Odd Lots - Goldman Sachs CIO on How the Bank Is Actually Using AI

Episode Date: August 8, 2024

There's a lot of hype around generative AI and many people have interfaced with ChatGPT, Claude, or Gemini at this point. It's fun to ask these large language models to come up with a song parody or to write a story, but most casual users of the technology probably aren't worried about things like copyrights, the sensitivity of what they're inputting into the platform, or even the accuracy of the answers being spit out. It's just fun to play around with the technology. For large companies, however, there's a lot at stake. And concerns over data privacy and output errors are even more pressing if you're a big regulated bank. In this episode we speak with Goldman Sachs Chief Information Officer, Marco Argenti, about how the bank is balancing risks and opportunities in AI. Argenti, who previously worked at Amazon Web Services, talks about the development of Goldman's own internal AI tools, what the new tech means for Goldman engineers and other jobs, what makes a good prompt, and much more.See omnystudio.com/listener for privacy information.

I'm June Grasso, inviting you to join me for the Bloomberg Law podcast. Every weekday, we help you make sense of the legal stories that shape the nation and the world. Listen for complete analysis of the biggest court cases, the latest actions from Congress and regulators, and the legal moves driving the markets, from corporate law to constitutional law, and from state courts to the Supreme Court. At Bloomberg Law, we go beyond the day's headlines. We speak with top attorneys, judges, scholars and policy experts to break down what the rulings really mean. We do this every weekday,

then bring you the best conversations in our daily podcast. Search for Bloomberg Law on YouTube, Apple, Spotify, or anywhere else you listen. On the East Coast, listen as you start your day, and on the West Coast, catch up in the evening. That's the Bloomberg Law podcast with me, June Grasso. Subscribe today wherever you get your podcasts. Bloomberg Audio Studios. Podcasts, Radio News. Hello and welcome to another episode of the Oddlots podcast. I'm Tracy Allaway. And I'm Joe Weizenthal. Joe, what's been your favorite chat GPT or Claude Prompt so far? You know, it's funny because I have a lot of fun with them and also I use them for serious things. So I'll like upload conference call transcripts and say, tell me what this company said about labor market

indicators or something like that. And that'll be extremely useful for that. Wait, do you actually find that more efficient than just doing a word search for like labor or working? I don't. I hate uploading stuff because you can only do it in like fragments. No, what? Tracy. No, let me, I'll show you how to prompt. No, I get a lot of professional use out of the various AI tools. But I also. so, you know, have a lot of fun with them. And there's even a song, and I'm not going to say which one that I wrote. I didn't use the lyrics.

No, I did not, like, because it's very good. Wait, what did you use? Did it give you an actual melody? What happened? No, so there was a song that I liked, okay? And the song title sort of rested upon a pun. Okay. And so I asked Chad GPT to come up with another song that sort of like had a similar

twist based on the headline of that song. I needed basically a song prompt idea. This opens up a whole can of worms. No, this is actually the perfect segue into what we're going to talk about today. Because for you and I using something like a chat GPT, we don't really have the same concerns that a proper company or large corporation would have. Like, it doesn't really matter to us if the answer is wrong. I mean, ideally we would like it to be correct. But if I'm just asking some silly question, it doesn't really matter what chat GPT spits out at me. And also, copyright kind of doesn't matter. So we don't care what it spits out in terms of who owns it.

And also, we don't care what we're putting in in terms of who owns that. That's right. But if you are a company, you are thinking about generative AI very differently. I just want to say one thing, which is that if I like... In your defense, okay, defend yourself. No, no, I'm not even trying to defend myself. If I upload, say, you know, the McDonald's earning. transcript and I say, what does McDonald say about the labor market? Then there's some quote.

I always go back and check that that quote is actually in there. So I do, you know, I'm not just blindly relying on it. I do also do my own work and everything. But yeah, it's very true. Like, so I can say, I get tremendous amount of use from chat UPT or Claude or whatever, and it is very useful to me. But it makes mistakes sometimes. And if you think about deploying AI in the sort of enterprise world, then maybe like a 1% mistake rate or a 1% hallucination or you ever want to call them is just completely unacceptable and a level of a risk that like makes it almost unusable for professional purposes. Absolutely. And of course, the other thing with AI is there is still this ongoing very heated debate about how transformational it's actually going to be. So you and I are

using it as, you know, a productivity hack in some cases or maybe to generate song lyrics or even songs in some cases. But what is the true use case for this particular technology? There's still a lot of debate about that. And so I'm very pleased to say we do, in fact, have the perfect guest. We're going to be speaking to someone who is implementing AI at a very, very large financial institution. We're going to be speaking with Marco Argenti, the chief information officer at Goldman Sachs. Marco, thank you so much for coming on all thoughts. Thank you for having me. Marco, tell us what a chief information officer does at Goldman Sachs. Whenever I see CIA, I always think chief investment officer as a sort of... Yeah, it's very confusing. Yeah. So what does the other

CIO do? So last week I was in Italy visiting my mother. She's 83 and she obviously doesn't know much about technology or banking. And so she said, what do you do at Coleman? And I said, you know, I just try to simplify. I say, make sure that the printers don't run out of paper. And interestingly, the CIO job has been traditionally associated with the word IT. And IT, I tell you, talk to any technologist, they don't want to be classified as IT. Right, because those are, you associate with those of the people who, like, see if the Ethernet cable is probably.

Those are the ones who tell you to restore your computer. I mean, I have a lot of respect for IT, but generally you go to the IT department when something doesn't work, okay? Yeah. And so it's very back office. And something that attracted me to this job, I've been here for five years, and this is the first time that I do like a CIA job before I was doing more like, you know, creating technology, et cetera, and service, I can talk about that. But is the fact that the role of a CIO has actually changed quite a bit. And now it's about really asking the question, you know, how do we implement technology in order to achieve our strategic objectives and actually to be differentiated?

And it's really sitting at the strategic table of the firm, okay? So today we live in a world where obviously a lot of the things that we want to do or every company wants to do are really kind of determined by how good you are at technology. And so I think the role of the CIO has changed quite. a bit. And now, you know, I would define it as, in general, defining technology strategy of a firm and also making sure that you have the right culture in the engineering team in order to execute on that. What's a day-to-day look like? Like, what's a typical day? You get into the office and then what do you do? Well, I mean, I get into the office and I generally, like everybody else, you know,

I talk to people every day and all day. And so I talk to people, you know, we have a bunch of meetings, one after the other, and I have teams coming to me with either regularly scheduled meetings or meetings that have been requested to discuss a certain topic. And, you know, we just go through... Is there a whiteboard? Well, right now in the age of Zoom, I guess still, you know, we have a globally distributed team. And so a lot of our people are not in the same office.

And so we use virtual whiteboards like everybody else. But I would say, you know, one of the things that I tried to do while joining Goldman, which was part of sort of the cultural agenda, that was emphasizing the importance of narratives and written word versus in our PowerPoint and talking. So which is kind of what I learned at Amazon over the years, okay? Oh, right. You were at AWS. I was at AWS. And one of the things you learned there, as soon as you join Amazon in any part of Amazon,

like the first few meetings are kind of shocking because nobody talks. Everybody starts reading. You start reading for like sometimes 30 minutes or 45 minutes. And if you're the author of the document, you're just sitting there basically. and you're just trying to look at people's faces and understand what they think about your document. And sometimes, you know, if you're with Jeff Bezos or others, you know, at that time, it can be pretty, pretty terrifying. And so this kind of shift from a culture of people talk, people comment on a PowerPoint and the discussion sometimes get, you know, driven by who has the stronger personality versus, you know, who has the greatest ideas. One of the things that I try to change is that a lot of the meetings that we do today, actually,

start the same way by reading a document. So I now read a lot of documents like I used to in Amazon, you know, I would say maybe 30, 40% of the meeting are starting that way. And I think people love it because it breaks the barrier of language. For someone like me, the English is obviously not my first language. It breaks the sometimes some of the people are more shy than others, etc. So people see that as a mechanism for inclusion. So back to your question, let's say 30, 40% of my meetings actually now start by our own. reading a document together and then commenting on that and making decisions. Can I just say, Tracy, I've always thought more meetings.

You should start with just reading because you hear like a quarterly call or a Fed event and someone just reads out a prepared text. It's like just let everyone read it and just jump straight into it. Let everyone do the reading first. You don't need someone standing up there talking about what's on a written piece of paper somewhere. Anyway, I agree that we could reduce the time of meetings. Yes. Okay.

So speaking of meetings and the decision-making process. Then talk to us about how Goldman Sachs decided to approach generative AI. What was the decision-making process like there, the development process, and we'll get to what you're developing, but like, how did you initially approach it? So I think our initial approach was really to realize that there were so many more things that we didn't know compared to the things that we knew, because it's a really new thing. And even if for companies like us that have been working, working on machine learning and traditional AI for literally decades, this felt like a very

different thing. What sort of timeframe are we talking about? Like was there a sort of like big realization that this is something that we need to focus on? Yes, because I was lucky enough that I got into the very, very early version of GPT, even before it was called Chad GPT. So the very first version was essentially completely, completely. a sentence. It wasn't even allowing you to do interactive chat. You would just paste

a text and that will just complete that text. And so I started to do that with a bunch of stuff and then I was saying that the quality of which this would continue was pretty much indistinguishable with the part that you actually put in that. And so we started to obviously talk between ourselves but also among other people in the industry and we all realized very soon that this would be, a, something very different, but be also something that could have a pretty profound impact in what we do. Because at the end of the day, we are a purely digital business. We don't bend metal. We don't, you know, like use high temperatures. We don't really have physics. And so it's all about how we service our clients. It's all about how

smart we are. It's all about how we can process incredible amount of information. It's all about, you know, how we analyze data in a very sometimes opinionated way. We form our own views on the market. We form our views of investments, et cetera. And so, So, given that this AI showed very early sign of being able to synthesize and summarize very complex set of information, but also identify patterns, we thought that could be something that we definitely need to pay attention to. So given that, one of the things that we decided to do very early on was to put a structure, and I can say more about that, put a structure around this so that we could experiment, but

in a sort of a safe and controlled way. Right. So you decided to develop your own Goldman Sachs AI model versus, you know, use a chat GPT or a clot or getting something off the show. Actually, initially we kind of thought about that, but then very quickly we decided that our time was spent much better with using existing models, which by the way, we're iterating really, really quickly. But then put them in a condition so that they will be safe to use. And also they would like, actually give us the most reliable information because taken as they are, you can't just drop a model in an environment like Goldman and then like, you know, to your earlier point

of a 1% inaccuracy, even at 0.1% inaccuracy is completely unacceptable. Plus, there are a lot of potential issues related to, you know, what data has been used to train and, you know, there is a lot of uncertainty with regards to, you know, like what are the boundaries between what you can safely use and what you can't. And so what we decided to do was instead to build a platform around the model. So think of that almost as if you had a nuclear reactor. You know that now you have invented fission or fusion, and there is a lot of power that can be generated from that,

but then you need to contain it and direct it in a certain way. And so we built this GSI platform, which essentially takes a variety of models that we select, puts them in a condition of being. completely segregated and completely secluded and completely safe from an information security standpoint, abstracts some of the ways to use the model so that our developers can use the models interchangeably, and then creates a set of standardized way to, for example, improve the accuracy using retrieval-agranted generation, access external

or internal data sources, applying entitlement so that someone that is on the private side, you know, it's going to see different information that someone is on the public side. And then on top of that, build a developer environment so that people will very easily be able to embed that AI in their own applications. And so imagine this. We got a great engine and we decided to build a great car around that. A lot of short daily news podcasts focus on just one story. But right now, you probably need more. On Up First from NPR, we bring you three of the world's top headlines every day in under 15 minutes because no one's story can capture all that's happening in this big, crazy world of ours on any given morning. Listen now to the Up First podcast

from NPR. What are you putting in the model? Because I have to imagine at a bank like Goldman, you know, you have a lot of data, but you must have just an extraordinary amount of unstructured data. There's conversations that bankers have with clients. There's other sort of meetings, the meetings you have, and there's words that are said during that meeting that could be synthesized in some way. In these early iterations, I upload a conference called transcript and I ask a question. What are you uploaded? What is the unstructured data that you have or the questions or the, yeah, what are you putting into it from your reams of knowledge that you must have internally? So one of the first things that we did was use the platform and the models to extract information,

from publicly available documents. That's kind of the safest way. Public filing, all the case and all the cues and obviously earnings and put our bankers in a condition to be able to ask very, very sophisticated, multidimensional questions around what was reported, cross-ref it with previous reports, cross-ref it with any announcement, any earnings call, transcripts,

all things that are out there, but just are difficult to bring together. And so that has evolved into a tool that basically we use and we're rolling it out right now as an assistant to our bankers so that they can service their client or answer client questions or even their own questions in a time there is a fraction of what he used to take. Even generate documents that then can be shared to clients and so on and so forth. And obviously we always have as a rule like when you drive a car that has some autonomous, capability that you always keep their hands on the wheel. Our rule is that there always needs to be a human in the loop, okay? And so the way that works is actually interesting, because we found out

that you can't just shove something into a model and then pretend that the model is going to give you the answer right away. Why? Well, because models by themselves, you know, they essentially apply a stochastic or a statistical way to understand what is the next word that they need to say. And so no matter how good is the material that you put in, there's always going to be some level of variability. There is almost like the intersection between the documents that you insert and what is, I call it like the shadow of all the knowledge, of all the things that the model has seen before. And so we really perfected this, you know, there are two techniques that are widely used to improve

the accuracy of the answers. One is working on the way those models represent. knowledge, which is called embeddings technically. And the concept of embeddings, by the way, everybody talks about embeddings, but then very few people actually, it took me a while to understand that well. And embedding is simply a way for the model to parameterize and create a description of what they're seeing. So if I see a phone, for example, in front of me, the embeddings of a phone could be, it's

a piece of electronic. Yes, one, it's definitely a piece of electronics. It's edible, zero. You can't really eat it. And then you have all these parameters is almost like 20 questions. I give you all these questions and you finally understand that is a phone. And that's what the embeddings is almost like the 20 questions of the reality inside of 20 is like 2000, 20,000.

And then you have the rag, which is the retrieval augmented generation, which is actually interesting because you tell the model that instead of using its own internal knowledge in order to give you an answer, which sometimes, as I said, is like a representation of reality but is often not accurate, you point them to the right sections of the document that actually is more likely to answer your question, okay? And that's the key. It needs to point to the right sections and then you get the citations back. So that took a lot of effort, but we're using that in many, many cases because then we expanded the use case from purely like banker assistant in a way to more like, okay, document management. You know, we process millions of documents.

Think of that credit documents, loan documents. confirmation. Every document has a task called entity extraction. So you need to extract stuff from the document and then digitize it and then model it in a certain way. And so the use of generative AI there does a great job at extracting information. And this is an interesting concept because you don't have to actually tell a fixed pattern. You can just say, give a lot of examples and then the AI will figure out from that pattern. One of my favorite example is the following. Let's say that my phone number is 555-321-3050. And someone writes in the document instead of a zero, writes an O. Okay? You can test yourself even with GPT. If you give a number

with an O instead of a zero and you ask GPT, what's likely wrong with this entity? GPD is going to tell you, Well, it looks like a phone number. That is an O, which generally is not in phone numbers, most likely this is the correct phone number. Now, nobody has written software to do a pattern match in there. And imagine if in traditional way of doing entity extraction, there were developers, they were writing rules. They were saying, okay, numbers, it needs to be 10 digits and blah, blah, blah. The AI figures out their own rules. Yeah.

There are the most likely. So this is the key thing. It has common sense. And that common sense, when you're dealing with millions of documents that contain all bunch of ways that you must might have written those things. And imagine the complexity of all the rules that you need to write that every bank has the same problem. This simplifies things tremendously because it's able to figure out what's most likely by itself.

And so that thing evolved into a tremendous. time saving for everybody in the bank that has to do with the workflow documents. And so that was a very interesting finding that we did early on. And so again, to summarize, models are raw material of intelligence. You need to somehow direct them. You need to guide them. You need to instruct them. You need to put them in an environment that actually gets the most out of that.

And that's what we've been focusing on. So going back to the analogy that you used previously, this idea of a nuclear reactor and sort of building the containment casing or the protective casing around it. I imagine one of the complications of being Goldman Sachs and working with AI is that you're a regulated financial entity. How does that added complexity affect your use of AI? Are there additional data considerations or additional infoset considerations? I think that's a great question because obviously we live in a regulated world. And in fact, I have to tell you that in this case, regulation actually helps us think through all the possible unknown,

of something that, as I said, is something that is still largely something that nobody really completely understands. So what we did was to put governance around the usage of the models and also govern us with regards to the use cases that we can implement on the models. Every bank has a function called model risk, which in the traditional sense, a model is any decision or any algorithm that is running, automatically to do, for example, pricing, or, you know, there is a lot of that tradition in every bank, risk calculation, et cetera. So that's the traditional model risk. We use that very well-established

pattern that has its own second and third line, like controls and supervision, also to validate what we do on the AI side. So there is a governance part, which we really set up very early on. We have an AI committee that looks at the business case, should we do this? And then we have have an AI control and risk committee that looks at, okay, how are we going to do that? And then the two of them need to actually come together before we can release a use case. And then of course we did a lot of work with regards to the, let's say, accuracy, lineage, and in a way the way you connect the output to where does the data come from? And who can actually see that, what we call entitlements?

And we did that in lockstep with the regulators. So that I think, you know, in a world. I think we put a sort of what we like to call responsible AI first since the very beginning. And it really helped us the fact that, you know, we embedded all those controls into a single platform. That is how our people use AI inside the inside GOM. This is something I'm really interested in just from a technical perspective. But can you talk a little bit more about that interoperability aspect? So you have a pool of data that is Goldman's that you presumably don't really want to share with

outside entities. So how do you plug that into? an AI model if you're working with chat GPT or Cloud or something like that? Yeah. So there are two ways that we do that. We use the sort of a large proprietary models in a way that we worked with Microsoft, we work with Google, we had very strong partnerships so that essentially there are controls that guarantee that nobody has access to the data that we put into the model, that the data leaves no side

effects, so it's not saved anywhere. It only stays in memory. The model is completely stateless, meaning that the state of the model doesn't change after the data comes through. So there is no training. There is nothing done on that data. And also that operator access, meaning who can actually access the memory of those machines,

is restricted and controlled and needs to be agreed with us. So imagine securing, putting a vault around those models. But even then, what's really, really sort of secret sauce, proprietary, etc., we like to use also different approaches. to use open source models that we can run on our own environments. Okay? And we like a lot of open source models. I have to say the one we particularly like is Lama and actually Lama 3 and Lama 3.1, especially... That's the one developed by Facebook.

That's right. Oh, META, yeah. So they recently announced Lama 3.1, which has a version that is 405 million parameters, so it's pretty large. And it seems to be performing, you know, the gap with those big foundational models is now very, very narrow. So for that, we run it in our own sort of a private cloud, call it that way, with GPUs that we own, and we train it with data that stays in that environment. So imagine that, you know, our approach is, okay, there is a sort of a rating of sensitivity of this data. Every data needs to be protected. Therefore, we use those safeties all

throughout regardless, but then for the super, super, super, super secret stuff, you know, we like to do it in our own environment. Since you're talking about building your own environment, and this is something we've talked a lot about on the podcast, hardware constraints, energy constraints, things like that. How does that manifest in your world some of these physical, real-world constraints to building out the compute platform at Goldman Sachs? Well, initially we thought maybe we can host those GPUs in our own data centers, and then immediately you round into consideration such as, A, first of all, the development

develop a lot of heat. Secondly, they consume a lot of power. Three, there is a decent chance that they might fail because of all those considerations if they're not properly addressed. And then they need very special, for example, interconnects and high-speed bandwidth between them. So the decision what we ended up doing is actually to have them hosted into some of the hyperscalers that we use, but use them in their own virtual private cloud. So those racks are basically only ours. And if you're asking me the more general question, which is, hey, where is the world going with regards of that?

Okay. So right now there are two really rapidly competing forces. One is pushing towards more and more consumption, and one is pushing for more and more optimization, okay? And I can talk about that for a couple of minutes. For the more consumption, I mean, really the two dimensions for scaling a model is one of the most important one is obviously the size of the size of the number. of the prompt or the context, okay?

And there is pretty good evidence that the larger the context, which is really like the memory of those models, and the more you can get out in terms of the ability to reason on your data. That has already gone up from thousands to tens of thousands to now millions, and there is a prediction, you know, you heard that some very prominent people saying that there could be the trillion prompt. And the power scales quadratically with the prompt. And so that points to a consumption of energy and GPU power, which

which is going to continue to raise exponentially. At the same time, we've seen great results with optimization techniques such as quantization, reducing from 16-bit to 8-bit to 4-bit precision, having even smaller models using what's called windowed attention, which means that you know that you can only pay more attention to some of the parts of the context instead of all of it. So maybe you need a smaller one. And so I'm seeing those two kind of going into two opposite direction. It's going to be very interesting to see how that evolves.

I would say for the short term, I see that definitely that trend is going to continue to go up. And one of the things that fascinates me the most is that from one version to another, the most striking difference is the ability to reason and the ability to actually come up with logical step-by-step instructions or step-by-step chains of thought of what the output is going to be. So we decided, okay, first of all, we need to get access to the most powerful GPUs. Secondly, we need to host them into an environment that actually allows for the most optimal functioning in terms of bandwidth, in terms of power consumption, etc. And then at the same time, we've been focusing a lot on optimizing the algorithms so that, you know, we can really go, we could really get the most out of that. Just to press you on this point, what are the conversations actually like with cloud providers at the moment when you're trying to get more compute or more. space, more racks, whatever.

Is it maybe different for you because you were at AWS? Maybe you can just call someone up there and be like, we would like some more space, some more servers, or have you found yourselves at times maybe limited in what you can do by the amount of power available to you? Well, I wish that would be the case, but I cannot just pick up the phone and get whatever I want. But I think so far, I mean, obviously because we are a really good client of those companies in general, but also because we've been very selective in the use cases that we put in production.

I have to say, like I said before, think about that. If you look at the consumption of resources today, those who consume more resources are people that actually do the training of their own models, okay? And initially, everybody was trying to do full training from scratch, which was taking like the absolute, if that's 100, we do fine tuning, which is adaptation of existing models. That could be one to 100 or less in terms of consumption or resources. So because of the techniques they were using and because of the fact that we decided to really focus on fine-tuning or rag versus full training, we haven't really hit any caps.

And also, I have to be honest, you know, we bought our GPUs pretty well early. So probably there wasn't as much craziness as there is today. So that's turned out probably to be a good idea. The news doesn't stop on the weekends. Context changes constantly. And now Bloomberg is a problem. place to stay on top of it all. Hi, I'm David Gura. Join us every Saturday and Sunday for the new

Bloomberg this weekend. I'm Christina Ruffini. We'll bring you the latest headlines, in-depth analysis, and big interviews, all the stories that hit home on your days off. And I'm Lisa Mateo. Watch and listen to Bloomberg this weekend for thoughtful, enlightening conversations about business, lifestyle, people, and culture. On Saturday mornings, we put the past week's events into context, examining what happened in the markets and the world. That on Sundays, we speak with journalists, columnists and key political figures to prepare you for the week ahead. Join us as soon as you wake up and bring us with you wherever your weekend plans take you. Watch us on Bloomberg Television.

Listen on Bloomberg Radio, stream the show live on the Bloomberg business app, or listen to the podcast. That's Bloomberg this weekend. Saturdays and Sundays starting at 7 a.m. Eastern. Make us part of your weekend routine on Bloomberg television, radio, and wherever you get your podcasts. You know, in video is huge. Everyone would like to have some of NVIDIA's market cap be their market cap. I have offering some cheaper product. We interviewed some guys who have a semiconductor startup that's just going to be LLM-focused

startups. We know that Google, for example, has TPUs, their own chips. Can you envision as a roadmap some alternative where GPUs are not the dominant hardware for AI? Well, that's literally like the trillion-dollar question. Yeah, well, that's what I'm asking you. Yeah, but I'm not an analyst. and I'm just a technology. Remember, I'm the guy that makes sure that the printage go run out of the thing.

I would say you're probably a better person to ask than an analyst because you're actually the one who's going to be making buying decisions. Okay, so then I'm going to answer to you. So you have to distinguish between, there are actually two dimensions that we need to consider. One is training and the other one is inference. Okay. That's the first dichotomy. For training at the moment, there's most likely nothing better than GPUs, okay?

Because when you train a model, the software, or the software, or, you know, PyTorch or whatever framework, needs to see all your GPUs as one, as a cluster. And it's not just the GPU itself, but what Nvidia has been doing a great job at is actually to make them work in unison with the virtualization software called CUDA, which runs on NVIDIA GPUs, which is a pretty extraordinary piece of software, and it became the standard for that. And also because, you know, the performance premium that you have on those GPUs, when you're trying to train those incredibly large models is something that you really, really want.

And so the training part, I'm pretty sure that it's going to be dominated by GPUs for a while. But then as those models get used, obviously the pendulum swings towards inference, which is the actual, now you have a model which is a bunch of weights and you just need to calculate a bunch of metrics multiplications. So on that, I think accelerators and specialized chips are actually going to have a really big role to play. So you may imagine that you go from a world where everybody builds the cars and not too many people drive the cars, and to a world where most people are going to drive cars.

And then there is another two dimensions, which is models that are hosted by the client and models that are hosted by a hyperscale. So as you know today, I can take a model like Lama, I can put it in my own environment. You can run it on a MacBook. I can run it on a MacBook or I can run it in my own data center and with my own GPUs. Given that I'm used to GPUs, given that those are the ones that we can buy, given that Kuda is what developers know, et cetera, I'm most likely going to use that. That's a good part for Nvidia for that.

But then there is another way to use those models, which is to have someone host them for me, and I just access them to an API. That's what services like Amazon Bedrock does. You basically choose your own model and then you serve it through them. When you do that, you don't really know what's underneath. You don't know if it's a GPU or if it is a GPU or if it is a product. an accelerator, if it is Amazon's own chips or Google's own chips, et cetera. So now the real question, that's why the trillion dollar question is, are most people going

to use those models through hosted environments where the hyperscalor will have a lot of freedom with regards to what they use underneath and most likely they will vertically integrate? Or are they going to use them, you know, themselves? More like, you know, in a self-service way. And in that case, it's less likely that those accelerators are going to. to dominate. We currently are a sort of a balanced way because we have our own that we use like I described and also we use the hosted models. And so where is this going to go? It's hard to say because I think it depends on the evolution of the models and it depends which

models are going to be made available as an open source that you can actually host yourself. And I think right now one of the greatest questions is are the open source models are going to be an absolutely on-par alternative to the hosted model, to the foundational proprietary models. And that, given Lama 3.1, that answer seems to be more likely a yes. I had a question about this, actually, which is, do you think Wall Street's attitudes towards open source have changed over time? And the reason I ask is because nowadays, it seems like a fact of life. Everyone uses open source, whether you're a Goldman or somewhere else. But I remember, you know, like back in as recently as like 2012, I remember Deutsche Bank had

like this open source project called the Lodestone Foundation where they were like, oh, we should all stop wasting our own resources, developing our own code and our own software. We should all pool our resources together and do open source. And they had to actually lobby. It was unsuccessful ultimately. But they were trying to get all the banks on Wall Street to work together for open source. Nowadays it seems like there's been this significant cultural shift. It's not even a question. So in general, my direction, my guidance to, you know, my team is don't build anything unless you have to. Don't think that just because you're a smart person, you can build software

better than anybody else. Maybe you can, but it's a good thing that we focus on building things that are actually differentiating for us. And then I think the use of open source software, which we very much endorse, is also a really good hedge with regards to which vendors to use because it really heavily reduces the vendor lock-in. Of course, open-source software, as you know, is a tremendous long tail. There's millions of that. And so I think there are best practices around the use of open source. And those best practices are, you know, like you need to run reviews on open source or

tech risk reviews or security reviews or anything as if almost you built it yourself. And then secondly, tending to concentrate on the larger, very well supported by the community type of open source. And so my philosophy is yes to open source, but then you need to own it in a truest way, because you are actually going to be generally the one that actually needs to support that. And so really building knowledge around that. And now you can ask AI to run the code for you and check it for box. Yeah.

So, okay. That, of course, leads to probably what if you ask everybody, Where did you get so far the biggest bank for the buck for AI? Most CIOs are going to tell you on developer productivity. And I think it's something that for us was the first project that we actually expanded at scale. I have to say that today, virtually every developer in Gomez-Sax is equipped with generative AI coding tools. And, you know, we have 12,000 of them. So we didn't enable yet the ones that are using our own proprietary language called slang,

but everybody else has an AI tool. And the results have been pretty extraordinary. How do you measure that? What are some numbers? Or how would you describe the results? So we measure it according to a number of metrics, such as the time that it takes from, let's say, when you start the sprint to when you actually commit the code

or when you complete your task, we measure it by number of commits, meaning how many times you actually put code into production. We measure it by a number of defects, which in this case is like, for example, deployment-related errors. So there are more like velocity and quality metrics at the same time. We have seen a wide range ranging from 10 to 40% productivity increase. I would say that today we are probably on average seeing 20%.

Now, developers don't spend 100% of their time coding. They maybe spend 50% of their time coding. So your question is what are they doing with half of their time? Where there is a lot of other activities such as documenting code, such as doing deployment, doing deployment scripts, doing a bunch of tests, et cetera, et cetera. So what's called generally the software development lifecycle. Okay? And so we see net of 10%, but then the cool thing is that those AIs and the things they were

building around that are starting to go beyond coding. They're starting to help you write the right tests, write the right documentation. They are even figuring out algorithms or even, for example, reducing or minimizing the likelihood of deployment issues, right in deployment script. for you. So as that expands, we're going to be closer to 100%, and therefore we're going to be closer probably to 20%, which, you know, for an organization of our side, is a pretty massive efficiency play. Can I ask a question about hiring developers? So I've probably read 100 articles over the years about Wall Street competing with tech companies to hire developers.

Like, oh, they got to have ping pong table. Lloyd Blankine used to say they're a technology company. Yeah, you got to have your ping pong tables and your free lunches and let people wear sneakers and I have all that stuff. But now it seems with AI, there was a number of, of people interested in it who are truly believing that within a few years, they might build the digital god that's 10,000 times smarter than any human, and that they approach the task with messianic fervor. And I imagine it, right, if you're at Goldman and you're trying to help a banker answer a question to a client about something in the chemical industry, like, maybe that's not like the thing that gets you out of bed, the way sort of like metaphysical realms

about what is the nature of consciousness and things like that, that AI people talk. Does that present any challenges or anything when trying to hire talented AI developers? I think developers love to solve real problems. And one of the things also that attracted me in the first place, not that it matters, but I'm saying, you know, I tell you my own personal experience is that working in a technology company is absolutely fantastic, but you're always like one step removed from the business or from the application. So I have to, you know, let's say you are the bank

and I'm the technology company. I need to sell you a tool that then you're going to use to run your business or improve your business. We are kind of one degree of separation less, i.e. we're right there in a digital business that is fast, huge amounts of data,

huge amounts of flows, immediate results, and that's kind of addictive. And so developers, especially when AIs are starting to do all those magical things that we're talking about, you know, they can see the impact on the business right away. And that I think is kind of attracting a lot of people. In fact, there is more and more people that are moving into the industries, oil and gas, transportation,

chemical, medical, finance. Because this is new and there's nothing more exciting than seeing it in action. So there is so much action going on that I think is actually really, really interesting. I think another question that maybe you haven't asked me, but it's kind of part of this question is what kind of developers, how is the professional? of being a developer has actually changed. Oh, wait, I had a related question. It's not quite that question, but you can certainly answer that too.

But, okay, to my knowledge, Goldman Sachs doesn't have a job title specifically with the words prompt engineer in it. So looking at the impact of AI on your business overall, is AI a net hiring positive or a net hiring negative for Goldman's employees overall? Well, meaning are we going to hire more or less developers? Yeah, does it lead to more jobs because you're doing more things and productivity increases or does it lead to fewer jobs because now you can automate a bunch of stuff? Well, listen, there is so many things that we would like to do if we had more resources

that I think this is going to be leading to more things that we can do. You know, some people tell me sometimes, so you're going to maybe hire less or have less developers? I don't know. I've been in IT, quote, unquote, for like literally almost 40 years, and I've never, ever seen that go down. But I've seen inflection points where you can actually get developers to do way more and worry about way less. There is not related to a business outcome. And so I think it's more like how the profession is going to change. In my opinion, we're going to be less low level and more, hey, I need to really understand the business problem. A, I really need to think outcome-driven.

A, I need to have a crisp mental model, and I need to be able to describe it in words. So the profession is going to change. And there are tasks that I think are so repetitive that the automation of those is actually going to help developers really kind of feeling really, really connected with the business and with the strategy, and that will attract people that are generally curious, that are generally interested in understanding what we actually do. So the focus kind of shifts from the how to the what and to the why, which is really kind of at the heart, I think of this evolution of technology over the years, from the back office

of IT, which doesn't even know what you're doing, but as long as your monitor is actually working, to, hey, I'm actually able to take a business problem and break it down into pieces that then even an AI can write code for. So to your specific question, I think this might maybe, potentially for some companies are going to try to realize some of those efficiencies by curbing the growth or even sometimes reducing it. For companies like us, they are extremely competitive, for companies that have lots of ambition, this is a race at the end of the day. And I think we're going to go for, you know, trying to get even more out of our developers and actually like, you know, trying to turn them more into something

that makes them feel super, super connected to the business. What about non-developer roles, non-tech roles? And, you know, again, I guess a company like Goldman doesn't have, you know, probably a lot of like low-level customer support things or in a window is like, oh, I need to change my plane ticket, et cetera. But, you know, a lot of modern work is essentially just answering somebody's basic question. Are there roles within a bank that are going to either fundamentally change or go away due to sort of agentic or generative AI? I think a lot of the work that is about content production or content summarization, will actually be streamlined quite a bit.

Like, for example, taking an earnings report and making it into 10 different sources in order for different channels of distribution. Here's the one for internal people. Here's the one for the client. Here's the one for the website, etc., etc. Imagine the creation of pitchbooks for clients where you take templates, you put a bunch of data,

you go out and do research, you take logos, you take this, you take that. There is a lot of that machinery and factory, which, you know, we have thousands of people doing that. I'm sure there's a lot of junior analysts who would be maybe glad to hear that some of making pitch books is going to be automated away. No, but I think that's a good thing. It takes away some of the toil. So I think at the end of the day, listen, right now, have you noticed that everything is kind of converging to words and concepts?

No matter if you're a developer, if you're a knowledge worker, those jobs are kind of colliding. And I'm absolutely, developers have seen that first. Why? Well, because it's a low-hanging fruit. The developers deal with the vocabulary that is not 50,000 words. It's like 200, 300 keywords for a language. So, of course, that works extremely well, and of course, that's the first thing to go. But I think eventually the knowledge worker is going to be the one that is really benefiting.

No matter if you are a developer or if you are working on a pitchbook or if you're working on a summarization of a meeting or the action items or you're working on a strategy, etc., etc. And I think overall, this will elevate the quality of the work, which then everybody says a happy worker or a happy developer is a productive developer. I think you're happy when you're actually doing something that allows you to do your best work. And I'm hoping that if AI allows all of us to do more of our best work, I think is going to be probably the biggest effect that we can have. I know we just have a couple more minutes.

So one very quick question, what makes a good prompt? Well, believe it or not, empathy. You need to be empathic and you need to be gentle and you need to be kind and you need to kind of, you know. Just looking at me. Like, I'm not empathetic in my prompts. I always say please and thank you. No, Tracy makes fun of me for how empathetic. No, I've said it's very sweet that you say please and thank you.

You need to take the AI literally by the hand and take it where you want to go. And I tell you that, you know, one of my interesting, more interesting experience with prompts is the following. You know how hard it is to get an AI to say, I don't know? Oh yeah. It's almost impossible. You're always going to get an answer. And so one time I decided I want to get it to the point. And so I had to navigate the prompt and the AI to understand that it was safe and okay

to say I don't know. And so then at the end I prompted it. I said, what's the capital of the United States or Washington, DC? Okay. And then I said, you know, what's the weather going to be like tomorrow? I got an answer and then I said, what's the weather going to be in a year? And it's simply, I don't know. And then at one point, you know, I even decided what to say. It's like, is there a role for humans in a world of AI? I don't want to know the answer. Dot, dot, dot.

Okay, well, everyone's going to be off on chat GPT now trying to get it to say, I don't know. Marco Argenti from Goldman Sachs. Thank you so much. That was good fun. Thank you, Jo. Thank you so much. Thank you so much. Joe, that was a lot of fun. And I have to say, I do not make fun of you for saying please and thank you to chat GPT. I'm going to repeat it.

I've said it's endearing. It's very sweet. And I've tried to follow your example. And I don't say thank you because I usually move on to the next question, but I do say please. I've heard this, though. It's funny that he said that because I actually have heard this, that there does seem to be quantitative evidence that words like please and thank you, et cetera, do actually improve the quality. Yeah.

Mad Busegan, who we've known on Twitter forever, has posted about this. So there's a good reason to do it besides just the habit, the all entities you talk to, you should be in the habit of polite. Oh, yeah, that was your argument, right? Yeah, yeah. Yeah.

Okay, well, I thought that was fascinating. Yeah. We've been talking a lot about AI and the sort of potential use cases and the chips that are driving the technology and things like that. But it was nice to hear from someone who's actually making the purchasing decisions and implementing them at a large institution.

Absolutely. That was probably one of my favorite AI conversations we had for precisely that reason, because it was interesting hearing him talk about this idea that right now, like these open source models, particularly like the latest version of Lama, is getting really close to sort of the core proprietary models. That was striking the fact that he sees perhaps particularly on the inference side of model usage, an opportunity for greater use of different types of hardware, also very interesting. That's right. And we're so used to talking about the massive amounts of power and energy that AI will consume. And we, you and I have had a lot of conversations about how we're going to power all these servers and things. But what's gotten far less attention is just optimizing the

way you use AI such that you don't need to consume as much power. So maybe doing less training, leaving training to the big, like, hyper-scalers or whatever, and then just doing the inference. In the end, it's going to be both, right? Because in the end, like, there's both that's going to happen. People are going to find algorithmic techniques, and Marco described some of them to lessen the sort of pressure and stress that you're putting on your hardware. But of course, that's just going to mean you're going to use it more. And then also people are going to have to solve the power consumption side, kind of like all of economic history in general, in which we're always finding new ways to get more out of the same, you know, gigajoule of energy, but also using more energy at the same time.

Yeah, absolutely. Well, shall we leave it there? Let's leave it there. This has been another episode of the All Thoughts podcast. I'm Tracy Alloway. You can follow me at Tracy Allo. And I'm Jill Wisenthall. You can follow me at the stalwart.

Follow our producers, Carmen Rodriguez at Carmen, Armin. Dashob Bennett at Dashbot and Kel Brooks at Kel Brooks. Thank you to our producer, Miser. And for more Oddlots content, go to Bloomberg.com slash Oddlots, where we have transcripts, a blog, and a newsletter. And you can chat about all of these topics in our Discord where we even have an AI channel. Great stuff in there, discord.g.g. slash oddlots. And if you enjoy Oddlots, if you like our continuing series of AI conversations, then please leave us a positive review on your favorite podcast platform. And remember, if you are a Bloomberg subscriber, you can listen to

all of our episodes absolutely add free. All you need to do is connect your Bloomberg account with Apple Podcasts. In order to do that, just find the Bloomberg channel on Apple Podcasts and follow the instructions there. Thanks for listening. You can get the news whenever you want with Bloomberg News Now. I'm Amy Morris. And I'm Karen Moscow here to tell you about our new on-demand news report delivered right to your podcast feed. Bloomberg News Now is a short five-minute audio report on the day's top stories. Episodes are published throughout the day with the latest information and data to keep you informed.

Yes, there are other products like this from a variety of news organizations, but they usually rerun their radio newscasts throughout the day. That's not what we do. We create customized episodes that can only be heard on Bloomberg News Now. And we don't wait an hour to publish breaking news. When news breaks, we'll have an episode up in your podcast feed within minutes, so you're always getting the latest stories and developments. Get the reporting and the context from Bloomberg's 3,000 journalists and analysts

we're all over the world. Listen to the latest from Bloomberg News Now on Apple, Spotify, or anywhere you listen.