more ways to do partnerships, brainstorming.

[00:48:10] We have a fellows program that goes between WorldQuant and Cornell as well. So people can go back and forth. There kind of be this nice exchange of ideas and expertise between the institutions.

[00:48:21] Interesting. You know, Cornell in Manhattan, like again I'm talking about 25 years ago used to have a computational finance. Their computational finance department was in Manhattan. I don't think it exists anymore there but I'm not sure. But anyway, thanks so much for coming on the show.

[00:48:35] Really, this is like my favorite topic. The age of prediction. Thank you so much. I hope you guys come on again. I really appreciate it.