Garrett Lord. Handshake.

Garrett Lord:

I remember calling Mamoon and Margo and Will and Megan on our board literally over the holidays. I laid out a whole investment thesis. We wanted to be able to target 25 or $50 million of run rate revenue. That seemed insane. I remember asking Will Reed, "What do you think good looks like in a year?" And he's like, "I think it's probably 100 million bucks." And I'm like, "Dude, you're absolutely crazy." I mean, we're starting from zero. We have one lab relationship. And now we've gone zero to a billion in a year, we're probably one of the fastest growing AI companies out there. The opportunity in front of us is so big that it just requires incredible execution. How can we be more ambitious? How can we turn around a month plan in a week? That's basically my entire job. I spend 0% of my time celebrating any of our success.

Nakul Mandan:

Garrett Lord spent 10 years building Handshake into the career network for an entire generation. 18 million students and alumni, 1600 schools. Almost every company in the Fortune 500 is a customer, and it's a company valued well north of $3 billion. Most founders would spend the next 10 years protecting that. Instead, over the course of 2025, Garrett did something almost no founder does. He refounded the company around a business that was barely a year old. Handshake AI went from zero to a billion dollars in gross revenue in 15 months. And to get there, he did layoffs, reset the entire company and bet the thing he'd spent a decade building. Today we get into the why and how of this transformation and how Handshake is shaping the future of AI. Knuckle Up. Garrett, welcome to the show.

Garrett Lord:

Yeah, thanks for having me.

Nakul Mandan:

So Handshake started in 2014 as a LinkedIn for college kids, and over the decade ahead you built it into a $3 billion business. Yet in 2025 you decided to transform it completely into a new thing, Handshake AI. Let's start with the genesis of Handshake AI. What were the early days like and what gave you this hunch that you need to go in this direction?

Garrett Lord:

It all really started at a Christmas party where I was talking to a researcher at one of the Frontier Labs. And so much of being an entrepreneur is I think trying to talk to customers, understand trends in the marketplace. I had seen quite a few recruiting companies really focused on AI and really serving the labs. And I was particularly curious. I mean, I think that's one of the main attributes of a lot of entrepreneurs, is just trying to figure out what's going around you. I used that opportunity at the Christmas party to dig in with this researcher. It was actually kind of funny. And normally at a Christmas party you meet 25, 30 people and I basically talked to one person all evening just trying to understand a little bit more about what they were recruiting people for. And what was clearly articulated from him was that there's really three stages of model building. There's pre-training, which is scraping the internet. There's mid-training, which is called supervised fine-tuning or feeding in high quality samples. And then there's post-training. And post-training is really around creating environments where models can learn from reinforcement learning, really codifying a lot of the knowledge and judgment that wasn't on the internet. It became clear in my conversation with him that obviously there continue to be continued gains in pre-training, but a vast majority of the gains that were going to be had in modeling was going to come from digesting what was in people's heads into forms of fuel in reinforcement learning that models could learn from. And he was articulating this kind of massive build out of post-training the amount of data that is needed, and really kind of summarized the conversation, like there are three stools to AI. There's algorithms, there's compute and there's data. And so I actually started our relationship two weeks later with that same exact researcher at that top tier lab where we started working with them on recruiting STEM PhDs. And so-

Nakul Mandan:

Just put a timeline to this. This was December '24?

Garrett Lord:

December of 2024. Yeah, exactly. Yeah. So much of being an entrepreneur is building a small strike team, taking some of the best people in the company and starting to explore this idea. We started to reach out to a bunch of other relationships they had with other labs and our real focus was on serving the top two model builders. We thought that if we could serve these top two model builders with these frontier capabilities and kind of evolve with them, that that was really an indication of what the future would look like. And so it started off kind of as a small, exploratory strike team and very quickly by January of that year, this was my full-time job at the company.

Nakul Mandan:

Were you looking in that moment or at that time, did Handshake, the business that had grown to a $3 billion plus value business, 250 to $300 million of revenue if I'm correct, were you looking for a new wave, a new wave to attach that business to? Was it slowing down or did this opportunistically come about as an extension that, hey, we have this network and we are very well-equipped to do this?

Garrett Lord:

We were serving a lot of the key players that were serving the model companies. So we were basically the number one recruiting source for some of the competitors that produce data in the ecosystem. And so really what does that mean? That means that we had the talent and the network. They trust Handshake, they use Handshake, they have it on their phone. And these companies were messaging talent in our network trying to hire them to do these type of gig jobs producing data. And we were hearing from our students that they weren't being paid on time. They thought it was kind of a messy process. And what became clear with the researcher is just how much more you can improve the experience and make this about something that people can learn AI skills, people can monetize their intelligence. And $55 an hour for a college student was like... You're not going to drive DoorDash and make $55 an hour. And so I really thought that we had this massive moat. In the human data industry, really, the supply chain is like you have to have access to an audience, and just economically you have a cost of customer acquisition, and then you have to vet the people, teach the people, and then put them onto your own platform. At the beginning, we had the people. We were vetting them and we were training them, but we didn't have the actual platform to be able to deliver the data. And so I thought that we could vertically integrate and move up the stack. And it's turned out, I mean, that advantage of having access to an audience, it's not just about... Some people would say it's about the economic cost of customer acquisition that's our advantage. And 89% of people that worked on Handshake AI last month came from the network. So obviously there's a massive customer acquisition advantage, but the bigger advantage is actually we're able to target the highest quality people. So a lot of the shapes of data today are focused on accounting or law or medicine or really the largest professional white collar domains in the country. And we're able to attract much higher volumes of high quality people. Say you want to target accounting, for example, like EY, KPMG, Deloitte, Accenture. These are all paid customers that spend a million plus dollars a year on Handshake. And so when we wanted to target accountants, we were starting to ping people in advisory, in tax that had five, six, seven, eight years of experience, that's big for accounting firms. And so we were able to turn around on the highest quality people much, much faster, obviously at a structurally lower customer acquisition cost, but that matters a lot for data quality because one of the primary inputs for data quality is like, what is the unique knowledge that person has and how good are they in their actual profession? And so that's actually the main captive advantage of Handshake. People think about Handshake as an early career recruiting network. And yes, we're the number one place that young people in America find and search for jobs, but we also have tens of millions of profiles of alumni in basically, we need to find farm supervisors. Turns out there's these large land grant institutions like Michigan State and University of Illinois who have a farm supervision and agricultural program. And so we're just able to structurally access really, really high quality talent and they trust Handshake. They used it in school, they had a profile built. So that was kind of how we thought the ecosystem.-

Nakul Mandan:

But did you feel it was a need to have or this was truly an extension of what you already had?

Garrett Lord:

Oh, it was a pretty radical pivot. I mean, just in terms of where we were operating the business, the old metrics of the business and still in a core SaaS business, you have top-line revenue growth, you have NDR, you have overall productivity rates, you have gross margin, you have contraction and new sales for existing customers. The way you run that business on a quarterly and monthly and weekly cadence looks radically different than how you run a human data business, and it looks very different than serving large enterprises with AI solutions that we're building for them as well. And so it took, on my side, I talked to a couple mentors and other CEOs and basically the consensus viewpoint... I love learning from others. I think that's so much of what makes Silicon Valley amazing, is being able to really call somebody up and ask for advice. But I started basically a separate team completely outside of the existing company that reported directly to me. And then over time we kind of grew a separate part of the office with a much different operating cadence and structural requirements to serve customers.

Nakul Mandan:

How early did you know that this was going to be a rocket ship and this needs all your attention as a CEO?

Garrett Lord:

From the moment of the holiday party, it became clear to me that although we were doing tens of millions of dollars of revenue by serving these human data companies, on the backside of that tens of millions of dollars revenue, they're making hundreds of millions of dollars revenue. And so we were just missing out on the entire LTV of what they were able to capture. And then what the researcher really described is, what he defined as this explosion of the amount of data that's needed to serve models. He kept it really simple in that he used this analogy that the reason that software engineering productivity gains have exploded and the job of software engineering has changed is because there's GitHub. There's tens of millions of human hours went into artifacting code bases. And so that's the perfect fuel for reinforcement learning. It's verifiable. All of the context is encapsulated in a single code base. Even more than that, an actual other software engineer has to review your PR, which is another way to generate rewards. Then you have unit tests. And so that productivity gain you got in software engineering was not going to end up happening in other white collar knowledge professions because there isn't a Reddit forum or GitHub equivalent. There's no Reddit forum of tax accounting in America. And so he was describing... He really recommended from the onset that we start working across multiple data types. So much of being an entrepreneur is like, what's your insertion point where you have a competitive advantage? And then how do you quickly bundle and work to other areas? So our insertion point was really around, what's called STEM reasoning or PhDs helping models in biology, chemistry, mathematics, physics. We then two months in the business then started moving to generalist domains, and that's multimodal, video, audio, personal shopping, personal finance, consumer tool use, things like Strava and other consumer apps. Then from there we moved into professionals because we had this massive moat of young alumni and all the way up to 34-year-olds in America where we were able to target people in... Basically we power 92% of schools in America. We basically powered the entire graduating class. We also power 60% of community colleges in America. So we were able to target basically every single profession. So we moved to professionals. Then we quickly spun up, what's called a safety team and that helps model builders make sure their models don't build bio weapons or say bad things to children. So we have a safety team. We have a robotics team internationally. So we basically work systematically across the different data types because inside these labs, obviously you're customer-concentrated. We now work with basically every frontier model builder today. But inside of these teams, they have a lot of different teams focused on different objectives. They're basically trying to hill climb in cancer therapeutics, or they're trying to hill climb in new material science discoveries where another team's trying to focus on better image generation or better conversational voice. And another team's focused on how do I actually call Bloomberg and FactSet and all these professional tools that are using this in finance? And so by working with the lab and building trust and producing high quality data, you can kind of move across the different pillars of the data that they're working on.

Nakul Mandan:

So there's a whole conversation we'll have on, everything you just described happened literally 18 months ago to now. And so how do you move at that speed? But before that, I want to also talk about like, this is not an easy change that you pulled off. You have an existing business, you have execs there, you have a full team there. How did you start communicating this to the team that was running that? Was that early in January itself or that happened over February, March, April? How did the board conversation go? Start with the execs. When did you start communicating to your team, "Guys, this is big and we are going full steam ahead on this."

Garrett Lord:

I think the story behind every one of these great companies is like, you have to have an amazing team. And I've just been working with a lot of these executives for, generally speaking, if you come work at Handshake, it's either not going to work and you're going to leave in a month as an exec or we're going to work together for years and years. And so got this gentleman, Jon Stull, who's our president and COO, Matt, our CTO, basically I started having conversations with them around, "Hey, I really think I need to start punching out of a lot of the core operating streams of the business."

Nakul Mandan:

That was in January?

Garrett Lord:

That was in January of that year. In December of that year, I started having those conversations. I remember calling Mamoon and Margo and Will and Megan on our board, literally over the holidays over Christmas, New Year's, and being like, "I want to make some of these amazing offers to new employees." I laid out a whole investment thesis, because so much of getting the budget to approve this was telling a story around what this could become. And fortunately, I think I have just an amazing board. So people were super on board. I think I started off saying I wanted to spend five or $6 million on this new business venture. And I said that by the end of the year, I think we wanted to be able to target 25 or $50 million of run rate revenue. And that seemed insane. And everyone on the board I think was like, I remember asking Will Reed, "What do you think good looks like in a year?" And he's like, "I think it's probably 100 million bucks." And I'm like, "Dude, you're absolutely crazy." I mean, we're starting from zero. We have one lab relationship. And we ended up in a year going for-

Nakul Mandan:

Like a billion.

Garrett Lord:

Yeah, a billion. So that's been the story of this journey all the way through, is my core executive team focused on running the core business, which is a ton of work. That's their day job. And in the first two quarters of the new business, I basically spun up a totally separate team that reported directly to me.

Nakul Mandan:

Was there pushback from your exec team? Was somebody worried, "Hey, Garrett, you're distracting yourself with something. We have a great thing going here or that landscape is already competitive and commodity," which we'll talk about the defensibility, differentiation, all of that, margins. Was there pushback at all from your team or the board?

Garrett Lord:

There was no pushback from the board. There was definitely pushback from the internal team. I think that, and this is a story of talking to a lot of CEOs, is incubating something new and different inside of a business is hard. Fortunately though, I had a lot of trust. I mean, you can frame this in a way that was like, "I'm going to go try this for a quarter and if it doesn't work, we're going to be just fine." And fortunately, our core business is very profitable and so there was plenty of money to go around. But yeah, I think it was a big uphill climb. I mean, I basically asked to assemble some of the strongest people in the core business that were being pulled from existing initiatives that were driving the core business. But I think you have to have courage, you have to be comfortable in ambiguity. I recognize I was getting asked a bunch of questions around like, what's our competitive advantage? What's our long-term moat? How is the market going to evolve? And it's like, I think every good entrepreneur, you just start like... I don't have the answers, right?

Nakul Mandan:

Yeah.

Garrett Lord:

But I'm going to start breaking it down. And so most of my week and months in the early business was just going on coffee chats and walks, basically calling every single person I could through my extended network. I went through every connection I had on LinkedIn and begged, borrowed, and stealed, for any friend they knew that was anywhere adjacent to the labs. Just try to learn, learn, learn, learn, learn. And every week in December and January and February of that year, it became extremely clear that this audience and the structural advantage we had with trust with tens of millions of people around America, knowing their grades, knowing what they did in school, knowing what they're doing after school, was this massive structural moat that [inaudible 00:16:05] accelerate. And that's really what we've done. I even make a joke, I mean, I couldn't spell reinforcement learning when I started this business. That was probably one of the more intimidating parts of this is like, what does it mean to post-train? What does it mean to mid-train? What does it mean to do evaluations? What does it mean to do an ablation? These are all terms I had no idea of. I mean, I think the story for other entrepreneurs listening is just like, I put my pants on the same way everyone does. Either left leg or right leg first. Everyone out there that's created anything started from the place that you did where you don't know what you're doing and you figure it out. And I felt super confident in our ability to go figure it out, especially once I realized just how deep this trust that we had built with tens of millions of people, how big of an advantage that will be for our business.

Nakul Mandan:

How did the communication with the rest of the team beyond the exec team go? Was there a all-hands at some point where you talked about this more openly? How did you manage the communication through the rest of the team?

Garrett Lord:

Yeah, I over time as an entrepreneur have really relied more on my team to do most of the communication. I mean, obviously I talk in all-hands, but I personally don't love being the one that's taking a lot of the credit and doing a lot of the storytelling. I mean, obviously I'm happy to tell the story at all-hands, but I really love the executives that run these teams to be the ones that are really driving the strategy and telling their managers and leaders and to extend to the leadership team what's important. So yeah, the two things I did very early on was like, we have an extended leadership team of 35 people. And so I basically briefed the core directors, senior directors, VPs, and C-level leaders, and here's what we're focused on. It also was not that cannibalizing to their existing core business. They had plenty of resources to be able to achieve what they wanted to achieve. But over time it became much, much more challenging to integrate these two businesses. At some points in this journey, we've turned away $15 million of work a month. We just cannot scale up fast enough. And so I think about mid-year in 2025, we had to make the call across the leadership team that we're just going to move over a way more resource from the core. Our entire recruiting team used to balance recruiting across the entire business. And it became a point where I was like, "Hey, we no longer can serve anything other than the AI business because we can't be turning away $15 million of work a month."

Nakul Mandan:

How do you keep A players motivated in the core team then or in the marketplace business?

Garrett Lord:

The key thing, and it's always hard to nail this as a founder, but there's just one story of Handshake, which is that we have a network and people come to this network to trust us to find a job, to learn the skills and up level, re-skill themselves to be ready for the jobs in the future. And then you also can come to Handshake to monetize your intelligence and knowledge and judgment in your profession. So there's a couple other benefits of using the network, like managing your professional identity and getting recruited by companies. But this one network, we wanted to drive, I really believe that we could just drive more reasons to engage with Handshake. The real durable moat of Handshake, because we have a million companies using it. We now have basically every model builder and now a lot of enterprises that are using Handshake's network to help improve their models and evaluations. And then we have a huge up-skilling and re-skilling business that's helping people learn the skills of the future, which I feel like is really, really important. A lot of people are super down on young talent. And I think that narrative is, there's a total bifurcation of the market on what people think. Some people are like, early talent jobs are going to go away. And other people, like Cloudflare, for example, Matthew Prince, he's doubling down. He hired 1,100 interns. He's seeing the fastest people inside of his organization pick up, they are actually young people. And so I want to equip the next generation of talent to be successful in this future. I don't think this whole narrative of we're all going to play board games for a purpose one day is a vision that I want for my kids and for tens of millions of Americans. So we want to help up-skill and re-skill and really teach people how to be ready for the jobs of the future. And then we are the largest network of young people finding jobs in the country. So basically all of higher education depends on us and then 100% of the Fortune 500 uses us to recruit. So this network, ultimately telling that story I think was confusing for people internally. They didn't understand what it meant to produce data for labs. They didn't understand the connectivity between how when you're producing data, are you learning skills? But fast-forward a year into this business, a little bit over a year, you can come to Handshake to learn, you can come to Handshake to find a job, you can come to Handshake to manage your professional identity, you can come to Handshake to monetize your intelligence. And our network's more than doubling this year and week on week utilization is over 50%. So the network's on fire because there's so many more reasons to come to Handshake.

Nakul Mandan:

So would it be fair, actually it doesn't feel any more like two different businesses and more like an additional revenue line item that has exploded, but the core elements of the product, which is the network and others have only, of course they've extended to RL environments and all kinds of post-training and mid-training offerings, but it still is built on the unfair access to the network you have. And so does it feel like two different businesses today or does it feel like multiple revenue line items?

Garrett Lord:

I think LinkedIn's a great analog. They have a network and they have marketing solutions, they have sales solutions, they have talent solutions, they have LinkedIn Learning. They actually have many different revenue lines on top of one network. And people come to it for different seasons and different reasons to use the network. And they also have a social graph. I mean, Handshake is very analogous to LinkedIn. We have a network and we monetize it by finding a job. We monetize it by helping you upscale and re-skill. We monetize it by helping produce in human data for labs. But if you take a great analogy, one of the world's largest healthcare companies, or in America, the largest healthcare company, or one of the top big four accounting firms, they're using our network to bootstrap their evaluations internally. So that insertion point in enterprise, that's also a huge monetization pillar as well. And so we think that all these businesses together are... We think that all these businesses together are, there's more reasons to engage with Handshake and get more value out of it and that's ultimately what users care about. I mean, I think this year our network will probably double, I mean, it's on track to double. I think by the end of the year we'll probably triple the size of the network. And then just the engagement metrics are so much more healthy because there's more reasons to come.

Nakul Mandan:

Did you have to reset the cultural elements at the company? Because this business that has grown to 250 million plus of revenue, it's a different kind of a beast. You're starting with something that is small. Within six months, it becomes a beast of a business, and then by the end of the year it's at a billion. The speed, the business model. Did you have to reset the culture, or it was almost for a while felt like two different cultures within the same company?

Garrett Lord:

I mean, it felt like two different cultures in the same company. If you look, if you just zoom back, Handshake was deeply uncool, then it was very cool. Then our growth slowed down immensely. And now we've gone zero to a billion in a year. We're probably one of the fastest growing AI companies out there, [inaudible 00:23:11] the investor. But what was required at the beginning is we definitely extended a lot of our core values. We have a principle called students first, which is in balancing a marketplace, you have to have one side that you prioritize over everything else. And so very early on, Handshake AI, one of the founding principles was fellows first. We're going to take on work. We're treating people really well. A lot of these human data businesses started out of basically the advent of labor arbitrage where there's international low cost country, talented folks that are basically drawing boundary boxes around stop signs. And it was very hard for those companies that were drawing a boundary box around stop signs with low cost labor to treat an accountant the way that they deserve to be treated or to treat a student the way they deserve to be treated. They were treating them almost like body shops. That we just felt like that was just not really ingrained in the way that we wanted to build this network. Fellow retention, fellow NPS, basically all the aspects of the fellow journey are something we tracked from the very, very beginning of this company because we believe if you treat people the really right way, this is just one stool of the network you're coming to. You're coming to find jobs, you're coming to upscale, you're coming to monetize your intelligence, you're coming to help work with enterprises. And treating people poorly was the wrong choice. But yeah, I mean I had to keep it real on what transpired internally. We went from a fully remote workforce with really primary hubs in San Francisco and New York, Denver, London, Paris, and Berlin to five days a week in office across two offices, New York and San Francisco.

Nakul Mandan:

Was it always remote first even before COVID, or COVID had forced you in that-

Garrett Lord:

COVID forced I think almost everyone remote first. We tried to focus our hiring on hiring around hubs so that we could bring people back to the office. But yeah, there was a massive initiative internally to get people all in the office five days a week. I mean, I just so believe in... There's no clipboard. So if your kid's sick, or you're not feeling well, or you have a hair appointment or a doctor's appointment, work from home. We're not going to be checking every single day. But the expectation internally is that we're all here pushing super hard five days a week, and you're collaborating with peers and you're part of this high energy culture. I made that call really midway through last year where it was like, this is the Super Bowl of our business. There's such a large opportunity to help people and to grow this business, and we got to be in the office five days a week. And that was not a decision that people liked. There are probably no fewer than 100 people that decided to leave the business. Attrition was very, very high as folks were trying to resettle in the DNA. We codified a value internally called Olympic pace, which was I think you have to be opinionated about how you codify your culture. And Olympic pace for me meant like Michael Phelps, ahead of the Olympics, was in the pool every single day for five years. I believe in stacking days. When you come in every single day, some days you have a 70% day. Some days you're on fire. Some days you had an issue with your kid and you're distracted. But if you just compound and stack day after day after day, and give it your best, great things will happen. That's been the story of my life and the story of this business. And so Olympic pace was a core value, and we're going to be in the office five days a week. And I think a lot of people didn't like that. And I empathize with that. I think so much of what was challenging for me as a leader in navigating that was like it's kind of unfair to change the expectations of the culture after you've joined the company. I recognize, I mean, that's a fault on me. I think if I were to do a retro on myself, I think I wish I would've codified the values of the organization earlier so that people understood what they're signing up for. But now we make it super clear. And I think as long as there's no surprises, the message that you're telling to the talent market, the folks you're interviewing, they understand what the expectations are, then it's a super fair trade. You're choosing to join this company for these values and for this mission. And that was an important part of change.

Nakul Mandan:

So let's talk about the speed at which you grew because this is a ridiculous amount. I think no business in history may have gone this fast, zero to a billion in one year.

Garrett Lord:

Certainly Anthropic and OpenAI, but yeah.

Nakul Mandan:

Anthropic and OpenAI. But you're in an elite tier. What started breaking? Can you walk us through last year of all the problems you had to face in terms of just scaling this at this speed? What started breaking first?

Garrett Lord:

I mean, I think we've hired 400 people, 300 people in maybe eight months, six, eight months. I mean, the number one problem in our company right now is we cannot meet the demand that the market's serving for us. We've cared deeply about our reputation and doing things the right way, treating fellows right, not over-promising the customers, paying people equitably internally. We try to do things in a focused way. I've just been through the growth at all costs, but yeah, I mean so many problems. I think the main way I think about solving problems as a leader is it's almost kind of fun to zoom back. I got to spend basically a decade learning how to build a business, and going through all the growth as a CEO and reinventing yourself, and maturing and understanding your blind spots and figuring out how to augment them. I kind of got to do that for a decade. And so now at 36 years old, being able to run a second chapter and run it all the way to the Super Bowl has been awesome. I still make a ton of mistakes as a leader, but I get to really learn from that pattern recognition. And the main thing I've learned, long-winded way of saying, is just so much of how I want to build this business is focused on hiring incredible leaders, and really building that extended group of leaders internally. We have 25 prior founders inside this business. We could have never built what we built without these founders. We could have never built what we built without hiring super high agency young folks. We believe a lot in growing talent internally versus hiring. That's one of the big lessons. I used to hire shiny executives. And the reality is I've done far better in homegrowing talent and giving people opportunities to step up and coaching them. I think the reason we have executed so well, and come from not being in this business to now being the number two player in the space, is by hiring incredible people. So I kind of systematically look at where do we want to be in six months from now, a year from now, two years from now? Really, it's hard to look two years, really six months and a year from now because things are on fire. And what are the capability gaps or problems that we really need to solve? And then I like to think about problems of red, yellow, green. What's red today but will be green in six months? And I try to not focus on those problems. A lot of the team wants to flag issues to me where there's a fire, and I'm like, red, yellow, green, is it going to be fine in a quarter? Is it going to be fine in two quarters? If it's going to be fine, if you guys have the ball, I'm not going to spend my time there. I'm going to try to focus on things that are red today that will be red in the future.

Nakul Mandan:

Give me an example of something that six months ago was red then and felt like if I don't step in, it'll be red again in six months.

Garrett Lord:

One of the biggest ones is international. We opened an office in Bangalore. We had a ton of talented people working in India remotely, and we were like, "This is insane." The level of intensity and passion and technical skill that's over there, the opportunity to work in a high growth startup where you're having a ton of ownership. I saw that very early on in this journey, and was like, "This is an absolutely wild unlock." But it's very hard to spin up a totally different team in person all around the world.

Nakul Mandan:

It's a different cultural sort of workbench.

Garrett Lord:

It's a different culture, totally. And so I focused with two of the leaders internally on how do we... And by the way, resetting teams to change their operating rhythm when you add in somebody, a team that's located all around the other side of the world, that requires a ton of force, just thoughtfulness and effort to get right. So we have gone from zero to 100 people in India in probably two months. And I think that capability as an organization is incredible because we basically can work 24/7 all the way around the globe. We have fellows all the way around the world. I mean, we need to hire now workers in Germany, in France. There's different cultural preferences when it comes to different shapes of data that we're producing. And also we didn't want to set up India as... A lot of companies set it up as this offshoring operation. That was not our ethos. This is going to be a completely equal team, no matter what office you're a part of whether you're in New York, San Francisco, or Bangalore. And you're going to work on high agency projects and impact this business. And we have seen an unbelievable success, but that was not something that we were just going to stand up without forcing it, completely punching it in. That's I think a great example of one.

Nakul Mandan:

How do you ensure quality control at this speed of scaling in terms of the output that you're providing to your customers and the internal processes? So many people at Handshake are probably barely five months in. How do you ensure the quality control is going all through those 400 people? Is it culture? Is it just daily communication? What does it actually take to ensure this?

Garrett Lord:

Onboarding Handshake was just like grab a laptop and sit next to somebody that knows what they're doing, and you just jump off the deep end. But that doesn't work when you start hiring hundreds and hundreds of people. So we had to build. We have two incredible leaders internally that focus on basically have built a bootcamp to make the skills, 80% of the skills that are required to do this job, at least you understand what you're getting yourself into and how to upskill yourself to be able to run these queues. So the training and learning program where you have dozens and dozens of people going through it at one time is absolutely critical. You have to codify a playbook. I think it's unfair to expect that every single employee that you hire is just going to be able to figure everything out on their own. And that doesn't work when you start getting hundreds of people and it's a more matrix organization. So building a training and learning program internally. Luckily I had seen that before in hiring a sales team. We had built a sales enablement team, and there was a bootcamp and QBRs. So I've taken a lot of the learnings out of the original business into how we speed scale a team. And then I think culture's huge. I think culture's really made in the hard decisions. There's only so much you can do to write it down, but you really have to role model it and incentivize people to role model it and talk about it in reviews. And I think it takes, when 85% of the organization is new to the company in the last six months, our culture really codified when we had the entire group of the first 50 or 80 Handshake AI people really hit eight months, 12 months in the program. And they had seen the way they were worked, they had seen the decisions that were made, and they started to represent and reflect what we expect out of the way that we're going to treat each other and treat our customers.

Nakul Mandan:

You've touched on hiring a little bit. So let's talk about that. How do you now ensure the hiring bar is high? Because the people you hire, you can give them good onboarding and training, but ultimately the people quality becomes the culture or direct input. So how have you ensured over the last year to ensure that the hiring bar is still high, and yet you are hiring this much in capacity of people that you need? Because you're probably constantly feeling behind on hiring.

Garrett Lord:

Yeah, I mean I think the story of hiring is, I'm maybe going to say something provocative. In Handshake, I've never worked with such a talented group of people. I think one of the important things I think I've learned in building this company is hiring people that you love to work with and you feel like you're learning from. I mean, the group of people that I get to work with at Handshake right now are, I think everyone that's here is kind of like ride or die. It's a group of people that have just been through such deep lows and such high highs that the shared experience of working together. And I feel like people feel super respected. Those relationships are what has gotten us here today. Folks like Young and Sahil and Matt and Vee and Stull. I mean, there's so many people on my team. I mean, I wish I could go on with just a name game here of people that have just laid it all out. And it's required. It's required laying it all out. The number of times that I'm in the office at 1:00, 2:00, 3:00 AM, and there's 75 people here is hundreds of days. I mean, it is entire weekends, family plans, vacations. There's a guy named Jake on our team, incredible. He started this business [inaudible] that I work with where Jake had a vacation plan with his girlfriend. And it was really important. He had been working hard and he just is like, "I can't go on a vacation. We have to get this delivery across the line." And so people have given up a lot to build this business. And I try to role model. I'm like a blue collar kid. I feel super thankful to be here. This is the greatest adventure and awesome adventure. I'm loving it. And I try to drive with a ton of intensity. We have huge ambitions on what we want to accomplish as a business. But when it comes to hiring, I really trust the people around me to hire well. And I think that interviewing's super imperfect. A lot of people on Twitter want to talk about how this perfect interview process, and the first seven people you hire are reflective of the culture in the future. I kind of think that's a little bit of BS. It's so hard to understand. We try to hire from friends in the company. You try to hire through referrals. We have super elite talent that's applying because the story is super fast-growing, but you don't really know how somebody's going to do until they're two, three months in. And then I would argue that, inside of a business, a lot of young people don't realize, I don't think you really are putting up points until you've been here a year, till you really figure out the full matrix. And so you try to hire really great people, you have to trust your leaders. I mean, underneath organizations, we hired an amazing guy that runs trust and safety, and it's like, poof, he hired 15 people in a month. It's like, how can we even vet that? We have to trust people. And then one thing I really like is that we run the leadership team and all the operating cadences, the data reviews, the daily syncs, the weekly syncs, in a way where everyone's kind of allowed to comment. I like to run a super flat company. Right now I probably have two dozen direct reports, which is very different than the way I ran a SaaS business, which was like functionally on leadership team, monthly business reviews. Literally the pace of this organization changes daily. And then the pace of these labs, I mean the labs move. They are so passionate and work so hard and we work for them. So if they're working at 2:00 AM on a Saturday, you better bet on our side, we probably have a delivery supporting their work. In some of these darker, harder moments, I really like to remind myself why we're working this hard. And it's like both my in-laws have cancer right now. They're both going through chemo and radiation, and it's like these custom cancer therapeutics and MRI cancer vaccines that are developing, AI is going to accelerate that. My entire cancer journey where I go to MSK and Cornell and all these meetings, and have all the doc reports, I'm using a model. I won't name models, but I'm using a model to break that problem down. And my friend that's starting a business who is a ski lift operator who is a really passionate guy, he helps immunocompromised and disabled people learn how to ski. He is not being treated super well in his job and he wanted start his own business. And he's using these models to figure out how to start his own business and work through employment contracts and permitting. And it's like I think these models are helping so many people around the world, and we are one of the primary inputs driving model progress. And so I think continuing to remind people of why we're working so hard and why this is important. And then long-winded around trusting people, but I put a lot of love and trust in my team. And I also am not one of these people that just fire, when people are off track, I'm not just going to fire them. I really believe that you can coach and grow and love on and be a hard ass and push people to grow. And yes, people, you can't teach them some certain raw innate skills, but the story of myself is I'm imperfect in so many ways, and I've just had to learn how to adapt and be the right person to lead this company six months from now. So I really believe a lot in coaching, codifying what good looks like, building playbooks, meeting the different personalities where they're at, and believing in people. And then yeah, of course, do we fire fast? We fire super fast. There's an onboarding expectation, 30, 45 days in onboarding, but it's generally like we're going to fire you right away or you're going to make it and we're going to push you and we're going to help grow you to be the leader you need to be in the company.

Nakul Mandan:

On leadership itself, you're probably outgrowing some people very fast. Have you had to upgrade execs mid-flight last year?

Garrett Lord:

I haven't upgraded a single exec.

Nakul Mandan:

Okay.

Garrett Lord:

I think it's maybe a hot take.

Nakul Mandan:

Yeah, let's double click on that. How? You just have to [inaudible 00:39:52].

Garrett Lord:

I think that's the old playbook. Firing executives, I just think that's the old playbook. A, I think you run teams in a way more... The story of Handshake is we actually have multiple businesses out of Handshake. So we have a research team inside of that research team. They have many different mandates. They're building benchmarks, they're doing loss buckets on models, they're ablating on open source models, they're building public research papers for the labs. They're building quality tooling that makes every single person better at producing high quality data for the labs. Inside of STEM, or in professionals, we're across 180 different professional categories. You have labor all around the world. So it's like inside of Handshake, it's very entrepreneurial on the fact that probably 80% of the platform and technology gets you there on data. We have a 30-person forward deployed engineering team, if not more, that is customizing the work for each one of the labs depending on the shape of data that they're producing, building that into a compounding technology advantage internally. And so yeah, I mean, to answer your question in a more spicy way, I was just talking to some fancy execs. They get introduced, you're on fire, business, meet them. And I talked to them and I was like, they don't know anything about post-training. They don't know anything about data. They don't know about the modern ways of working. Most of our employees log into a coding harness every single day they do their job. Almost the entire [inaudible] operation, we believe that the way we're producing this data, we want to agentify the entire process. And so we have probably made every single person inside the company four or five times more effective than when we started this business. And systematically, we're just automating every single part of the problem we do. We have these forward deployed engineers that are basically codifying skills, MCPs, building their own harness, basically to help automate the work that we're doing internally, and put the human, because there's a lot of annoying work we have to do, to put the human in a more and more advantaged position. And so it's like, am I going to get that person into it?

Nakul Mandan:

Is it partly, and I'm also trying to grasp this, that if you have such insane product market fit, of course, talent quality matters a lot still, but if you have such insane product market fit, the talent also is able to grow to that faster because the talent is executing against the product market fit versus making product market fit happen, which requires a different level. And that's why people have to constantly upgrade. Because one would imagine, maybe just to double click, the person who was managing a FD team even nine months ago was managing a much smaller team. They're managing a much bigger mandate now. So typically the reason people upgrade execs is because just the management level management skills are different.

Garrett Lord:

Yes. So good point. So I would just say that in situations where people get leveled, the pie has grown so much that they're in a job that was 5X bigger than they thought it was going to be. So yes, was there a forward deployed engineering leader that ran forward deployed engineering and then our CTO started helping? Yes. But that person thought they were going to be running a two-person team, and is now running a 30-person team and it's getting split. And now they're responsible for forward deployed engineering on robotics and internationalization, and onboarding tens of thousands of people in countries with a totally new payroll stack and legal stack that what we have internally. And so yeah, the pie is growing so immensely. I kind of split up my team into core operations and zero to one. So we're constantly launching new businesses inside the company. We're moving, across each data type there's new what's called reinforcement learning environments, which is the longer shape of data that really is indicative of where the future's headed. You have the 70% of the portfolio is what we know how to do. And then 30% of our time, if not more, is spent punching in completely new businesses. Because we believe that in the future there'll be a lot of consolidation in the space. It used to be you could start a human data company if you grabbed 15 doctors, and you jump ball them to write hard prompts and they built verifiable answers or rubrics on the backside. That used to be sufficient. You no longer can raise $2 million seed round and build a human data company. The amount of work you have to do internally to be able to produce high quality, long horizon data is... So I believe that, we already see this today, we've bought eight companies in the last three months. We've basically been picking up a bunch of super talented teams that- ... we've basically been picking up a bunch of super talented teams that believe that when they bolt on their capabilities to Handshake, we can commercialize that and scale up plus they want access to our research, plus they want access to our platform, plus they want access to basically unlimited free fellow acquisition at scale. And so we've been on a huge acquisition spree, which is I had never bought any company up until Handshake AI. That's been a huge part of our ingredient. And then the 25 founders internally too, they're doing... But we don't manage people as tightly as you used to. You got to give people scope and rope and a big goal and then we're pretty real with people on where they need... Like data acquisition, this is an initiative internally that is insane now. I mean, we're acquiring hundreds of thousands of documents. We're buying data from businesses. We're building full simulated businesses and reinforcement learning environments where labs want to have basically an entire operating business that they can train off of. And that was a capability that did not exist three months ago. And now we have 15 people on acquiring data, synthetically enriching that data, anonymizing and de-identifying that data. It's like the pace in this company is, or the pace in AI alone is just, it's not SaaS.

Nakul Mandan:

And so your own job has also changed a lot over the last 15 months or 18 months, I presume. Where did you have to get better dramatically fast over the last 18 months as a CEO?

Garrett Lord:

Oh, getting a lot less sleep. I don't think I've ever... I mean, it's just one fun joke with the new child and then this business. No, but the serious answer, where and how to get much better? Learning how to operate a business like this, throwing away the old patterns and approaching things with first principles, leading a team of 20 direct reports versus six or seven. I mean, I spent all my time launching new things on problems that won't get solved naturally in the next six months where leaders are struggling to figure it out and on hiring. The number one way I can get leverage as a leader is hiring. I mean, I've never thrown myself harder into hiring the next generation of talent. And also because the business has gone zero to a billion dollars, the type of talent we're able to recruit right now, I mean, Handshake was a great business. Super impactful. It's a three and a half billion dollar company, doubling revenue every single year. We acquired pretty good talent. We have never been able to acquire the level of talent across all of our business, upskilling jobs and the labs business and enterprise business like we are right now because the best people want to work on generationally transformative businesses. And the progress we have made, you kind of can't recruit that level of talent early on. It was very hard to do because you have to take so much risk. But now with the revenue scale of the business, we're just able to get... More and more people are learning what we're doing.

Nakul Mandan:

Are there people around you, founders and CEOs who've helped you level up on hiring and 20 reports versus six reports and moving fast? Are there people you talk to when you are running into an operational issue?

Garrett Lord:

Yeah, I actually mostly talk to my team. I mean, I'm a student of the game. I love talking to advisors. I mean, great people are like Christopher Payne at DoorDash, who's prolific and incredible. I've got our board members that I talked to, Margo. I talked to several people that have run at scale businesses, but the vast majority of coaching, and maybe this is different than how I... Originally when I ran the SaaS business, I was relying on external coaches, "How do I build a demand generation events muscle? How do I create a category? How do I think about pricing and packaging? How do I think about what did LinkedIn learn in terms of marketplace attribution over time?" I was learning from them. In this new AI business, not a lot of people know what's going on. And the patterns that worked in the past are like-

Nakul Mandan:

Maybe they're anti-patterns now.

Garrett Lord:

They're anti-patterns, yeah. And so most of who I spend all my time talking to is the core leadership team of Handshake AI. And keep it super real. We're super honest with one another and are constantly pushing each other and arguing and trying to debate, getting to the right point. And I actually run... I believe one thing I constantly am doing is you need to have everyone understand the story of where you're going. So we very regularly have 30 to 50 person meetings where everyone is brought into the folds. I'm like, "Here are all the problems, here are all the goals. Here's why we're making what decisions we're making. Ask any question you want." But when we leave this room, there's not enough time for each person to describe what they're doing and debate it out. But when we leave this room, everyone needs to lock arms. You might not have the goals perfectly in your head, but at least you understand the story of everything unfolding across the company from training and learning to fraud to robotics to STEM to professionals, to reinforcement learning environments, to research. Everyone, I try to bring along 30 to 50 people and then their job as leaders is to bring along the hundreds of people underneath them. But that's very different than my six-person executive team that I used to have in the core business.

Nakul Mandan:

Let's shift towards the business model itself. I'll still man the bear case on the company, so labs keep needing data to train themselves on specific frontiers. So code gen is where it started, with accounting, other things. But the need keeps shifting because once they get to a certain critical mass of training, they're moving on to the next thing. So one is where you're collecting the data that's constantly shifting. And so it's not a business that can stand on a platform like the traditional software businesses you build once you sell for 10 years. The other part of it is ultimately maybe synthetic data is what all AI is trained on, so this network might not be as yieldy in the future. How do you think about why is this a company that has some defensibility over the years to come versus a window of opportunity that you massively capitalized on?

Garrett Lord:

So I mean, if you just look at, I think the backdrop is if you look at data companies, people for the last five years have thought data companies we're going to go away. And every year the market is almost doubling it from a CAGR perspective. The lab budgets for data are growing, we know some of the lab details, but they're all almost doubling year-on-year. So there's more data that's required and that's one thing. I'm super long data. The other point I would make is if you think about what's driving a lot of model progress right now, it's scaling up compute and scaling up data. Most of the gains sucked on pre-training and most of the actual gains are coming from the post-training side of the house. So if you look at the latest Microsoft paper that were released where they actually broke it down or you look at Qwen or Kimi or the open source papers, the way they're driving a lot of the gains is by more data and more compute. It's very clear that... And then it's also fun in the fact that we now know that when models scale up with enough data and enough compute, you have what happened in software engineering, which is a complete reinvention of the job where you went from helpful copilot to models doing a 3-hour, 6-hour, and 18-hour task. And so the constraint on driving that productivity gain in accounting or the constraint on driving that productivity gain in oil and gas or logistics or finance more broadly is the fact that there is no GitHub. This work is fundamentally unverifiable. And so you make a point around synthetic data. Synthetic data is obviously driving a lot of gains, but most of the gains are actually being driven by making what's unverifiable in human heads into verifiable content that models can actually learn from and reinforce in learning.

Nakul Mandan:

Do you believe there will be an end to the window though or the frontier will always exist? Unverifiable data will always be needed.

Garrett Lord:

I think recursive self-improvement is something people are talking about. And I think that there will be human data for as long as humans are needed in the economy. For as long as humans are needed in the economy, when RSI fully exists, humans are no longer needed. So really you're kind of thinking that... Realistically, the simplest way to answer it is the day that humans are no longer needed is the day that-

Nakul Mandan:

Humans [inaudible 00:51:58].

Garrett Lord:

... humans are no longer needed. And I think a great analog of this is Waymo. Waymo and software engineering I think are indicative of where the future's headed. So Waymo's been having, they had 3,000 people driving 10 years ago, they have 3,000 people driving today. They used to be human in the loop around, are you staying in the right lane? And now it's human in the loop around a dump truck backing up and a stroller coming across the street. It escalates to a human, it gets synthetically enriched, it gets pumped into reinforcement learning, and then humans never need it and they got in that situation. So if you think about what's happening in software engineering, software engineering is, it's incredible in the fact that you're basically putting a human in the loop in a very verifiable outcome. And so it's compounding at an incredible rate plus all the open source and repositories where all the context surrounded. If you start talking to accountants around how helpful these agents are, we just released a public paper called BankerToolBench, at associate level investment banking tasks, which is just literally the first year investment banking job, the top model right now gets 23%. It's not able to automate a task that people are doing. I think there's a huge disconnect between what people are saying is all the jobs are going to be eliminated a year from now and the reality of productivity gains in AI and enterprise today. And a lot of the road, and that's why I think Handshake's important and all of our great competitors are important, is the road to carving that out until RSI exists is codifying the knowledge, judgment, decision-making of humans in these job disciplines into training fuel to learn and automate from.

Nakul Mandan:

Coding also worked because of GitHub having verifiable outputs and there's a lot more corpus. How are you guys, you don't have to spill all your secrets, but how are you guys finding the full trace to outcome on accounting? And is that buying private data? Is that simulating environments for people to act as accountants for tax prep or other things? How are you actually doing accounting or marketing?

Garrett Lord:

Yeah, I think you nailed it with accounting. Let's just pick banking as a great analogy. The shape of data that we're producing today is first you have to understand the taxonomy of how people are spending their time. So what percent of the time is going into leverage buyouts versus discounted cash, DCFs? What does the actual job look like? Then you have to understand, you have to break that task down into a prompt. And then you have to build a completely realistic environment called a reinforcement learning environment with the actual tools itself. Each one of these tools ideally have real world data that mimic the messages, the real world. They actually have parity and functionality with the real production applications that are used. And then you have a set of verifiers on the backend and you have to generate sparse rewards across when you're rolling out, you have to make sure it's not reward. You can prevent reward hacking or minimize reward hacking, but you have to make sure that the way that you actually design the reward structure works for different model builder's techniques and different model builders have very different techniques, generally speaking the same arc of what they're trying to accomplish. But yeah, the new shape of data is real professions, long horizon tasks, real tools and then rewards. And they roll those out in post-training runs and they hill climb to the reward of the objectives that humans have defined. And so small things you might not realize, but you'd say that models today have completely solved financial modeling. Well, the reality is if you look at what humans care about, they care about the formulas and the way that they're written. A formula has to be malleable. You can't be logging into an Excel spreadsheet with dozens of tabs and it's referencing tabs across each, or a cell is referencing across multiple tabs. It's a mess. You puke all over it. So a fun study from us, 522 investment bankers in BankerToolBench, 70% of people would rather start over from scratch than use the model that AI has produced because they can't pick up that model and malleably use it to improve it. So it's like when you start really breaking down just a single function, let alone rolling up the performance marketing budget, which is a finance and a marketing exercise, we're not there yet. It's so unverifiable and these are long horizon tasks that that is the future of training, codifying what people know into environments that a model builder can train off of.

Nakul Mandan:

How does differentiation play out between the various platforms that you compete with right now? Is it based on people are specializing on certain verticals, you might be specializing in accounting or medicine, somebody else might be doing something else or is it a battle for the margins?

Garrett Lord:

We're not saying it's a battle for the margins. I mean, there's so much demand. It's really about quality. Quality trumps everything else, quality and reputation. I think it's something we feel really fortunate about by, we're the number one or number two vendor at the top two model builders and most of the other labs and it all comes down to be able to produce high quality data. And that's really, really hard. By the way, so it's quality, it's volume, and it's speed. You got to be able to produce high volumes of quality data. Ideally, you can turn it around really fast, but cost is really not the main factor. There's an insatiable demand for data from every one of the research teams. And by the way, back to the point around RSI, when you talk to researchers, all of them... There's a big difference between what the MBAs say and what the researchers say. The researchers are like, "This is an amazing business." And the MBAs are like, "I thought the models are producing all the data themselves. Aren't models able to coach themselves?" And the reality is there's a huge disconnect between people that actually understand what's happening in these model buildings and these labs and people on the outside.

Nakul Mandan:

In your view, given the margin structure, it doesn't seem like it's a fight for margins [inaudible 00:57:36]. Do you foresee in the next four or five years a data training platform going public?

Garrett Lord:

Yes, we aspire to go public sooner rather than later, that would be... We have the scale to do it, both businesses are very profitable. Yeah, we would love to be a [inaudible 00:57:52] company.

Nakul Mandan:

Five years out, how much of your work is more about frontier labs versus open source and versus enterprises directly buying from you for their own post-training?

Garrett Lord:

I think our labs business will get to $10 billion run-rate revenue sooner rather than later. I mean, there's going to be that amount of money spending. And then if we can continue on on the share, if we continue producing high quality data and maintaining our reputation and working across all the verticals of data that they're trying to acquire, the most likely market structure is just two or three very large players that control most of this work. The barriers to entry to compete as a small player, it's kind of like Neolab versus labs. It's very hard to compete as a small player. Obviously on the fringes, there will be great small players. In cancer therapeutics research, it's very hard. We're constantly facing verticalized companies trying to bite off one chunk of it. And that's amazing for the ecosystem. And I like that there's competition. I like for the labs, there's competition. It keeps everyone on their toes. But I think the long-term market structure is probably two or three large players that are building the tools, building the environments, having... This will become a full-time job. Millions of people around the world will be doing this. I mean, I think the future of a lot of employees in enterprises actually looks more akin to being a Waymo driver in your job. The idea that an enterprise, say in insurance underwriting has 25 people that are the most elite underwriters, the idea that they're even doing the underwriting right now, I think in 2026 makes no sense. They should be creating the data that actually powers this elastic underwriting judgment engine that just compounds over time. I don't know if you read the Satya post that he talked about, this compounding-

Nakul Mandan:

I did.

Garrett Lord:

But yeah, so I think our labs business will be massive. I think our recruiting upskilling business, it should do a billion dollars revenue very easily. There's $14 billion a year spent on LinkedIn. There's $9 billion a year spent on Indeed. The recruiting upskilling jobs to future, helping people adapt to what's about to unfold is, that's a multi-billion dollar run-rate business. Our labs business, we love supporting the frontier. I think it's so cool. And then we have our enterprise business where we're helping large enterprises really codify their knowledge and judgment and evals and there's a whole stream of work we can talk about there, but that's probably a multi-billion dollar revenue line item. And ultimately, it comes back to the network. We want the network to continue to grow and drive value for users and have more reasons to come to Handshake and get value out of it.

Nakul Mandan:

Do you foresee a world where enterprises start adopting more and more open source and they just own their own post-training environments themselves? There will be an inflection point for you where the enterprise business will be larger than the labs business or that's just too far out?

Garrett Lord:

I think that large enterprises will leverage frontier models and they'll leverage open source. I mean, I think there's still so much gain to be driven out of the frontier models. For example, if you look at, you're going to call one model Gemini for multimodal reasoning, which is the best at it. And you're probably going to call Opus or Codex for code. Opus is better at some aspects, Codex is better at others. Some companies will go all in on one model company, that's okay too. But I think over time you'll want to be calling the right model for the right use case and I think the frontier will continue to stay at the frontier. Also, you'll see players like... Other amazing players have different distribution, different advantages. Take Google, for example, the distribution advantages they have, Microsoft the distribution advantages they have. But I think all these labs, we're still in such early innings. I mean, the story is clear. What's happening in law and what's happening in software engineering is what's going to happen in every single profession. I don't think people really realize that as they capture data like they have done in GitHub and as they scale up their models, what we've seen with software engineering where our software engineers are two or three times more productive and they're managing a series of agents that do the work is going to happen in every single knowledge worker discipline.

Nakul Mandan:

Maybe pointing to a specific prediction that if you have a prediction in mind for open source adoption versus OpenAI, Anthropic or those frontier lab models continuing to be used by enterprises, do you foresee open source adoption driving up?

Garrett Lord:

I think OpenRouter's data is probably the most indicative of this right now. I think that early signs... I think though model builders will also make it... The thesis of open source right now is that you can build this compounding flywheel. I think fine-tuning models has not made much sense for the last four years. People that were fine-tuning, it very rarely... It doesn't make sense. But if you look at super sophisticated vertical application AI companies, like one of the large support companies that you know of, these are friends of mine that run these companies, 85% of their queries are run by open source, but that's a very, they know the distribution. They're super good at mid-training and post-training open source models. I don't predict enterprises will be doing that over call it Codex or Anthropic or Gemini or Microsoft or Copilot. But yeah, I do think in certain areas they will leverage open source models. I think the frontier model builders will probably confront this idea of mid-training and post-training on your own data and building environments to support... I mean, I think that both ecosystems will prosper, but I don't think that the only answer to your data, your judgment, you're not going to give it to the labs will be, all you can do is use open source. These organizations are too sophisticated and they will come up with answers for how you can train on your own data as well as riding the overall frontier intelligence curve. And I think that's the big thing for enterprises is a lot of our core of our enterprise business is evaluating agent performance and helping you pick the right models for your use cases and to know where you can build a compounding advantage. Obviously we'll talk more about that later, but yeah.

Nakul Mandan:

One thing we haven't talked about is physical AI. Interestingly, Scale AI, which was the first data labeling business. I think it was for-

Garrett Lord:

Self-driving.

Nakul Mandan:

... self-driving. I don't know if it was for [inaudible 01:03:27]. Are you guys making inroads there? Is that a big part of the future of Handshake AI?

Garrett Lord:

We do a lot of egocentric data capture. I mean, if you think about what they care about doing, they care about diversity of data. They care about high quality data. They care about diversity of environments. They care about egocentric video capture. You want enough data in a certain distribution and then you want the next thing and the next thing and the next thing. And I think a lot of robotics models are still in the pre-training phase. And so yeah, there is so much demand to leverage... And we basically, if you think about our audience, 25 million people who would love to be making 55 plus dollars an hour doing egocentric video capture. Yeah, we actively have people right now, I won't disclose the exact size of it, but capturing egocentric video for many of the labs. And this is where the vertical player, horizontal players, there's other robotics-focused data companies, I just think that trade doesn't make a ton of sense. The reality is a network is going to own most of that. And yes, there are certain niche data types that these players will own over time, but I think by the end of this year, I think we'll be one of the largest robotics data companies, if not the largest robotics data company in the entire space. And we entered robotics in Q1 of this year. The ability to get to scale, the ability to leverage your research team, the ability to get to... It's going to be harder and harder to compete as a Y Combinator company building a robotics data company.

Nakul Mandan:

What about bio? The kind of stuff that AlphaFold and others are doing around coming up with new proteins and stuff, are you on that frontier also or that's at least niche enough that vertical specific players will have to start with that.

Garrett Lord:

I think vertical, some projects that we're thinking about kicking off, I think there will continue to be... If you talk about biology medicine, we have a ton of projects in that discipline. We have thousands of PhDs and professionals across all the STEM disciplines, probably tens of thousands of people in the STEM disciplines producing data that they want to use to accelerate science. I think it will become harder and harder for Handshake to participate in super deep vertical, specifically you meet the real world. Wet lab biology is hard for us to crack. I mean, we are doing wet lab biology work today, but if you think about what that would mean as an extension for Handshake, that would mean that we're actually needing to physically build out footprints. We're going to need robots. We're going to need to build a full cemented environment. There are several areas around Handshake where I do believe vertical players can do a great job, but if we can do our job right and we continue to... But if we can do our job right and we continue to grow this business and produce high-quality data, I want to compete in that. I think it's great for the world that people keep us on our toes. And whatever serves the lab's best, we're down with, we're going to compete as hard as we possibly can. But I think that in wet lab physical biology, we'll probably have other great competitors, and that's good for the world.

Nakul Mandan:

So it sounds like you guys are at the frontier of every aspect of data training, whether it's again, pre, mid, post, every vertical.

Garrett Lord:

Not as much the pre-side, more on the mid and the post, mostly on the post. And then yeah, I think we've been the most ambitious company out there in terms of bundling and moving across the value chain. Yeah. I mean, I'm not going to talk about what we're spinning up 'cause all of our competitors listen to these podcasts too, but on every major shape of data we are seeding bets. And then I think we've gotten really good with 25 founders of the company. We have a whole strike team. And they stand up the business and they roll to the next one and to the next one to the next one to the next one. I mean, our coding practice right now is on fire producing high-quality coding data in environments, and that's a business that we entered at the beginning of the year, right? And so, we want to be a trusted player to all the frontier labs, and we want to work across all the main shapes of data they care about.

Nakul Mandan:

What does it all mean for knowledge workers?

Garrett Lord:

I think the accurate answer is you'll see disruption, frontline customer support workers. That's a job that people are needing to get upskilled and re-skilled to be ready for the jobs of the future. And a lot of jobs will be eliminated, right? I also think a lot of jobs will look a lot more like software engineering, where we're hiring more software engineers. I think that the fun thing to think about is how elastic it is and will you do more of it if it's cheaper? If it's lower for us to build great product, and there's opportunities to grow revenue, we'll probably build more product, which means we're hiring more software engineers, not less software engineers. And we're probably two to three times more... We're shipping two to three times more value to our customers every year than we were in the past. And so, is that the case with accounting, for example? I don't quite know the answer to that. I think in the case of law, more small businesses will be leveraging lawyers because it became cheaper to interact with lawyers. So I think it's very job-dependent, but I will say I believe in humans' ability to adapt and I do believe in our ability to... I think that's why I'm so passionate about upskilling and re-skilling. And we've done work publicly with Google and with OpenAI, and really proud of the partnerships we've had there, and working in other labs as well. And we want to help be a place that people can turn to to figure out what skills they need to be ready for the jobs of the future. I think there'll be a lot of job displacement, but I want to live at the intersection of helping the people adapt and find these new jobs and upskill themselves. I mean, even my uncle was messaging me. There's an AI course at the company that he's being asked to learn, and it's, "Come to a Handshake and figure out the tools you need to do to be 10x the accountant inside of your company."

Nakul Mandan:

So let's shift to the inner game in this transformation. So, you have a 10-month-old at home, you have a second baby you guys are expecting-

Garrett Lord:

In October, yeah.

Nakul Mandan:

In October. Congratulations. You mentioned your in-laws have both been diagnosed with cancer recently, so you're moving to New York. All this while this company has blown up in a way that very few in history have blown up in a good way, but still it's all intense. How are you holding this all together?

Garrett Lord:

Step one, have an amazing partner and wife. I mean, she's like the rock in our family, and usually the best thing in my entire life is our family. I think having a son and a family provides me a ton of fulfillment. I think my whole world has gotten much smaller. I don't have a ton of time. The time that I do have, I am focused on hanging out with my family. Basically, every hour of the day that I'm not working is spent hanging out with my family. I try to work out. I think you asked about managing the inner game. So the biggest things that help me manage the inner game is working out in the morning. Just doing something light or heavy. Just getting the body moving is really important for me. I don't sleep nearly as much as I want to, but I do try to prioritize getting five or six hours of sleep every single night. Weekends for me, I mean, I'm on a bunch of calls and meetings, but I generally try to spend Sunday really focused on the family. I don't like interspersing the calls all across the day. I like having a chunk of family recharge time. I love routine. I mean, the older I get, the more I realize how important routine is. I basically don't drink alcohol anymore. I try to eat super clean. I think working out, eating good, being recharged by family. And then work's fun. I mean, it's a suffer fest at a lot of times. And have you heard of Type I, Type II or Type III fun?

Nakul Mandan:

I have not. So double click on it, actually.

Garrett Lord:

So Type III fun, they use it in mountaineering and I think it applies to business. Type III fun is like... I love to mountaineers. We're going up Mount Shasta. It's freezing, you're tired, it's cold, sucks. Sucks the entire time when you wake up at 3:00 AM, it sucks when you're coming down. Two weeks later, you're like, "That sucked. I never want to do that again." That's Type III fun. It's miserable. Type II fun is like climbing up the Grand. Type II is like, it's a huge suffer fest and you really don't want to be there while you're doing it. But then a week later, a day later, a month later, you're like, "That was awesome." It's like, "The struggle was worth it" and Type I fun is just where it's fun all the time. I think a lot of doing anything great in life is Type II fun, where it's like as you compound and as you put in more work, and as you push yourself, in the moment it can be seen as really hard, and afterwards it feels really gratifying and you grow and you learn from it. And I think work for me is a lot of Type II fun. There are moments of Type III fun. It's rarely ever always fun. It's not Type I, but I like work. I like making impacts. I think my wife and I... She is a very ambitious. She's got a huge job at Andreessen Horowitz. We push it super hard. We want to live life to the fullest. That's one other trade-off. You asked me about lessons learned. I want to work with people that I find fun. You asked about executive hiring and firing people. I'm not going to work with somebody that drains energy from me. I want to work with people that are pushing equally as hard and I learn from and I grow from. And so, yeah, I think it is possible to have it all, but it's not super possible to have a ton of friends and you can't balance everything.

Nakul Mandan:

Yeah, you have to give up on Type I fun, I guess.

Garrett Lord:

Yeah.

Nakul Mandan:

You have to use your framework. Have there been moments of doubt through this journey, or the product market fit has been so insane that it has always felt right in these last 18 months?

Garrett Lord:

Oh, it's felt super existential all the time. I think it's more self-imposed. Yeah, I mean in the early stages of building this business, one of the scary things is just signing up for projects and needing to figure out how to do them, and being resource-constrained. I mean, there's a certain point in our organization, I think especially in Q1 of this year, where the biggest mistake I made this entire year was not hiring another 300 people. And because of that, we did not have enough. We have great reputation producing, we have a lot of opportunities where labs want to give us work, but our overall service level and how we're able to serve them went down dramatically because we just did not have enough bandwidth internally. And so, I think that's felt pretty existential. I think the more existential parts of the company were, quite frankly, far earlier in the journey, like when we had no money, we're living out of a Ford Focus, driving around school to school. We've been very well capitalized and profitable. And now, most of the pressure is self-imposed. I think our leadership team, one of the things I like to say is, "No rearview mirror." It's maybe not the most fun way to live life, but my job is to drive standards and constantly pushing people to do more. If they think they're doing a great job, then it's like, "How can you do more? How can we do it faster? How can we be more ambitious? How can we turn around in a month plan in a week? How do we drive intensity and excellence internally?" That's basically my entire job. It's not a great way to live, but I spend 0% of my time celebrating any of our success. The opportunity in front of us is so big that it just requires incredible execution, and I'm just trying to push it in every single area of the business.

Nakul Mandan:

In the first decade of Handshake, the mission that you were on, I remember meeting you very early on in that journey and you had this mission around democratizing opportunity, right? And that's why Handshake started as this career platform for college kids and all of that. What is the mission in this new phase of Handshake?

Garrett Lord:

I want to help people figure out where they fit in the new economy, learn the skills and find the jobs of the future and adapt. I think the narrative that's out there right now, that humans will not participate in the future, is just a narrative that I want to fight against and help. If you think about you zoom back into Michigan or you zoom into college students graduating, or you zoom into my uncle who's an accountant, people are scared about AI coming. And they're trying to figure out what to do and they're trying to embrace it, and they're trying to figure out how to figure out how to get the productivity gains out of that. I want Handshake to be the place where people turn to do that, and is a network of hundreds of millions of people around the world that are using this to upskill and find jobs in the future. I mean, that's kind of one part of our business. The other part of our business I really want to drive is I want to pave the path to AGI. I want people to look at what's going to happen four years from now when a new custom cancer therapeutic's discovered or a new energy source that helps. And in particular, what I care about is also the rest of the world, not just America. I think a lot of the productivity needs that are going to come, and science acceleration's going to come and medical discoveries that are going to come, are going to impact billions of people. I feel pretty empowered and excited about the fact that most of that progress coming from data and compute. So we have our friends building out power infrastructure and buying GPUs and doing inference. And we have our friends at the labs building incredible algorithms, the smartest minds of the world. And then we have players like us, who wake up every day and grind our faces off to try to produce the highest quality data that goes into the algorithms, that leverages the compute to power model progress. And that vision of helping labs do this, and then also enterprises, I think it's pretty cool to think about. The future, the way they're talking about it, I don't like it, like every business is going to be just nuked. The entire Fortune 500 is just at a massive disadvantage and every job's going to get automated, humans have no role in every company. I don't like that future. I want companies to succeed with technology. I want them to get the most out of it. I want them to accelerate earnings. I want them to hire more people. That's what excites me.

Nakul Mandan:

As you think about the next couple of years for Handshake AI, beyond recruiting and meeting the demand that's coming to you, what genuinely scares you about the future of Handshake AI?

Garrett Lord:

I think it's very hard to continue to balance all sides of our business. And man, that's really hard. We have 15 different businesses inside of Handshake AI. We have a huge learning and upskilling and jobs business. We have an enterprise business that is exploding. I think probably what scares me most is my ability to grow, to be able to hire and manage and drive a vision, and drive standards, and build a cohesive, high-functioning group of people that can manage such a wide-breadth business. But there are great analogies of this in the past, like Google and Amazon. And I mean, I think Handshake can be a company of that size one day. I think right now in AI, it's go big. Fully go for it. I think the coolest jobs in the world to do right now, one of them is to work at the labs. If you can get a job at the labs, I would work at the labs, specifically on research and product and engineering. That's the coolest job in the world. The second-coolest job in the world is working at a data company. We're powering all of model progress. I would 100 times rather work at a data company than work at a vertical AI application company because a lot of those companies will find out how they evolve over time. I think many of them are durable, many of them are not. But sitting at a data company, we basically see the vertical AI companies being built out and we see the labs being built out, and we see enterprises changing. And so I think on our side, managing the bets, managing how we go big in a way that fully takes advantage of all the opportunity in front of us, from a place of abundance and excitement, is probably what scares me the most.

Nakul Mandan:

Do you have a hot take on application layer surviving as an independent layer? Do you have a hot take on all these vertical AI companies will ultimately get obliterated by Anthropic and OpenAI?

Garrett Lord:

I have a couple fun thoughts. Let's pick Procore.

Nakul Mandan:

Yeah.

Garrett Lord:

Do you know what Procore is?

Nakul Mandan:

Yeah, of course. Yeah. Yeah.

Garrett Lord:

So Procore, for those who don't know, it's a system of record for builders. And my good friend who's a builder and my dad who's a builder used the product. So my question would be, is my dad more likely to learn how to use a Frontier model with extended thinking in a coding harness? Or is he more likely to use the invoice AI management system of Procore that's bundled into the way that he pays his employees and he manages all of his customers? What do you think the answer is?

Nakul Mandan:

Is? Frontier Labs.

Garrett Lord:

You think you'd rather use the Frontier Labs?

Nakul Mandan:

Yeah.

Garrett Lord:

Yeah, I'll take the opposite bet on that front.

Nakul Mandan:

Okay.

Garrett Lord:

Yeah.

Nakul Mandan:

But isn't it easier for even your dad, as much as he may have been a different generation, just to use a chatbot who gives the answers faster?

Garrett Lord:

That's a good point. I mean, I guess structure and answer a different way, is I think where there's complex workflow that is ingrained in the existing operating process, it's harder to just wipe out all that middleware, right? So maybe where I'm bearish a little... I mean, I think these legal vertical AI companies will be big companies one day. I'm not saying they're going away, but I do think that the incremental switching cost to... And by the way, where they're probably most sticky is in the Walmart's supplier onboarding legal procurement process where there's heavy workflow. Where there's heavy workflow and compliance, and meeting the standards at buying box of Walmart, they will probably use these vertical AI companies. But I think 90% of legal AI usage will fall onto the latest Frontier model and not land on one of the legal companies. So will the legal vertical companies be super-high enterprise value? Yeah, I mean I still think they're big companies, but the company I like a lot more is OpenEvidence. I think OpenEvidence is really cool. They have a whole data flywheel, basically, where they're able to... If you think about how a doctor uses OpenEvidence, and I'm not a doctor, but they're basically sitting at this incredible distribution of how people are using it. And then they're able to systematically swap out Frontier models. They're able to fine-tune when they need to. They have a constantly evolving evaluation set that helps them, and they have distribution. So they just continue to compound and compound and compound and compound. And they're able to customize that product in such a unique way for each one of the different use cases that doctors use it. And I think model builders, it's hard for them to do that across every single category of work. The story the model builders would say is that the models will get so good and good enough to the point where they'll be able to build and optimize that experience for each one of the doctors. And I kind of buy that argument too, but I would say that I like OpenEvidence more than I like legal, and I still think OpenEvidence is an amazingly large business.

Nakul Mandan:

Yeah. Yeah.

Garrett Lord:

Yeah.

Nakul Mandan:

Cool. All right, next up is our quick-fire round. Red flag you never ignored in recruiting?

Garrett Lord:

The gut.

Nakul Mandan:

Gut.

Garrett Lord:

Gut feel on a candidate.

Nakul Mandan:

Most overrated piece of advice in AI right now?

Garrett Lord:

I think your diet should be mostly research, like researchers talking to the labs or people at the labs, and not getting too stuck on what's happening on Twitter. I think a lot of people at Twitter don't understand what's happening at the forefront of AI.

Nakul Mandan:

One thought that you disagree with from the lab researchers when it comes to human data?

Garrett Lord:

I think I continue to want to push researchers to care more about how we work with humans, and how we train them, instruct them, and the turnaround times we expect people to work. I continue to think that we as an industry can get better and better at the way that humans interact in producing data. And I will continue to try to push our lab customers to drive an experience that's great for humans.

Nakul Mandan:

A job that will definitely not exist in five years?

Garrett Lord:

Frontline customer support.

Nakul Mandan:

A job that will survive the next five years in knowledge working job?

Garrett Lord:

Software engineering.

Nakul Mandan:

A career skill you'd recommend every 18-year-old to pick up given where the future is going.

Garrett Lord:

I think trying to be high agency and operating in ambiguity, specifically taking risk. Learning, growing on your own, putting your own agency and growing and learning for students, specifically picking up AI skills. Young people have incredible opportunities in these large companies to drive an impact, and they need to embrace AI and push themselves to learn more about it.

Nakul Mandan:

So last question for you. A young person watching this podcast wants your life, wants what you've accomplished. What's the one thing you tell them that journey actually requires?

Garrett Lord:

Over time, as I've gotten older, I've become a lot more confident in just my abilities. And I think specifically what that's meant for me is you have to push yourself to continue to grow. Everyone that's ever done anything cool in the world puts their pants on one leg at a time. And anything you want to make happen in the world I believe is very possible. That's the fun thing, when people come to work here, sometimes they're just like, "Wow, you can just do things. You can just figure it out." And I think stacking days day after day after day, and giving it your absolute all, great things happen. The power of compounding is really powerful in life. So what I would say is if they want to make an impact in the world and build a great software business, I would say just find a great team and just put one foot in front of the other every single day, and give it your all. And what it requires, I think a lot of people try to start businesses just 'cause they're so excited and thirsty to start a business. And I really think that you should try to do something that you're innately passionate about by. Because a decade long journey, if you don't have something that fuels you through the hard times and gets you excited to drive a motivated team and work really hard, you're not going to make it. I think all the best founders, most of them that I interact with now, they have some edge that drives them, and that generally it's connected to the passion they have for the business that they're building.

Nakul Mandan:

Garrett, this was amazing. Thank you.

Garrett Lord:

Thank you. Appreciate it.

Nakul Mandan:

Thank you for coming on.