- 4 hours ago
Category
🗞
NewsTranscript
00:00So we've been talking about AI strategies and data strategies. The last panel talked a lot about data too. They
00:06got a lot into it. Neema is actually Goldman Sachs. So he's the chief data officer, also head of data
00:11engineering. So he's the right person to talk to you about this. You've been with Goldman Sachs for about 20
00:17years. You're a partner. So if you could just take us through for the audience here, take us through your
00:24journey, like how you started and how you got to where you are today and what your day includes.
00:30Like what you do day to day. I'll try to summarize 23 years in one and a half minutes. Let's
00:34see if I can do it. So I started at Goldman. I studied computer science. I'm a Cali kid. I
00:42studied computer science. I'm right out of college, joined Goldman in technology in the deep, deep back office, in the
00:50bowels of the back office.
00:55And from that, like that, that first experience of being like an analyst learning, like what an enterprise is, like
01:02what's going on. And then in 2008, what happened, this was five, five years. And I was still a kid
01:08in my twenties, five years, financial crisis hits.
01:13And everyone's like, oh my God, what's going on? Like, are any of these banks going to exist anymore? Are
01:17we going to exist anymore? And so I give a lot of credit to the leadership there.
01:23We actually took a data driven approach to the financial crisis. Um, we had a group called core stress. This
01:30is the front office quant teams who were basically put in charge of like, how are we going to figure
01:34out our exposure to all of these, um, bankrupt companies.
01:38And, and, and we got, we got very lucky where other banks were started going through the filing cabinets trying
01:46to find their ISDAs. We actually had all of our OTC derivative trades in one place. It's called SecDV. Um,
01:53and so these quants were coming around, around the whole firm, sort of scouring from the middle and back office,
02:00like what data we could bring together to sort of understand our exposure to Lehman or what, or, or other
02:05banks.
02:07And, and, and, uh, they actually tapped me from the deep back office. They're like, oh, you might know something
02:13about liquidity risk, uh, because at that point I had been sort of on this learning journey about what the
02:19front office is learning about banking, learning about. And so I was like, this curiosity sort of put me in
02:25a position where they, where they're like, okay, you seem like you could help us. And I was like, yes,
02:30I can help you.
02:30And so from that point on, we actually built this aggregated database to tell us all of our counterparty exposure,
02:37which again was sort of a novel idea in 2008, where everyone was sort of focused on market risk. Um,
02:44and funny story, that database actually won a, uh, a prestigious award at Goldman that was usually given to the
02:50bankers for the best M&A transaction.
02:52The database won that, that award that year, not, not even our team, just the database. Uh, and that sort
02:58of started, uh, uh, that sort of started our, uh, my personal sort of data journey at Goldman.
03:06And this was like when I, so it sort of clicked with me, wow, like financial services, the power is
03:11in the data, bringing it together, giving insights.
03:14And it turned out the really, really cool thing was not only that this database won this award for saving
03:22the firm, uh, not going bankrupt, but that, um, after, after that project, after we had weathered the storm, people
03:32started finding more business use cases to use that data.
03:35So the traders and the quants and the salespeople would come to us and be like, Oh, we didn't know
03:40that if we could join settlements data with this data, with this data, with this data, we could actually help,
03:46help our clients trade better.
03:48And so it turned out that like this thing that was meant to like, you know, as a risk thing
03:54ended up being a business idea generator and a client, uh, idea generator.
03:59So that was like, that was sort of, uh, a journey that myself, like I was on, but also the
04:05firm sort of realizing that, Oh my God, like data's not just this weird exhaust.
04:10Right.
04:11From like what we do, but actually could be a differentiator.
04:15Now you talk about data.
04:16Um, a lot of companies want to use AI to gain this edge.
04:20Um, so why is data strategy?
04:22I mean, so important.
04:23Why is that where people have to focus first before you get to?
04:27Yeah, that's a great question.
04:28Yeah.
04:28So everyone, I think everyone sits around and says like, Oh, data is like the fuel of AI and Oh
04:33my God, garbage in garbage out.
04:35But like, I don't, I'm not sure anyone actually explains like why, why is that?
04:39Why does it matter?
04:40Like, why?
04:40And here's my, here's my personal mental model.
04:43This is the, this is the way I've taught myself.
04:45So maybe, maybe a learning moment.
04:47Let's see if hopefully I'm not being too, uh, too basic here, but for, for the last 60 years or
04:5440, whatever of computer science, it's basically been.
04:57You hire someone like me and my team.
04:59We sit in front of a keyboard and we write rules into the computer.
05:02We like type rules or like, ah, if you click this button, send an email.
05:06Like if you push this button, it like starts a YouTube video.
05:09Like if you type these keys, like I'll, I'll work a word document will appear.
05:15And like, those are rules codified in the computer.
05:18Like we actually write that code that then does those things when people interact with the computer.
05:24Like that's what we call like rules based computing, right?
05:28The whole AI thing has sort of shifted that.
05:31And what, and, and here's my mental model of AI.
05:35My mental model of AI is instead of me typing rules into the computer, right?
05:41What AI is doing is it's actually learning by example.
05:45It's learning the rules by example.
05:48So like the first version of AI was like, you take a million pictures of cats in like different poses.
05:55And you feed all those million pictures to the computer and you're like, this is what a cat looks like.
06:00This is like what a cat looks like in a tree.
06:02This is what a cat looks like on a street.
06:04This is like what a cat looks like here.
06:06This is like the cat, the back of the cat looks like this.
06:08And so you have like a million pictures.
06:11And, and then what happens from there is that the computer sort of starts learning what a rule of what
06:17a cat might look like in other settings.
06:20So then you give it a different picture of a cat that's never seen out of them.
06:23You put the million to the side, you show a new picture, and it's like, oh, that's a cat.
06:28Because I've learned from what a cat might look like from these million things, right?
06:32That's it.
06:33It's just learning by example, learning the rules by example.
06:37Now, what are the examples?
06:40The examples are data.
06:41That's what the examples are.
06:43So that is the reason why everyone talks about data, data, data, data, data, because you need those examples to
06:50basically teach the AI
06:52what you're trying to do.
06:54So if you give a bunch of wrong examples, it's just like, it's just like raising a kid.
06:58I have a four year old and an eight month old.
07:00Like you give them a bunch of wrong examples.
07:02They're going to draw the wrong conclusion from those examples.
07:05Same exact thing with the AI.
07:07You give a bunch of data about what you want it to learn.
07:10If you give the wrong data, it's not going to, it's not going to learn the right thing.
07:15I appreciate that because you explained it so perfectly.
07:19And like, like you have young kids, like, you know, but it makes sense.
07:23So all that said, what is your strategy?
07:27Like when, when it comes down to it, what is your strategy?
07:31So a lot of people, when they talk about data strategy, and especially if you talk to technologists, they'll start
07:36talking about the infrastructure and, oh my God, I have to pick this vendor.
07:40I have to get to this cloud or I have to put all my data in some infrastructure.
07:44We have like, at Goldman, we have like a totally different thesis.
07:48We actually start from the business concepts.
07:51We start from the business workflows and the business concepts.
07:54And this is sort of like the way we've been doing data for the last 10, 15 years.
07:59You start with like, what is a trade?
08:01What is a counterparty?
08:03What is an OTC derivative?
08:05And you sort of start building this mental model of finance through data.
08:10So you say like a trade, okay, it has a buy-sell flag.
08:12It has a quantity.
08:13It links to a counterparty.
08:15It links to like a firm account.
08:17It links to this OTC derivative product or whatever.
08:20And so we started on this journey of building basically a data graph of what the concepts in finance are.
08:31And that is actually a joint venture between engineering.
08:38We call ourselves engineers because that's like the cool way to say technology now.
08:41Tech geek.
08:41Tech geek.
08:41Yeah.
08:43Uh, like, uh, that is a, that is a, a joint effort between engineers and subject matter
08:50experts in that whatever data domain we're actually trying to model.
08:55So if we're trying to talk about our HR data, we actually sit with the HR people who are experts
09:05in HR, not the HR engineers, the actual HR experts, and we're like, well, what does a person mean at
09:12Goldman Sachs?
09:13What is an employee, and how does that work, and what does it mean to be a lever, and what
09:17does it mean to join, and how many join dates can you have?
09:20And so you actually sit together, and you build this shared understanding of the data, and this shared ownership of
09:27the data, and we do that across all of our lines of business, all of our investment functions, all of
09:33our non-investment functions, all through the whole stack, and you start building this sort of knowledge graph or this
09:41understanding of all your data, then us as engineers, I run the data engineering team as well,
09:48then we sort of help the firm then project that into whatever infrastructure people want.
09:55So it could be a really fast database you need for this thing, but you really need a deep analytics
10:00thing for this thing, and so from those concepts, then we could do the sort of engineering work in the
10:06background,
10:06you know, but everyone has a shared understanding of the data, and then they could query that data in that
10:11shared understanding.
10:12That's like the, that's the pretty cool part that I think sets us apart from what other people talk about.
10:19So what are some of the challenges that you might face?
10:22Because I know we've had, you know, different questions from the audience come in.
10:25One was about cyber security, like how do you address the issue of data security, keeping everything secure?
10:32Yeah, look, we, again, we have a, we have a thesis on security.
10:36First of all, we have like the best CISO in the world, so that's like, helps a lot, and the
10:42team, and our security team is the best.
10:44But we sort of have this thesis about, about, about, about cloud infrastructure that maybe, that maybe is contentious,
10:53but basically at these hyperscalers, you know, they, they basically can spend billions of dollars on security, and data security,
11:02and cyber security.
11:03And so our thing is like, we like that.
11:06We like that you could spend $10 billion on making sure everything in the hyperscaler is secure, and we'll just
11:11leverage that.
11:11We make sure that all of our data is encrypted, that all of our entitlements are super precise, that like,
11:17you can't see what I could see,
11:18because we sit in two different functions, we have info barriers, and things like that.
11:22So those are all like rules that we built, like business logic rules we build on top of the infrastructure.
11:27But from an infrastructure perspective, we think like that's the best sort of security way to go.
11:32Now, when you're talking about implementing AI, let's say you were talking about data, is it the same with generative
11:39AI?
11:39Is it the same with agentic AI?
11:41Is it the same strategy that you use?
11:43Yeah, so the AI, again, I'm going to talk my own book here, because I'm the data person.
11:48So I think it always starts, I think it always starts with the data.
11:52We just, again, our thesis is humans needed to understand the data, and now AI needs to understand the data.
11:58Same difference.
11:59The agent, it's actually even worse for an agent or AI.
12:02Like, the agent or AI can't go ask another human in ops, like, wait, what did you mean by this
12:07field?
12:08Did you really mean that this account, or did you mean this custodial account?
12:11So, like, a human could go to a human and ask, like, what that data means.
12:14Agents can't do that.
12:15So, again, this shortest shared understanding, this graph understanding, this explanation of what a trade is, codifying that in the
12:23computer.
12:23What a trade, and how is a trade linked to a counterparty, and how is a counterparty linked to this,
12:27and this.
12:27Like, that building that sort of knowledge, understanding, what we call a semantic layer, or whatever, that becomes the power
12:36of how an AI could understand the world of Goldman Sachs, and the knowledge of Goldman Sachs, and the understanding
12:42of Goldman Sachs.
12:42So, for us, we've been doing AI for a long time.
12:46It started with traditional ML and things like that.
12:48It's always been the same thesis.
12:51Like, get your data in an understanding for the humans and the machines, and, like, great things will happen from
12:59there, basically.
13:00Can you give any specific use cases?
13:02Like, how is Goldman Sachs using AI?
13:05Yeah, I think a few of these panels have talked about some of the things.
13:10Like, first of all, I would say, like, if you have an engineering organization, a tech organization, and they're not
13:18100% all using agent coding tools, like, you're way behind now.
13:26And, like, way, way behind.
13:28Like, way behind.
13:29Like, and, again, I'm going to tell you guys, I've been an AI skeptic for a very long time.
13:34If you ask me, if we were on this panel a year ago, you know, I told Marco, who is
13:38our CI, like, this stuff doesn't work at all.
13:41Like, this just doesn't work.
13:43It just doesn't work.
13:44Like, and I spent, like, six hours a day writing code.
13:47I was like, this doesn't work.
13:49Like, six months ago, the whole world changed, and the stuff is, like, superhuman.
13:54It's, like, incredible.
13:55It's, like, blowing my mind how insane it is right now.
13:59So from a developer thing, like, your team has, like, even if you're outsourcing your tech, you better tell those
14:07people that they better be using, like, coding agents.
14:10Like, it's just the world has made some crazy step function here that, like, is going to leave everyone behind
14:19who's not doing it.
14:19So that's, like, a slam-dunk productivity one, like, you just got to do that.
14:23There's some other productivity things people have talked about, like, yeah, you got to roll out, you know, co-pilot
14:29for your emails and chats and blah, blah, blah, whatever, whatever.
14:32Like, the real stuff, like, we're thinking, like, I really think about is what gives Goldman Sachs an edge, what
14:39gives Goldman Sachs an edge in the future, like, when knowledge is sort of commoditized, when everyone sort of has
14:45the same knowledge, like, what is left of Goldman Sachs?
14:48And I think a few of our key business leaders, I'm going to attribute to them that this was sort
14:55of their idea, but there's an importance, we think there's an importance of, again, a shared understanding and a shared
15:03knowledge of what happens at Goldman Sachs.
15:07There's a tribal knowledge of people sending emails, and I had this meeting, and I, oh, I did this, and
15:12we need to sort of codify that for the whole organism, for the whole Goldman Sachs organism.
15:18So building sort of this, what a colleague of mine, Chris Chirshava, called, like, an intelligence layer for Goldman Sachs.
15:25You could ask any question and get any answer, but not just, and not just humans could ask questions, but
15:31agents could ask questions.
15:32Agents could ask questions and then derive some insight from that question and then sort of push the answer to
15:40that insight back into this intelligence layer.
15:42So that's, like, I think, one of the most fundamental, like, things we're working on.
15:47And then I'd say something that we're thinking about, but this is sort of like the end dream state, is
15:57taking real trading and PM workflows
16:02and sort of thinking about how to decompose those into humans and agents becoming superhuman.
16:09So, like, an agent helping you build your thesis, an agent helping you execute your thesis, an agent then helping
16:15you look at the outputs of your thesis
16:18and then feeding that back and doing the whole loop.
16:20So I think thinking about workflows across sales and trading, banking, whatever, like, that's where the real value, I think,
16:29is going to accrue.
16:30Yeah, you're leaning into my next question.
16:31Where do you see the most value?
16:33Is that?
16:35Yeah, I think reimagining workflows so that the humans are superhuman.
16:44Like, I have, again, I have a, I'm an eternal optimist about humanity.
16:49So I have a huge, I have a huge thesis that, like, this is not going to destroy the world
16:54and it's not going to take everyone out of jobs.
16:56Like, sure, there's going to be some jobs that, but, like, I actually think all of this is, is, like,
17:02giving us superhuman abilities to do whatever we want to do better, faster, cheaper.
17:08And so I think the value is going to accrue to those people who are, like, super interested in learning,
17:15super analytical, think, like, outside the box,
17:19and then let this AI help you become, like, a super intelligent version of that.
17:26So, like, whether you're a trader, like, okay, I could, you know, it used to take me hours and days
17:31and weeks to come up with one hypothesis,
17:33but now I could come up with 10,000 in an hour, or I'm a salesperson and I used to,
17:39you know, I used to have to read all the stuff and I used to have to do this and
17:42this to tell one client, like, this is maybe what you should do.
17:45But now I could do that personalized across thousands of clients and I could help them personalize it to that.
17:51So, like, it's really about imagining, reimagining, I think, personally, the killer use cases, reimagining these workflows of how this
17:59stuff makes you a superhuman in your role.
18:03And the last few seconds we have left, I can't have you here without asking, how do you use AI
18:11on a personal level?
18:12Like, when you're home, like, how are you using AI?
18:15Yeah, so, I put my four-year-old to sleep at 8 p.m. and from 8 p.m. to
18:202 a.m. I'm writing code, agent code.
18:22I'm not writing the code, I'm telling the agent what to do with my code.
18:25So, that's the first thing I'm doing about six hours a day.
18:28The other thing is, when I'm not doing that, I'm building tools for my four-year-old at my eighth
18:33month old.
18:34So, I built, like, an interactive story generator for my four-year-old.
18:40So, he could pick the character, he could then pick the setting, he could also pick the style of the
18:45drawings, and then he could pick the sort of theme, and it sort of builds, like, a nighttime story for
18:50him that's personalized to him and what he's picked, which I think, that's been my favorite thing to do.
18:57That is awesome.
18:58Who knew you could do that?
19:00Nima, thank you so much.
19:02All right.
Comments