Scaling AI Efficiency at LinkedIn with Erran Berger, CTO of Engineering

(Karthik Ramgopal at 00:00:00) Got one of the most at-scale verified identity platforms on the Internet today, and I think we're pushing 100 million people, members who are verified, if we're not already there. And pretty amazing, especially in a world today where abuse is rampant. It's almost hard to know, am I talking to a real person or a bad actor? Am I talking to a real person or AI?

(Joel Beasley at 00:00:30) I've gotten to talk to many people at LinkedIn, from Mohak, all different types of people. Everyone that I've met has been brilliant, kind, fun, intelligent. It's just been an absolute pleasure. So the bar is high. But I'm curious, what role are you in at LinkedIn?

(Joel Beasley at 00:00:45) What are you doing there?

(Karthik Ramgopal at 00:00:47) Yeah. So I'm CTO of our engineering organization. We pretty recently restructured software development at LinkedIn into two organizations. One that we call engineering, one that we call infrastructure. I lead the engineering organization. That is the teams that build all of our user-facing applications, from our flagship application that you probably have on your phone to our enterprise products that we sell to recruiters and salespeople to our advertising business. It also includes a bunch of the platforms that support those businesses. So things like our trust platform and how we keep people safe and customers safe on the platform, messaging, search, those kinds of things that are really foundational to building any of these products. It includes our AI foundations team. Obviously, so much of the value that we create for our members and customers is from AI, and so we have a team that's developing foundational capabilities that our product teams use.

(Karthik Ramgopal at 00:01:44) And it also includes some productivity teams focusing on employee productivity. In particular, a huge part of our employee base is go-to-market. And so much of how we go-to-market with our products is tied to the features that we build as well as how we enable people. And so that's kind of end-to-end. And so that's my scope.

(Karthik Ramgopal at 00:02:08) And I've got a peer, his name is Raghu. Brilliant guy. Somebody I've partnered with for a long time. I don't know if you've had a chance to meet him. He runs infrastructure, and so, you know, we depend on a lot of what his team produces to build amazing products for our members and customers.

(Joel Beasley at 00:02:25) Yeah. Absolutely. We did... His last name starts with an H, right?

(Karthik Ramgopal at 00:02:29) Yeah, it does.

(Joel Beasley at 00:02:31) Yeah. We did something, I think, back a few years ago. But, yeah, every... Is Mohak still there?

(Karthik Ramgopal at 00:02:36) Mohak's still kind of within the larger Microsoft ecosystem. He's working on some pretty fun projects around the future of work, and obviously we still spend a lot of time together. And I reported to him for, I think, at least a decade before he kind of moved on to bigger and better things. And Raghu was my peer reporter for basically just as long.

(Karthik Ramgopal at 00:02:58) And, you know, we kind of took over the reins from him and learned a lot from him. So it's been fun kind of watching him scale his impact.

(Joel Beasley at 00:03:07) Yeah. After a thousand episodes, there are a handful of stories that I'll remember forever, and him riding around on that tricycle is gonna be one that just sticks with me until the day I die.

(Karthik Ramgopal at 00:03:18) There's probably some other stories, but I'm not sure if he'd want me sharing them.

(Joel Beasley at 00:03:21) So, yeah, we'll...

(Karthik Ramgopal at 00:03:21) Stick with the tricycle story, maybe.

(Joel Beasley at 00:03:23) Off the mic.

(Karthik Ramgopal at 00:03:24) Off the mic. Exactly.

(Joel Beasley at 00:03:26) So when you iterated through that list of products, I mean, I was familiar with about half of them. You guys do way more. You know, I saw that safety one. Is that the one where it's telling me... It's asking me to verify my identity? I saw that come up recently.

(Karthik Ramgopal at 00:03:40) Yeah. I mean, we have a lot of products, and I think there's a lot of features even within any of those products. The one you just talked about was verifications. You know, we've got one of the most at-scale verified identity platforms on the Internet today. And I think we're pushing 100 million people, members who are verified, if we're not already there. And it's pretty amazing, especially in a world today where abuse is rampant. It's almost hard to know, am I talking to a real person or a bad actor? Am I talking to a real person or AI?

(Karthik Ramgopal at 00:04:19) Having that verified badge on a profile, in contacts when you're messaging with somebody, seeing them in your feed, or they're reaching out to offer you a job—are they a real recruiter? It is so important, I feel, and more important than ever. And so, you know, a few years ago, we decided we are going to give free access for people to verify themselves, who they are, where they work. And I think our members really appreciate it. Customers really appreciate it. They know that the people that they're talking to on the other end are real.

(Karthik Ramgopal at 00:04:55) And so, you know, it's amazing to see. And, you know, I've been here for 17 years. Seventeen years later, we're still finding new ideas that really make our ecosystem better, the world of work better, and verifications is just one example.

(Joel Beasley at 00:05:14) All right. Well, I want... I've got questions about the 17 years. But before that, I want to talk about this moment of the splitting of the function. You say you reorged to infrastructure and this other part.

(Joel Beasley at 00:05:27) What was happening? What conversations were occurring where you all got together and you said, this is a moment where we need to do this? What was happening?

(Karthik Ramgopal at 00:05:39) The truth is that, actually, you know, Raghu and I both reported to Mohak, and we had a few other peers before the organizational change. But the majority of engineering, even at the time, was reporting to us. And he was running infrastructure, and I was running most of my current scope, actually. And so I think on paper, it felt like a really big change. But, actually, in practice, it wasn't that substantial.

(Karthik Ramgopal at 00:06:04) I was running all of our product experiences and the platforms that supported it. That was already the majority of my scope. He was already running infrastructure. So Mohak kind of moved to take on that bigger role at Microsoft. It felt really natural.

(Karthik Ramgopal at 00:06:19) Our teams work and have worked really well together for a long time. Raghu and I, like I said, have been working together for a long time. And, you know, down to day-to-day, leaders on my team, engineers on the ground in my team, collaborate across the aisle with folks on his team. So while it feels like a big split maybe outside looking in, internally, it felt really seamless and not really like a big change besides maybe some title changes for the two of us.

(Joel Beasley at 00:06:50) Right. And 17 years. So how did you get hired originally at LinkedIn?

(Karthik Ramgopal at 00:06:56) Well, it's a funny story. My wife will tell you that all of my career success is entirely due to her. I was living in San Diego at the time. We both met right after college. We both went to UC San Diego. She decided to go to law school. She went to law school up here in the Bay Area. We did a long-distance relationship, as you know, people do, for a year. And then I think basically it was, hey, if this relationship is going to continue, maybe you should think about moving up to the Bay Area.

(Karthik Ramgopal at 00:07:28) And so I started looking for jobs up here, knew somebody. It was a pretty classic LinkedIn story, but IRL. I had a second-degree connection who worked at LinkedIn, gave them my resume, they handed it to a hiring manager. And one thing led to another, and I was interviewing to be the tech lead for profile. This was back in 2009. LinkedIn was basically a pretty simple product back then.

(Karthik Ramgopal at 00:07:50) You signed up for LinkedIn. You created the profile. You sent out some invitations, and you waited for something good to happen, maybe. Thankfully, the product has become much richer and more useful since then. But I was like, oh my God. I get to be the technical lead for the face of LinkedIn, this emergent company.

(Karthik Ramgopal at 00:08:12) And so I took the offer. And, of course, none of that would have happened—full credit to my wife—if we hadn't had the conversation about where the relationship was going. And we got married, and we have a wonderful family and kids now. But the start of that journey was really that conversation that we had back in 2009 that led to me applying here.

(Joel Beasley at 00:08:33) Dude, shout-out to wives. That's how this podcast got started.

(Karthik Ramgopal at 00:08:36) Big shout-out to wives. I'm big plus one to that.

(Joel Beasley at 00:08:40) I was writing this book, and I was like, here's everything I learned from developer to, you know, first-time CTO. I just wanted to capture that moment. And I wrote it, and I started sending copies out to other technical leaders and saying, hey. I don't want to get ripped apart in the Amazon reviews. Will you just take a look at this and tell me if you've had similar experiences?

(Joel Beasley at 00:09:00) We got on a couple calls. I was having so much fun. I was like, this is so much fun to talk about all these topics with these guys. And my wife's like, you should take the recordings from the calls. You should release them as a podcast.

(Joel Beasley at 00:09:10) I said, no. I was like, that's silly. There's a million podcasts. She's like, you should do it. We did it. We're a thousand episodes deep now, and it's been my full-time gig.

(Karthik Ramgopal at 00:09:18) There's the origin story of Modern CTO. That's wild.

(Joel Beasley at 00:09:23) Yeah. Yeah. And so it's been almost 10 years. But seven years full-time, it replaced my income from being a software engineer and building businesses and stuff. And so I just said, this is what we'll do.

(Joel Beasley at 00:09:36) And it's been great. And now we get to talk about all sorts of cool things with great people like you. One of the things I was interested in talking about is AI and efficiency. There's a lot of conversations about the next phase of AI. What are your thoughts?

(Karthik Ramgopal at 00:09:50) Yeah. Definitely top of mind across the industry and certainly here. You see, you know, a new... Feels like a new personal agent being launched by some company, you know, every day now. And so we see both what's possible and what people are starting to want and expect out of AI, which is it's not just something that I, when I have a question, I go to it and it gives me an answer and move on with my day. It's something that is a true assistant to me and it's almost maybe always on and doing work on my behalf. And, you know, I can just tell it to go take care of something and it goes and takes care of it, tells me when it's done. And with that evolution becomes increasingly complex workloads. And two things: bigger models that are more capable, obviously, and therefore take more compute, and workloads that require much more inference and many more turns. And all that means more compute and more expensive end-user features that companies like ours are shipping. And so if you assume that this is a new normal—and in fact, usage of those workloads will go up, people will have even greater expectations in terms of capabilities of what these things can do—the cost of the products we're shipping is just growing exponentially. And so at the same time, we still have to run businesses.

(Karthik Ramgopal at 00:11:22) We still have to make sure that the value that we're creating for our end users, be they consumers—we call them members—or customers, enterprises, that the products we're shipping them are creating value, but also that they can more than cover the cost of building those things. And that's where efficiency comes in. And so the strategic bet that CTOs like myself have to make now in technology is, how do we enable these amazing, rich, and, you know, ambient, anticipatory user experiences that are very compute-heavy and build them in a way that enables us to run a highly profitable business? And so that's kind of the essence, both, frankly, internally, but a lot of what I'm hearing from my peers outside the walls of LinkedIn.

(Joel Beasley at 00:12:20) I'm gonna dig as much as I can. You just tell me when I hit a wall. Okay. So how do you budget that? You want to make this great product, and obviously there has to be the business. How do you go about saying... Do you start with the bucket of money?

(Joel Beasley at 00:12:35) Do you say, okay, this is the bucket of money. You can make it as great as you can within this? Or is it the other way around? Do you try to come up with the craziest thing and then go pitch the bucket? How does that actually happen on a feature that's directly tied to revenue?

(Karthik Ramgopal at 00:12:50) So, I mean, actually, let me even back up before I even answer that question. One of the most important things is knowing where your costs are going, right? So many companies—you'd be surprised how many companies at scale couldn't tell you, hey, for this particular feature, how does it wind up resulting in compute spend across all the infrastructure systems that we run: CPUs, GPUs, storage, streams, indices, and of course, inference for AI? That's a really hard problem, especially when you've... If I'm a startup, frankly, even if I'm a high-growth company, and I've never really had to worry too much about efficiencies because in some ways, you know, with scaling laws and software, pre-AI, this wasn't really an issue. So people didn't invest too much in that type of telemetry, and now you've got this sprawling architecture.

(Karthik Ramgopal at 00:14:05) And you have to go now retrofit instrumentation from a page view all the way down to your databases and everything in between—it's a hard problem. And so we spent years building out the telemetry and the analytics to be able to say truly, for any feature that's running in production, how do we attribute the costs? So foundationally, we had this capability. So then we could say, okay, we want to ship the following features next year. That's everything from a GenAI feature. It could be just a user-facing UX feature, but it's gonna grow traffic in a particular way. It could be the next rev that we want to do on our feed AI model that's gonna be more compute-intensive, but will have higher quality recommendations for people. Any number of these things, we then say, okay, these are the big, big rocks that we want to ship, because you don't... It's not practical to do it for a 12-month roadmap for every little thing. What are the big things that we're going to ship next year?

(Karthik Ramgopal at 00:15:07) As best as we can tell based on what we think we're going to have to build, what are our expectations for usage, and then, of course, how much does every compute unit that this thing is going to use cost? You kind of roll that up into a budget. And then you say, okay, well, how much can we afford based on what our revenue forecast looks like next year? And then you try to kind of fit all those big things into the bucket budget that you have, and you're trying to figure out, okay, well, maybe this thing is gonna drive up a little more revenue than we thought. So let's figure out how to fit it in. And maybe this thing is not gonna drive revenue at the kind of margin that we want. So maybe that's something that we defer until we can figure out a way to make it more efficient. So those are the conversations. It's a little bit of the art of how it works.

(Karthik Ramgopal at 00:15:54) But, again, most importantly, if you don't have the telemetry, you can't do any of that. And so that was a big investment for us.

(Joel Beasley at 00:16:03) Let me recap so I understand properly as an engineer. So foundationally, you built a proprietary system to measure the cost of features running in production based off of this unit of compute. Right? So you have a feature, a unit of feature consumes a unit of compute in production. And when you're even at the drawing board of the feature, you can kind of estimate the potential compute usage.

(Karthik Ramgopal at 00:16:26) That's right. And then, I mean, the reality of software engineering is that you draw up what you think, and then you have to go build it. And so sometimes we, just based on our history of being able to build new things and then doing the hard work of craft and software engineering, we believe we can get it to a particular cost profile. And so, you know, it's interesting. Tomorrow I'm going to be doing a deep dive in our company All Hands, talking about all of the work we've done in efficiency in feed, in our search stack, and then some of our enterprise products and job seeking products.

(Karthik Ramgopal at 00:17:09) And, you know, over the last year, every one of those stories starts with, we were at, like, one QPS per GPU, and then we were able to find 100 or, like, a thousand x the efficiency as we went from an initial implementation of that AI product to what we finally shipped and scaled to our members. And I don't think that when we started those projects, we knew exactly what all the levers were going to be that were going to drive that two or three orders of magnitude efficiency gains. But we had a toolkit and we had, you know, creativity and obviously a bunch of really talented innovative engineers. And so we kind of projected...

(Karthik Ramgopal at 00:17:53) We knew a priori roughly what our cost envelope was going to be, and then we fit it in. In a lot of cases, we even surprised ourselves how cheap we were able to get those things to be and still preserve quality, which is, again, like, you've got to balance those things. It's not helpful if you make it, you know, really high throughput and cheap, but then it's not a product that members or customers love.

(Joel Beasley at 00:18:16) Absolutely. Well, if you're on my team, we're not even having that conversation. We're only making greatness here. That's all we need.

(Karthik Ramgopal at 00:18:23) Of course. You've got to start with that, right? You've got to start with value, or else why are you doing it?

(Joel Beasley at 00:18:28) That's right. Okay. So let's bring this down. There's a lot of people listening to the show that have between 20 and 1,000 engineers in that frame. Obviously, that's way smaller than LinkedIn. And so for people who are in that frame, should they be thinking about efficiency? If they are, are they trying to replicate what you just described, or do you have different insight for them on where they should be paying attention?

(Karthik Ramgopal at 00:18:55) I think everybody needs to be thinking about it.

(Joel Beasley at 00:18:58) Okay.

(Karthik Ramgopal at 00:19:00) You know, AI is the product now. The model is the product, Satya likes to say. And those models are... And, like I said, expectations of how those models are going to be used to create end user experiences are increasing. And every company, at some point in their lifecycle, if they want to succeed, have to figure out how to become profitable.

(Karthik Ramgopal at 00:19:27) And not just, like, a little profitable, but, like, very profitable. And there's not a world where the experiences themselves are so compute hungry and therefore so expensive to run. The only path to profitability is efficiency. Now, we have the luxury of running our own data centers. We own the full stack end to end from the apps that we ship all the way down to the bare metal GPUs and, frankly, the kernels on those GPUs. We get to optimize the whole thing. So not everybody has the luxury. Frankly, not everybody needs to do that.

(Karthik Ramgopal at 00:20:04) But even for companies built in the cloud, even for companies that are purchasing their, standing up maybe their own small data centers to do their own training and inference, having an understanding of a path to efficiency that leads to profitability feels like a non-negotiable.

(Joel Beasley at 00:20:27) Yes. That's what people... That's what you... That's the focus right now, efficiency.

(Karthik Ramgopal at 00:20:31) You've got to get to product market fit. Don't get me wrong. Right? Like, there's no point in working on efficiency if you don't have a product that people love that are going to use that you're going to be able to sell. Right?

(Karthik Ramgopal at 00:20:40) So, like, I'm kind of skipping that step, which is a really hard step. Right?

(Joel Beasley at 00:20:45) Well, good luck finding your way to 200 engineers without that product market.

(Karthik Ramgopal at 00:20:49) Fair. Totally. Exactly. So if you're taking that for granted based on your audience, obviously, then, yeah, I think. And, like, you know, all of this is the tough task of being a CTO. Right? Like, you've got to understand the business and where it's headed and the trajectory. And then you have to do this balance between how much... How many of my engineers am I putting into new features and growing the business and creating more value, and how much am I putting toward efficiency.

(Karthik Ramgopal at 00:21:22) And so that as we grow and as we have more success, we do it in a way that's profitable. And if it's not profitable now, it eventually becomes profitable on the timeline that, you know, the exec team feels like they... You know, the executive team and the shareholders expect. And so it's just... It's the job. I mean, there's a lot of tough parts about being a CTO, but, honestly, that's the core of it, which is, like, how do I balance these competing priorities to make sure that the technology that runs our products helps us achieve the greatest success.

(Joel Beasley at 00:21:55) Now when you... With the big push for efficiency, are you... Do you have dedicated teams just looking for efficiency, or are you having the lunches and sort of, like, the Friday afternoon meetings where people are just showing what they're doing with efficiency to other engineers? Like, how do you actually tactically get this efficiency ingrained in a large organization?

(Karthik Ramgopal at 00:22:16) I mean, honestly, it's all of the above. Obviously, the infrastructure teams at LinkedIn carry a big load there because they have the greatest leverage to find opportunities for efficiency. So a team... You know, we've got a team that's focused on kernel optimizations for our GPUs, and that means that any time that we can improve the way that the kernel operates or the utilization of those GPUs by offloading work onto CPUs that doesn't have to run on GPUs, that benefits every workload that's running on a GPU at LinkedIn, so some of our most important key product features. And so the greatest leverage for efficiency happens at the bottom of the stack.

(Karthik Ramgopal at 00:22:55) On the flip side, vertical teams also have to think about this. So I give the example a little bit of the feed team, for example. We launched a new sequence-based model, an agenda recommender for feed, and it's been doing amazing in terms of helping people find the best posts in their feed. And they were able to find by context compression and all sorts of other examples, some of which we haven't talked about, but we'll probably post soon. How do we pack more inference into one GPU?

(Karthik Ramgopal at 00:23:34) And so this happens at all layers of the stack. And I will say it's been a cultural shift for us because for the first fourteen, fifteen years of my tenure here, everything... Like, compute was not a constraint. The constraint was, how do we ship faster, better, higher quality, more performant products to our end users to create more value because compute was cheap.

(Karthik Ramgopal at 00:24:05) And that's changed over the last couple of years, and I think that's therefore changed the culture. And now engineers have to think about it more. And frankly, I think these are more interesting technology problems because you still have to ship the really amazing product features, the next gen AI models for click prediction in ads or whatever it is. But how do I do that efficiently and with a mindset of understanding the full stack and no longer just saying, well, there's an abstraction over inference and the inference team and the AI platform will take care of that. No.

(Karthik Ramgopal at 00:24:42) Like, you actually have to understand how your feature and your workloads thread all the way through the stack. That's a really interesting technology problem. And I think that that cultural change made engineers realize that this mattered. And I think they're finding it really interesting too.

(Joel Beasley at 00:25:00) Quick side note on this. You brought up the feed and the recommendation. So this has been a recurring conversation over the past decade on the podcast. Right? And about how the algorithm works and all of that stuff across multiple different of the major players, LinkedIn included. One of the things that I started thinking about and started noticing a trend toward maybe five years ago, I started noticing small steps towards giving the end user... And I'm talking, like, in the social aspect, like the scrolling of an individual, more control over the algorithm. It looks like everyone's taking small incremental steps towards exposing how the algorithm works and allowing people to say. Have you seen that trend as well?

(Karthik Ramgopal at 00:25:46) Yeah. I mean, I think the kind of most obvious manifestation of that is the shift toward what we call out-of-network content. But it's, you know, I think TikTok really was one of the first platforms to do this really effectively at scale, which is okay. I'm both going to infer what your interests are, but I'm also going to ask you what your interests are. And then I'm going to adapt the feed to that.

(Karthik Ramgopal at 00:26:14) And, you know, the magic of it is... I mean, it's like that we can explore new things, emergent things, things that are new interests of yours, and then the feed adapts to it as close to real time as possible. And so I think that, you know, that's almost implicit control, but it feels like the products now really get me and are serving the things that I want and that I'm interested in. And then, of course, I think we have, you know, there's all sorts of variants of what control looks like. There's, you know, kind of a network-based feed, a chronological-based feed. Like, all companies are experimenting with this. Being able to basically spin up a feed based on a natural language prompt, which I'm sure is something that is going to emerge. I mean, you see it on Spotify.

(Karthik Ramgopal at 00:27:09) Right? You can basically say, hey, I want to spin up a playlist. And so I think we're going to see more patterns like that—

(Joel Beasley at 00:27:13) Yes.

(Karthik Ramgopal at 00:27:13) —where you can come to a place like LinkedIn and say, you know what? Today, I'm really interested in what's happening with data centers. Right? And, like, the politics around it as well as the ground truth of it.

(Karthik Ramgopal at 00:27:29) So, like, show me a feed curated by industry experts in the data center space and in electricity space and in the regulatory space. And show me a feed of that. And I want to go deep on it. But maybe that's an interest for this week and then next week you have something else. And, you know, shameless plug for LinkedIn.

(Karthik Ramgopal at 00:27:48) And since I work at LinkedIn and I've been working on the feed forever, one of the things I think is uniquely interesting about LinkedIn in this world of being able to actually go and say what you want to consume today is we talk about verifications, real people talking about these things. You have their professional identity, right, associated with the content they're posting. And so it's... You can see... You can imagine, you know, an expert in, I don't know, like urban planning or the electrical grid, having a take on this stuff.

(Karthik Ramgopal at 00:28:28) And you know that they're credible because you can click through their profile and you can see that they're an expert in this thing and have been working in this field for a long time. And so I think that, of course, every company is going to have their version of this, I think, and every product will have their version of this. But I'm pretty excited for ours because that professional identity really is going to stand out when people want to go deep, especially in professional topics.

(Joel Beasley at 00:28:53) Everything's kind of becoming an API these days. Right? Like, it's... My mind is having a hard time wrapping around the short term. I feel like I'm pretty clear on the long term.

(Joel Beasley at 00:29:06) I feel like if you and I are having this conversation in ten years in the future, most of us are going to have personal agents and our world... Our digital world will filter through that agent.

(Karthik Ramgopal at 00:29:15) Mm-hmm.

(Joel Beasley at 00:29:15) Right? But there's this gap between now where, like, I still log in. I still use these products directly, but more and more every day, I'm using... Like, I used to use Apollo all the time.

(Karthik Ramgopal at 00:29:28) Yeah.

(Joel Beasley at 00:29:29) But now, I use Apollo through Cursor's MCP, right? So, I used to use all... Like, my tools are slowly building themselves into my agent where I go to my agent and I sit down, I want to get this work done, or I want to see a certain type of content. And so that transition where, like, do you see a... Am I wrong? Like, do you think that that world in ten years isn't going to be what it is? Let's start there.

(Karthik Ramgopal at 00:29:52) No. I think that's right.

(Joel Beasley at 00:29:54) With what I just described?

(Karthik Ramgopal at 00:29:56) I think everybody will be using personal agents for lots of things much sooner than ten years from now. I think that we are... I think it is really early, and you're seeing early adopters like yourself and me and other people, especially in tech and in Silicon Valley who are really throwing themselves into this world. I think that adoption at scale is going to take some time. There's a learning curve with how to use these things and how to use them well.

(Karthik Ramgopal at 00:30:29) They'll get better. There's kind of people just getting comfortable transitioning out of the way that they use technology and products. I think it's not unlike the desktop web to mobile transition where, like, everybody could see that the future of Internet usage was going to be on these devices. But there was... And I don't know. I don't have the data offhand, but I think the actual transition took far longer than we like to think in retrospect. From the day, for example, that the first iPhone came out to where the majority of Internet usage was on phones, it took a while. And so I think we're going to see kind of a similar transition. And then I think that there's, like, really interesting parallels to what you just talked about, if I'm kind of jumping ahead to the follow-up question on it, which is, in some ways, this is like a new operating system, just like mobile was a new operating system, and there were apps within those mobile OSes. And so, like, I think that there's some manifestation or thought of these apps within the personal agent ecosystem.

(Karthik Ramgopal at 00:31:41) And there's probably going to be multiple OSes, multiple personal agents, maybe some professional, maybe some personal, maybe multiple personal that are kind of case specific. And then I think it's possible entirely that, you know, while there are now these new OSes, these portals to the Internet, there will still be apps that can deliver really differentiated capabilities for a specific job to be done that exceed the way that that kind of general purpose personal agents can deliver. And they remain a destination despite the fact that the portals to internet might be these personal agents. So that's my take.

(Joel Beasley at 00:32:28) Yeah. No. You're helping me really open up my mind to this. It's like I still pay for Apollo. Right?

(Joel Beasley at 00:32:33) It's just the interface in which I use it has changed. And, you know, it's a... I don't like recommending content to people. So I'll give you the ten-second summary so you don't have to listen. I interviewed Sir Tim Berners-Lee, creator of the World Wide Web, maybe seven years ago, six or seven years ago.

(Karthik Ramgopal at 00:32:50) Amazing.

(Joel Beasley at 00:32:51) And he... Yeah. He was brilliant. Or he is brilliant.

(Joel Beasley at 00:32:55) But he explained to me... I asked him where the future of the internet was going, and he explained it to me at this time where I could understand what he was saying, but it didn't make complete sense to me. But when you introduce the concept of the technology we have today, it is mind-blowing how accurate this guy was about the future. And so the way he described it in short was the example he used was, like, right now, all of your data is at the bank.

(Joel Beasley at 00:33:26) Right? You have all your banking data at the bank, and then you have a bank application, and then you can read your bank data with your bank application. He goes, well, then you'll be able to have other applications, which I was like, okay, API. That's cool.

(Joel Beasley at 00:33:39) Yeah. And then he's like, then there'll be local data sources where the data is local and you read and write and those read and write sync up. And he just kept going on and on. I'm trying to remember from seven years ago, so I'm probably messing it up. But it didn't... There was a lot of gaps in my understanding of how this would work. But now with the way that the AI models are working with MCP. Yeah. That to me is what that is.

(Karthik Ramgopal at 00:34:04) For sure. Yeah. And so, again, we'll see how this plays out, how companies integrate with the personal agents. I think that there's obviously business questions to navigate on both sides there and partnership questions to navigate there. But I think you can...

(Karthik Ramgopal at 00:34:22) You know, a good mental model is a personal agent's OS. The connector is the app, in the same way that, you know, pre-mobile, there was a particular UX and post, you know, iPhone and Android, there was a new UX. The way that you access the app was through that UX. But the actual application itself was the platform that powered both of those experiences. This feels like now a new UX that's text and voice and dynamic.

(Karthik Ramgopal at 00:34:52) And the new OS is the agent, and it sits alongside mobile and sits on desktop. And by the way, I think while we're making predictions, I think there still will be certain experiences that people will want to have where text and voice and personal agents maybe are not the best way to use those apps. You know? It's interesting even today, despite 90% of our usage on LinkedIn being mobile workflow, like people getting work done, recruiters, it's all primarily desktop because it's just so much better to do work, you know, on the web than it is on a little tiny device, a big screen, like all that stuff. And so I'm not suggesting that that recruiters...

(Karthik Ramgopal at 00:35:44) That's an example that will stay. It's entirely possible that personal agents will change that. Of course, we've launched our own hiring assistant for recruiters. But I think that there will be examples of internet usage that are still better, you know, maybe on a mobile UX or on a desktop UX than through personal agents. And so this is the period of sorting through all that that we are going through, and it's going to be super interesting to see how it evolves.

(Joel Beasley at 00:36:14) Well, I fully agree with you on that. The work associated with the screen size is definitely a thing. I like to look at my own personal habits and there are things that I like to do on the desktop. Like that's desktop work. I need to be at the computer for that.

(Joel Beasley at 00:36:27) And then there's things on the phone, you know, and then that bridge is kind of changed. There are little changes down there. It's like, oh, it used to be a desktop thing.

(Karthik Ramgopal at 00:36:34) Totally.

(Joel Beasley at 00:36:34) But yeah. I want to give you some credit, though. You nailed that connection between the iPhone and today. It's been, like, twenty years? It came out in 2007 or something, and it's almost...

(Joel Beasley at 00:36:45) Yeah. So about twenty years. Right? And at the... And you and I were both building within that time frame, actively building.

(Joel Beasley at 00:36:53) Building J2ME and—

(Karthik Ramgopal at 00:36:54) J2ME applications at a startup, trying to get navigation to work using J2ME and being so frustrated. And then the iPhone came out. I'm like, what am I doing? Like, all that stuff is totally irrelevant now. It was just this crazy moment.

(Joel Beasley at 00:37:10) That was brilliant. And we all could see it coming. We could all see the shift from the desktop to the mobile. But there's not one specific moment that you can point at and be like, that was it. It's just this blurry...

(Joel Beasley at 00:37:20) So I think that's a really good way to take in this transition from using the applications directly to using them potentially through agents. It's like, yeah, there will be a lot of situations where those blend, but we just don't know exactly when it's going to happen.

(Karthik Ramgopal at 00:37:34) And I think this is just... The best things that companies like ours can do right now is really focus on learning and being agile and trying things, because I think the companies that tend to get ahead in these moments are not the ones that necessarily wait for the patterns to emerge, but they're exploring different ways to adapt. And then, you know, most of those experiments that they do are not what we call one-way doors where you try it and there's no dialing back from it. You know? And so if you assume that most of the things you might try are these two-way doors, you can try it.

(Karthik Ramgopal at 00:38:15) If it doesn't work, you just sunset the thing and try the next thing. You are going to figure out... You're going to be amongst the people that figure out the patterns in this new world of personal agents. And so that's really our mentality. And so from a technology perspective, as a CTO, I'm like, okay, how do I make it as easy as possible for us to experiment with as many of these things as possible, as quickly as we can so that we can figure out, okay, there's the winner of what it means to do job seeking in an agent-first world.

(Karthik Ramgopal at 00:38:52) Let's go.

(Joel Beasley at 00:38:54) Yeah. Right? And you know what? That actually brings up another thing I'm curious about. So...

(Karthik Ramgopal at 00:39:01) From...

(Joel Beasley at 00:39:01) From an engineering perspective, and if I'm getting too detailed on a specific product, you know, we can just talk about that and you can tell me. I'm trying to think about bots. You know, huge part of the infrastructure in the system is to stop the automated bots from doing stuff. You know, all the big companies I talk to talk about it and I get it. But then, like, how are you going to differentiate bots from personal agents?

(Joel Beasley at 00:39:27) Like, that might be interesting. Have you guys run into that yet?

(Karthik Ramgopal at 00:39:32) Yeah. I think that's a good question. I mean, certainly, you know, especially with browser use, somebody's using, you know, a browser use agent that's taking control of your own desktop and issuing a command that... It looks like, you know, a human using that browser with your own user agent, your own IP, your account, but it also... You know, it's directed by the person.

(Karthik Ramgopal at 00:40:05) So I think this is fundamental questions about, like, it's... You know, if the person's... If the person is trying to get value out of LinkedIn, there may be even... I mean, the whole premise of LinkedIn is it's a social network, network effect. As people do more on LinkedIn, that's generally beneficial for the network overall.

(Karthik Ramgopal at 00:40:21) Right? And so I think, certainly, there's interesting questions here from a product strategy standpoint. But I think it is something that is happening and something that, in some cases, is totally fine and value accretive to LinkedIn overall. So I think we're 100% thinking about it. I think that the...

(Karthik Ramgopal at 00:40:43) As I think about our own strategy in a world where people will be using personal agents, some of which maybe we can detect and some of which we can't. I think that we start with... I think if we start with member and customer value and we think about where do people... How do people want to use LinkedIn? Do they want to use it through an assistant?

(Karthik Ramgopal at 00:41:05) Do they want to come to LinkedIn? In what circumstances might they want to come to LinkedIn to do the same things? I think we can figure out a way to deliver the best product that ultimately meets our members and customers' jobs to be done. And then from purely from a technology perspective, I think that there is a place for task-specific agents that can outperform on quality and cost and latency when they're purpose-built, especially when the comparison is computer use, like browser use agents. Like, imagine an agent that we build for job seeking, purpose-built using our internal APIs, and somebody trying to use a personal agent that's navigating our browser to try to accomplish the same task.

(Karthik Ramgopal at 00:42:01) Like, it's very intuitive which one's going to be more efficient. Probably which one will be higher quality because you're doing inference based on, you know, the data models themselves and the APIs themselves versus the lossy translation from those APIs into some UX. Right?

(Karthik Ramgopal at 00:42:21) And of course, the latency. Right? It's just slow to do browser use relative to an API invocation, inference, and then a response. And so I think that there is, you know, like, if the comparison is browser use versus purpose-built agents, first-party agents, I think with the companies that can build awesome first-party agents, I think there's a place there. And then I think there's still a question about, well, like, let's say you have that agent.

(Karthik Ramgopal at 00:42:44) Should that agent be integrated via an API with the personal agent? Or is there a reason for why people would come directly to that agent to use it? And so I think that there's another interesting strategy question. Maybe the answer is both. Right?

(Karthik Ramgopal at 00:42:57) And this is one of the things that companies like ours need to experiment with and learn about what their members and customers and users want and, you know, and figure it out from there.

(Joel Beasley at 00:43:07) You know what I love about this generation right now is that we all grew up building this stuff and there being limitations and us wishing there weren't. And now that whole generation is the generation in charge of things. And we're living that out still. It's been beautiful to me to see all the engineers be open to that. They're not trying to hold and protect and hide away.

(Joel Beasley at 00:43:35) They're like, all right, this is what people want. Let's give people what they want. And to me, that just makes me excited for a bright future because I think we'll innovate faster. I think we'll get to better quality of life and improvements for everyone on every different vector.

(Karthik Ramgopal at 00:43:49) Yeah. I think it is, you know, a little bit of a value creation renaissance for sure. I think a lot of some of the constraints have been released. But I think these things go in cycles too. Right?

(Karthik Ramgopal at 00:44:04) So, you know, the other part of this transition is that companies need to figure out how to adapt their business models from web mobile to web mobile and now personal agents. And maybe with those adaptations come... Certainly with those adaptations come incentives to create a ton of value, but also business models that arise that may introduce new constraints that we don't even really... Can't even really predict or understand because they've got to be innovated on in the same way that the advertising business needs to figure out now, okay, what does an advertising business look like in a personal agent world? And so that adaptation is going to come.

(Karthik Ramgopal at 00:44:42) And so all new patterns of the way that you do, you know, selling and purchasing and advertising, that's all going to change. And so I think that these things all, in some ways, are just cycles. And, hopefully, you know, we get past... We don't get... Constraints don't get reintroduced that then curtail value.

(Karthik Ramgopal at 00:45:01) But I think, ultimately, that's part of what needs to be figured out.

(Joel Beasley at 00:45:05) I love it. I love the energy. I love the... You know, at the bigger companies, it's rare when I get really, really brilliant people with lots of energy on their teams. And everyone that I've got to meet from Mohak to Raghu to Kevin. Did you know Kevin Scott? Of course. Because I know he...

(Karthik Ramgopal at 00:45:21) Okay.

(Joel Beasley at 00:45:22) Yeah. Yeah. Everyone I've gotten to meet over there just... Do you still interact with him at all or is he... Because he's over at...

(Karthik Ramgopal at 00:45:28) He's still...

(Joel Beasley at 00:45:28) ...Microsoft?

(Karthik Ramgopal at 00:45:29) Active on... He was one of my first bosses. He actually... I reported to him before I reported to Mohak, and we're still in touch with him.

(Joel Beasley at 00:45:36) Oh, did you?

(Karthik Ramgopal at 00:45:37) Yeah. He's got an awesome, you know, personal IG that he... Where he's like, it's not about tech. It's about his personal hobbies that's super interesting. So...

(Karthik Ramgopal at 00:45:48) Shout-out to Kevin if... I don't know. I... You know? Yeah.

(Karthik Ramgopal at 00:45:50) You should reach out and follow him if you're interested in tech people doing craft and making things with their hands and not in tech. So...

(Joel Beasley at 00:45:59) That's so great. Yeah. We... He came on years ago to help... We helped him launch his podcast that he has.

(Joel Beasley at 00:46:06) I know he's still doing it. But I don't know how he is. Behind the Tech. Behind the Tech. Yeah.

(Joel Beasley at 00:46:11) Yeah. It was great. So that was... It was fun getting to hang out with him. As we wrap up, just one thing on team building and then a leadership question. Is that okay with you?

(Karthik Ramgopal at 00:46:20) Yeah. Absolutely.

(Joel Beasley at 00:46:21) Okay. There was something in the notes about you guys moved to a pod mix instead of siloed teams. Can you explain the structure change on the teams and why you guys did that?

(Karthik Ramgopal at 00:46:31) Yeah. I mean, I don't think that... I mean, this is something that's become pretty commonplace, I think, across the industry now, which is realizing that with AI tools now, one individual can do so much more. And so by virtue of that, we can have smaller teams of engineers and cross-functional people working together and achieve more. And, obviously, with less people involved, I think that's less overhead.

(Karthik Ramgopal at 00:46:57) And so teams move faster. And then secondly, functions are blurring. And I'm sure you've had people on this podcast, and I listen to lots of podcasts, and everybody's talking about, hey, well, because of the AI tools, you know, a product manager can submit a PR, a designer can submit a PR, an engineer can draft a PRD. And of course, there are still, you know, what I think are custodians of crafts who are real specialists who ultimately have a sense of what great looks like in every functional discipline. But those disciplines are blurry, and people are kind of spanning more of the kind of end-to-end idea-to-production life cycle.

(Karthik Ramgopal at 00:47:38) And so that's the basis of these pods. You have smaller teams who are cross-functional, but expectations that every individual within these teams are kind of reaching out of their... Out of what would have been their traditional lanes and shipping faster. And we're seeing pretty awesome results from that. So I think that's kind of the new normal, so to speak.

(Joel Beasley at 00:48:02) Exactly right. And it makes it even... It amplifies it to make it a little bit even murkier when you start talking about... Like you said, you listen to a lot of podcasts. A lot of people are talking about this.

(Joel Beasley at 00:48:12) It's like, okay. Yes. But also, it means different things at different levels. If you're at a healthcare company making heart valve pumps, that's a different conversation than if you're at a startup on day one with SaaS or if you're on an enterprise B2B where there's billions of dollars of revenue. So I always tell people when they ask me stuff...

(Joel Beasley at 00:48:28) Remember that you're listening to this show and we're talking to this massive spectrum of leaders. Find people that are close to your lane so that when you're taking their advice, it's actually helpful.

(Karthik Ramgopal at 00:48:38) Great call. Great call. It's totally different to be a company operating on our scale versus a small startup, for lots and lots of reasons.

(Joel Beasley at 00:48:49) The single biggest mistake you've seen another engineering leader make, not necessarily at LinkedIn, just like in the open world, when trying to move fast with AI. What's that mistake?

(Karthik Ramgopal at 00:48:59) I don't know, single biggest mistake. I don't know if I can point to one thing. But I think that there's a little bit of a false premise right now that pushing really hard on AI transformation is somehow at odds with what's good for your employees.

(Karthik Ramgopal at 00:49:21) And what I mean by that is, the future of building products and doing software is with AI at the center. And by pushing our employees and being bold about adopting these tools and reimagining how we work is both good for the company and also good for the employee because it prepares them for what software engineering is going to be like in the future and what the jobs are going to be like in software and product building in the future. And so I don't know if it's a mistake, but people hold back a little bit or are worried about, am I pushing too hard? Is this making people uncomfortable? Is it creating anxiety?

(Karthik Ramgopal at 00:50:02) And all those things may be true, but I think that the conversation with our employees can be, hey, this is an investment that we're making both for the company and for you, and we're going to push you really hard, but we're going to support you. And hopefully, you wind up being at the cutting edge of what it means to build and ship software as a byproduct of how hard we're pushing on transformation.

(Joel Beasley at 00:50:33) And of all the leadership advice you've ever received throughout your entire career, what's the one piece of advice that happened, you tried it, it stuck with you, and you've kept it long term?

(Karthik Ramgopal at 00:50:45) So this is a Kevin Scott one. It's funny that his name came up. I remember having a one-on-one with him. I was pretty early in my senior management career, I think. I was dealing with some super stressful situation, which probably in retrospect was trivial, but felt really big at the time.

(Karthik Ramgopal at 00:51:05) And he was like, listen, you're going to have a long career in running organizations, leading people. And the most important thing to remember is this is going to be a roller coaster. There are going to be some moments where you're on top of the world. You just shipped this product you've been working on forever and it's crushing it, or you made some breakthrough. And then there are going to be times where you're in the deepest, most depressing, stressful moment.

(Karthik Ramgopal at 00:51:32) And the key to having a successful career and a long career as a leader is not getting too high on those highs and low on the lows and knowing it's a roller coaster. And actually, as you get more senior, the highs are going to get higher and the lows are going to get lower, and you just got to stay kind of in the center. And if you can be mindful and conscious of that, you're going to be both more successful. The way that you lead will be more authentic. You will inspire more confidence, and you will probably sleep better at night. And so that's something that's definitely stuck with me, and I think it's totally true and something I now share with my leaders in my organization.

(Joel Beasley at 00:52:19) Well, it's brilliant advice.

(Karthik Ramgopal at 00:52:21) It feels very obvious, and it's actually super hard to do. So, you know.

(Joel Beasley at 00:52:26) It is. Because you want to go—you want to get high. When it's going up, you want to go crazy. But then if you do it enough—so that's the thing. That's the beauty of the experience. You have to do the roller coaster enough to realize I'm going to step off the roller coaster.

(Karthik Ramgopal at 00:52:40) Totally.

(Joel Beasley at 00:52:42) That's the experience. That's the wisdom, right? And then every once in a while, you do a loop-de-loop. You're like, oh, that's right. That's right.

(Karthik Ramgopal at 00:52:48) Unavoidable. Human nature for sure.

(Joel Beasley at 00:52:52) Oh, this has been fantastic, man. You have made my day. Today is a great day. Thank you so much for doing this.

(Joel Beasley at 00:53:03) And if you found this episode useful, please share it with a friend or colleague who you think would get value from it. And if you have topics that you'd like to hear discussed on the podcast, either add me on LinkedIn or send me an email: [email protected]. Every time I get an email or LinkedIn message, it absolutely makes my day and inspires me to keep going.