Why AI's Next Problem is Data | Garrett Lord on Training Real-World Models

And what that will look like is like a massive race between companies where unless you're investing in agendifying your overall experience, you're just going to fall behind your peer group in the public markets. And so I think like open source models and frontier models will both win in this world.

Hey everyone, welcome to the Founders in Arms podcast with me, Immad Akhund, co-founder and CEO of Mercury.

And I'm Raj Suri, co-founder of Lima and Tribe. And today we're welcoming Garrett Lord, co-founder and CEO of Handshake. Welcome, Garrett.

Thanks for having me. Excited to be here, guys.

Garrett, this is the first time you and I are meeting. I guess you've known Immad for a while. Tell me more about Handshake. I was just looking it up today, so I'm not too familiar.

Totally. So, I mean, the first decade of the business is very, very different than the last 18 months in this business. But the first decade, I started the company after working 2 internships at Palantir, really out of a personal pain that I had in college of finding my first internship and job. And I started with a group of friends from this rural school called Michigan Tech, trying to make it easier for really any student in the country, regardless of kind of what their parents did or what school they went to, navigate and find great internships and jobs. And that business does over $200 million ARR, and there's a million companies that use it. And most students in America have profiles and think about it as like a verb for the way they launch into the workplace. And really kind of proud of that business and We are investing in a bunch of things around like helping students be more prepared for this new economy and learning AI skills and kind of excited about kind of the core business. But the vast majority of how I spend my time right now has been really building a new business side of Handshake, kind of leveraging the asset we have, which is a ton of trust and engagement on the core side of the platform, to help model builders improve their intelligence of their models, specifically improving their models in scientific domains like biology and chemistry and drug discovery. Discovery, all the way to professional careers, like making agents actually work better in real workplace scenarios. So helping agents get better in oil and gas and in retail and in manufacturing. And obviously, you know, we know where software engineering agents are at today. They're able to run for 10 days. The rest of the economy is not at that level of acceleration. And that's largely what we build for model builders is reinforcement learning rewards, environments, and the shapes of data that help push models forward.

How do those 2 businesses link together? I think it's probably obvious to you, but probably not obvious to the listeners.

Yeah, so if you think about the network that we built over the last decade, there's over 30 million people that are on the product. And what we built with them is most of these people used Handshake in college, used it to find their first job. We know where they went to school, we even know where they went to work after they graduated from college, and many of them think fondly of Handshake. And so we're able to re-engage the alumni audience at a And that's a pretty incredible rate compared to our, our peers. Like, that, that really has been the moat that's enabled us from— to really accelerate so fast. I mean, we've gone from $0 to $1 billion revenue in the last just a little bit over a year. And the main reason we've been able to accelerate so fast is because the network. So 89% of people that worked last month came from Handshake. And so we're just able to structurally turn around on the best professionals much faster at higher volumes than our peers who are using like TikTok and Instagram to find accountants.

And these So people are creating data that the models are using to train their model, right? You find someone in the oil and gas industry and you're like, hey, can you explain all the material science behind oil and gas or something like that? And they train the data, something like that?

Yeah, something like that. I mean, I would describe it more as they're working inside of video games. That's the easiest way I described it to my dad. They're working inside of video games with the full context of what they have in their job. So they have access to files on their computer, they have access to files in the cloud, they call WellView and IPS Bolt, like all these tools that they have access to on their computer and in the cloud. And we build, essentially, to think about it simply, is we build like video games where they execute tasks connected to all the things they have access to in their job. And then we build all— the easiest way to think about it is like the rewards or infrastructure to score how agents are doing. And that ultimately gets encapsulated in like a Docker container. And we send off that Docker container to be used in training.

I kind of imagined all of this training is like text in, text out, but you're saying it's actually like computer use, like tool use, like it's much broader than just like text going in and out.

Totally. If you think about the shape of like how models were trained, for some of the engineers, it's like they go through 3 stages. It's like pre-training, which is scraping all the tokens available on the internet. They go through mid-training, which is called supervised fine-tuning. And that used to be the shape of data that people built. Like it was, you know, you probably heard of this term reinforcement learning with human feedback, right? You probably— essentially the idea there was, you know, doctors Yeah. So, like, users were providing preferences between, did I like A more or B more, right, to train the model on how to answer questions. They'd also get better at, like, mimicking. But when we're in 2026, like, 70% of the compute that's spent on, like, training a model is actually spent on reinforcement learning. And reinforcement learning is the shape of data that we really focus on, which is, like, helping models actually do— helping accelerate jobs and not just, like, mimic responses.

So one question I have just generally is like, okay, so is there like diminishing returns to training these models, right? Like where does that limit reach? You mentioned 2026, you know, a different type of training than before. So what's that future gonna look like? How is this gonna evolve, do you think?

I think the easiest way to answer that question is like, if you think about software engineering as kind of indicative of where the future's going, The job category of software engineers is like maybe 2, 3, 4 times the level of productivity that a software engineer used to have. Agents are able to execute autonomously for like 10 days, intervening to human when they need input. But, you know, we'll pick Mercury, for example. Like your finance team is not able to like close the books or manage a performance marketing forecast like autonomously. It's not able to call the tools or really do the job of other professions beyond software. And so when you look at like agents and their ability to run for multiple hours, like in oil and gas and manufacturing and finance, these agents can run for 2, 3, 4, 5 hours, but they're not putting humans in this significantly advantaged position of like managing multiple agents that are doing their job. And that we know, you know, just talk about 2026. We know that as you scale up enough compute, against enough rewards, you can get what happened in software engineering. And the reason software engineering accelerated in such an incredible fashion is like everything's been written down in a way that's very verifiable in code. And so they have so much code data, and when you throw enough compute against that, you get what happened in software engineering. And so right now they have the compute, but there's a huge bottleneck in the supply of high-quality data or rewards to accelerate every other profession. And then the last thing I would say that's interesting there Whereas like the direction of travel again is like for a model builder, there's like hundreds of billions of dollars, trillions of dollars of revenue, right? They want to automate more and more of the economy and put more humans in advantaged positions. And in order to drive the revenue really gains that they want and they aspire to do, CEOs want to see other teams be enabled like software engineering. And they are enabled today, but just not at that extent. And so there will be, there already is today, like tens of billions of dollars being spent to capture data or rewards that helps improve models in other economically valuable disciplines. We didn't even talk about sciences like material science or drug discovery. They aspire to invent new materials, new battery technologies, cure cancers. There is no Reddit forum or GitHub equivalent on the internet.

What is the actual flow here? Is it that you, Garrett, or someone on your team is like, hey, we need more data around engineering, and then you recruit people, and then you collect a bunch of data, and then you sell it 5 times to 5 different labs? Or is it more like a lab comes to you and they're like, hey, we need this very specific data, can you go get it for us? And then you then go collect the people and the data and then sell it just to that one lab?

Yeah. So the way the conversations typically start is a researcher at one of the labs will say that we're— let's pick oil and gas. We're struggling in oil and gas. I'm getting feedback from my forward-deployed engineers and my sales folks that when I deploy a model inside of Shell, or we're hearing that our models aren't good enough. Can— and these researchers are incredibly great at Math Olympiad, software engineering, brilliant algorithmic minds, but they don't know a lot about oil and gas. Like, oil and gas is a $5.5, $6 trillion market, right? There are hundreds of thousands of jobs, tens of thousands of companies that participate in that. And it's actually like upstream, midstream, downstream. There's a whole ecosystem around it, right? And so where we'll start— long-winded way of saying— where we'll start is we'll help them evaluate where their model is working well and not working well. And the way that we do that is— to walk through the full cascade— is like, we are finding experts in oil and gas in, for example, like downstream finance roles. We are kind of teaching those experts, equipping them with an AI copilot. They're going about like creating tasks, and also helping us build the tools that they use. Like, what is the work context of what happens in a finance upstream role? What is the job that I'm doing day to day? What do I want AI to do? And then we have to then build dozens or in some situations hundreds of tools, we have to populate their entire world in this video game or environment to actually look reflective of what they're doing in their job. We then design the— we create tasks that are not too easy that models already beat and not too hard where models can't learn from. And then we go about kind of taxonomizing and sharing with the researcher in a very clear, quantitative way, here's what your model is good at doing in terms of agentic performance, here's what it's not. And that's what's called an eval set. An eval set is a series of tasks and basically a grading infrastructure that allows you to score that all the way down to like, you're bad at these tools or you're bad at these reasoning steps. Then from that point, we'll then iterate with them to say, like, where do you want your model to get better? As you move your models, you know, as they're releasing every single quarter, every single month, where do you want it to get better? And we'll design hundreds, thousands, tens of thousands of tasks structured in an environment that they call on these multibillion-dollar training runs that improve the targeted capabilities of their model in oil and gas. And so that's generally the flow that we're building. And then The platform is basically an accumulation of all the projects that we've worked on to push model performance, and we actually have to like train our own open source models internally. So the best way to generate a credible response to researchers actually say we took Quen or we took Kimmy or we took GLM, and we trained it on our data against against this kind of holdout set, and we're able to beat you in this one area. And that way they know that the data actually drives performance and gains. So we're kind of like a full stack. Model builder internally. Yeah, we don't pre-train models, right? We generally speaking are doing RL or post-training models, but we actually are building models internally that are really good at one particular area. And then they then, you know, spend tens of millions of dollars to drive these capability improvements.

I was talking earlier to someone trying out these models. And it seems like computer use has finally got pretty good. If you're using these models, I feel like a year ago, computer use kind of sucked. We're finally getting to the level where a model can go use a mouse and navigate a website and do pretty well. Is that what you're seeing, that computer use tasks are just much better than they used to be?

Yes, for sure. And I also think that the deeper-cut way to answer that question is where there exists data in the pre-training distribution, you can get a lot of gains out of RL. One of the problems that we're actually seeing in a lot of these professional domains is that there's an insufficient amount of data available on the internet. So they're actually, you know, in order to improve gains in RL, they're actually trying to collect more data upstream on the pre-training side of the house to improve downstream performance. But yeah, computer use is a super important part of it. Also, like, verification for a lot of these SaaS tools and long-horizon agentic performance requires computer use in order to be able to, like, grade the actual agent and how it's doing. So, uh, excited about the progress in computer use.

I mean, computer use is a complete game changer, right? Like, I think it really unlocks like a massive number of— 100%. I think like, you know, the reason coding works really well is like it's all in text files, right? Text files in, text files out, right? But most jobs are not like all text-based. So I think computer use, we've seen, I've seen a lot of startup pitches that like really relied on computer use. And when like, when they were doing it like a year ago, you were like, it's like, it's kind of not really working, guys. But it's kind of fun to see like, These startups that start early and then they try to bet on where AI is going, and sometimes that plays out. I think computer use is one where the startups that started a year ago relying on computer use getting better, I think, are finally delivering results. What do you think models are bad at today that in a year or two's time they're going to be great at?

I think they're going to be great at everything. I think the primary bottleneck is just like, where do they want to focus? Right? Like, you probably see these startup pitches too. But like, what you're seeing right now is people like Open Evidence or Harvey, right? Or Andrew Dai that's basically picking a particular vertical and just focusing on capturing a large amount of data and spinning up a data flywheel in a particular vertical. And many startups are able to do that with much more focus and faster than, than the large model builders can chase everything, right? And so I think we'll continue to see improved generalization. But in a lot of these verticals, I think there will be huge companies built that are targeting a particular domain and going super deep. And the way that they're going deep is pretty simple. They're taking base open source models that are pre-trained, and they're post-training them on You know, a larger supply of high-quality data. Data really is the real moat of these large businesses. I mean, obviously, there's these research labs are predominantly like algorithmic process is what's driving a lot of how they want to design their cultures and what they work on. But really, what what what drives capabilities is high-quality rewards for reinforcement learning. Like 70% of the money that's spent on training a model is spent on reinforcement learning, and we know that.

That's a good point.

That if you can capture the right amount of data, you can hill climb pretty much anything. All right, uh, it's hard to hill climb unverifiable domains. Um, like, I think the way to look at automation in the economy, it's a, it's a deeper cut, but in any process that you can get like 8 people to agree that the outcome is correct, no matter how long it is, you can generally make an incredible amount of progress in hill climbing. And like, so if we, the 3 of us, were all designing like Our board deck, right? We all like started a company. We're doing a board deck. All three of us would have a pretty tough time on agreeing what a good board deck is. It's like so unverifiable, semi-verifiable. Like it's so it's so driven by personal preferences. But like how you know what is a good response to a support ticket? Or even better, what's a what what is closing the books look like? Or what is rolling up the marketing forecast look like? Generally speaking, the professionals that industry can like align. on what good looks like. And any chance you can align on it, you can score it. And if you can score it, and keeping it pretty high level here, you can hill climb it, right? And so I think we'll see, that's the way that I look about this permeating the economy is like, where is there a large amount of economic value, a lot of jobs, a lot of money to be made that is, you know, most people would agree there's like a consensus right answer. And that's where we'll see hill climbing, uh, I think over the next 1 or 2 years. Um, that's how I think about it.

What do you think about robotics? I feel like robotics is one of those areas that maybe it's not that great at, but like, are you also creating kind of post-training data for kind of robots, or you're not doing too much of that?

We are, but it's mostly pre-training data. There's like 2 types— I mean, there's many types of data, but there's 2 predominant categories. One is what's called egocentric, and the other is called teleoperations data. Teleoperations is essentially, uh, professionals engaging in actually driving a robot through a scenario, right? And then egocentric data is just essentially capturing with sophisticated cameras the point of view from a human, oftentimes with gloves or other assets that they use depending on their techniques. But I think robotics is— I mean, we're still in just like, there is no corpus. There isn't a large enough corpus on the pre-training side of the house, and they're just focused on like essentially as much diversity of use cases targeted at the domain that they're trying to target, whether it's like unboxing or supply chain and logistics or truck unloading. But I don't think we're going to see— I'm not like a futurist in robotics, so, you know, but I don't think we're going to see this like general-purpose humanoid that's able to do everything in the next 3 to 5 years. I think we're going to see incredible acceleration in particular domains in robotics where they're able to spin up a data flywheel. But I think robotics is still pretty early.

Yeah, that's interesting. I wonder if there's a coding equivalent in robotics. You know how coding was a huge market, but it also allowed reasoning to just get better and be cross-applied to other domains? I wonder if there's something like that that becomes the core robotics unlock, if that makes sense.

Robotics is already getting very good.

Where do you say it's getting very good? Out of curiosity, what proof points do you put up?

He's got a robot vacuum and he's really into it.

Robot vacuum. But I mean, self-driving cars are robots too.

Oh yeah, that's true.

And so, yeah, so like, uh, but, but, and then I was reading an article a few days ago, um, and by the way, I run a futurists and robots group chat and there's 2 million robots working in China. So like there's, um, 2 million robots working today in China.

What are they doing?

Yeah, like they're, they're in factories, right? And, um, And some of them are purpose-built robots. Some of them are humanoids. You probably saw the World Games, you know, all the videos online, you know, with robots running and stuff like that. I think the future of robotics is already here. We're just not seeing it in the US yet. And that is actually true for many things. You know, I talk with other founders about this is like China's ahead of the game in many different areas that like, that we're only just starting to see right now in the US. I mean, they were ahead in, in like, you know, some social trends like TikTok and things like that as well. Um, and, uh, you know, certainly most hardware and manufacturing trends, you know, EVs and stuff, now they're way ahead. Um, but, um, so I think robots are one of those things. And so, uh, the, the use— the actual building of the robot and the training of the robot is there. It's, it's, it's like already, uh, happened, and China can do it at cheap enough rates to be viable. Now we're just seeing an explosion of use cases. And I think, Garrett, to your point, we're seeing this kind of niche training happening, right? Like, I have a friend now working on a robot specifically for fruit picking, um, right? And like, so he's developing this robot just, just for this one specific use case. Um, and so we're seeing the same explosion of, you know, letting a thousand flowers bloom just for, uh, for robotics. It's a very, very exciting time in robotics too. I don't think humanoid robots are the solution necessarily, but I, I think there's going to be If I was to pick an area to invest in, like, or like to work in as someone, you know, starting their career, I would pick robotics for sure. There's, there's, uh, it's going to be a 10 to 20 year sort of very exciting arc coming forward.

I agree. The only point I would just make on top of that, I think there's, there's this idea of like generalization versus specialization, right? I just don't think we're at the point where robots are generalizing well across domains, right? They're— yes, when you, when you can close the loop in a particular process, robots robots can be extremely effective, right? And I think that the, the thing I'm excited about long term is like robots and robotics in general being able to like generalize across domains in, in a, to a much greater extent than they're able to today.

Yeah, yeah, I think, um, the generalization— yeah, so that's what I guess these world model companies are working on, right? Like, it's kind of, um, the generalization of Uh, is it— is that right? Like, these world model companies are focusing on both the physical world and, um, you know, the digital?

Yeah, a lot of the labs have aspirations to do both, which is fun, yeah, uh, to see, right? I think there's a lot of— yeah, I mean, TBD on how that all plays out in terms of specialization versus, you know, generalization and quality of talent, the compute infrastructure they have. Like, I think increasingly so what is happening in model land is You know, in order to be competitive, you need to— you know, used to be— used to think— I think people used to think you could be competitive raising $1, $2, $3 billion. Uh, I think nowadays, like, where things are headed, it— you need to have access to the best researchers. You need to have access to an incredible amount of compute. You need to be— you know, you need to be able to actually train and serve. It's, it's really becoming, uh, challenging to take a big leap in the industry without an incredible amount of capital.

How are these open weight models doing it? Because they don't really monetize as directly as Anthropic or OpenAI. Is it just because distillation works so well that they're able to keep up?

Well, I think credit's due. They're extremely clever. I think most of the researchers I talked to at the top labs are extremely impressed with what What they're— what a lot of these small teams are able to pull off. I think that there is a lack of transparency on how many resources they actually have and what the compute infrastructure actually looks like underneath the hood. I think that, you know, at least the gossip among circles is that they're actually, you know, training on more compute than they're actually talking about. But people are— I mean, so that's one. Point 2 is, yeah, an incredible amount of the gains are coming from distillation. But they're clever with it. And they are buying data from, you know, American data companies. They're buying a lot of the Frontier queues that get sold to, you know, a lot of the top labs, which I think, you know, I'm personally not a fan of. I think that, you know, I'm like very—

But how does the— I just don't understand how the business model plays out. Like, if, you know, if an American lab is willing to spend, I don't know, $2 billion on like data and pre-training and post-training and all that. How can they spend as much without a business model attached to it? Because it's just open weights, right?

I think it's all being subsidized by Alibaba and the government and a lot of the large capital sources in China. I also think, to pivot a little bit, the American open source community is also really exciting, like Thinking Machines, Reflection. Google has an open source model, right? Like, I think we'll continue. I'm like, I'm really excited about really what, like, America and like what can happen over the next 12 to 18 months. Like, we'll get Nimotron with what Nvidia is doing. I mean, Jensen's comments, he's making a huge investment in improving the base models. Um, and The long-term supply chain, I think it's quite clear enterprises will call many different models. There'll be an ensemble of models underneath the hood. They'll train their own models on their own data. They'll try to leverage open-source models to lower their inference costs. They'll still call frontier models. I'm actually incredibly bullish on Frontier Labs. But I think that the modern enterprise is not going to be tied into just one provider, kind of like they're able to use GCP or AWS or Azure. They're going to be calling the right model for the right job. And they're also going to probably be training their own models on their own data to achieve frontier performance on their internal use cases and benchmarks.

By the way, I'm actually doing an Open Weights AI Summit next week. Aaron Levy is our keynote. And this is such a new area of Open Weights focus that a lot of people still don't know what it is and how best to use it. But I think, to your point, that this is going to be a big area of of focus. Where do you see that ecosystem in terms of like, um, um, do you, do you see the US has a chance to, to lead here? And, um, do you think like if a typical enterprise, would they be, would they be more interested in taking an open weights model, training it on their own data and doing that work themselves or using a closed model? Like, how would they think of that decision? You know, someone, someone like Mercury maybe. Yeah.

Yeah. I answer the question quickly. Like, yes, America is in an amazing spot to win. I'm like super thrilled about the progress that American model builders are working on. And there's a huge investment across the ecosystem, largely spearheaded by NVIDIA, helping support and champion American open source models. There's also a ton of venture money flowing into that, which is incredible for just us in general. And the teams are super high quality. So that's kind of point one. I think America is in an amazing spot to win and be on the lookout over the next 6 to 12 months on what we can build. And we're supporting almost every one of the American open source model providers with the data that's helping build these base models. Then what's happening in large enterprises? So I think you answered the question like it's either or. You're either going to use open source models or you're going to use funded models. I think that's the wrong way to frame the question. I think that people will use both. I think you use Mercury as an example. I think Mercury will 100% be calling GPT, you know, one day 6.5, and we'll also be calling Opus, you know, 7. Like, it's 100% the, you know, like, you will pay for frontier models in use cases that you, that you see value in. You'll also most likely be using, like, if you look at GLM outside of software engineering, like, a lot of the knowledge worker tokens outside of software can be served by open source models today, right? And that's why you see, I think, inference costs from these large providers coming down. They're trying to make it more economically attractive to not leverage open source. And we'll see this token battle, I think, unfold in 2027. Right now, Handshake and Mercury are very focused on driving adoption and getting our employees to use these tools more. But as inference costs— we spend $14 million a month on inference costs in our company alone. And I would love to spend $20 million a month on inference costs. And then I want to then think about what are the economics underneath the hood of how I can optimize those costs. And I think that will really unfold over 2027 in terms of like the cost optimization. And yes, like if you take Mercury, for example, like in know your customer, you know, or like, you know, trust and compliance and money wiring use cases, Mercury has a ton of longitudinal data. And he knows his business better than I do, but a ton of longitudinal data that really he's not going to give to anyone else. Like that is his advantage in, let's call it, Fraudulent money transfers, right? And I don't know if that's the right use case, but let's pick a process that's highly proprietary, where you have a ton of longitudinal data. Mercury will 100% be taking an open-source model and training it on their own data to achieve better outcomes than a frontier model in that particular use case. And most importantly, the reason they'll do it is because they'll want it to be recursively self-improving. They'll want the model to get better every single month and week they're in business. And in order to do that, there's only so much you can get out of optimizing context and optimizing your harness without actually controlling the weights yourself. So I think the, like, 10 years from now, every large business, you know, the large, large public companies, they'll look more like AI labs, where they're thinking about their data and their assets and their flywheels where they drive margin. They'll look more like AI labs than the enterprises today. Because the reality is, like, pick insurance underwriting, like, pick a large public company, I'm not going to name one, even though we're talking to one, like, you know, they have tens of thousands of people that do insurance underwriting. And realistically, what that job looks like in the future, you mentioned Waymo, it looks probably a lot more like Waymo, like where agents are actually doing the job. And in use cases where the stroller's coming across the street, and the dump truck's coming, like in Waymo, where the Waymo doesn't know what to do, it will escalate to a human, and the human will drive that scenario. And then the next day, the model will get better, just like it works in Waymo, right? And so insurance company that actually ends up really agentifying a lot of the experience, we'll see, you know, significant basis point improvements in like, you know, their overall cost structure, they'll see basis points improvements in their underwriting overall performance. And what that will look like is like a massive race between companies, where unless you're investing in agendifying your overall experience, you're, you're just going to fall behind your peer group in the public markets. And so I think like open source models and frontier models will both win in this world.

Oh, cool. Imad, are you using any OpenWeights models? Curious.

No, I mean, right now we're very much focused on getting as much adoption as possible for our command agent. And we're using the best reasoning models. I would say one thing that you want, it doesn't have to be open source, but smaller models for is faster responses. Some of these more complex frontier models take forever to answer things that are actually really simple. Sometimes it's like, how do I send a wire? And it's just like you want an instant response. and like a quicker, faster model is good. We kind of skipped over this point, Garrett, but like $0 to $1 billion in one year is insane. I don't know, there's probably not that many companies that have ever done that on like a new product line. So I guess like what breaks, like, I mean, this is a very operationally intensive business as well, right? It's not just like purely like, you know, just software, which like it scales quickly. So like, you know, how did you manage this?

Yeah, I mean, it's crazy right now. Like, we we can see two billion right in front of us. We're just trying to climb towards it, which is you know a great cool position to be in. It has been an incredible. I mean, we have an incredible team, and we've I think hired unbelievable people who are I think working harder. I mean, this level of growth—it's like it has been an incredible. incredible amount of work with a great group of people. I think the team we pulled together across every aspect of what we're building is, I think, the leading team in the market. We also have this massive structural advantage that's enabled us to really speedrun multiple categories of data and do so in a way with, you know, how would you build a data company in 2026 from the, from the ground up if you would really approach things with first principles?

I guess, did anything, did anything break? Like any big learnings where you were like, wow, I wish you'd done this differently in the last 12 months?

Yeah, I mean, you can't hire 400 people in 8 months and things not break. I mean, it has been— everything has broken at some point. I mean, I would say everything is breaking every 90 days, and then we're having to rebuild it, right? Maybe main learning lesson. It's been really— I think I wish we would have invested in having more offices to be able to work 24/7. Around the clock. Like, we just built a huge office underneath an amazing leader in India, and that's been an amazing accelerant for us to be able to have, you know, literally where we're able to chug 24 hours a day. I wish I would have done that earlier. Another thing I wish I would have invested in earlier is really like codifying the difference in cultures between both companies. Like, it is, it is one company, but it's also 2 companies. And there's a very different like rhythm of the business and operational cadence required to scale for AI labs than serving large enterprises and universities and consumers. And I think, you know, we came out with one of our core values. I wish I would have kind of like gone through that exercise earlier with the team. Like our number one value is Olympic pace. Olympic pace is the idea like Michael Phelps was in the pool every single day for 5 years up into the Olympics. And like, that's what it takes to win at this level. And you have to have passion and, you know, You, you don't have to give it your best every single day. Like some people don't, you know, you have a day where I did last night, my kid was up really late, I didn't get a ton of sleep, like, I probably have like 85% juice today, right? It's like, but you just compound incredible day after day after day where you're giving it your all over 5 years. And, you know, you win gold medals, right? And that's really the attitude we want our team to have. And I think is embodied by a lot of the top leaders in the company.

It's a crazy time, you know, and Well, it's hard to predict what's going to happen in 6 months rather than, you know, versus like even 3, 4 years from now. But I'm going to ask you to try. Like, what do you think this like AI data economy looks like in 3 to 5 years? Um, what are gonna be people be paying for on like a, a, you know, um, a regular basis? Or will there be, you know, like the same type of data economy, or there's going to be some other use cases that people will need 3 to 5 years from now?

I mean, we're already seeing thousands of people quit their full-time jobs to do this. I think we'll see in the future millions of people quitting their full-time jobs to do this. I think this will become a totally new job category on the internet right now. If you want to work part-time, like the gig economy, you know, you could, you could be a DoorDash driver. I'm not disparaging that in any way, but like, you know, there aren't many like white-collar gig opportunities to work around your schedule and make $100 to $400, $500 an hour. And I think I think that there will be, in the not too distant future, tens of billions of dollars a year. I mean, we're paying out, I think, north of like $3 million a day to— I don't know the exact number, but that number's just, you know, started at 1, now it's 2, now it's 3. It's just like, soon that number will be paying out $10 million a day, or paying out $20 million a day. You know, you asked to talk about the future, and this will— I think companies will be leveraging their human capital and their people to also be participating in creating these flywheels internally that play to their advantages and their margin. So I think it's a totally new job. We're investing deeply in like upskilling and reskilling, uh, and like jobs of the future. So we bought an amazing company called Uplimit with an incredible leader that was kind of an advisor of ours named Julia. But we are trying to participate in like being the job network of the future, the job network that really powers the AI economy. And that's like finding amazing jobs as jobs get reinvented. Learning and upskilling and reskilling and positioning yourself, whether you're at a company or trying to find a job, monetizing your knowledge and kind of proving what you know by making money doing it to be, you know, to get picked off by companies. And there's like a lot of thirst for understanding like how to accelerate my job in accounting, for example, with AI, right? So I want to invest in more and more reasons to come to the network. And ultimately the prize is like beat LinkedIn, Uh, you know, I think there hasn't really been an incumbent that's challenged them. Like, they're the largest job network in America. I think they're, they're ill-equipped for like where the future's headed in the economy. And, um, I want— yeah, I think there's incredible flywheels between businesses that we're building.

But I mean, do you think like in 5 years, uh, like people will still be doing this type of work? Like, because at some point the, the— I mean, once you quit your job to train train the, the AI agent and you've done that, like, you're not going to have anything new to train, you know, the agents anymore.

Yeah. Question for you, do you think there's more money spent on software engineering data today than a year ago, or less?

More, for sure.

Yeah. And do you think there'll be more money spent on software engineering a year from now than there is today?

On training?

On, on data to improve. Like, is software engineer— like, Waymo still has 3,000 people driving cars, right? They've been doing self-driving for a decade, right? The number of exceptions and the explosion in expectations as you get agents to run longer and longer, yet alone, we haven't even talked about multiple teammates interacting, like the finance and the marketing team. Like, there will be, there will be tens of billions of dollars spent per year improving agents to like—

So you're basically, yeah, getting from like, you know, 3 nines reliability to 5 nines reliability, basically all those edge cases We literally—

an agent in oil and gas can run for like an hour and a half. It like can't call the tools, it can't do the job. Like, it's great with PowerPoints, great with emails. But we are seeing, we are seeing such little work automation and acceleration happening in every other area of the real economy. And every white-collar knowledge worker wants an agent that can run for 10 days and do their job and help them do their job and put them in a more accelerated position. And you're not going to find—

Oh, for sure. There's no doubt about that. Yeah. Uh, and, and I agree with you, like, there's probably still lots of edge cases for Waymo that they haven't addressed, but there's also— I mean, there's tons that they've already addressed. So like, the training that they're doing is very different now than it probably was a couple years ago, right? Uh, yeah. Um, so, um, and I think a lot of the training with Waymo in particular is like, yeah, new cities, right? New—

like, you know, you go to a new city, they're driving around, they're not accepting passengers, they're collecting all the edge cases, and then they're going back and they're They're capturing their words to do that with humans. Then it starts to spin and they open up the city, right? And like that whole— I mean, we are— how many— it's going to happen, right? But it's probably a decade away for Waymo. I mean, I'm completely speculating on Waymo, but I do know in White Collar Knowledge Work, like, we don't just aspire to like help a wet lab biologist like with their work. They want to like discover new drugs.

I feel like one thing that's missing in AI, and I would hope in the next 5 years we could crack, is continuous learning, right? Ideally, even with this Waymo thing, ideally the Waymo fails in an edge case. And instead of it being like, hey, now we need to go post-train the model, blah, blah, blah, it would be like someone takes over and says, hey, actually when this happened, you should have done this, or demonstrates it. And then it's like from that point onward, that model has learned this new situation. So I wonder if in 5 years, let's just imagine we have continuous learning and maybe a lot of it is actually how you train a junior employee. Hey, you know, you're new here, let me tell you what to do. And whenever you get stuck, like, let me teach you. And then it just keeps getting better. Like, that would be actually kind of a, uh, I, I think where models need to get. And that would be kind of like an interesting kind of world where, like, you have these models that have been, like, continuously trained for, like, you know, very specific jobs. Or, like, or even, like, within engineering, like, you know, not every company has the same type of engineering culture. Like, you might want In this job, actually, you want to code like this and these are things we care about. So you might have a lot of per-company models that are continuously trained.

A fun line I heard from another founder in this space was, likely if continuous learning is cracked and models can teach themselves, the last company that will ever really exist is a company providing the humans to teach the models. I think the world we're describing where models, there's truly a takeoff of continuous learning, Like, you know, civilization and being a white-collar knowledge worker as we know it today is, you know, we can all speculate what will happen. But, uh, you know, I think, I, in my opinion, we are, we are not close. Like, and I don't think a lot of the researchers really, when you start asking them over dinners and not reading their tweets, think we're close to like cracking. Like, if you look at actual agentic performance today and what's driving the gains of moving these models forward, progress, like from version to version, it's more and more data and a bit of algorithmic process, right? And it's all being done in RL. And so there needs to be like several prize-winning breakthroughs to crack that. And, you know, I'm excited about that world, quite frankly, like with 2 in-laws that have cancer and like, you know, as I get older and have kids and like, it's exciting for a world that we're all trying to build towards. Um, but I think we are a long way out.

You know, um, do you guys read this Bill Gates essay?

I have not.

Came out a few days ago.

Huh? No. What did it say?

No. So actually it just came out this week. Uh, basically, you know, he talks about the future of AI. Um, like, uh, the turbulent AI era is here. The choices we make now are critical. And, um, you know, this was a very interesting essay where he talks about how AI is going to cause, you know, could cause massive job loss if we don't don't, um, you know, make the right choices now. And he basically paints this future of like how AI is already doing this now. He talks about robotics as well. Um, I just recommend you guys read it. It's a really interesting essay. Um, and, um, you know, I, I think—

How do we, how do we avoid massive job loss according to—

I mean, he wants to tax tokens. He wants to like, you know, um, uh, you know, regulate the, the development of AI, uh, which is not something you'd expect a tech person to say. Um, but, uh, you know, he actually, uh, is proposing that as well. Uh, but, but Garrett, this is an area you are a specialist in. I mean, you know, you've started this company to give people more opportunity. Um, and so how do you think those opportunities will evolve over time? Um, you know, uh, you know, we all want people to feel purpose in life and, and want to have, you know, rewarding jobs. Um, you know, I think training agents will be there for a while and then there'll need to be a next phase of this, right?

Where we're really trying to take the company is Trying to create a world where more humans participate in the economy, not less, and more people are advantaged and accelerated, not less. And when you start to zoom into, like, you know, even employees at Handshake, let alone employees at General Motors or, you know, large companies in the middle of America, like, there's— I was talking to my uncle who's an accountant around, you know, he manages a huge team of accountants and, like, he barely knows how to use AI. Like, he doesn't understand the skill file or what an agent is or— If I mentioned computer use, he would be mind-blown. He wouldn't know what— There's like 300 million Americans. We still need to teach the people how to learn these tools and put them in a position of acceleration. I think that you talk about job disruption, there'll be a lot of job categories that change immensely. There'll be a lot of people that will need to upskill and reskill and reposition their experience. I still believe human beings will play a very valuable role in the economy. And I think if you are not getting with the program of picking up how to actually accelerate yourself in a role, like as an employer at Hinchcliffe, I don't want to hire you, right? And I think like that will start to disseminate out to every company where the prerequisites and knowledge you'll need to have of managing agents that are doing multi-day jobs will become the status quo to get a job, right? And where I really want to take this is like helping people understand the jobs of the future, helping people understand how to upskill and reskill them for the jobs of the future. And really, in a lot of ways, like helping the agents actually get better and better at accelerating their jobs, and helping them really spend their time on the most valuable use of what they're trying to accomplish at the company.

Fair enough. What do you think?

It's really hard to predict the future. Like, you know, we've seen— almost every company has seen this massive acceleration in engineering. And, you know, I think like engineers are at least outputting like 2x more than they were a year ago, but we're still hiring more engineers. So, and actually funny, funnily, we were talking about this and the problem is like actually all of our competitors are also seeing 2x more engineering output. So like the only way you can like keep up is, you know, you have to go like, okay, you know, I have to go 4x. But even that, like, you just have to— like, the requirements for what a company has to do in terms of, uh, in terms of product are just getting higher. Like, more product velocity, more, more personalization, more kind of, you know, unique features. Uh, like, it's— so it's like kind of a treadmill, right? Like, I think this is why I'm always skeptical of these, like, one-person billion-dollar company, because the problem is the one-person billion-dollar company is going to get outcompeted by the 20-person billion-dollar company. Uh, there might be a moment in time where you could capture like a biggish company, and I do think companies will get more efficient, but, but there's just this like, uh, competitive pressure that forces like, uh, you know, you using all the resources that are like beneficial, and humans are still like a beneficial resource. Uh, I think the interesting thing, and like, you know, this gets into more on the sci-fi stuff, it's like, yeah, when we get to continuous learning and these models basically just like an employee. At that point, that's full AGI to me. And all bets are off at that point. It's very hard to predict a future where you have human-level intelligences that can learn on the job and do almost everything a human can do. I don't know where we go there. Then maybe they're compute constrained and we can still compete against them in our own way. Or maybe everyone's doing human-to-human services, right? Everyone's like doing massages and like, you know, other things, like being a masseur, like other things that only humans like will do, or being an artist, right? I don't know, or sports, right? There's a, there's a lot of things, uh, that humans are like probably uniquely suited to that we other humans would only want humans to do well. Uh, but yeah, very hard to predict the future, but right now I would say it's an accelerant.

I mean, just, just, just, just to make a joke around that, like, I, I'm not super excited for my son To, to, uh, if he wants to be a masseuse, so be it. If he wants to be an artist, that's amazing. But like, the worst— like, I, I think just to be provocative, like, unless he wants to do those jobs, I'm not excited for him to be a masseuse or an artist. Like, I want him to be able to make an impact. Like, I— right? Like, I think that the conventional answers are like, my son's going to play board games for purpose, or we're going to all live in like— communities and draw art all day. That's not the future that I want for my 1-year-old son. So I do think the industry needs better answers on what jobs of the future look like, how to get people there, how to put them in an accelerated position, how to really enable all classes of people in the economy to prosper. But I don't know, just to be spicy about your exact answer, do you agree?

Yeah, I, I think there's like a— I don't know if you're quite getting to this, but there's like some value. And like, we probably want to be builders and we want to like push science forward and things like that, which like, you know, I guess the world we're painting with like AGI, maybe like we can't do that. Have you read the Culture book series?

No.

All right, it's got— it's like a post— Elon Musk is really into it, but I did read it before him. Uh, but it's very into this kind of post-abundance world and like what might look like. And like, you know, one of the ideas in it is like, yeah, maybe humans are like given resources, right? Like, it's like, okay, you know, this— yeah, if they want to, they can go like shape a whole planet, uh, because this is like super post-abundant. So, uh, but like, they're, you know, so, uh, like maybe we find a way to allocate resources. That's cooler than playing board games.

Yeah. Uh, yeah, I, I think humans are naturally drawn to challenges. So like, they will— we'll just find other problems that we want to evolve, um, you know, and there'll be some, like, you know, it will seem like to, to like the human, um, populations like 100 years ago, it will seem totally unnecessary. Like, why would you want to go shape another planet, right? But for us, it'll be like, oh yeah, there's nothing else to do, this is what we're going to do now, you know. It's the same thing as like we have a lot of humans working on video games, you know.

I think 100 years is probably not far enough, but like if you go like 1,000 years ago, right, like most people found meaning by like hunting And like having, yeah, like all these things that like we don't do at all now. So yeah, maybe the things that we think give us meaning today, like sitting in front of a computer to like, yeah, write some code or something, will like not exist in 1,000 years.

Right, right. No, I actually do think like these job categories, these knowledge work categories will be gone. Like I think if we won't be doing that much of these knowledge work stuff, so it will, but there will be other stuff that humans want to do and that will seem very incremental to us now. But will seem very interesting to us in the future once the knowledge work has all been solved. All right, and and so like I think we we can't even envision them yet because we don't even know what that world looks like. But it's gonna it's gonna happen. And so I I think one point that Paul Buheit mentioned to me that stuck in my head is that like we're already living in a world of abundance now. And and you know if you compare what our world looks like today. to like, you know, 100, 200 years ago, right? It will seem abundant to those people. Like, a time traveler from the past coming here would be like, you know, what do you mean I can buy fruit from any country in the grocery store? You know, you know, what do you mean? You know, all— what do you mean I get electricity in my house for like, you know, you know, um, for cheap? I don't have to go outside to a bathroom. Like, like, this actually would seem abundant to someone else in the past. And I think if we were to time travel into the future, You know, things that would seem difficult today would would be free, you know, essentially or close to free tomorrow. So, but then there'll be another set of problems that we can't even envision right now. So, you know, and so yeah, I think as long as we don't destroy ourselves in the human race for whatever reason, but but I think we'll figure it out and we'll adapt as we always have. You know, so it's it's an interesting theoretical question. Not theoretical, actually. It's a practical question. It's but it's It's, um, I, I think about a lot.

I do think that the tricky part is like, I think we'll get there, but society needs to change. And often like these societal changes are like quite disruptive. Like the Industrial Revolution was like clearly great, but like extremely disruptive, like led to major war, probably led to communism. Like, yeah, there's a lot of things that came from like the Industrial Revolution, uh, that, you know, we think now amazing, but like the society changing to like match that level took a long time and was like very destructive. So yeah, how do we do this one a little smoother is probably like a good question.

For sure. Gary, this has been super fun. Really appreciate you joining us and entertaining our wild questions about the future. Hopefully it wasn't, you know, we didn't ask you too many tactical questions, so I hope you enjoyed that.

No, it was great. I loved it, guys. Thanks for having me on.

Thanks for joining.

And what an insane story for Handshake. Like, I'm really impressed you've kind of guided the company through this time. Congrats on the success.

Thanks, buddy.

Thanks everyone for listening. Please leave us a review on whatever channel you're listening on. Makes a big difference to the algorithm and to the success of the show. And see you all next week.