AI – A History of Investing | AI Part 1

19:36 Oscar Pulido: Welcome to The Bid, where we break down what's happening in the markets and explore the forces changing the economy and finance. I'm your host, Oscar Pulido. Artificial intelligence can be described as the simulation of human intelligence by machines. As we heard in the previous episode in this two part look at AI, we heard how AI has been evolving for decades. But today we're going to take a look at the implications of AI for the investing landscape.

20:01 Here with me today is Brad Betts, Global Equity Researcher for BlackRock Systematic, an investment team that emphasizes the use of data driven insights and cutting edge technology in their approach. Brad focuses his research efforts on using artificial intelligence, Machine Learning, and Natural Language Processing to Generate Alpha. In other words, he uses AI to attempt to outperform the markets, and he's been researching these technologies for decades. From his time as a scientist at NASA, to the pivotal role he's played in bringing AI innovations to BlackRock.

20:33 Brad will help us understand what AI means for investing and the opportunities that lie ahead. Brad, thank you so much for joining us on The Bid. Brad Betts: Thanks for having me. Oscar Pulido: So Brad, AI is seemingly everywhere right now. We're hearing about it on company earnings calls. We're hearing about it in news headlines. But you've been researching and using these technologies really throughout your career. So I'd love to hear more about the evolution of AI in terms of how you've seen it.

21:03 And how it pertains to the investment management industry. Brad Betts: Yeah, my pleasure. I was an intern in the summer of 1989. I was working in a military research lab and one of the senior members of the lab had gone to a conference and came back and gave us all a briefing. And it was the first time that I had heard about artificial neural networks. And these neural networks are what is key architecture behind things like chat GPT.

21:24 And it was a particular demonstration of a truck, a big rig, you know, an 18 wheeler being backed up. Here I was seeing a computer learning how to do it. It wasn't being instructed explicitly by humans, it was learning how to do it. And this fascinated me. And it could do it even from a position that you call jackknifing a big rig, which is a really tricky position to get a truck out of. And it was riveting.

21:48 And so that was my first exposure. Back in those days, we had nowhere near the computing power, nowhere near the volume of data, nowhere near the accelerated computing hardware that we have now, and nowhere near the sophistication of the algorithms. But it captivated me. And it in no small part led to the great fortune I had in that next step of my career, which was to do my doctorate in something called the Information Systems Lab at Stanford.

22:13 That lab was buzzing with ideas of information theory and signal processing and cryptography and optimization and compression and medical imaging. It was just a wonderful time. And there were giants of the 20th century that were there, but there was another powerful force in this information systems lab at Stanford. And it was a gentleman named Bernie Widrow, Bernard Widrow. And Bernie Widrow had developed the first practical use of a neural network. If you made a phone call for many decades in the U.

22:45 S., you were using artificial neural networks. It was being used behind the scenes to do things like cancel echoes on the line. Anybody that remembers older phone calls and you could sometimes get a really annoying echo. Well, that was a technology that had been developed by Professor Widrow and others, something called Adaline. Professor Widrow and others built Adaline in 1960, before I was born. So the history of this goes well back into the 1950s. And it was the first practical use of a neural network.

23:15 And again, these neural networks are what are underpinning so much of the progress and excitement. I then had the good fortune to use what I had been trained and taught at places like NASA, at startups in Silicon Valley, and then, close to 16 years ago now, at BlackRock. The evolution has been, for me, one of extraordinary excitement and privilege to see. The changes in these technologies to see the advent of more data, of more computation, of increased sophistication in algorithms has led to this remarkable excitement that we're seeing now.

23:49 Oscar Pulido: Well, it's interesting because for many of us sort of common folk we think of AI as a recent development. I mentioned the recent headlines, but I think you started your story with an internship back in 1989 and you mentioned a couple of terms, by the way, you mentioned neural networks, machine learning. I know we talk about like large language models, natural language processing. How do you think about those terms? What do they mean? How do they apply to the investment industry?

24:17 Maybe we can go there. Brad Betts: The field is definitely filled with jargon, acronyms, and terminology. Artificial intelligence is a very broad and ambitious field of endeavor, which is obviously looking to try to mimic human levels of intelligence. Machine learning is another term. And you can think of it perhaps as a sort of a subset of artificial intelligence that tends to be more focused on performing a particular task. It's using computers to learn from historical data, to take advantage of historical data, to make predictions about the future.

24:49 That's a common use of machine learning and its roots are in the signal processing and computer science and applied mathematics, the applied mathematics of optimization community. You mentioned natural language processing. Another term that people will use for that is also text analysis, but text analysis, natural language processing, NLP, is trying to have computers understand and take advantage to perform some tasks, human speech, human writing. So you mentioned large language models. Chat GPT is an example of a large language model.

25:21 We call them large language models because they, as rule of thumb, are trained with large amounts of text data. And also they have a very large number of parameters. You can just think of them as dials. You have one dial on your thermostat at home. Imagine having billions of dials. So how you set those dials is critically important. The last one that you asked about. Neural networks, maybe the slightly more correct, people usually just shorten it to neural network.

25:52 We really say artificial neural network to distinguish it from the biological neural networks in our own human brains. These artificial neural networks were meant to be very crude approximations of our human brains, of the neurons and synaptic connections. And it turns out that These techniques, pioneered by people like Bernie Widrow and many others, have proven over many decades now to be staggeringly successful. Under all of those terms you've mentioned, natural language processing, machine learning, artificial intelligence, large language models, under all of it is large amounts of data.

26:34 Large amounts of computing power to take advantage of that data. And then also this accelerated computing. It's not just the standard computers people tend to have at home. There's specialized hardware that is well suited to machine learning, natural language processing, artificial intelligence tasks. In particular, it actually grew out of graphics processing, but it turns out it was well suited to the mathematics of this, of this area. So I hope that gives a flavor and we've only touched some of it.

27:02 Oscar Pulido: Well, one of the added terms that comes to mind when you think of AI, and you mentioned it, one of the most widely publicized headlines around AI when ChatGPT was released at the end of 2022. And I think you put it in the category of large language models. So how does the large language model that ChatGPT uses compare to what you would use in investment analysis. Brad Betts: We use large language models for investment at BlackRock, but we use them very differently.

27:31 So we take these large language models and we tune them to make forecasts for the returns that companies will generate in their stock. So in effect, you can think of this model being tuned and becoming a real specialist. Big piece of computer code powered by massive amounts of data and computation, and becomes a specialist in making forecasts of what will the return be to Apple over the next week, over the next quarter, over the next month.

28:01 And the process by which we do that is one called fine tuning. So when people are interfacing with something like ChatGPT, they tend to be putting in text and they get out text. Whereas when we use them for investment, we tend to be putting in lots of text. And what comes out are numbers, forecasts of returns. And we use those forecasts of returns to conduct trading. Oscar Pulido: And Brad, one of the criticisms perhaps of ChatGPT is you put in text and the text you get back may or may not be accurate.

28:35 It might have provided a lot of convenience. There's a bit of like, how do we know what it's telling me is true? How do you then assess that with the fine tuning that you describe when you put in text that gives you numbers? How do you know how believable that data is in order to make your forecast? Brad Betts: You put your finger, of course, on a very difficult and challenging part of our job, which is exactly what you said.

29:00 How do you trust? How do you tune the neural network properly, appropriately, to do what it is that you want it to do? So, this fine tuning process, think of it as picking up an initial base model. That sort of understands a language, say English. It understands English very well. It can generate tokens, maybe not quite as well as chat GPT, but it can do it pretty well. And then what we do is we feed in new text and we give this text a label.

29:30 And this label is used by the neural net to understand if it's done something well or poorly. Say you take the transcript of an earnings call, or say you take a broker report from some sell side firm. This is a phase where you're training your neural net. Then you're going to use the neural net for new data. But initially what you're doing during this fine tuning is using history to make the net much more specialized. And so you put in a piece of text and you give it this label.

30:00 You can use the future returns that a company will receive over some horizon. And then the neural net takes that information, feeds it forward through its architecture, and it produces a forecast of what the return it thinks will be achieved. We reward or penalize the neural network on how well it does. And then the neural network, the algorithm, I'm describing it. So I'm personifying it. Of course, it's just math and code, computer code. But what happens is the neural network then adjusts its weights.

30:31 I mentioned earlier that notion of like maybe having a billion dials. How do you set those? This part of the algorithm is now called backpropagating. The neural net says, okay, I. Took that piece of text? I think that the firm in the future, over the next, say, month or so, isn't going to do very well. Oh, but you're actually telling me it did do very well. Thanks for giving me that feedback via that label. I'm now going to go and adjust my weight slightly to try to close that gap.

31:02 But you're doing this with huge amounts of information, and you're doing it very rapidly. And it's not just one piece of text, it's taking in billions of these items, and it becomes... Very skilled at matching what happened. And now what you can do, once you have trained it in this fashion, now you can use new pieces of information that it hasn't seen. So nobody knows because now you're talking about a real prediction. You're talking about the future.

31:26 The neural net can take that piece of text and say, based on everything that I've learned, that I've been trained powerfully. is a brilliant algorithm called ADMM, the Alternating Direction Method of Multipliers. Under the covers, a lot of this comes down to doing optimization well.

31:57 And this ADMM technique was brilliant. And we have developed and tuned an ADMM implementation for so many years, that under the covers allows us to do this fine tuning, because I described it as it's just matching returns. We actually get it to match exactly the investment problem or as close as we can make it to what we're really going to do. So that is to say we want the neural network to learn to maximize returns while minimizing risk and minimizing transaction costs and minimizing borrow costs.

32:33 And we're able to do this because of the brilliance of ADMM. Oscar Pulido: Can I go back to something you mentioned? And that was fascinating how you painted that image of the neural network learning being given feedback. Isn't the data that you are providing historical data and we're in a new investment regime. The Great Moderation is over. That period of low interest rates and low inflation. And growth being more stable and predictable. Well now we're in a new investment regime of more volatility.

33:03 So how do you think about the predictive nature of these models when maybe we're at an inflection point and some of the real time data that you are feeding it, it might not be trained for that? Or does the data go back so far that it has plenty of history to have picked up on various economic cycles? Brad Betts: Yeah, a huge change was the pandemic. As humans, we hadn't roughly seen a pandemic of this scale for around a hundred years, but think back to March of 2020, how many of us really predicted how sharp the rebound would be in markets and whether you're a computer or whether you're a human, a shift of that kind of magnitude is an incredible challenge.

33:45 The markets are a very dynamic system and the nature of a dynamic system is it keeps changing. It will continue to be true. Algorithms, humans have to be nimble. They have to be able to take advantage of what information they have available. To do the best job that they can for prediction. And we'll see, it's a fascinating time to, to be alive, to see who is able to adapt more effectively to these changes that, that will always be there in the markets.

34:11 Oscar Pulido: And that race between man and machine brings up something that we touched on in our previous episode, which are. Some of the ethical concerns that arise from artificial intelligence and what the impact is going to be on industries and society. And there seems to be two schools of thought here. One is very fearful about the rise of these technologies and there's another side of the camp that's a little bit more irreverent. So what's your view in terms of the business impact, the impact on society from artificial intelligence?

34:40 Brad Betts: It's my life's work. I don't go to bed worrying at night around a machine takeover. The opposite. For me, I'm excited by the opportunities that these present. The opportunities. You know, I feel that sometimes the conversation gets a bit focused on the risk. And that's human. But sometimes it misses the extraordinary opportunity. The opportunity that will come. You look at algorithms like AlphaFold, a neural network that learned the 3D folding structure of proteins. The opportunity that will afford for drug discovery is wonderful.

35:08 I fall in the camp of Being more excited. I think that the impact business will see is around automation. I think a lot of the ways initially and unsurprisingly businesses will see these impact will be in productivity enhancing tools, for example, on things like email. In code generation, in the translation of old computer code into more modern languages, to reduce operational risks, to put them into a more modern framework. How many of us really go to bed at night achieving inbox zero?

35:36 Imagine if you had an agent that was sitting there that was summarizing the emails for you, was summarizing presentations, was drawing your attention over time to the ones that It's starting to learn that you're more interested in, that it feels are more worthy of your attention, that are able to summarize. I think that is a way, and firms like Microsoft are working very hard and highly incentivized, of course, to bring those kinds of capabilities to many enterprises.

35:59 So I think these opportunities for automation are a path where we're going to see a flurry of activity around the adoption of these technologies. Oscar Pulido: I think the term inbox zero would resonate with a lot of resonates with me. It seems like an impossibility unless I delete everything. And then surely I'll have missed a lot of important things. But while you work in the investment industry, it sounds like a lot of the examples that you gave were actually just about productivity.

36:24 Automation. You mentioned healthcare as well. Are there other opportunities for artificial intelligence in, in other industries or other just parts of our life that we're not thinking about, that you think about when it comes to artificial intelligence? Brad Betts: You see them more and more in things like chip design, the layout of chips on silicon, on car, aircraft, spacecraft design, improvements and advancements in simulation, environments that, that come from these technologies. There's a lot, and I don't want to give short shrift to investment, my goodness.

36:53 There's so much opportunity that we have taken advantage of at BlackRock and will continue to take advantage of at BlackRock through these technologies. Our allegiance, Oscar, is to whatever is efficacious. Our mission at BlackRock is to do right by our clients. It's to do right and to bring them those investment results that they want and need. What really excites me, the opportunity for these technologies in an investment context, is that opportunity to continue to improve the outcome ultimately for our clients to execute the mission.

37:22 And also I mentioned that it It feels very personal for me. My mom grew up one of seven kids. They were of modest means and she put herself through high school and she put herself through nursing school and, and she raised three boys and she and my dad built a home for us and saved their money and bought me and my brothers our first computer, a Commodore 64 that took savings from them that took effort. And I've been obsessed with computers and math since I was a little boy.

37:49 And Oscar, if I can put my life's study of math and computer science, if I can put that to use. To help our clients to, then for me, that's a life well lived. That, that idea that maybe I can try to help a family that has never had a vacation. There's an extra basketball as a birthday gift or something, all the things that, that my mom never had. Then for me, that's, that's life well lived. Oscar Pulido: I love that story.

38:20 I'm sure a lot of our listeners will say the same. Brad, I know you're based in San Francisco. I'm just thinking about your comments. I've been through the office, but there's a quote somewhere that says something like the art of investing has become the science. I'm wondering whether you put that up there when you first joined? Brad Betts: Well done, Oscar. Good for you for remembering. And that comes from one of the most brilliant researchers I've had the privilege to work with, Ron and Richard grinald.

38:44 Oscar Pulido: Brad, thanks for all the insights and most importantly, thanks for joining us on The Bid. Brad Betts: Thanks for having me, Oscar. It was a real privilege. Oscar Pulido: Thanks for listening to this episode of The Bid. If you enjoyed this episode, check out part one. Where we took a look into AI's history and its inflection points for investing. Subscribe to The Bid wherever you get your podcasts.