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

Thank you for having me.

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, 'But what do you do at Goldman?' And I said, you know, I just try to simplify. I say, 'Make sure that the printers don't run out of paper.' So, 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 are the people who like see if the Ethernet cable... those are the ones who tell you to... that you know, that... 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 CIO job, before I was doing more like, you know, creating technology, etc., 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.

Well, I mean, I get into the office and I generally, like everybody else, you know, I talk to people every day, 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 a part of sort of the cultural agenda, was emphasizing the importance of narratives and written word versus, you know, PowerPoint and talking. Okay. So, which is kind of what I learned at Amazon over the years.

I was at AWS. And one of the things you learn 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 gets, you know, driven by who has the stronger personality versus, you know, who has the greatest ideas, one of the things that I tried 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 meetings are starting that way. And I think people love it because it breaks the barrier of language. For someone like me, that 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 us reading a document together and then commenting on that and making decisions.

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 on machine learning and traditional AI for literally decades, this felt like a very different thing.

Yes, because I was lucky enough that I got into the very, very early version of GPT even before it was called ChatGPT, right? So the very first version was essentially completing a sentence. It wasn't even allowing you to do interactive chat. You would just paste a text and that would just complete that text. And so I started to do that with a bunch of stuff and then I was seeing 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 B, 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, etc. And so given that this AI showed very early signs of being able to synthesize and summarize very complex sets 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.

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 will 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, over 1% inaccuracy, at 0.1% inaccuracy is completely unacceptable. Plus there are a lot of potential issues related to, you know, what data has it 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 GS AI 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, abstract some of the, you know, ways to use the model so that our developers can use the models interchangeably, and then creates a set of standardized ways to, for example, improve the accuracy using retrieval augmented generation, access external or internal data sources, applying entitlements so that someone that is on the private side, you know, is going to see different information than someone that 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 the great car around it.

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, 10-Ks and all the Qs, and you know, obviously earnings, and put our bankers in a condition to be able to ask very, very sophisticated, multi-dimensional questions around what was reported, cross-referenced with previous reports, cross-referenced 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, you know, service their client or answer client questions or even their own questions in a time that is a fraction of what it used to take. Even generate, you know, documents that then can be, you know, shared with 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 the 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 in technical terms. 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. An 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 electronics, yes, one, it's definitely a piece of electronics; it's edible, zero, you can't really eat it, you know. And then you have all these parameters. It's almost like 20 questions, you know, I give you all these questions and you finally understand there is a phone. And that's what the embedding is, almost like the 20 questions of the reality, instead of 20, it's like 2,000, 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 section 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 the credit documents, confirmations. Every document has a task called 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 examples 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?', GPT is going to tell you, 'Well, it looks like a phone number that has 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 matching there, right? And imagine if in the traditional way of doing entity extraction, there were developers that 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 that 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 humans 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 of 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 know, 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.

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 unknowns of something that, as I said, is something that is still largely something that nobody really completely understands. And so what we did was to put governance around the usage of the models and also governance 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, etc. So that's the traditional model risk. We use that very well-established pattern that is also, you know, 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 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 entitlement. And we did that in lockstep with the regulators. So that, I think, you know, in a word, 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 our people use AI inside Goldman.

Yeah, so there are two ways that we do that. We use the sort of large proprietary models in a way that, you know, we worked with Microsoft, we worked 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 source, proprietary, etc., we like to use also a different approach, 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 Llama, and actually Llama 3 and Llama 3.1 especially.

Yeah. So they recently announced Llama 3.1 which has a version that is 405 billion 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 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 secret stuff, you know, we like to do it in our own environment.

Well, initially we thought maybe we can host those GPUs in our own data centers. And then, you know, immediately you run into considerations such as, hey, first of all, they develop a lot of heat. Yeah. Secondly, they consume a lot of power. Three, there is a decent chance that they might fail because, you know, of all those considerations if they're not properly addressed. And then they need very special, for example, interconnect and high-speed bandwidth between them. And 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 ones...

The size of the prompt or context is obviously important, and there's pretty good evidence that the larger the context, which is really like the memory of those models, 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's a prediction that there could be the trillion prompt. The power scales quadratically with the prompt, which points to a consumption of energy and GPU power that's going to continue to rise 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 window attention, which means you can only pay more attention to some parts of the context instead of all of it. So maybe you need a smaller one. I'm seeing those two kind of going in two opposite directions. It's going to be very interesting to see how that evolves. For the short term, I see that trend is definitely going to continue to go up. 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 chains of thought of what the output is going to be. So we decided, 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, power consumption, etc. And then at the same time, we've been focused a lot on optimizing the algorithm so that we can really get the most out of that.

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, 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. Initially, everybody was trying to do full training from scratch, which was taking like the absolute 100. We do fine-tuning, which is adaptation of existing models, that could be 1 to 100 or less in terms of consumption of resources. So because of the techniques that we're 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, we bought our GPUs pretty well early, so probably there wasn't as much craziness as there is today. And so that turned out probably to be a good idea.

Well, that's literally like the trillion-dollar question.

But I'm not an analyst, I'm just a technologist. Remember, I'm the guy that has to make sure that the...

Okay, so then I'm going to answer you. You have to distinguish between two dimensions that we need to consider: one is training and the other one is inference. That's the first dichotomy. For training, at the moment, there's most likely nothing better than GPUs. When you train a model, the software or PyTorch or whatever framework needs to see all your GPUs as one cluster. It's not just the GPU itself, but what Nvidia has been doing a great job at is actually making them work in unison with the virtualization software called CUDA, which runs on Nvidia GPUs. It's an extraordinary piece of software and it became the standard for that. Also because of 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. So the training part, I'm pretty sure, is 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 matrix multiplications 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, to a world where most people are going to drive cars. And then there are another two dimensions: models that are hosted by the client and models that are hosted by a hyperscaler. As you know, today I can take a model like Llama, I can put it in my own environment, I can run it on a MacBook, or I can run it in my own data center with my own GPUs. And given that I'm used to GPUs, given that those are the ones that we can buy, given that CUDA is what developers know, etc., I'm most likely going to use that. That's a good part for Nvidia. But then there is another way to use those models, which is to have someone host them for me and I just access them through 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 an accelerator, if it is Amazon's own chips or Google's own chips, etc. And 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 hyperscaler 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 themselves in more of a self-service way? And in that case, it's less likely that those accelerators are going to dominate. We currently are in 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 open source that you can actually host yourself. And I think right now one of the greatest questions is: are the open source models going to be an absolutely on-par alternative to the hosted, foundational proprietary models? And that, given Llama 3.1, that answer seems to be more likely a yes.

In general, my direction, my guidance to 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 them. And so I think there are best practices around the use of open source. And those best practices are: 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 the 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. But now you can ask AI to run the code for you and check it for bugs.

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 Goldman Sachs is equipped with AI coding tools, and 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.

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, 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 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, right? They maybe spend 50% of their time coding. So your question is, what are they doing with half of their time? Well, there is a lot of other activities such as documenting code, doing deployment scripts, doing a bunch of tests, etc., etc. So what's called generally the software development life cycle. And so we see a net of 10%. But then the cool thing is that those AIs and the things that we're building around that are starting to go beyond coding. They're starting to help you write the right test, write the right documentation. They are even figuring out algorithms or even, for example, reducing or minimizing the likelihood of deployment issues, writing deployment scripts 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 for an organization of our size is a pretty massive efficiency play.

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, 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, 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. And 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 this is kind of part of this question, is what kind of developers, how is the profession of being a developer actually changed?

Well, listen, there are 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. 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 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 that 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. I really need to think outcome-driven. I need to have a crisp mental model and I need to be able to describe it inward. 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, they're going to try to realize those efficiencies by curbing the growth or even sometimes reducing it. For companies like us that 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 trying to get even more out of our developers and actually trying to turn them more into something that makes them feel super, super connected to the business.

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 pitch books 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 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. But I think that's a good thing. It takes away some of the toil. And 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. Developers deal with the vocabulary. There is not 50,000 words, it's like 200-300 keywords for 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 pitch book 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 it's going to be probably the biggest effect that we can have.

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, take Joe's looking at me like I'm not empathic with my prompts. I always say please and thank you.

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 one of my more interesting experiences with prompts is the following: you know how hard it is to get an AI to say 'I don't know'? 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 and said, what's the capital of the United States? Washington DC. Okay. And then I said, what's the weather going to be like tomorrow? And I got an answer. And then I said, what's the weather going to be in a year? And it simply said, I don't know. And then at one point I even decided what to say is like, is there a role for humans in a world of... Yeah, I don't want to know the answer, dot dot dot.

That was good. Thank you, Joe. Thank you so much.