The Goldman Sachs Exec on AI, Wall Street & the Firm That Went Bankrupt in 40 Minutes

[00:00:00] Marty Chavez got his AI PhD back in 1991. How many AI jobs did he have to choose from when he graduated? Exactly zero. So he went to Goldman Sachs and spent decades building the machines that took over Wall Street. Now he's at 6th Street, sits on Alphabet's board, and helped launch isomorphic labs. Few people have watched more AI hype cycles come and go. Fewer still have profited from them. This is my conversation with Marty Chavez from the Silicon Valley AI Summit in London. I'm Eric Newcomer, author of the Newcomer

[00:00:28] Substack. Let's get into it. When you think about institutions like Goldman Sachs, they're certainly more cutting edge than many. So when you think of the big old guard, powerful industries, where do you think they are right now on this AI journey? Are they fatigued where it's been oversold? They're true believers? Or what is your sense of these mass enterprises and their relationship to this AI

[00:00:57] mania? Well, this will be a little bit particular to Goldman Sachs. So I'm old. So I like to think that this is the most exciting time ever to be alive and to be a computer scientist. And also, there's nothing new under the sun. I was crazy before. All those two thoughts, right, at the same time. And so at Goldman, we've been working on the frontier of AI

[00:01:23] for a long time. It's just that the names keep changing, right? So why did I end up at Goldman? Because I got a PhD in AI in 1991. Do you know how many AI jobs there were in 1991? Not a matter. Yeah. Exactly zero. So instead, you were like, I'm the smart guy. I guess I go to Goldman Sachs. I just got a random letter from a headhunter. And the letter said, I've been instructed to make

[00:01:50] a list of entrepreneurs in Silicon Valley with PhDs from Stanford in computer science, and you're on my list. And I had bills and student loans. And so that's how I ended up there. And at that time, Goldman, I wouldn't even say I got a PhD in AI or machine learning. Or I'd just say computer science. Why? Because it was embarrassing. Because AI couldn't do anything then. And so I ended up at

[00:02:17] Goldman and they had this crazy idea. Let's build a digital twin of the trading business. Let's build a piece of software that models everything that happens in the trading business so that we can lose a lot of money and then say, ah, it was only a simulation. Okay? Right. As opposed to losing it in real life. But then around 2011, we started to see something different happen. And it's kind of cute. We called them algos. Does that name?

[00:02:47] Algorithmic trading. Algorithmic trading. But it's the same thing exactly as agents. It's humans out of the loop making trading decisions. There were pieces of software that put orders to buy and sell stocks into the exchange. What could possibly go wrong? Right? Everything went wrong. There are many companies that no longer exist because they bankrupted themselves. There's a legendary example of a

[00:03:14] company that thought it was trading in a simulated test environment, except the trades were being routed to the actual change. And in 40 minutes, they were bankrupt. That was night trading. Wow. So this stuff has been around for a while. We just didn't call it agents. And so at Goldman Sachs and many other places, this is just the next iteration of something that's been around for a

[00:03:38] long time. And I think we already know how it's going to go. Right? So there's a lot of concern about jobs, for instance, job loss. Well, so here's what happened in the trading business of Goldman Sachs. Over 15 years of bots or algos or AI or whatever you want to call it. There are, would you guess if there are more people in the business? I assume way more people. Way more. The business is way bigger.

[00:04:05] Complexity gives people a lot of stuff to do. It makes way more money. It does very much more complicated trades. But if you were to list all the activities of the people 15 years ago and the activities of the people now, they're completely different. And did they, do you think the same people kept their jobs? I guess one of the issues with, and I want to get into sort of the public policy stuff more after we get through this period. But

[00:04:35] like, do you, but just given you brought that up, like, do you think the same people kept their jobs or new jobs were created in some people? So when, when, so there was one day in 2011 where I became co-head of the equities business, right? So it was a little Goldman Sachs drama. If you were a quant like me, you thought this was the best thing ever. If you were a trader who wasn't a quant, you're like, not the champion, not that guy, you know, not that guy, right? I want the guy who's like, you need gut.

[00:05:03] Well, I'm going to get to that. Right. And so, so I, I did what you do at Goldman. You have a town hall, you bring everybody together. And so I said, you, you all have three strategies. One, tell the computers what to do. It's been a great strategy for me. It's not for everybody. Two, collaborate with the computers and the people who tell the computers what to do. I love this strategy. I urge you all to adopt it. Everybody can adopt this one. Strategy three is

[00:05:32] idiotic in the name of what you think of as your job security, stand in the way of progress, stonewall the people who tell the computers what to do. If I catch you executing that strategy, I will accelerate the end of your career to right now. And afterwards, my boss said, Jesus, my, that was dark. I said that I thought it was, I was being helpful and people still to this day,

[00:06:02] say, I was listening and I went for strategy too. From this anecdote, like I, and just the experience of Goldman Sachs, one, one takeaway, if they've already been doing it is that they don't want some external company doing it for them. Right. It's like we've not generally, right? Yeah. So what is their opportunity, I guess, for all these startups that say, oh, we're going to do all this AI stuff for you, or are they just proving to incumbents that realize, wow, if this is our core competency,

[00:06:30] ultimately we're going to have to do it ourselves. So when I started at Goldman, 93, we had in our group a mantra that we, that we thought was really clever and cute, which was the only thing crazier than writing all your own software is not writing all your own software, but it was 1993. So we wrote our own object oriented, transactionally protected database. We wrote our own programming language. If you squint,

[00:06:55] it looked a lot like Python, right? But, but this was all happening in 93. And that was a strategy that served us really well. But when I became chief information officer, we changed it and we said, we're going to, we're going to do it a different way. First, we're going to see if there's open source. And if there is, we're going to use that. And then only if there isn't, are we going to go out and look at vendors with a preference for open source packaging by vendors. And then as a last resort,

[00:07:24] we're going to write our own software. And they're still doing that at Goldman. So you can sell software to a huge company like Goldman. The, the exciting part of it is it's probably going to be a big contract if, if it's going to be material for Goldman, but the sales cycles brutal. The last thing they want to do, it's like, all right, only if you can convince us it's better than any alternative. That that's, and we would say that up front that you had to do that. But this is

[00:07:50] just, this is very particular to something on the scale of Goldman and my firm. Now it's a private capital firm, Sixth Street. We spun out of TPG. We have $140 billion under management. We buy a lot of software. We're huge users of AWS. There is an AI company that we actually are using. It's called Abacus.ai. And you can think of it as a data wrangling plus smart order routing to a bunch of

[00:08:19] different LLMs and it's really on fire inside the company. And we would not do that on our own. On the other hand, we were big believers in this moment of the model surrounded by a harness, right? That's the, very much of the moment. And so we have built our own harness for the investment process. We think it's very specific to us. To switch between closed source models or to make open source models

[00:08:46] or models that you control work well? We use a bunch of models underneath it, but it's a harness for the end to end process from we first hear about a company, we find its virtual data room, all the way to what are the work products we've got to generate so that we can get to our investment committee. And then what kind of write-up actually gets investment committee to say, yes, let's do this.

[00:09:12] So that whole process, we're finding that with the right harness over a set of models, and we're using a bunch of different ones, we can greatly increase the aperture of deals that we look at and the quality of them with the same number of investment professionals. What is your view? And this is such a big question, but on the market sentiment to AI, I mean, we had this sort of Asia freak out, we, you know, SpaceX is trading down a little bit.

[00:09:42] I didn't get the latest. It's only been a week. I know it's only been a week. It's priced to perfection. I don't know, but what is, yeah, I mean, do you think, you know, one way to put it directly is like open AI, Anthropic will be received positively by the markets or what can you tell us of your view of how the markets are going to continue to read the AI moment? So there's obviously the excitement that I think all of us share and we can see what AI is good at.

[00:10:10] And then there's also old school finance professionals who look at these companies and think, hard to see the path to profitability. Even opening AI Anthropic? Yeah. I mean, maybe it's there and they're going to have an opportunity to talk about it and we'll

[00:10:34] learn a lot more, right? It's not completely obvious to me that it's there. And maybe it doesn't need to be there, right? Maybe there's a long time that investors will continue to fund that gap. And that's the thing that's unknowable, right? It's one axiom of finance is that investors will keep funding until the moment that they don't and nobody can predict it. It would be a fool's errand for me

[00:11:02] to attempt to predict when that would happen. Well, you know, right ahead of the SpaceX IPO. I think this was undercover. We talked about it, some newcomer, but Alphabet went out and raised a ton of money. You noticed that. Yeah. Berkshire. How much was that? We want the money versus we want to remind the investment community you could invest in these very speculative businesses or you could invest in our company, which has one of those very speculative businesses and the old school. And profits. And profits, yeah.

[00:11:32] And distribution. I don't know. What do you make? Why raise all that money? Well, it was something like, I don't know, it was like $85 billion. It ended up with the green shoe, it ended up around there. Well, it was $85 billion because it was possible to do it. And we're always looking at the right way to fund this incredible moment, right? You can see the CapEx of all the companies and Alphabet's is public.

[00:12:02] And the leadership has been saying a lot to the market about what that trajectory looks like. And so as a finance person, I will just look at that and say, well, there's equity and then there's credit and then there's everything in between. And so what makes sense? Which is a better deal, equity or debt? I mean, in some ways, is it wrong to think, the market right now is rewarding ambition in AI. So if we take on this money through equity,

[00:12:30] we are only going to get almost rewarded for leaning in, even though we're raising money. Is that too dumb of a way to think of it? No, that's it. There are many ways to think of it. Old school finance. It's just, there's that spectrum of ways that you can finance this ambitious and appropriate and exciting CapEx plan. And this is one way to do it.

[00:12:54] And why do you think, assuming you do, Alphabet is well positioned right now. I mean, we surveyed people on products that they thought might get disrupted. Google search was certainly not their top answer, but it was an answer. I sort of made the case for you guys that, well, to the extent Google search is getting displaced, it is getting substituted with Gemini, your own product. But what, I don't know, but there is clearly a threat of disruption to that profit generator that is so key to the Alphabet

[00:13:23] story. What gives you confidence in what Alphabet is doing today? So things change really fast, right? So I've been on the board for four years and there was a moment when people would say, oh, oh, Marty, I'm really like, you're on the board of Alphabet. That must be really tough. And I know, I remember when you guys were so beat down. And don't you know that search is dead and it's going to be completely, like,

[00:13:53] it's interesting and cute that you think that, right? That was not going to happen. That was, people said it, that doesn't make it true, right? So Google has been working on AI for a long time, as we know, and it's been in search and it just, there just keeps being more and more of it, right? I'm old enough to be able to say things like, AI is just more software.

[00:14:18] Or to quote my friend Astro Teller, it's linear algebra on steroids, right? And it's great and it's exciting. And of course we're going to use it in search and it can't be a huge surprise that the search box is morphing to do more stuff. And one thing that I've always said since the beginning is that, you know, chatbots, like fun, but a CLI from the 70s, is that how we're going to all interact with

[00:14:47] AI? It seemed unlikely to me, right? And then we moved beyond the chat interface. And then now, actually, maybe it's back to the future. I'm using a CLI all the time. It's called anti-gravity. Right. Right. So what, how much, I mean, I think there was this story of like, I mean, obviously, you know, the attention is all you need paper. We literally started today with In Gomez, uh, yeah, exactly. So you have cohere who was a coauthor on that paper,

[00:15:14] obviously not at alphabet anymore. Um, but I mean, the story was sort of like, oh, they, they were the geniuses, but then they weren't hardcore enough. And then they sort of like got hardcore, maybe they, you know, Sergey came in or whatever. Like how, how much do you think it just took time? Or how much do you think there was this sort of like awakening? I will give you my, my, my personal perspective is Sundar said 10 years ago, it's, we're going all

[00:15:43] in on AI. And there's a lot of people I know you don't think this, but for maybe people out there in the general public, AI began with chat GPT in 2022, but there's been a much longer history of it. Right. And, and if your mission is organizing the world's of in world's information and you see

[00:16:07] this incredible technology and you see that it has this particular human characteristic, which is, it makes things up. That seems at odds with organizing the world's information. Right. There were really things like, oh, that doesn't, that's not what we wanted. It's not, it, it was, it took a while to make it part of the mission, right? But I totally get,

[00:16:30] if you're a startup, you, you can have a different plan and why not, why not release it? That makes sense. That wouldn't have been my view, the right move for a company whose mission is to organize the world's information. You know, we're among futurists. I know you're, you're ground, you're grounded person, but I mean, I, we first met at a bio, uh, in better bio conference, you know,

[00:16:59] Alphabet's exploring, I think interested in, uh, data centers in space. I mean, everything from, yeah, disease research, there's spaces in the, so I guess of the many sort of like moonshot ideas out there. Which ones are you sort of most energized about, or really a believer that our world will be changed in the next, I don't know, five to 10 years? I have to start with Isomorphic Labs. Yeah. So that I was just, I wake up every day excited about doing disease research.

[00:17:28] Demis Sasabas co-founder. I think I have a particular, a particular angle on it, right? So I, I went to college. I hadn't done any due diligence. I thought I'd major in computer science. They didn't have a computer science major. So the science professors are recruiting and there's a biochemistry professor. And he says to me in 1981, like little 16 year old me, the future of the life sciences is

[00:17:55] computational, which was a crazy thing to say in 1981. He said, if you sign there as a biochem major, you can take computer science, you can take physics, you can take chemistry, you can do it all and join my lab. And we're doing something truly insane and wonderful, which is we're crystallizing proteins, one protein at a time. And we're taking all the atomic coordinates and we're putting it in

[00:18:21] a database. And I thought that sounded incredibly cool. That was the genesis of the protein data bank. So as a 16 year old, I'm working on crystallizing the capsid protein of the tomato bushy stunt virus. We never forget a name like that, but this is what led to alpha fold. And I remember saying to ourselves and hearing from the people who really knew what they were talking about, the professors,

[00:18:48] this is going to be a 50 to 100 year journey. Right. And these timelines are short. We got to start somewhere by putting all the coordinates in a database. Right. And then Demis and team came along and they, they use that. And then they did all kinds of other wonderful things for which they got the Nobel prize, of course. Right. Including one of the things that Demis talks about that I find so counterintuitive and exciting is at some moment,

[00:19:16] and I'm oversimplifying, they decided we've taken all the 50,000 proteins that have been laboriously crystallized by grad students. And that's all we got. And now we're going to guess the structures of 75,000 other proteins for which we do not know the structure. And then we're going to feed it back in. Now that could have sent the model out into the weeds, but instead something amazing happened. It leapt

[00:19:42] to this new level of understanding how proteins fold. And then after that, it could predict hundreds of millions. Right. So is your, like we just saw, you know, mid journey is doing these sort of sauna scans. Do you feel like your health today has been improved at all by the AI industry? Or you think a lot of these, I think it's on the come. Yeah, I think it's on the come. And one of the,

[00:20:06] and this is not going to be surprising to anybody. I think it's well accepted at this point that AI is really good at finding molecules given a target. And that is an incredibly important part of inventing new drugs, but it's just one part. There are many, many other parts and they're complicated and

[00:20:30] they're complicated because many of them are legal, social, and political judgments, right? Like a clinical trial. If a randomized control trial is the only acceptable evidence of a drug's efficacy, then those are expensive and those are slow. And so to my mind, there's many levels of simulation

[00:20:52] that we still have to figure out. And alpha fold is a linchpin of that, but maybe to really do it, we're going to need to be able to simulate not just proteins, but cells and tissues and organs and bodies, and then whole groups of people and the psychodynamics of whether you take the pill or not. And at the level of a society, is a society going to pay for this treatment, right? All of that has

[00:21:20] to go into the simulation. I think that's going to be a long journey. I don't, I don't see that happening in the next couple of years. We have, we might have this technology to have this amazing idea or development, but then so much needs to play out for it to actually be realized. Yes. I wanted to, we started with the experience in big organizations. And so I just wanted to ask you a sort of, you know, there's so many startup founders here who want to figure out how to

[00:21:45] sell to big institutions or partner with companies like Alphabet. I don't know what advice would you give like the small startup for, for going after those like big enterprise sales or striking a big partnership? Um, so I've been on both sides of this. I've been, I've been an entrepreneur as well. And, um, so,

[00:22:09] okay, here, here it is. This is something I learned. Well, it's something I learned the hard way. So in one of my startups, it was a dot com era startup. It was an early software as a service company. Give us your trades. We'll give you the risk and analytics, but we didn't even have the term SaaS at that time. This was, this was 2000 and we had this product. I heard someone say earlier, we had this product

[00:22:36] and it turns out that there were features in the product that nobody actually cared about. And we worked really hard to build it and we were getting some deals and it was kind of working, but it was painful. And we started the company two weeks before the dot com bubble burn. So timing was interesting.

[00:22:56] And eventually our investors gave us a head of sales. This guy's name in the market was Dr. Payne. So let's start there, Dr. Payne. And he came in and he said two things that, that froze my blood, basically, as a new head of sales. He said, if you, Marty, don't get a call from a very angry customer

[00:23:24] screaming at you because I, your head of sales have crossed the line in my sales tactic, then I'm not doing my job. And then the second thing he said to the sales team was customers buy a product when they have unbearable pain and you have convinced them that only your

[00:23:50] software product can put an end to their pain. Everything else is just getting lucky. You're in a hype cycle, right? And so it really comes down to that. You have to understand the people in the organization, not just the organization. You have to understand in a large organization, there is always someone who tells you I am the ultimate signing authority. And that is always

[00:24:16] false, right? There's always someone higher who has to be convinced. And if you're not talking to that person, you probably don't have a sale. And if they're not experiencing unbearable pain, then you're probably not going to have a sale. And unbearable pain, what I learned in that era, unbearable pain is having to restate your financial statements, going to jail because there were

[00:24:43] falsehoods in your financial statements, right? So if my product could mitigate that risk, then people were going to buy. Otherwise, getting lucky. Marty could talk all day. Hopefully people, never have I felt looking back has made me so optimistic about the future while sort of setting limits on how ambitious to be. Well, human nature is a constant, right? That is true. Thank you so much for joining us. All right. Such a pleasure, Eric. Thank you. Thank you.

[00:25:12] That's our show. This is the Newcomer Podcast. Thank you so much for listening. Please like, comment and subscribe. You can follow more on the Substack at Newcomer.co. We publish every talk from the Cerebral Valley AI Summit on our Newcomer AI Summit's channel. And you can find my conversation with Cerebral Valley co-hosts, Max Child and James Wilsterman at our Cerebral Valley show channel. Lots of stuff going on here at Newcomer. Thanks so much for your support.

[00:25:38] Please leave a comment, suggest guests. Thanks so much for your time. See you next week.