Invest Like the Best with Patrick O'Shaughnessy - Avi Goldfarb - The Economic Impact of AI - [Invest Like the Best, EP.321]

Starting point is 00:00:00 This episode is brought to you by Teegas, the modern research platform for leading investors. I'm a longtime user and advocate of Teegis, a company that I've been so consistently impressed with that last fall, my firm, Positive Sum, invested $20 million to support Teagis' mission to expand its product ecosystem to unify and streamline investor research processes. In addition to the library of 55,000 transcripts, Teagis now combines at-cost, on-demand calls with a full suite of financial workflows, whether it's quantitative analysis, company disclosures, management presentations, earnings calls, TECIS has tools for every step of your investment research. They even have over 4,000 fully drivable financial models. TECIS's maniacal focus

Starting point is 00:00:41 on quality as well as its depth, breadth, and recency of content makes it the one-stop end-to-end research platform for investors. Move faster, gather deep research to build conviction, and surface high-quality alpha-driving insights to find your differentiated edge with Tegas. As a listener, you can take the TIGS platform for a free test drive by visiting tegis.co slash Patrick. Before we transition to the episode, I want to highlight the Founders podcast, which is part of our Colossus network. David Senra, who hosts founders, has devoted his life to learning from history's greatest entrepreneurs, and every week he distills the lessons of a different founder.

Starting point is 00:01:16 If you want an entry point, I highly recommend starting with episode 136 on Estée Lauder or episode 288 on Ralph Lauren. I hosted David on Invest Like the Best last summer, and it's hard not to walk away, insanely energized after listening to any episode with him. You can find a link to founders and those episodes and the show notes of this conversation. You can also search all past transcripts on our website, join colossus.com. Hello and welcome everyone. I'm Patrick O'Shaughnessy and this is Invest Like the Best. This show is an open-ended exploration of markets, ideas, stories, and strategies that will help you better invest both your time and your money. Invest Like the Best is part of the Colossus family of

Starting point is 00:01:58 podcasts, and you can access all our podcasts, including edited transcripts, show notes, and other resources to keep learning at join colossus.com. Patrick O'Shaughnessy is the CEO and founding partner of Positive Sum and the CEO of O'Shaunacy Asset Management. All opinions expressed by Patrick and podcast guests are solely their own opinions and do not reflect the opinion of Positive Sum or O'Shaunacy Asset Management. This podcast is for informational purposes only and should not be really. relied upon as a basis for investment decisions. Clients of positive sum or Oshonnessy asset management may maintain positions in the securities discussed in this podcast. My guest today is Avi Goldfarb. Avi is a

Starting point is 00:02:43 professor at the University of Toronto's Rotman School of Management, the Rotman Chair in Artificial Intelligence and Healthcare, as well as the co-author of two bestselling books on AI and its economic impact. His most recent book, Power and Prediction, is probably the best piece of content than I have read in explaining how AI may reshape business models, systems, and products. We recorded this before GPT4's release last week, which, if anything, makes Avi's ideas on AI's impact all the more poignant. Please enjoy my conversation with Ollie Goldfarb. Avi, we were just chatting before we hit record about chat GPT. We have to start here. I'm so curious about your reaction to it because it came out almost immediately after your book

Starting point is 00:03:24 was published. So we're going to talk about all the ideas in the book in great detail. But first, just give me your unvarnished reaction to when you saw that come out and what you've been thinking about it since. My first reaction was wow. I'd seen large language models before. I'd see the earlier versions of GPT, including GPT3 and stuff from OpenAI and they were doing great things. But chat GPT showed me to advance faster than I'd expect. First is wow. And then, along with my co-authors, we started to think through, well, what does this mean for the future of AI?

Starting point is 00:03:53 What does this mean for AI in business? And at the time, when we heard people talk about chat GPT, everyone was focused on, oh, no, we're going to have to change the way we write final exams. Or, oh, no, there's a handful of people who make a living from writing text. They're going to be toast. They're going to lose their job. And those worries are real. I don't want to belittle them. They're real worries.

Starting point is 00:04:17 But for most people, it struck us as this huge opportunity. And what do we mean by that? Well, there's lots of people who have amazing skills that are not getting ready. And so writing limits their opportunities for work, their opportunities for income. And if they have better writing, then a whole bunch of new business opportunities grow, a whole bunch of new work opportunities grow. And the organizations that figure out how to take advantage of the people who are good at lots of things, but not writing, are going to be the ones that get the first pass and take advantage of chat.

Starting point is 00:04:50 GPT. There was this story on Twitter. I'm not going to speak to the veracity of it, but it was a compelling story. There was this landscaper whose English language skills were terrible. He did not know how to write a customer service email, saying, hey, I'm going to be there on Wednesday, and I'm going to do whatever I'm going to do in your backyard, and then I'll send you a bill. Getting a coherent note out of this guy was really hard, but the story went that he was a great landscape, but he struggled to find business. A mentor, one of his clients, a little bit ambiguous who this person was, built an interface with GPT3 to help the guy to turn his quick notes, hello, arrive Wednesday, 9 to a nice coherent email to a potential customer

Starting point is 00:05:28 and seemed to work really well. The story went that the person's business thrived to herself. That kind of a narrative, I think, is going to happen a lot, which is someone's got to reinvent the interface. Something hasn't happened yet, which is somebody has to invent the interface that allows millions of people to take advantage of what this technology can do to write more easily and whether that's going to be Open AI that does it or Microsoft that does it or a startup, who knows. But once someone writes that interface and figures it out, then the potential to upskill millions of people is amazing for the economy. Not to ignore that thousands of people will be worse off.

Starting point is 00:06:02 ChatGPT is emblematic of something that we see in lots of emerging AI applications, which is, yes, they automate something that, humans do, and there's reasons to worry about that late market consequences. But the hope is that what they're automating is something that the people who really earn a lot do. So you think about in medicine, what's AI doing, it's automating diagnosis. And diagnosis is the domain of doctors, and doctors earn more than just about anybody else in medicine. So if you could automate diagnosis, then what happens, well, nurses and pharmacists and all the other millions of medical professionals can now provide much, much better medical service.

Starting point is 00:06:43 And they can take advantage of their skills. There's hundreds of thousands of doctors in the US, and their special skill and diagnosis is going to go away. They'll have to retool and figure out how to deal with that. But there's millions of other medical professionals who are now going to be able to do their jobs much better, be much more productive. And that upskilling provides a lot of what we see as the hope and opportunity for AI. If you were to just back up to your first book, prediction machines now, and talk

Starting point is 00:07:10 about what happens when the price of something falls, in this case prediction, and the impact that that has compared to other similar examples of some valuable thing and the price of it falling, whether it be electricity or something different. Talk at a high level about the importance of that simple concept, that the cost and friction of something matters a lot in terms of how much it's actually used by humans. You ever took an economics class? The very first thing you learn is that when the price of something falls, you do it more. So if coffee gets changed, cheap, you're going to buy more coffee. The reality is coffee can only get so cheap. But with technological change, the fall in the price can be extraordinary. Think about your computer. It feels like the

Starting point is 00:07:49 computer does all sorts of things. But really under the hood, your computer just does one thing. It does arithmetic, does it really, really, really well. And so what we saw is the first applications of machine arithmetic problems. The really first applications were effectively World War II and Trubby after we had cannons and they shot cannibals. And it's a difficult arithmetic problem to figure out where those cannibals are going to land. And so we had these teams of humans whose job was called computer who figured out the trajectory of those cannibals. Then machine arithmetic came along, computers, and they could do that arithmetic better than the human computers. The first applications were, we were already doing arithmetic, and now we have machines doing it. Then I think about accountants.

Starting point is 00:08:28 Yes, I'm an accountant from the 1950s or 60s or even into the 70s, but they spent a lot of their time doing. It was arithmetic. But then machine or arithmetic got better, and we no longer had accountants spend a lot less time adding up columns and numbers. They spend a lot more time thinking about tax policy and strategy, etc. But then arithmetic got even cheaper and started to realize that lots of things that we didn't use to solve with arithmetic can be solved with arithmetic as an engineering problem. When arithmetic is cheap enough, mail can be reframed as arithmetic, music can be reframed as arithmetic, even pictures could be reframed as arithmetic.

Starting point is 00:09:01 Kodak solved pictures with chemistry, chemical engineering. They were fundamentally a chemical engineering company. But machine arithmetic has gotten cheap. that by 20 years ago, we started to really think about pictures as an orthotic problem, no longer a chemical engineer. So as something gets cheap, we do more of it. And when that change is several orders of magnitude, then we can identify all sorts of extraordinary applications with today's artificial intelligence.

Starting point is 00:09:27 You can usually think of it as prediction technology. By prediction, I mean in this statistical sense, taking information you have and filling in information you don't have. The first applications of machine prediction were good old-fashioned prediction problems. You walk into a bank, you want a loan and a little officer at the bank looks you up and down and decides whether they trust you and predicts whether they're going to pay back. Over the course of the 20th century, that became more and more rigorous. And now increasingly, banks are using machine learning tools, using AI to predict whether you're going to pay back. I already talked about medical diagnosis.

Starting point is 00:09:56 Medical diagnosis is prediction. You take in data about your symptoms. You feel in the missing information of the cause of those symptoms. That's prediction. In writing, turns out, a lot of writing is prediction. How do you write in response to a query? Well, you're filling in this information about where words should go and what words belong next to other words in response to a particular query. Free form writing, maybe not, but if the chat CPT is doing fundamentally, is it's taking your query, write a five-paragraph essay on how AI will affect investing, and it will look at other five-paragraph essays that exist on the internet. We'll look at discussions about AI and investing that exist in its data set, and it will fill in the missing

Starting point is 00:10:39 information of what should those five paragraphs look like. We never imagined, frankly, even when we were writing prediction machines, that writing essays would be a prediction problem, but here we are five years later and very clearly it's. And so that's the essence of it, which is as something gets cheap, we do it more, and when that change in price is exponential, or is at least several orders of magnitude, the way we think about problems can change, things that we didn't used to think of as prediction in the context of AI, really dark. I love the computer example so much that you start with good old-fashioned X problems and then you start to migrate into more interesting areas that are hard themselves to predict.

Starting point is 00:11:14 You can't predict what prediction will be used for other than to say it will be used for lots of cool things that we can't imagine. Another favorite example that I've seen you write about, which I think really nails this point home more on the business side, is the transition between, say, a steam-powered factory. I'm interested in the deep details here, like how one of these factories is set up and construct it to one that is fueled by electricity. And the story between that transition is a great opportunity to talk about point versus application versus systems solutions. I'm happy to spend as long as we can here because I think it's such an incredibly powerful way of understanding how the business world and the entrepreneurial world might start to imbibe this technology. I love that question because it is the core thesis of our new book, Power of prediction, unpacking that.

Starting point is 00:11:57 It's what I've been spending more time than anything else thinking about it in the last couple of years. Let's start a little bit on motivation, which is we wrote prediction machines in 2018, thinking the revolution in AI was about to happen. And then three or four years later, we felt like we saw some cool uses of AI, but the world hadn't changed. That led us to think about the history of technology generally and electricity in particular. Edison's patent for electric light bulb was 18. So if you're paying attention in the 1880s, it was clear that this technology was going to be transform it. The patent office was inundated with new patents. The newspapers were filled with

Starting point is 00:12:33 people coming up with new ideas and how electricity might be used in all aspects of society. But if you look at the adoption rates of electricity in factories and households, how many people were actually using it. It wasn't until the 1920s that half of U.S. workplaces and half of U.S. homes were electrified. It took 40 years from recognizing, hey, this technology has extraordinary potential to most people being affected by electricity at home or at work. Why didn't it take 40 years of wandering to get from, hey, this is a big deal to it actually making an impact in most people's lives? Well, in the 1890s, if you ran a factory, here's what your factory would have had a steam engine,

Starting point is 00:13:14 sometimes a watermill, particularly a steam engine at the center of the factory. And that steam engine powered everything. It was one big power source. All the machines in the factory were connected by belts to the steam engine. And if you remember your high school physics, the more distance that the energy has to travel, the more energy gets dissipated. So what you want to do is you want to locate your most power-hungry machines as close to the steam engine as possible. The logic at the factory was built around the power needs of the machines. To the point where in order to make sure that you're using energy efficiently, you'd put machines above and below your steep edge.

Starting point is 00:13:49 They'd had these tightly cluster multi-story factories typically where the logic of the workflow was determined by the location of the power source. And so you'd have people moving pieces up and down and around because of where the power sources were. And then once you did that, that's really inefficient. So you'd even have workers doing a lot more at each station because they'd put a lot more pieces together. They were much more specialized workers because once you were located somewhere using a machine, having to then move it up to the next machine and then back down was really costly and inefficient. With a steam engine, that was the logical smart way to build it. Then electricity came along and a few people said, you know what, this is cheaper power. We can save 5, 10, 15% on the energy cost, perhaps because we don't have as many belts where energy gets lost or we happen to be near a good source of electricity.

Starting point is 00:14:33 So they take out the steam engine, drop an electric motor at the same point, but not mess with anything else. This is what we call a point solution, where you take out the old way of doing things, drop in a new process, but don't change the workflow at all. And those factories, they did save a little bit on energy costs. But for the most part, it wasn't worth the bottom. For most factories, saving 5, 10, 50%, percent wasn't worth trying to figure out how to get ready to steam engine and drop in the electric motor, find that the way to get electricity into the factory, all these different kinds of changes. And so they just kept doing things the old.

Starting point is 00:15:04 Starting around 1900, a couple of people recognized that electricity wasn't just cheap power, but electricity was distributed power. It allowed you to decouple the location of the power. source for the location of the machines. And once you can decouple the location of the power source and the location of the machines, you can build an entirely new kind of factory, where the logic of the workflow in the factory is determined by the production process, inputs and outputs, rather than by your powering. And so the quintessential 20th century factory that you might imagine, which is in cheap land and suburban areas, huge factories, inputs come in one end, outputs come out the other with

Starting point is 00:15:42 modular production. That exists of a reinvention of how factories worked after the recognition the electricity did is it decoupled the power source from the machines. So it created a whole new factory system. Once people started to figure out, hey, that's much more efficient. It's a much better workflow. It can allow us to produce much more efficiently, much more effectively, and maybe even different kinds of consumer products, like the mass market automobile that were not possible before under a steam power check. That's what we think about as, system change. It's not just taking out the old process, dropping in the new one, not messing with the workflow. That seems easy, but if taking out the old and putting the new is costly,

Starting point is 00:16:21 then it's probably not going to be worth the bother. It's only worth the bother to electrify your factor if you can do something totally differently and deliver a new kind of value. And if you look at the history of technological change, these big picture impactful technologies called general purpose technologies, also called GPs, but it's a different GPD. That kind of reinvention process where in order to get value out of the new technology, you have to do things differently, occurs over and over again. So it occurred with electricity. It actually occurred with a steam engine, 100 years before electricity. It occurred with the impact of computing on the way business is operated through the 70s, 80s, and 90s. And now we expect the same thing to happen with artificial

Starting point is 00:17:03 intelligence over the next decade. Can you say a bit about the midpoint between a point solution and the system solution, which is this idea of new applications also being built in the middle, what's going on there? So what was different about electricity is that machines could be switched on and off and can be taken in and out of the overall workflow relatively. But if something's powered by a steam engine, turning on and off the steam engine was really, really hard. Everything is powered by the steam engine, to connect the steam engine by a belt, by a thing moving fast. It was costly and risky to turn on and or off a machine because you'd be,

Starting point is 00:17:38 moving one of these fast-moving belts. And so typically you just kept your machines on all the time. With electricity, even if you took out the steam engine, dropped an electric motor and didn't change where things were in the factory, you could start to do different kinds of machines because you could have them flip on and off. And once you can have the machines flip on and off, you can take advantage of electricity within the existing workflow,

Starting point is 00:18:00 but to allow different production processes. Application solutions are at a high level. There's somewhere between the point solution and system solution. So they allow you to build new processes without having to mess with the entire system. And often when you have these technological changes, a lot of the successful entrepreneurs in the early days are building new applications that can usefully fit into the old system, but allow the companies, the organizations do things better, faster, cheaper, more efficiently in various ways. In the internet, we can think about applications. We call them applications. In the 1990s version of the internet, it was new ways to look through information, new ways

Starting point is 00:18:44 to connect buyers and sellers, but without inventing a whole new company. Electronic communication led to this technology called EDI, electronic data interchange in the late 80s, 90s that really helped supply chains be more efficient, like retailers connect with suppliers more efficiently, but it didn't really involve totally changing the workflow. It was an application that allowed them to do. what they were already doing, but better, that actually led to real value. That was like the Walmart and Amazon standard.

Starting point is 00:19:11 I remember Amazon trying to change it, and it was so deeply embedded that they couldn't mess with what Walmart had originally built, which is fascinating. If you think about these three ideas, the tendency, I can't remember who originally coined this idea, but maybe it was Steve Jobs or Chris Dixon or someone like that. But the idea is when a new technology comes, you do point solutions. They call it skeuomorphism. You tend to take the tech and apply it to something we did in the prior era. So the Notes app on an iPhone looks like a notepad, a digitalized version of the old thing.

Starting point is 00:19:42 And that's the point solution as you described it. As you think about prediction and machine learning, prediction models, artificial intelligence, whatever you want to call it, talk about that same progression as you've seen it so far and as you expect it may unfold in the future. What do those same ideas of point solution, application solution, system solution feel or seem like to you in the world of prediction and AI? I'm sitting in Canada. My favorite example of the point solution is company Darifin. And Verifin is predicting financial fraud. So they do.

Starting point is 00:20:12 It turns out banks need to predict financial fraud, and they have all sorts of processes to do that, a combination of human and almost symbolic logically, a bunch of fifthend states. Derafin actually was in that business already in predicting financial fraud, but not using machine money, not using AI. They were fine. And then they realized that given the data they had and given what they were trying to predict, they could use machine learning to do the exact same thing. It was just cheaper and better, but no one had to change anything.

Starting point is 00:20:38 In fact, many of their customers were already using Gerafit. It was a perfect point solution in the sense that there's already a workflow, you're already making predictions, it's already intermediate, so it's not even replacing a human in that particular context, but we can do it better and more efficient. At the point solution, they took out the old machine process, dropped in an AI process, workflows stayed the same, and they took off. This small St. John's, Newfoundland-based business suddenly became. became a tech darling and was, at least for Canada, our first AI unicorn. I think about application solutions. Another company we've worked with is called Ada Support. It is a startup for AI and customer service.

Starting point is 00:21:19 They were small. They stood in startup mode, but the real moment came in March 2020. In December 2019, they signed up a new customer called Zoom. In December 2019, no one knew who Zin was. And what they were focused on was automating parts of the customer service work. In particular, when people emailed Zoom for password resets or a handle of other requests, what Ada figured out is those requests are pretty standardized, and you can predict what kind of response to those customers are going to want.

Starting point is 00:21:48 You can add an AI into the workflow that automatically at least writes the email for the customer service representative to look over and send, and eventually could totally automate that response. And they did, and they helped Zoom. But then March 2020 came and Zoom's demand for these customer service services, queries went up by well over a factor of 10. And they could keep up because as they were scaling, most of the customer service queries were pretty straightforward. And they could have this application layer, this application AI, which is not quite doing what they were doing before, but it plugs enough into their existing processes that it can work to allow Zoom and others to scale. So

Starting point is 00:22:26 Ada was an application, which lets use AI to support customer service representatives so that one customer service representative can serve many, many more customers. And application solution doing things a little bit differently, but plugged into an existing value chain. System solutions for AI are harder to find. You can think of two. One of them is in prediction machines, we talk about, we call the knowledge. If you're a taxi driver, especially a taxi driver in a complicated city like London, it could take a long time to learn your way around the city. So the city of London, they call that ability to navigate the city called the knowledge. And it actually takes about three years of school to know enough about the city to be licensed as a tax driver. Then a prediction tool came

Starting point is 00:23:10 along called GPS and some maps that allowed you to get from point A to point B with real-time traffic reports, even if you didn't take three years of school. If you went to the city of London, you have to get used to driving on the other side of the road. But notwithstanding that, you could get from point A to point B not quite as efficiently, but almost as efficiently as a professional on a taxi driver who had three years of school. What that meant was the first applications, I should say, were point solutions. So many people, who was taking advantage of real-time traffic at GPS, they were professional drivers. Whether they were truck drivers or taxi drivers or others, people who made a living as a driver, became more productive if they knew real-time

Starting point is 00:23:47 traffic reports and had these predictions and could use the GPS, Google Nax or something else. The eventful of companies realized if we take digital navigation, so navigational AI, navigational predictions anyway, combine it with digital dispatch and combine it with a prediction tool about where people are going to want to get picked up by taxis to go from point A to point B, you can build an entirely new kind of business. Uber and Lyft and other ride-healing services created an entirely different system. They use that same technologies that taxi drivers could use to drive more efficiently, figure out how to find customers better, but in the process, they upskilled anybody who could drive could now effectively be a professional taxi. So they created a whole new system for

Starting point is 00:24:27 transportation. The other example of prediction leading to an entirely new way of doing business is in digital advertising. Advertising in the 1990s wasn't that different from the advertising industry in the 1960s that you might have seen, say, on Madden. I don't know about the sobopper part of it, but the way the business worked where there was a lot of charm and a lot of fancy dinners and big sales was a lot of how Madison Avenue worked back there. And when digital technology came along, when online advertising came along, the way people did it in the 1990s looked exactly like the way used to happen. If you wanted to advertise in a magazine in the 70s, 80s, 80s, you'd go to the magazine owner and say, hey, I want to advertise in your magazine, and they'd say, okay, here's our rate

Starting point is 00:25:05 card, here's our prices. It'd be some back and forth in negotiation, but fundamentally, there was the prices. And you would advertise. So if you want a full page out of people magazine, it's going to cost you out a thousand dollars, certain years, that's what you get. And they tell you their circulation. And the internet came along. You wanted to advertise online. That was the model they did. They said, hey, the internet's like the magazine industry. So you reach out and say, hey, yeah, who I want to advertise on the ad. And they say, great, here's our rate card. If you want to advertise in real estate, it's going to cost you $20 per 1,000 views. If you want to advertise for small businesses, it's going to cost you $50 per thousand views. If you just want to advertise in the

Starting point is 00:25:39 search engine, that's a generic advertising. It's not targeted at all. It's going to cost you $10 per thousand views, something like that. The industry just took the old model and applied it to a point solution. Then some people will realize, well, online advertising is different. Online advertising is different because there's this one-to-one relationship between the server sending you the ad, in the user. And that allows them to know not just whether the user's seeing an ad, but to predict a lot about the user who they are in their particular context of that particular moment. So what's different about online advertising is recognition that we can target ads based on the prediction about who the user is. And so over the next 20 years, an entirely new industry arose

Starting point is 00:26:17 that still called itself an advertising industry. But with completely different players, completely different technology, a completely different way of doing things, there's an opportunity to show somebody in an ad, and now the industry has this real-time auction that happens in fractions of a second for that slot. That was inconceivable in the old advertising world, but because we have this prediction combined with other complementary technologies around digitization, we have an entirely different advertising industry that now has players that we never even heard of before, demand-side platforms, fly-side platforms, and measurement agencies, data providers, all these types of players, a lot of them are just Google, but all these players are only it exists because of the AI.

Starting point is 00:26:55 as an industry is an industry that's experienced system level change through prediction technology. Can you talk a little bit about the transition from a world governed by rules to one governed by decisions? To say more about this concept of why a decision governed world might be more interesting than a rule governed one. More interesting and more complicated. A lot of the way we do things now, because we don't have good information, so we just follow the instructions. Because if we follow the instructions and everyone else follows the instructions, we know things will coordinated well. And what we mean by decisions are if you have a prediction, then if you know the state of the world, you don't have to do the same thing every time. You can start thinking creatively

Starting point is 00:27:34 and doing things different. After our experience over the last few years, the rule that many of us experienced was stay home. Warranty, we have no idea if you might be infectious. The easiest thing to do is to apply a rule to everybody. Just stay home. Joshua, my co-author on the book, recognized pretty early something that most people in the public health community weren't talking about, which is that for most people, COVID was not a health problem. At any given time, more than 99% of the population, this was summer 2020, did not have COVID. And if you don't have COVID, COVID's not a health problem. If you don't have COVID, what COVID is is an information problem, because you're worried about all the other people you might interact with whether they would have COVID, which would then turn it

Starting point is 00:28:17 into a health problem. Once you understand that COVID's an information problem, well, now we think, Here's an opportunity for AI. Here's an opportunity for prediction machines. Because what does AI do? It fills in missing information. That's what prediction is. Once you say, Cohen, for most of us, it's not a health problem. It's an information problem.

Starting point is 00:28:33 We see an opportunity to overcome the rule of stay home. Because if we had the information about who was infectious, then we don't have to follow the rule of stay home. We can now make a decision based on our predicted likelihood of being infectious and interacting with another infection. So in the presence of predictions, we could actually have gone about our business. We had a good AI for predicting whether somebody had COVID, then there would have been no grids.

Starting point is 00:28:56 And effectively, we would say, hey, you know what? You 1% of the population stay home and the rest of us, we can go about our business. We didn't have a video. Actually, there were all sorts of people trying to build that over the course of 2020 and AI to predict whether somebody had COVID. And there were these apps that you cough into your phone and to have some prediction. They worked all that well. Retrospecting seems kind of funny.

Starting point is 00:29:13 Now that worked. The best tool we had for predicting whether somebody had COVID was a rapid test. It's not perfect, but at least to figure out to predict if somebody was in fact, it worked pretty well. Unfortunately, that tool for a long time wasn't useful. Because we were so focused on rules, in order to use that prediction to make a decision, to move from rules to decision making, all these other things that happened in companies had to change. If you wanted to use rapid testing to test people coming into your business to make sure they didn't have COVID, and then open up if you were one of those businesses that were shut down, well, it seems easy. You just have a

Starting point is 00:29:47 decision about whether people can come in or not. But there's all these other decisions that coordinate with it, then you'd have to start figuring out. If you're doing it for your employees and they test positive, are going to give them sick pay. Well, lots of companies didn't have sick pay. Now they have to think about it. If you don't give people sick pay, they're not going to want to test. You need a whole bunch of decisions around health data because health data is private and how do we treat them. Turns out you need decisions around toxic waste disposal because this was considered biohazard because it was a medical test. We had to develop new processes around that. There are all these different pieces that in order to be able to move from the rule of, hey,

Starting point is 00:30:20 has to be shut down. We are, for that matter, our business is just going to stay open and the rules are open or closed. The decision is we allow our employees to come in to work if they test negative. To make that decision happen, there's all these coordination challenges that require a whole system. If we had control or whoever is trying to overcome the rules and say, hey, I want to make a decision, if they're also making all the other decisions, that it's easy. You don't have to worry about coordination. The problem is, in most companies, and in most contexts where you're trying to really move from rules to decisions, you need to coordinate between lots of people within the company and even between others outside the company.

Starting point is 00:30:54 All sorts of other people are involved and it becomes a much more complicated process. As you're talking, I'm just picturing the world moving from integers to decimals, a binary decision, which is a rule of thumb, treat everyone the same homogenously, to now what everyone's going to need to understand about the world is that it's probabilistic. There's some percentage chance that some set of things happens, and that's what a decision is. Investing, for example, you're thinking about expected value. A whole bunch of things could happen. what are the weights or the odds that certain things happen? When those things happen, what do I get as a return? You're building a probability tree. Is that the right simple way of summing this all up that what the power prediction does is move us from a binary to a decimal? I like that way to think about it. The world was always in decimals. If we're trying to fill in missing information, there's always some probability. But without a prediction machine, you have a lazy way of thinking, which is we tend to think it's either this or that. And so therefore, we're going to round up or down. We're going to round up or down. We're going to round up or down.

Starting point is 00:31:49 And what prediction machines do is they give us an explicit number. Oh, there's a 36% chance this is going to happen. And then the hard part is most of us aren't that good at figuring out what to do if there's a 36% chance of this going to happen as opposed to a 28. There's a whole learning process in getting used to making these decisions in the face of statistic. People who are probably best in the world of this are investors. Some investor audience is very familiar with. Let's think through the probabilities right out the decision tree or do our Monte Carlo simulation and figure out how this all plays out. But for many decisions, even investors when they're not investing, it's hard to think through

Starting point is 00:32:23 what's probability. Give an example of this of where it gets really hard. You've seen a movie I have a robot? Yeah, of course, and read the book. It's a great science fiction. In the movie, there's this flashback scene about why he hates robots. Here's why he hates robots. The protagonist, Wesmith, and his little girl, they're in a car accident, and their cars are sinking into a river,

Starting point is 00:32:41 and it's clear that they're both about to drown, and that a robot comes along and saves him and not the goal, and that's why he hates robots. by his aside, he then figured out, hey, it's a robot, I can figure out why it made the decision. The human you can never know, he audits the robot, and he figures that the robot predicted that he had a 45% chance of my, and the girl only had an 11% chance. That's why the robot saved the adult man and not the girl. And then he says, well, the 11% was more than enough, and a human being would have known that. But that's not a same about the predictions, that's the same by the decision. The judge is the relative value of an adult man's light versus the little girls.

Starting point is 00:33:13 The protagonist in the movie, Will Smith, says that that girl's life is worth more than the decision. than four times his life effectively. 11 versus 44 or 45, then you get them more than four. I don't know that we humans all agree on that, but I think we agree that that's hard. It's uncomfortable, and it's not something like doing. There's lots of decisions where it's useful to us. It's convenient that the prediction and judgment are bundled in our head. And hiring decisions is the same thing.

Starting point is 00:33:36 Why don't you hire this person and not that person? Well, maybe it's because you predicted that they had a better chance of succeeding in the company versus the other person. or maybe it's because your tastes are toward people who look like that or act like that or speak like that. It's convenient for us that our predictions about success are bundled with our human biases so that we don't really have to face our biases explicitly. But once you have a prediction saying, okay, this person's 90% likely to succeed and that person's 85% likely to succeed, which one are you going to hire? Now we have to face our biases in a way that we do.

Starting point is 00:34:08 That could be good. A lot of my optimism around AI is partly around how bad humans are and all sorts of things. things, including bias, but in this particular example, it's just that decoupling of the prediction from judgment. So the prediction for the rest of the decision is hard. And it requires a different kind of skill because most of us don't think probabilistic. I know a lot of your audience are a vestor. So another way to think about this is that skill you have at investing and thinking probabilistically will permeate all aspects of life because when you're trying to decide what to do, rather than just relying on doing the same thing every time, you'll now have information on

Starting point is 00:34:44 the relative likelihood of different things happen. I can't have a good discussion on AI on a business show without talking about concepts like disruption and sources of power in businesses. And I think there's a really interesting concepts when there's a GPT, the general purpose technology kind, regime change, which is probably at the peak of the steam engine era of a factory. The biggest factory was really hard to compete with. They had perfect. lots of efficient movements and structure and planning and blah, blah, blah. It would have been foolish decision to go up against them. I'm going to build a better steam factory.

Starting point is 00:35:18 But if you came as an electric factory, you could compete because of this transition. Talk a bit about that through the lens of prediction. There's lots of big technology businesses, let's say, there's lots of great data businesses. I'm trying to think of the existing dominant steam engine equivalent of today. Say a bit about what you think will happen, how disruption will occur, where power will shift and accrue I love thinking about those dynamics, and I think that's what all investors are, at least of my type, are thinking about now. Who's going to win in all this? Is it going to be the incumbents? Is it going to be the disruptors? Is it going to be some combination? Where all the great businesses? Where all the bad businesses? Some of these businesses that have taken off probably are not going to be good in the long run because they're so easy to copy, et cetera.

Starting point is 00:35:59 I just love these dynamics of power and disruption. In any gold rush, the lack of a better metaphor, people benefit by providing the equipment. There's going to be huge opportunities for companies that are providing the compute power underlying A. Right now, that looks like it's the traditional big tech companies, and they're providing the underlying compute power, and as the demands for compute power grow, there's going to be opportunities. Basically, everyone's buying cloud that's all running Nvidia GPUs. Exactly. If you make chips or if you provide cloud services, then for you. There's going to be a handful of opportunities there for sure. Second point is, maybe more interestingly, there's a whole bunch of business opportunities that are constrained by bad bridge.

Starting point is 00:36:44 My favorite example is thinking about pharmaceuticals. So pharmaceutical companies have patents, which means they have a monopoly on certain drugs over a period of time, and they can make a lot of money. But it's often a challenge to identify people who need their drugs, particularly for those that are relatively rare tree or relatively bare. There's a push in that industry toward Blockbuster that, hey, if you hit 65 or need this drug pretty much no matter what, we don't have to worry about the pretty much. It's not quite world-based, but really mass market.

Starting point is 00:37:09 That's where a lot of the opportunity is. Now, imagine we have an AI that does better diagnosis. And in particular, a diagnosis a disease that it's not that rare, but it's hard to identify. And so you have a prediction machine that identifies it. You have a drug that helps cure it or at least manage that disease. In the presence of a good prediction machine, the pharma company can now sell a lot more, increasing the profit, and also help a lot more consumer. This isn't a pharma company exploitation.

Starting point is 00:37:34 This is, hey, because we have a better prediction technology, we can diagnose. we can diagnose people at scale more efficiently for this previously seemingly rare disease, any production capacity on that kind of drug becomes much, much more profit. There's a lot of this kind of opportunity, which is once you have better prediction, you can do things that you couldn't do before. And if you have the assets that are a complement to prediction that could take advantage of it, there's huge money-making opportunity. You can see that in pharma, can see that in some kind of personalized goods and services.

Starting point is 00:38:04 This is because I sit in education, but I think there's huge opportunities in education because there's rules everywhere in education. We have people go co-work by cohort. Everybody in the second grade kind of learns the same stuff. Everybody in the third grade and stuff. Historically, we haven't had a really good way to deal with that because everyone in the second grade in terms of social emotional benefits by being with other second graders. And so how can the same teacher teach in a class of 30, teach 30 different lessons.

Starting point is 00:38:27 It's impossible. Add in a good prediction technology. We can now think through personalized education in terms of what they're learning the classroom. That's the best way to get this particular kid to do better and to reach their potential while also potentially keeping the class at the social emotional needs altogether with your cohort. I described that in the second grade, but I also think that same thing applies to my MBAs. Doctoral students largely get personalized because two doctoral students per professor, but the rest of the university, they don't get that personalized education, but they could.

Starting point is 00:38:54 In the presence of good prediction tools, we could deliver an entirely new kind of education. So that's piece one. Now, let's talk about disruption. There's two different concepts of disruption in the academic literature and they get mixed together. There's the Claytonson kind of disruption, which is that you start up creating a worse product effectively, but it serves a particular niche in a way that's not of interest to the core customer. So the plaintiffs and kind disruption was talked about disc drives and each new generation of disk drive was cheaper. Maybe didn't have as much storage in the first waves as the older ones, but could be applied to a wider variety of applications. The established discarized companies were selling into the same big enterprises

Starting point is 00:39:33 and the big enterprises didn't care about the digital storage that was going to be for small-scale application. So they ignored it. And a new company came along. And as they got better and better and better, they ended up disrupting and being able to do the stuff for the big ones as well, who called demand side disruption. Rebecca Henderson developed, well, you might think of a supply side disruption, which is companies have a way of doing things. This is around SOP, state operating procedures. And it's very hard to break them. If a new company comes a lot, new business idea comes a lot, new business model,

Starting point is 00:39:59 if that's going to change who does what in your organization, you just might ignore it. And so it may never happen. The example we talk about Empower and Prediction is in the context of Blockbuster Video. I was a kid. It was a big event on Saturday. We'd go, we'd go to Blockbuster video and pick our video. After the activity was going to Blockbuster and picking a video. Such good memories.

Starting point is 00:40:19 This generation will never know the joy. That just not a thing anymore. The story we often hear is that Blockbuster didn't see digital distribution coming. But if you actually look at the Blockbuster documents, they saw it coming. they just couldn't manage their organization to move it in the right direction. They were largely franchised, and the franchise owners, they didn't want digital distribution. There wasn't really a way to incentivize them the way the company operated to enable digital distribution. This is the shirky problem you talk about in the book.

Starting point is 00:40:48 Exactly. It wasn't that they didn't see it coming. Given the way the organization was set up, there wasn't much they could do better. The failure of Blockbuster in the long run, maybe there could have been some way to hive off a separate company. or do something like that, but it wasn't a failure of vision of recognizing what was happening. It was a failure of not knowing how to take the organization they had and turn into the organization. That's when we see disruptive opportunities, which is there's an incumbent, and they do things the way they do things. And what prediction enables is a totally new way of operating.

Starting point is 00:41:20 And while a totally new way of operating requires new humans and a different set of humans who are doing different things, a different set of assets, and that could just be too heavy a lift for a company. One of the hilarious things that's happened predicting how this will all play out was that we would replace a bunch of easy to replicate human jobs first and eat up this chain. The truck drivers were screwed and then eventually maybe 20, 30 years from now will get to the white collar artists or something. And it seems to have happened almost exactly the opposite. The stuff that's getting created the magic of Chad GPT or of Dolly or the stable diffusion models or whatever is sort of giving us stuff. by a prediction, and that's all these things are. Pixel prediction or word prediction or letter prediction. It's all these things do, which is why they hallucinate and do weird stuff sometimes

Starting point is 00:42:09 because they're just making the best prediction they can. And we know it's wrong. Back to your point on judgment. What do you make of all that? What do you make of the ordering in which this technology seems to be attacking existing functions or areas of the world or economy versus what we thought they might? I just find it so funny that we can't predict what the prediction technology will do. we absolutely can't predict it. It's prediction technology, so it relies on data. It's something totally new. We don't have data. So we have no idea what's going to happen. Partly because it's prediction technology, and if we think about, especially in generative models, what makes them so useful is they make the process much more efficient as long as you know what good output looks like.

Starting point is 00:42:51 So if you think about, I don't know if you play with JetGPT, presumably you have, or Dally, too. You know what you're looking for. And that's when it's most useful. If you know the anxiety, I do you see around chatchipt ads giving all these wrong answers. There's a lot of gotchas. I'm an expert in this and I asked to write a five-barragraph essay on that five-paragraph essay was wrong. Of course it was, but it's a great starting point for you to now turn it into a great five-paragraph essay. You use Dolly for graphic design. There's lots of people, including me, who are terrible at drawing.

Starting point is 00:43:18 But I often want to embed to graphic design and things I do. I can now create images at scale in a way I couldn't before. You ask for an image and it'll give you 10 and maybe only one of them. is what you're looking for, but you know what you're looking for. You know what the right answer is. And if you know what the right answer is, these models are incredibly helpful. The hype and excitement is running away from the technology and the anxiety is running away from the technology. Because this so far is a technology that's great if you know what you're looking for. That means when you're trying to identify the business opportunities from general of AI,

Starting point is 00:43:51 AI, more generally, you need to think through. You need to recognize this isn't going to be able to allow me to take humans out of the loop and processes that happen. But instead, maybe you have someone who spends all day writing an article. And now that's going to take them five minutes. They're going to know what the right query is. They're going to look at it and read it and say, oh yeah, that's great. I need to edit these six words. Done. And then they write a lot more content. They can embed imaging in that content through other tools. And somehow, I don't know what the example is because it has happened yet. Some creative entrepreneur is going to take that and figure out how to create a new advertising industry, a new consumer products industry, a new entertainment industry, etc. Earlier, you talked

Starting point is 00:44:32 about personalization as like a function. We should think about that function of personalization as something that gets drastically cheaper. Are there other functions? And again, I'm thinking about now I'm shifting my advice to businesses. If I'm an entrepreneur out there, how I should be thinking about the world differently. And it's very helpful for me as an entrepreneur, my myself to think what's like the more hyper-personalized version of this thing? Because that's going to be a lot more feasible or easy or cheap or whatever because of this technology. What other functions like personalization would you urge entrepreneurs to put as arrows in their quiver as they think about their products and their features? So start with what does it mean to serve customers in

Starting point is 00:45:12 particular industry go? And then you think through how do we actually serve customers in that industry and go to, okay, let's take a company. Let's go through their process. and try to separate the things they do into two categories. Category number one is actually delivering on your mission to serve customers. And category number two is the things you do to compensate your customers for your failures. And the more things that fall into that second bucket, the more you worry about disruption, and the more entrepreneurs should see a real opportunity.

Starting point is 00:45:41 As an example of this plays out, think about airports. It often airports in the world, Solention or Singapore, they're spectacular. They're beautiful. They're multi-billion-dollar. structures, great shopping and restaurants and all this other stuff. But how did the super rich fly? What are the airports of the super rich look like? Their sheds. Private terminals don't have great shopping and great restaurants and all those things because the ultimate in service in air travel is to ensure smooth air transportation. That's actually the mission of Solentian Airport often voted

Starting point is 00:46:11 the best airport in the world. But most of what they do is about a failure to deliver on that. Most of what they do is to compensate customers for the fact that they have to spend hours and hours' vehicle. But no one wants to spend time at year. That's an industry where you can think through, well, here's all the things they do that aren't about serving customers well. In lots and lots of industries, you can identify serving customers well sometimes means personalization. Sometimes it means efficient processes. Sometimes it means no waiting, depending on the context, it can mean all sorts of different things, getting you healthy quickly or preventing you from getting sick in the first place. But think through what does serving customers well mean and then go through the core companies in that

Starting point is 00:46:48 industry and identify how much of what they're doing is about serving customers well, and how much of what they're doing is about failing to serve customers, compensating customers for the fact they fail to serve them well. And then the last step, because we're focused on AI disruption, is among those, can you use prediction if you had better information? Could you then go directly to serving customers well and skip all that architecture that you have, all those SOPs that deal with the fact that you don't serve your customers as well as you could? I love the airport example. I love the category example. It's exactly the kind of thing I was thinking about. And it raises an interesting question about information and data. We've talked a lot about the importance of those two things and why if we had perfect information in many cases, we would do things differently. The COVID example is a great example of that. How should people think about the management of generation of information and data? It just seems this is an afterthought probably for a lot of, even now, even in 2023, the vast majority of data that gets created. never gets used. And the fundamental lesson of AI in the last five years is that the scale and

Starting point is 00:47:52 quality of data is everything. The way these things get better is by feeding them infinitely more data, not by structuring them better or imputing human knowledge into them. It's just more data. Say what you've learned about data and information and how people should think about it. So I love the way you put it's scale and quality because it's not just scale. You actually need both. And both of those need to serve a particular purpose. So if you don't know what you're trying to accomplished. Lots of companies say, we're going to try to organize all of our data. They spend millions and millions and millions of dollars creating a easy-to-use data interface, but they have no idea what the data is going for. They don't know what they're predicting, and ultimately,

Starting point is 00:48:28 much of that money ends up being wasted. When you're thinking through scale and quality of data, your starting point really should be, let's say we're going to do system level change in our company. Let's say we have identified a way to deliver a much better product through a new system-based of better prediction. What does that ultimate prediction look like? what are we trying to predict? And then you go back and think through what data do we have already that's going to help that and what data do we need.

Starting point is 00:48:53 Now, in second stage, what data do we need? Well, now, how do you go up on collecting data? Strategy number one is you can buy it. Sometimes it might be out there. But more commonly, you can create it. And you can often create it by launching products into the market that don't represent the system level change,

Starting point is 00:49:09 but that represent either a point solution or application solution on the way. In the auto industry, we think a lot about autonomous vehicles. The company that launches the first autonomous vehicle, and they can start collecting data at scale from that, is going to have this beneficial feedback loop, and they're going to collect more and more and more data. It's going to be hard for anybody to compete.

Starting point is 00:49:27 But in order to actually launch that autonomous vehicle, you need to have enough data in the first place that you overcome the regulator and you're creating a safe and not dangerous. And so far, that's been a meaningful barrier, hasn't it? So what's the strategy? The strategy is you embed sensors, into all the cars you have on the road, even if you're not as an automotive company, an autonomous fuel. You don't have autonomous driving yet. And we've seen Tesla do this. They have cars with all

Starting point is 00:49:52 sorts of sensors that are driven by humans and are trying to create data at scale through the sensors. And the other car companies are doing the same thing. If you bought a car in the last little while, the privacy policy on your car is incredibly complicated. Basically, it's because they're sending all the data about what you're doing, probably to help you now, but a lot of it is to help them build the car in the future. Some of the AI is already in there, like a warning for what the other drivers might be doing, but some of the AI are not there yet,

Starting point is 00:50:19 but they're collecting that data strategically in order to build what they want. Category 2 is to proactively collect the data based on actually going out and doing things with customers. There is a third category that's really intriguing, which is what we used to call simulation, but now the word that people use is digital twin, fundamentally the same thing,

Starting point is 00:50:38 which is you create a simulated version of the world that your product gets embedded in. And then you try different things and see how the system reacts. The challenge there is making sure the simulation is good enough to do what you want. Singapore has a digital simulation in the entire city. If they're going to build a new building, they have a sense of what that's going to mean in terms of extra density, extra drivers, etc. And they can simulate if we build this new building, what's going to happen to traffic? And do we need to change the roads?

Starting point is 00:51:06 And they can simulate building the building and changing the road this way versus that way or adding it want versus not and see how it all plays out. A third way to collect data is if you know enough about the situation, you can simulate it and then try hundreds, thousands, or in millions of different possibilities. And the underlying technology there is another kind of AI called reinforcement learning. You can think about it as a way to strategically collect data in order to help your predictions. You mentioned I robot earlier. One of the scenes that I distinctly remember in that movie is when Will Smith is trying to

Starting point is 00:51:36 trying to take over control of a car, and it's incredibly hard for him to do so. All the cars are not driven by humans. It makes me think of this interesting dichotomy I'd love your opinion on, which is, in the one sense, some stuff's going to be like farmers. We all used to be farmers, and now none of us are. Technology killed the profession. And then the other end of the spectrum would be, well, maybe Dolly makes more designers, not less, because the barriers to being a designer are far lower and everyone's imaginative and people like design output. Maybe there's latent demand for design that we just can't fulfill because the frictions are too high.

Starting point is 00:52:12 How do you think about those two categories? Because you start going down the professions, doctor, lawyer, accountant, are these going to be like farmers? There's just going to be no lawyers left because law is code-ish and we still need some farmers. So we'll still need like some lawyers, but we've got 97% less of them or something. This seems to be a really important question of those technologies. make a category blossom or does it decimate it? And I'm curious if you have ways of thinking

Starting point is 00:52:37 about that applied to specific professions or industries. Farming blossom. Just be clear. So there's a lot fewer farmers. But a lot more efficiency. Right? My hope, and I'm using the word hope on purpose, is that all those other professions turn out to be like farming, where we don't have to worry about the stress and cost of dealing with the legal system, whether it's for taxes or criminal or whatever else. That friction mostly goes away, but we still have a few lawyers. That sounds like an amazing world. The transition is stressful to be managed, but the future, we all can deal with the state efficiently without having these very expensive intermediaries between us and the law of being lawyers. That sounds great. Medicine too. A lot of the stress in the medical system

Starting point is 00:53:19 is around the social aspects and the personal aspects of dealing with bad health news. Right now, doctors don't have a lot of time for that. Other medical professionals might. So a world where we have fewer doctors, but 10 times as many medical professionals, sounds like a wonderful one. That's an upskilling one. You describe those as a dichotomy, and I kind of think one's almost a necessary condition for the other, especially since we saw what chat GPT can do, the thing that I'm most excited about for AI is this idea that millions of people can now do the things that historically only thousands could do. That's going to be stressful and disruptive for the thousands, and that's a important policy question. But it's

Starting point is 00:53:59 going to be amazing for the rest of us. And thinking through the transition from the 19th century, through the 20th and farming, yeah, we've a lot less farmers, but food shortages are, aren't really an issue with North America anymore. That's amazing. If we can do that across many industries, that's the real opportunity for AI. We've obviously covered a ton of ground here, conceptual, specific, et cetera. In closing, what are you most excited about and what are you most worried about? If you just think about the spectrum here of the future, what sits at the end of those distributions in your mind. I've covered the most excited, which is I think this opportunity to take things that only

Starting point is 00:54:34 a handful of humans can do, they're in sort of an outside living because of it, and allow everybody else to do it, or not everybody, millions of others do it. It's a very exciting world to imagine, whether it's writing or graphic design or diagnosis or aspects of financial services or other things. What I'm most worried about? The direct consequence of that is there are people who will be hurt. And I worry for them and try to think through, how do we mitigate the pain that they're going to experience? Whether that's for income or training or other,

Starting point is 00:55:03 else there's a whole set of questions around there that are a thing one to order about. Perhaps bigger thing to worry about is, yes, the technology cannot skill many, many people, but the underlying technology might be owned by a small number of companies. If this AI future leads to a reduction in equality in terms of skills, but leads to a massive increase in a handful of people who essentially own capital, who own the machines, and they become much, much better off. But leading to massive inequality and exploitation of their work of power, that's a real thing to worry about.

Starting point is 00:55:33 Maybe the best way to put it is, I worry about the concentration of power through AI. What a fascinating conversation. I really encourage everyone to go read the books, especially Power and Prediction, which is the more recent of the two, just again, lots changed since the first one came out, But just conceptually, it's been the best piece of content I've encountered on applying these ideas to business and systems and products. And I just think it's so valuable as people can tell from this conversation. Whenever I interview anyone, I asked the same closing question, what is the kindest thing

Starting point is 00:56:02 that anyone's ever done for you? Early in my academic career, I had a lot of failure. First, maybe 10 papers I submitted to journals, got rejected. It was brutal. Our vice dean at the time, I had charged a faculty. He seemed to believe in the underlying idea of understanding the economic technology. We would be clear. At the time, I had no idea he was doing this. I only know this with the benefit of hindsight, but he essentially structured my early career in terms of who was

Starting point is 00:56:28 responsible for promotion decisions around what I was doing and what my responsibilities were at the university to set me up for success despite what on the service was failure after failure after failure. He believed in me in what I was doing and really more than anybody else career-wise sent me up for, I like to think of it as success anyway. I'm really appreciative. It's a nice magical example, especially because you didn't know what's happening behind the scenes. There's something about that that makes it all the more special. I only found out when he retired. Well, this has been a really great conversation. I'm excited to share it with more people. Thank you so much for your time today. Okay, thanks. Take care.

Starting point is 00:57:07 If you enjoy this episode, check out join colossus.com. There you'll find every episode of this podcast complete with transcripts, show notes, and resources to keep learning. You can also sign up for our newsletter, Colossus Weekly, where we condense episodes to the big ideas, quotations, and more, as well as share the best content we find on the internet every week.

Invest Like the Best with Patrick O'Shaughnessy - Avi Goldfarb - The Economic Impact of AI - [Invest Like the Best, EP.321]

There aren't comments yet for this episode. Click on any sentence in the transcript to leave a comment.