ChatGPT took the world by storm, reportedly hitting one million users within five days of its launch. The technology is now estimated to have over 180 million monthly users. In this episode of Trading Tomorrow - Navigating Trends in Capital Markets, we look at the reasons behind the burst in popularity and what the future might hold for AI and MLs.
Tune in for a fascinating conversation between Host, Jim Jockle of Numerix and expert Adam Hyland, currently a PhD student in the Computer Supported Collaboration Lab at the University of Washington. From why Hyland didn’t see this trend coming, to if prompt engineer training will become the norm, to other tools that have garnered similar attention; this episode is a treasure trove of AI knowledge.
ChatGPT took the world by storm, reportedly hitting one million users within five days of its launch. The technology is now estimated to have over 180 million monthly users. In this episode of Trading Tomorrow - Navigating Trends in Capital Markets, we look at the reasons behind the burst in popularity and what the future might hold for AI and MLs.
Tune in for a fascinating conversation between Host, Jim Jockle of Numerix and expert Adam Hyland, currently a PhD student in the Computer Supported Collaboration Lab at the University of Washington. From why Hyland didn’t see this trend coming, to if prompt engineer training will become the norm, to other tools that have garnered similar attention; this episode is a treasure trove of AI knowledge.
Welcome to Trading Tomorrow navigating trends in capital markets, the podcast where we deep dive into the technologies reshaping the world of capital markets. I'm your host, jim Jockel, a veteran of the finance industry with a passion for the complexities of financial technologies and market trends. In each episode, we'll explore the cutting edge trends, tools and strategies driving today's financial landscapes and paving the way for the future. With the finance industry at a pivotal point, influenced by groundbreaking innovations, it's more crucial than ever to understand how these technological advancements interact with market dynamics, from the transformative power of blockchain and secure transactions to the role of artificial intelligence and predictive analytics. We're here to ensure you stay informed and ahead of the curve. Join us as we engage with industry experts, thought leaders and technology pioneers, offering you a front-row seat to the discussions shaping the future of finance, because this is Trading Tomorrow navigating trends in capital markets, where the future of capital markets unfolds.
Speaker 1:Chatgpt took the world by storm, reportedly hitting 1 million users with five days of its launch. It is now estimated to have over 180.5 million monthly users. It is a topic nearly impossible to avoid seeing in the press, in pop culture and through social media. Over the years, we've seen many companies where ideas explode in popularity and sharply decline. For example, blockbuster Video, once valued at $3 billion, now only one of their 9,000 stores is left. On the other hand, abba's music gained international popularity once released in the 1970s, but it has continued to permeate pop culture through the years, now arguably as popular as when they started. So here's the question Will ChatGPT become the next Blockbuster or the next ABBA? Will it disappear, giving way to more popular options, or will it become a staple in our future, permeating many industries?
Speaker 1:While it's impossible to predict the future, as we in the modeling world know, clues to the future may lie in the past. So here to discuss this is Adam Highland. He's a graduate student at the University of Washington in their Department of Human Center, Design and Engineering, where he studies how engineering communities coordinate around standards, also studying how communities build things that look like standards but aren't. You can follow his work at adampunkcom. Welcome to the show, adam.
Speaker 2:Oh, thank you.
Speaker 1:It's a pleasure to be here. Let's jump right in. Chatgpt has taken the world by storm, yet it's not necessarily that new. So did you see this ChatGPT trend coming?
Speaker 2:So I'm going to go ahead and say no, and I think that there's a lot of folks that did not see it coming, because if you weren't in this world and you were just looking at the performance of how a chatbot worked that was based on some natural language programming, you might think, oh, they're getting a little better, they're getting a little better, they're getting a little better, but not this explosive change in capability that we saw when these models first started coming out and especially that we saw like last year with GPT-3 and GPT-4. So I did not see this coming. This was really impressive how much it blew up.
Speaker 1:And, arguably, how quickly it's integrating itself into the work and life, if you will, of so many individuals, whether it be on the phone or on their PC. I mean that, I think, is that adoption rate has got to be ultimately surprising.
Speaker 2:Yeah, and the adoption and, like you said, I think integration is a really important. Distinguishing feature of how much this sort of blew up on the scene is that not only were people adopting it, but people were linking it into other processes, like connecting the chat GPT to a text-to-speech generator right or the reverse right, connecting it to voice-to-text and then sending that to chat GPT, where all of these multimodal components coming together is something that we have not seen at this speed, just ever.
Speaker 1:But you could also argue that there's like an emotional component with chat GPT. So maybe you could talk about that a little bit and how that's connecting to this popularity explosion, if you will.
Speaker 2:Yeah, yeah. So we think in like the financial world, you're working with technologies of like big servers and very powerful models and as they get faster, you don't really see a like someone doesn't treat it differently. But with something like chat, gpt, when it seems to speak in language, suddenly the advance that causes is like whoa, whoa, this is talking to me, I can say something to it and I can seem some sort of affect coming from it, and people behave really differently to tools like that when there's a perceived emotional component.
Speaker 1:How would you compare the advancements and capabilities coming from GPT-2 to 3 to 4, and arguably 5 is right down the road?
Speaker 2:Yeah. So this is, I think, connected to the why I didn't see it coming, because I was really interested in working in this space when GPT-2 came out and that was made available and that was a much smaller model. And one of the distinguishing features about these large language models which are powering these chatbots, which are powering image generators, is, as they grow in parameters. You know, you add more and more and more parameters to this distribution. You can think of it like if you're tracking a trend and you fit just a regression line to it, you've got two parameters, you've got the slope and the intercept. But if you have something that you know is tracking a lot of it, you might have five or six or seven or so on and so forth.
Speaker 2:For these you've got billions of parameters and a model that's got about eight billion parameters. It can do some stuff. It can do sentiment analysis, it can see you know is this good or is this bad. It can complete short sentences. It maybe can do if it's trained on code do a little bit of code completion. But then you get a model that's got like 50 billion parameters and that starts to be able to do new things, things that are more sophisticated things that the smaller model wasn't even doing poorly, it just wasn't capable of. And then you have a larger model. You get 10 times that parameter space, like 500 billion parameters. If you can start to see novel generation of text, you can see strong summarization capabilities, you can see ability to pass reasoning tests like the LSATs or the GRE, you can like. All of these new things come from just increasing the parameter space. So one of the fascinating things about watching this was, as they're adding parameters to these models, they're becoming not just more sophisticated but capable of new stuff.
Speaker 1:Well, I wish I had that when I took my LSATs. So what about? What does this say for the future? Right, you know? What could someone expect from a GPT-5?
Speaker 2:This is kind of interesting because in you know, there's a little bit of debate about this.
Speaker 2:There's some folks that argue okay, you're going to start to see diminishing returns as you as you add more and more parameters.
Speaker 2:And you know, in terms of like energy costs, like adding more and more of these parameters for the compute time makes it much, much more expensive to compute these models and sometimes it makes inference more expensive.
Speaker 2:But like just building the model requires an enormous amount of time on these computers and as you add these parameters it just gets more and more expensive to do. And then there's another school which is starting to see some research show up in their favor that we will continue to see more novel capabilities as we add these parameters, that the kinds of understanding and reasoning that we expect may just be possible through the additions of scale. So you and there's like other routes, like there's so much research in this, like some people are adding scales, some people are changing how the models are trained and behaved in the first place. So really focusing like deep mind is working on like building, like geometry and algebra models that are not really large language models per se but they're something like a harness of a different kind hooked up to it, and so they're training different things. And that's not about the parameter space, but even just in large language models building these chatbots, it's possible that increasing this parameter space is just going to continue to increase in capability.
Speaker 1:Well, you know one question that I always come back to, right and thinking about financial institutions and banks, a lot of them have banned the use of of chat GPT in office and obviously you know you're interacting with a public model, right, and you know it's something I'm mindful of, of making sure I'm not utilizing any proprietary data in my interactions and things of that nature. But not everybody is mindful of that. What is the implication in for chat GPT five when you know, you know arguably maybe 20% of the interactions are including proprietary data at this point?
Speaker 2:Right, and so, like one of the one of the things is absolutely like everything you type in there can be used for training. Everything you use for embedding documents to use the document API can be used for training, and there isn't a human being at open AI saying, oh boy, this looks proprietary, we're not going to use it. Right, the ingestion of these materials and then the use for training is done by machines. And also we've seen that people have been able to get chat, gpt and other large language models to divulge some of this training data. So if someone knew, for example, that your financial institution was heavily using you know a good chat GPT and they also knew a way to to exfiltrate that training data, they might specifically target you by targeting open AI and attempt to recover some of that. Now, that's pretty sophisticated but, like Lots of these things are not outside the realm of possibility, especially for a really concerted actor and so for for people who are privacy focused, there are.
Speaker 2:You might not be able to use open a I. You might have to run a local large language model. You might build one on your machines and make sure that you, your organization, controls it. It will probably be less sophisticated than the one that open a is able to build, but it also can be tailored much more closely to your interest, and it's something that I do with my students, where we build, like we have a small large language model they interact with and that allows them to do things that would be difficult to do with the content policy. We're trying to get it, to get it to say like okay, help me, help me develop this, this drug, that's bad, right, just to show that it's possible that, even though these things have L Ethical filters, they can. They can be, they can be breached and you know. Other elements of like trying to crack these models open are hard to teach when we're just dealing with open AI, but if you have your own machine, you can teach these and, conversely, if you have your own machine at a financial institution, you can control a lot more of what it generates, how it learns and how people interact with it.
Speaker 1:So so you know it's interesting to so obviously open. Ai took the world by storm, right, but now it seems Microsoft, google, you know Microsoft, with co pilot and it's integration with Office 365, which a lot of institutions use. I mean they, they come out right when you open that little panel. It's, you know your, your information and your organization's information is protected. Is this creating, you know, a launch pad for, let's call it, the tech incumbents to kind of corner the market?
Speaker 2:Yeah, so that's one really powerful way to do it and I think Microsoft is really smart in investing really big in open AI and integrating it into their tools, into Visual Studio, in Dove C 65, like you said, and that provides a strong ability for lock in. And a lot of these firms have tried to do lock in a little bit. When they started working with large language models a little bit differently, they tried to say, okay, the cost of training and the cost of adjusting these models with human feedback is a moat for us. Right, we spend Twenty million dollars to train this model. We spend fifty million dollars to, you know, adjust it with human feedback for people that we pay to rate responses, and that is hard to replicate.
Speaker 2:And I think one of the things that some of these firms are discovering is that that's not enough. It's not enough to build that moat with really big model and really expensive human feedback training. You need to be able to integrate it into services that people expect to see it, like no one else except Microsoft is going to integrate something into Office 365. And so if Microsoft has, you know, a strong ownership interest in open AI, they can, with that partnership, they can provide, like you said, some incumbency, and if someone, if some other firm wants to be able to achieve that kind of lock in, they're going to need to take an approach like that. It's not going to be enough just to make a bigger and bigger and bigger model and hope that it doesn't get leaked and hope that it doesn't get replicated, because those are two really useful strategies for competitors.
Speaker 2:You know there were two this is not about chat generators, but for image generation. In 2022, when they released stable diffusion, the, there was one company, stability AI, that released it, and another company, novel AI, that had a similar model that didn't intend to release it. They were going to put it behind an API and they were going to make money off use of it, and within a week of their launch, the model was leaked to the Internet and their whole business model based on that mode was just gone in a week. And so you know, some of these companies, I think, are discovering that they need different strategies in order to maintain some sort of market power.
Speaker 1:You know. It's interesting that you mentioned that. I was reading something recently that said about 85% of all AI related projects fail as compared to IT type projects, and I guess you know part of that is not just the business model itself but the training, the bias. I guess there's so many components for large language models to be successful. Would you agree with that?
Speaker 2:Yeah.
Speaker 2:So I think there's a couple of things Like first, this is new, this is a really new space and I would expect a lot of projects to fail in this space just due to the novelty.
Speaker 2:But I think, specifically because people look at chat, gpt and imagine the capabilities that it's got, their eyes get bigger than their stomach and they start to imagine like, oh yeah, we can just plug this thing in there and it's going to be able to solve this customer service call problem, it's going to be able to generate, you know, like efficient text or something. And they're not working in really in the reads with the problems that these models have, and like realizing, oh, you know, in order to connect this to what my exact business needs are, like how I've articulated this specific need for this I have to be able to solve problem A, problem B, problem C, problem D. And when you're imagining that, oh, this model is just so capable, it's producing this text, it understands my responses, and you kind of think about it in those terms, it's easy to forget the implementation details that are really going to drive whether or not you're successful, and I think that's like the big element that's causing a lot of these to fail.
Speaker 1:I'm starting to fancy myself as a prompt engineer. But you know where, would you say. You know the general world is in terms of understanding prompts, getting there and getting the most out of a GPT model.
Speaker 2:Yeah, so this is a really. I think this is a really fascinating conversation, because prompt engineering is kind of this very new thing and it came about because the models that we're working with are trained on large swaths of text and so they operate by what seems like natural language. So when you interact with it, you're writing English in a sense, and it seems like it's producing English, and so that has created this moment where we can interact and structure the responses of this model using what feels like English. And we've tried to do this in programming before. You know, like SQL which lots of people might be familiar with this listening to was originally designed so that, like you could write, but it felt like English If you write it. And now it doesn't really feel like English but like prompt engineering kind of does.
Speaker 2:And one thing that I'm interested in seeing in the sort of years ahead is whether or not this persists or whether or not we get sophisticated enough tools to layer on top of these large language models so that we may interact with those tools with a more programmatic element. Like maybe it becomes much more efficient to just someone's written a big Python library that talks to a large language model and then the people writing it right in Python and being a prompt engineer is something that lasts for like three years or five years. Alternately, that's a possibility. Alternately, it could be that the capabilities of these large language models to do work, to write code, all these things are so sophisticated, and we preserve an ability to work with them in natural language, that everyone becomes a prompt engineer. There's very few programmers in the future and we're all prompt engineers. I don't really know which way that's going, but I think that both of those are distinct possibilities.
Speaker 1:One of the things I think about is in cloud adoption, right was kind of the big thing I wanna say in 2011, especially within financial services. It's only come into fruition today, so it makes me think of that classic kind of Gertner hype cycle the high expectations, the trial of disillusionment and getting to kind of adoption. So we've seen all this hype around chat, gbt, but is this really the first time we've seen this kind of hype around AI or an AI parallel type of technology?
Speaker 2:I think no, and I'm gonna point out two times in the past. And the first one is like right when electronic computers were invented, for two of the earliest general interest books about electronic computers, written in 1949 and another one in 1953, one of them was called Faster Than Thought and the other one was called Giant Brains, colon machines that think, and the even had just having built these computers, just having started to use them for general purpose stuff, where we were starting to think and imagine their possibilities for replacing thought, for expanding thought, for extending thought. And when we came face to face with the fact that actually working with the computer didn't really seem like thinking and that the kinds of things that were solving problems were very far away from people, that hype kind of died down like a little bit. And in the 60s you saw it again among people that weren't technologists, with actually with a chat bot called Eliza, and this was developed at MIT in 1964 and was something that was meant to.
Speaker 2:It was a very, very simple program, very simple program. It was just doing very simple pattern matching and it had a list of sort of responses and it could replace elements in the responses when it replied, but to someone interacting with it. It had that same sense that we feel with chat GPT, where they're, oh, it seems like it's understanding something, it seems like it's responding with feeling and with sentiment that humans have, and this became really, really popular, like it just blew up in popularity. Nothing kind of came of that because, unlike you mentioned, chat GPT is getting integrated into a lot of stuff and being hooked into things. You couldn't really integrate Eliza into something else. Like you could use it and you could play with it and then that was it. But that became very, very, very popular.
Speaker 1:If you're listening to this podcast, either you know the power of chat, gpt and AI, or you're on the path to discovering that. What doesn't AI do well and where do you see areas of near term improvement?
Speaker 2:So it's really hard to answer this question in the longer term, because I think if you look at the history of technology, you will always find sort of the humanist saying well, technology cannot do thing A, and when thing A falls, it becomes technology cannot do thing B, and when things B falls, you know. So on and so forth. And if you were a gambler you should bet on the machine eventually being able to do a thing that we imagine as human. So I have a paper that I wrote last year trying to talk to educators and artists about understanding machine learning, and it was titled Hands Are Hard and the idea of the paper was if you were looking at these image generators like stable diffusion when it first came out, and you asked them to generate like a human face, it would do a pretty good job, pretty decent job, get a good job. You know you'd look at that face. You know I had nothing wrong with that, Maybe the ears are weird or something like that.
Speaker 2:If you asked it to draw a hand, it was a horror show. All the hands were bad and we were wanted to introduce people to this concept and get them to understand, like what these machines fail at. But we wanted also to say that this is not going to last. Even now, if you go to the current version of stable diffusion or something like mid-journey or Dahle three, they're much better at generating hands. So all of the kind of things that we said, oh, these are, you know, elements, the reasons why it doesn't draw hands.
Speaker 2:Well, some of those has kind of evaporated in just the span of a year, and so it's quite difficult. You know, like Yogi Berra says, predictions are hard, especially about the future, but it's quite difficult to look ahead and say it's not going to be able to do this thing, and so I almost am hesitant to say that there's sort of a redoubt of human performance that is not going to be eclipsed by some of this I want to remember is it's not just like the machine doing it by itself. When we say that the human performance is being eclipsed, these are teams of human beings working together with these machines, with incredibly sophisticated tools, learning how to use those tools, making new capabilities possible. So you know, these are systems. This is like.
Speaker 2:This is Microsoft, the corporation is is achieving superhuman performance in some of these capabilities. Right, we can think about. It is like oh, the large language model is, but like that's really the work of just so many researchers and so much you know, like so many people working together to train these things, to find sophisticated new ways to work with them. Like that's, that's what's building these capabilities and and that is not something to be underestimated.
Speaker 1:Do. Any podcast that quotes yogi bearer gets an A plus in my book. One final question that I'm going to get to our trend drop. You're at the department of human centered design and engineering and you're studying engineering communities, right. So what is the depth? How are you defining communities? Physical communities, like outside my window?
Speaker 2:Network virtual one of the things that we work at my lab I study under professor charlotte lee, and one of the things that we work our communities that are not just like one team, one firm and my research. Right now I'm studying the history of how we standardized computer arithmetic, and the floating point arithmetic standard was one that was put together by academics, by people from industry, all working together in a distributed fashion, and so we can kind of, in that context, imagine the community of people who were concerned about reliability of numerical programming on computers and that shifted over time. You know, at the beginning of computing it was folks working in government, like physical laboratories, like our national laboratory or the national physical laboratory in the UK, and then it was like aerospace and nuclear engineering, and then it started to move out into different areas and there were different stakeholders who came in and there were different people who were sort of capable of contributing to these discussions.
Speaker 2:So it becomes a little amorphous and you have to very carefully track how that, how that happens over time, like the community of people who care about floating point arithmetic on computers today is really different. Then it was in 1985 when the when the first standard was created and you know different again from you know when we started doing this kind of arithmetic on computers. And when I think about the context of AI, it's also quite difficult to figure out what the community is, because some of these there are communities and on GitHub around particular kinds of software connecting these things. There are communities that are just building up around certain creators on discord when they built a tool and the community is kind of central. The central version of the community is the people who show up to that discord server and talk about the tool and try it out and change things and those folks might are building stuff that changes.
Speaker 2:You know how these models work for lots of people, but the community that's doing it might be very small and so like having to like zero in on that and have to figure out how do we determine membership? How do we determine who's in charge? How do we follow what's going on Is a tricky problem.
Speaker 1:You know it's interesting because that kind of speaks to a little bit of the DNA here at numeric. So we have a whole bunch of people that are came out of Los Alamos, yeah, so you know, in terms of high speed computing, quantitative finance, the whole bit. So we've got to our final question. We call it the trend drop. It's like a desert island question. So if you could only track one trend in chat, gbt or more widely, if you prefer, in AI, what would that trend be?
Speaker 2:So I think for large language models, something people are sleeping on is search. We think about these as generators, text generators, and I want to look at and I think people listening to this podcast should look at companies that are training local large language models to do document searches for large organizations. So you feed in a great corpus of information from your organization and it learns, in a sense, the structure and the nature of the kinds of documents produced by that company and allows for a much more sophisticated search to determine for audit purposes, to determine perhaps, like, oh, what might be missing from this set of documents, what might look doctor from this set of documents or, just for you know, recognition. So like I think they're probably a few startups in stealth mode that are interested in document search for large language models, like for corporations, and I think that's a really interesting spot to look.
Speaker 1:Well, adam, I mean, this is fascinating. I have a feeling we could probably talk for another hour, but I want to be mindful of your time and I really appreciate your insights. And again, if you want to learn more about Adam and his work, you could follow him at Adam punk dot com. Adam, thank you so much, absolute pleasure, thank you.
Speaker 1:Coming up next week on trading tomorrow, navigating trends in capital markets. We've talked about the future. Now it's time to talk about the present. As I continues to dominate headlines and corporate to do list, the call for a bus transparent regulation grows louder. We speak with Professor Michael well made, one of the most important, influential AI regulation and ethics voices. It's an episode you do not want to miss.