The Grammarly–Superhuman Megadeal, plus TWiST 500 chats with LabelBox and Apptronik’s founders | E2146

27:43And it starts to look awfully like an email client at the end of the day. Email actually is nothing more than a list of things to do. And that is what we're building. It's what we've always been building. I'll add two quick things. And Alex, I know we'll get you onboarded to the product again soon. I think just first off, that the task of dealing with email is its own interesting task. This seems like an odd thing to brag about, but I've been a superhuman user for years and I had Rahul pull the stat for me. So I have currently 144 week inbox zero streak, hitting inbox zero every week for two and a half years.

28:20And that is largely credit to superhuman. So if you view this as a chore, we can do much better than that. I think there's a next level, which is now do you communicate better? Do you communicate more effectively? how much faster do you respond to emails, what's the quality of the responses. And then you start to think about it the way Rahul was just describing, that it's your workbench for working with your agents. And part of our view of it was, why do we care so much about documents and email as being two surfaces we really wanted to have the best first-party experience in, is the places you work with humans are the natural places to work with agents as well.

28:58That's how we think our teams are about to expand. It's a very natural surface for us. On the agent that follows you around, I love that idea. I love the idea of having an assistant that knows my context and can help you wherever I am. Is that something that you can build with technologies as kind of basic as a browser extension? Or do you need a different vehicle to bring that to market so that way it can interact with me across my workflow? Because not everything I do is inside of a browser. We're all familiar with AWS, Amazon Web Services. That's the cloud platform that powers so many of your favorite brands.

29:33But do you know about AWS Activate? That's their program for startups, where they provide up to$100 ,000 in AWS credits for all startups. Whether you're backed by an investor or you're bootstrapping, money is time to keep innovating, time to delight customers, and time for you to keep gaining traction. You need runway, and AWS's Activate is going to help you with that runway. We hear this story from so many of our founding university companies. They're finding product market fit, the words getting out about their product, just a little boost. Finding some savings here or there to get a little extra time can be the difference between getting traction and bringing in revenue, or, hey, let's call it what it is, running out of money, okay, and shutting down.

30:13AWS knows this. That's why they've created the ultimate toolkit for early-stage startups looking to boost growth. With AWS Activate, you're going to get up to$100 ,000 in AWS credits, hands-on support and training, plus exclusive discounts with some of our favorite companies and tools. So start getting the support you need at every stage of your startup journey. To learn more, visit aws.amazon.com slash startups slash credits. That's right, aws.amazon.com slash startups slash credits. Yeah, that's right. Right. I mean, I think the first off, the way Grammarly works, and if we think about Grammarly is really I think it's a very misunderstood product because I think people understand it as as a proto agent, as that OG agent.

30:57But actually, it's two separate products. You know, there's one we think of as the agent platform that brings AI to you everywhere. And the other we think of as your agent. That's the that's your high school grammar teacher in your pocket. And the platform is actually where most of the technology has gone. So Grammarly currently works in your browser. It works on your desktop. There's a desktop application that works on it and on your phone. So there's a, both on iOS and Android, a version of Grammarly that can work across all your applications. So across that, we work across 500 ,000 different applications.

31:30And yes, that's everything from your browser apps, Superhuman, Gmail, so on, but also your desktop apps. Slack is a great example, but people use Apple Notes. People use all sorts of things that are desktop applications as well. And we've done the hard work to be able to read what's on your screen, be able to annotate it in a way that is unobtrusive to you, and then be able to take action on your behalf. We call that the AI superhighway. That's the term we use for it, bringing AI right to the end user. And our view of it is that right now we're only running one car on that highway. We're running your high school grammar teacher.

32:06That's the one car. and we're expanding that so that anybody can build an agent that works across all those different surfaces. So that is a key part of what we're doing. So then Grammarly's kind of back end today is the platform on which many other agents can be built, filling out this AI highway with other cars, to your analogy. That's right. And you're going to see a big shift in the product in the next couple of months and that'll become much, much more obvious. We've been working really hard on it and then we'll start opening up the platform so anybody can do it. I mean, that could be, like we described a couple of examples earlier, it could be things that connect to a system.

32:39Imagine your Salesforce agent wasn't like a Salesforce app that you had to go to, but rather an agent that was with you everywhere and had all your sales data at your fingertips and was just going through and marking up your, whether you're writing an email or a document or so, and marking it up as you go. You can imagine agents that represent people. So I want an agent for, I'm a big fan of Kim Scott's Radical Candor and Kim's building an agent that is, hey, you really want to follow this? Don't read the book and stick it on your shelf, but really have this person that's sitting next to you saying, hey, you're kind of not following the principles of the book right now.

33:08Here's what you should be doing. So I think that market of the set of people I want next to me as I'm working is sort of limitless. And we want to be the platform for doing that. So just to be clear, you're saying that Grammarly is eventually going to have an agent marketplace where other people can build things and deploy them onto the foundation that we discussed that Grammarly has already built that interacts across a user's computing environment. Yeah, that's right. And it's being built on top of a marketplace we built in Coda called the PAX Marketplace, which is already one of the most thriving marketplaces in the world.

33:39There's about 8 ,000 or 9 ,000 of those PAX that have been built, there's 800 or so in the gallery. And so there's already many of those agents already exist. And we're taking those and bringing them over to the Grammarly platform. All right, Rahul, I want to get back to you because when this news dropped, first of all, I was a little bit shocked. I've since kind of figured out what you guys are doing. Frankly, I quite like it. Over on social media, Some people were a little bit confused, expressed concern that their favorite product is going to be taken into a larger company. So for folks out there who are current superhuman users who may not be Grammarly customers, what assurances can you give them, if any, that their experience is not going to become, to use a common phrase, Salesforce?

34:18Well, without commenting on Salesforce, anyone who's watching this has mine and Shishir's absolute assurance that whatever that is, it's not going to happen. Shishu, as he mentioned, has been using the product since 2017, credits his success in his email to the product, and the entire team is moving over. This is an acquisition not just of a product, not just of a team, but of a business, an ecosystem, a brand. It's something really, truly very special. And we have every intention of fully accelerating what it is that we're building and what we're doing here. As a founder, one of the most exciting things about a deal like this is that you have access to significantly greater resources.

35:00So we're going to invest more than ever in AI, as we've talked about. We're going to invest more than ever in our core email experience. Even after building it for 10 or 11 years, we still have a laundry list of things that we want to make better. We're going to build out calendar, and we're actually very close to that. Anyone who's in Superhuman today will see that. Tasks, we're going to connect calendar and tasks and email beautifully together. We're going to reimagine chat, redefine collaboration, pulling on everything that we've learned about work communication for the last 10 years. And then, of course, as Shashir said, we're going to connect all of this up with a new way of working with agents, which we think is going to free all of us up to be more creative, more strategic, and more impactful.

35:39So I get it. It's reasonable to be concerned in a time of massive change. But one of the things that really appealed to me about this particular pair up and Shashir as well being a genuine user of the product for so many years now is this goal to build, that he mentioned to me in the past, to build a stable of some of the best AI brands in the world. And that's very different. He gave me some examples, for example, of when Facebook acquired Instagram or when Facebook acquired WhatsApp or how Google has treated YouTube. That is how we want to think about superhuman at Grammarly. so a crown jewel the one that retains its own identity and will be a piece of a larger puzzle i like that a lot and i'm sure that a lot of folks are going to be very glad to hear that their superhuman experience is not going to change uh now rahul you sold reportive after just a couple years of building it now you've been working on superhuman for a very long time you had a pretty big round back in 2021 i'm curious just from the founder's chair why was is this the right time to sell?

36:46That's a great question. I think there's a few things that people often say, oh, like, was it a growth thing? Or like, is it a runway thing? And, you know, mostly for the insider baseball and for the other founders, I'll share the following. We were actually growing faster than ever. We grew twice as fast in 2024 compared to 2023. We had runway for four plus years, we would have been break even, probably early next year by this point. So it truly is none of those things. It really is that technology is changing faster than I've ever seen in my whole career. And towards the end of 2024, I think a number of companies realized the magnitude of the apps and agent opportunity that we just talked through.

37:31And also they're looking at their own products portfolio. They're realizing they have an email size whole. And I think one of the skills every founder needs is to spot the zeitgeist, to realize when that kind of thing is happening, and then to be able to go and prosecute that opportunity. And so it was obvious to me at the end of last year that there were a number of companies that had that thought. And it makes sense. There's, as I said, north of a trillion hours of time going to email every single year. There's no other work app that we spend more time in. Superhuman is almost actually definitely a one of one asset.

38:05Like we spent 11 years grinding to get to this specific point. And so that was why I knew it was the right time. All right. From the buyer's perspective, this idea of an email shaped hole in everybody's productivity suite, I think is absolutely correct. I think there's going to be a race to build that new AI needed productivity suite. And I felt very lucky to be able to work with Rahul and, come up with a deal that I think will produce great impact for our overall company, for the current superhuman users, but is really a one-on-one product. I mean, there really is, the next best new email products on the market are almost unheard of.

38:43And so I think it's a great partnership to be able to bring that to a much broader group of people. Yeah. And it's not like Gmail has gotten a lot better in the last 10 years, let's be honest. Now, Shashir, you recently secured a billion dollars worth of growth capital with General catalyst. Pretty interesting way to fund growth of the business. I really liked it. You did note that some of that was earmarked for strategic acquisitions. I'm curious, did you use that capital for this deal? And did you have Superhuman already in mind when you secured that capital, or did this come afterwards? Rahul and I have been talking for a bit.

39:14So yes, it was in mind as we went through it. The specifics of the deal aren't being shared, but it was an all-stock deal. So the capital allows us to take bigger bets here. But in general, we're working through how to expand our offering as fast as we can. And having access to an extra billion dollars is a pretty good way to do it. Rahul, was that large pool of capital that you'll have access to now to probably use to grow superhuman an incentive? Or are you more interested in the engineering resources that Grammarly can bring to bear on your, quote, laundry list of to-dos? Well, definitely those two.

39:54Distribution, greater resources. I mean, there are so many reasons to join forces here. It was almost a no-brainer. Pretty clear, actually, in our first conversation that Shashira and I had. Wow. Okay. So, Shashira, then just to wrap us up here, because I think everyone wants to know how this is going to play out in the real world, but your company's enormous. 700 million ARR, making a lot of big deals. So, IPO plans this year, next year, people are talking about everyone getting finally out. Are you one of those companies? Nothing to announce now. I think we're pretty busy building. I think it is quite the gold rush right now as the world adjusts to AI changing many things that we just didn't think were possible.

40:34I mean, the idea that we're going to have agents writing and rewriting all our emails for us just seemed kind of crazy. And now it's completely possible. And, you know, we're very focused on growing that. and when we're ready, we'll take it out to the public markets. All right, well, here's hoping that you guys can eventually take these agents to become so smart that they can, instead of writing the email for me, simply contact someone else's agent, handle the communications and bring back the context I need because then we can just cut out the whole middleman and short circuit email for the future.

41:03But in the meantime, Rahul, Shashir, thank you both so much for coming back on the show. Congrats on the deal and I guess go build. Thanks, Alex. Bye.

41:16Now, you may have seen a major news item in the last couple of days. Meta and Scale AI have come to a$14-15 billion deal. That investment threw an enormous amount of light onto the realm of data labeling. And in the coverage of that enormous news event, a company called Labelbox said that they might generate hundreds of millions of dollars in new revenue this year. Naturally, that got us thinking. We looked into the company. We absolutely had to learn more. So please join me in welcoming CEO and co-founder of Labelbox. It's Manu Sharma. Manu, welcome to the show. Thank you for having me, Alex.

41:46Really excited to be here. And if you're watching the video version of this, I will note that Manu is not using a virtual background. He is on the Stanford campus, which just goes to show, I think, the connection between AI work and academia is holding strong. But Manu, just to start, what is data labeling? And can you tell me a little bit about the difference between human-led data labeling and machine-led data labeling? Yeah, Alex. So nearly every AI breakthrough that we have experienced since AlexNet, all of these models have actually been trained on human-generated data. Back in the days, you know, when around maybe, let's say, 2017 to 2020, supervised learning, which is a technique to build computer vision models.

42:29You may have seen lots of self-driving cars work starting popping up at that time. They were all trained with human-labeled data. And kind of labeling at the time used to be fairly trivial. You're kind of looking at an image and saying, hey, that's a car, that's a pedestrian, and things like that. And that fundamental technology or technique was widespread for all kinds of computer vision applications. and after kind of the transformer kind of innovation or breakthrough appeared, a lot of the focus went into kind of making that unsupervised learning stack work. And turns out, even in that paradigm, you know, these models are learning from, let's say, all the web data, which, of course, is human-generated for all these decades on the Internet.

43:17And then there's also a technique called post-training where the very experienced humans are teaching these AI models to be better assistant. And now the state of the art is that these models obviously are very capable across the board. However, they still lack reliability with agents and performing day-to-day tasks. they're getting better, but even those kind of skills have to be taught by, through human data, believe it or not. And now, you know, there is sort of perhaps a misconception that, you know, like it's kind of machine-generated data or human data. The state-of-the-art or the frontier is essentially hybrid data.

44:08Like there are, like most of the latest cutting-edge data sets that that are being generated that improves these model capabilities are fundamentally kind of AI assisted. So these experts, humans across wide range of domains in human knowledge, coding, mathematics, physics, you name it, they actually use AI systems to kind of produce the very best signal that is then used to enhance the capabilities of the models. So we've had machine-led data labeling. we've had human-led data labeling, and now we have AI-assisted human data labeling. Do you think that's going to be the state of the art for a long period of time, or is this just one more iteration on a cycle of constant change?

44:54So I believe that there is a long road to make these models highly capable and kind of be fully integrated in the economy. So when you look at kind of the adoption of these AI systems, They are obviously amazing and really changing the way we do work. However, these agents are still not as reliable. These agents still need to be taught how to kind of fully automate a task, you know, whether it's maybe even software engineering, like, you know, you're coding with humans in the loop, essentially. But I think, you know, when will be the time where AIs can actually write code for days? or weeks, and humans are sort of like steering or sort of managing them.

45:45And these examples kind of are across the board. You know, when will the AIs be able to be 100 % act as a customer service on your business behalf and sort of like resolve all the issues? And, you know, these kind of capabilities are very lucrative to solve, and they require a ton of new kinds of data. And you have to generate this data somehow. And one of the best techniques that we are pioneering is essentially using technology, AI, and software, and workflows with human experts to very cost-effectively capture human expertise in the data, which then is used to train these models through reinforcement learning.

46:38All right. Now, when I think about you guys, I think that what you described falls into your kind of data factory concept. Can you explain that for folks who might not be familiar with what you guys are saying there? Yeah, for sure. So in the data labeling business, or let's call it like data generation for post training for these models, there is a number of kind of ingredients that goes into it. So first of all, you have to have a network that is attracting and assessing really high-end talent who exhibits certain knowledge that is superior than the AI. It could be just sort of like a base kind of IQ, like humans can have high agency to learn new things and kind of do things that AIs can't do.

47:30But then also there is kind of deep expertise, you know, whether that's, let's say, an aircraft design, it could be medicine and healthcare and so forth. And so you have to have that sort of network and bench. And that's really just a starting point. And so you may have seen like a number of companies popping up as a expert staffing agencies model. And all they're really doing is helping some of these AI labs to hire talent. Now, we take that idea further. We call it Data Factory because you kind of need a workforce to obviously operate in some sort of a setting to actually produce human data.

48:09So Labelbox has been in business for over like seven years building that technology platform. So we build all the tools, the workflows, and techniques to actually produce the data. And so we have a network. And then on top of that, there is a kind of fairly complex and sophisticated technology platform, which uses AI and software and things like that to allow these experts to work in a manner that produces the very best data at a very cost effective kind of parameters. and that verticalized stack, what we call is data factory. And sort of this analogy kind of really holds where, you know, when you're building a factory or you're operating a factory, you've got to have infrastructure, machines, assembly lines, the workflows, the practices, but then you also need the workforce.

49:01And LabelBox is fully verticalized in the sense that we actually produce the data and offer that to our customers that in a manner, you know, They can actually get the data very quickly in a matter of hours, if not days. And it's really the data that these AI labs want ultimately to continue to win in the race. Yeah, it's interesting because if you go back to the SaaS era, there was a lot of discussions about software versus services and how you wanted to be a majority software business. You might sell services at cost simply as an enablement tool for your software business. But it really feels like the way you're describing the data race for AI, there is a strong human component in it, at least at the expert level.

49:47How does that impact how you price and sell label box services? Because I'm curious how you think about the two different sides of the data factory. Yeah, totally. So it's a fantastic question and observation. So, you know, for the first five years of our company, we primarily sold SaaS. We basically built the world's leading data labeling platform, and our platform is used by over 30 % of Fortune 500 companies. And in this context, our customers are bringing their own workforce and operating the entire data labeling operations on our platform. And in those contexts, it would be sort of like a SaaS-based pricing systems where it's based on how much consumption or data customers are generating, and they're paying some kind of a correlated cost to us on that.

50:39And when we decided to fully verticalize and get into kind of data services business, you know, just over a year ago, now we are pricing essentially on the data, like how per example, perhaps. And sometimes it's the value that our data is bringing among our customers. So, for example, just a couple of weeks ago, we enabled one of the leading AI labs to increase the state-of-the-art performance by 15 % on a data set that we created for them. And that could be extremely valuable to that company in the order of hundreds of millions of dollars of value generated. The model just bumps up in a leaderboard, and they are able to onboard more workloads and things like that.

51:27And so that is fundamentally like how we are pricing. It's based like on the value of the data we are generating for the customers. Is it hard to price? Because I feel like value from data is squishy. And you guys, if you're pricing that way, probably want to charge more because you think it's more valuable. And they might want to say, oh, you know, it was only so good. So how do you handle that tension? Yeah, there's always that kind of a discussion around the data and the cost and the value. and it is fundamentally hard to price and it's a new frontier for not just us but many other companies, the native AI apps companies who are essentially like a very similar model where they're kind of automating some workflow and so forth.

52:11And in our business, particularly the challenging part is that every few months, the state of the art is kind of at a much higher end. And so, for example, a year ago, a lot of the data we were generating was around kind of basic assistance, like teaching English skills, like how to write great essays. Now we are producing oral environments for these labs to kind of do this reinforcement learning with verifiable rewards. Very different technical skillset needed, very sophisticated data sets. And, you know, ultimately data is actually becoming more costlier. And the reason is that the expertise needed is becoming more kind of more complex or more nuanced.

53:00However, with that said, we are proud to be saying that we are actually one of the leading companies in the sense that when you look at the data and the cost, we are able to, we are the unmatched in that performance or the empirical curve. And we are able to do that because of a very strong technology stack that is operating on top of the kind of the expert network. I really appreciate that answer. That explains it kind of top to bottom. Mano, I want to go back in time. You guys last raised, the publicly known round was$110 million led by SoftBank, and it was announced in very early 2022. That was pre-ChatGPT's launch and what you might call the starting gun of the Gen AI era.

53:44I presume the market moved towards you in that year. So how much did the release of ChatGPT speed up demand for your company's services? And then if you can, how much did you guys grow last year? Yeah, so the demand has increased substantially. I mean, it's really all engines on fire, you know, since actually like when we decided to enter the data services market. So kind of a bit of a history. Just about up to less than a year ago, we decided to enter data services market. Up until then, we were essentially selling enterprise platforms across the world, across these enterprises. And I'm proud to say that we are powering most of the AI labs now.

54:33And so we've grown from basically entering the market to now powering substantial workloads for nearly all the big AI labs. And, you know, we've been growing very fast since then and basically in hyper growth kind of trajectory. And I think this latest news, kind of the market dynamics that over the last two weeks has further catalyzed or accelerated that growth. And, you know, so it's been really exciting times at Labelbox. And, you know, there are challenges around like how do you scale that quickly and, you know, meet customers and all that. And so, you know, I call it still to be a challenge, but at least it's a fun challenge.

55:17I do want to just verify something. You did say that the sale of scale to Meta and its customers leaving could actually generate hundreds of millions in new revenue for LabelBox. Is that a 2025 metric or is that a total contract value over kind of a multi-year period number? We expect substantial revenue over the next 12, 18 months for sure. We're experiencing it already right now at a much higher steep curve. and it gets harder and harder in larger numbers, but we are very, very excited about what's next in the next 12, 18 months. I appreciate that. Now, with the exit of scale, it could kick off some consolidation in and around your space.

55:58You mentioned some of the startups that are kind of popping up in your domain, admittedly targeting kind of a fraction of what you're doing. But I'm curious with your history of capital raises and success of the company and growth, are you guys interested in going into the market and maybe buying some smaller companies? We do get quite a few interests or inbounds almost every other week from smaller players or companies. And we are always assessing when the right fit is and when it makes sense. So certainly, we are very, very open to it. A lot of it really comes down to the complementary strengths and weaknesses and sort of that match.

56:42But we are actually seeing a lot of M &A activity and interest kind of boiling in the market across within our industry. Just because I have you here and I get to ask you this, by M &A activity bubbling up, is anyone trying to buy you or is this all companies that are reaching out to you as a possible acquirer? It's the latter, yes. Okay. Any names that... It would be wise to say that here. It wouldn't be wise to say those things, but, you know, but let's just say like many, many players are interested in joining forces with sort of leading companies and and so forth. So, you know, I think like any other category, right?

57:27Like there's a lot of enthusiasm. There's a lot of interesting players that have different techniques and different bets that they make. And at some point you expect like, you know, there will be some sort of consolidation. And I think now that the markets are more opening up for M &A, I think we are feeling it. We're seeing that through just like all kind of movements. Absolutely. Now, one thing that I was surprised by, just given how well the company appears to be doing from where I sit, is that, at least publicly, you guys haven't gone back to the capital markets and raised more. Is that simply because you've been running the company pretty efficiently?

58:04Yes, that's right. We've generally ran the company fairly conservatively when it comes to capital and efficiency and things like that. And we really wanted to be very sure and intentional when we go into data services market that we are going to enter to win and have all the key components and capabilities needed for us to demonstrate that. I mean, just over a year ago, we were not doing anything in this space entirely. And so in just like last four or less than four quarters, now that we are plugged into all these labs and we are winning the market share from kind of the other providers, it's really exciting.

58:52And I think we might do something in the future. So one thing that struck me as interesting about your background is that you've worked with drones before. And kind of a tangentially related point, I was also surprised to see that Uber is getting into the data labeling game in a more serious way. What's going on with people that deal with real world hardware systems getting into the data side of AI? What should we take away and learn from that? Alex, so I've been very interested in technology across the board. And so when I was in school, I actually was doing a lot of research in neural networks.

59:30In fact, I remember my earliest times was I was trying to make flight control system in airplanes, like powered by neural networks. And you don't believe like at that time, the neural networks were like three or four layers, and you could count the neurons in MATLAB and Simulink. And, you know, it's just a testament on how fast the technology has moved just from like a decade ago. And so in engineering, of course, you know, a lot of my work was in genetic algorithms and like these algorithms to optimize things. And so a lot of my interest kind of, I suppose, when I was in software technology, startups in Silicon Valley drone deploy and Planet Labs, I saw computer vision starting to take off.

1:00:12And then I sort of like all these things kind of came to me and said, like, hey, you know, I would love to build a company here. And I think there's an opportunity to build some enduring company that kind of sits in between this what we call AI and humans. I do think that, you know, over the next few, like, we are moving into a world where we will be steering multiple, if not millions of AIs, and we would still want to be able to manage them and kind of keep them aligned. And I do believe that to do that, you still have to talk to the AI in some fashion. And the data interface is essentially what LabelBox is trying to build.

1:00:54And the techniques and all these details will continue to evolve. Just how the labeling has changed over the last six years, I think it will be very different in the next few years too. But I think the intrinsic need to continue to manage the AIs will remain the same. And so we are kind of at that frontier and always want to build awesome tools and services there at that layer. Manu, if I'm tracking you correctly, you're saying that the way to speak with or interact with AI models is essentially through bringing them data. Is that correct? That's right. That's right. 100%. The way you ensure AIs are behaving the way you want it to do, or it's representing your businesses, or it's performing certain agentic tasks, you still want some checks and balances.

1:01:45You still want some evaluation data sets to test it before you roll things out. Or when things go wrong, when exceptions happen, you want to be able to have some sort of a human in the loop to make that sort of judgment, like, hey, why it got wrong and how to teach it to get better. And, you know, I think we are very early in the innings of, you know, in AI adoption. I think it's going to power a ton of the economy over the years. And there's going to be just insatiable appetite for new forms of data that ultimately represents human preferences and human intent to translate that to AI systems.

1:02:29Manit, thank you so much. What's the URL and where can people find you on the internet? So, labelbox.com. I'm actually not as kind of present on the social media, but you could probably find me on LinkedIn, old school guy. And on Twitter, I have a handle ManuAero. Yeah. All right. Well, please enjoy Stanford in my stead. I miss California. Manu, we'll have you back on in a couple of months. Appreciate you. Awesome. Thank you for having me, Alex.

1:03:00Today, we have another Twist 500 interview for you. Today, we're not talking about just LLMs up in the cloud. No, we're going to go back to planet Earth and talk about physical things, robots. But not just robots. My favorite subcategory of that group, it is humanoid robots. There's a lot of companies working on this space, everyone from Tesla down to a number of startups, and they're making really big progress, moving us towards a world in which we are not as reliant on human labor as we are today, but instead can ask our friendly robots to do some stuff for us. So please join me in welcoming to the show.

1:03:30It's Jeff Cardenas from Aptronic. Jeff, how you doing? I'm doing pretty good. Thanks for having me on. I want to point out that I did ask about the cactus and his background before we started. And he said, quote, it's a Texas thing. So I didn't know that. But Jeff, I do love to learn. And I'm glad you've taught me about, I guess, Austin-based botany. Is it actually a Texas thing or is that more of a you thing? Well, I think what I said was I'm from Texas, so we have to represent. But no, it's a joke. Texans are very proud of being from Texas, though. So I like to represent. I have noticed that.

1:04:01I have noticed that. All right. So Aptronic is a company that's been on my radar for years now. I think pretty much since you guys announced Apollo, that was back in 2023, it is your humanoid robot. I think it's five foot eight inches, about 160 pounds, can carry about 55 pounds and has a relatively smiley and happy demeanor to it. So just to start, what did I miss there and what has changed in the Apollo model since it was first announced? Yeah, I think I mean, I think you covered the basics pretty well there. We've made a lot of robots, particularly a lot of humanoids since we got started. We're on our 15th iteration of robots since we got started.

1:04:41We've done about 10 iterations of humanoids. I think, you know, the key thing for us is really a much more robust version. So the first version that we showed of Apollo 1 was really an engineering prototype, not designed for mass manufacturing. We only made a handful of those systems. We have the new version of Apollo. Apollo 2 is now built and online in the lab, much more robust, much more scalable, really wins in every sort of performance category we care about. Still roughly targeted at the same payload, a little bit larger battery to achieve that four hour runtime that we're targeting. A lot of times there's a caveat in robotics where if you look at runtime of the robot, it depends on what it's doing.

1:05:24So if you're lifting peak loads the whole time, you're going to get one runtime versus you're doing something light duty. So we've learned a bunch from Apollo 1 and then poured that all into Apollo 2. But we'd already done a lot of humanoids. We had a pretty good perspective of what we wanted to build. And Apollo at the time was the robot that we always wanted to build and really just continuing in the same legacy that we started with. So you said the word robust a couple of times. I'm curious what that means in a robotic context. Does that mean just literally more durable or does that mean just more capable of a wider array of tasks?

1:06:01I'd say both. You can think of the hardware as you don't want to be hardware limited, right? So you need a stable platform that you can train and build models on top of. So if the robot's breaking every day or whatever your mean time between failure is, you can only make so much progress. There's a huge overhead cost of maintaining unstable hardware. So sort of the foundational thing you have to get right in this space is you need a stable platform that you can build on top of. And once you have a stable hardware platform, you can start to build more. You can make them more affordable and cheaper.

1:06:38But kind of that's the base case thing that you need. And these humanoids are pretty complex as you really dive into it. So that was the key thing we wanted from the next version is a really stable, robust platform to build on. So going from Apollo 1 to Apollo 2, is Apollo 2 sufficiently robust and market ready to kind of go past the testing stage with partners? And we can talk a little bit about with whom you're working in a second. But is Apollo 2 a robot that you're ready to kind of take to market in numbers? I mean, I think we're probably going to build, you know, in the hundreds of Apollo 2s.

1:07:13So I would say we're still going to be piloting. I've tried to be very honest about where I think the humanoid market is at overall. You know, the thing I like to point out is humanoids are something that humans have been thinking about for thousands of years. You know, we conceived of humanoid robots before we conceived of computers. The idea that we could build, you know, mechanical machines that could do things we didn't want to do. And we're at the very front end of the humanoid market overall. And so we're still going to be piloting for all intents and purposes for the next year. And we're going to be learning a lot from these pilots as they advance.

1:07:51And I think it's important to sort of define what we mean by a pilot because everyone defines us differently. But we're not at the point where we have a turnkey humanoid that you can drop into your environment and you can just pay it to do work 24 hours a day, seven days a week. I would argue that nobody has that yet, though I think there's several companies that are really on the path to get there. So I still think of Apollo 2 as the next generation of system that is out in the world. It's doing meaningful work, but we're still learning a lot from it. We don't have it dialed in. We're ready to start building tens of thousands of these things just yet.

1:08:29Yeah, well, there's a range there between can drop it in and it can do anything and needs to be hard-coded to lift a box. So can you just, for me, who's not spending every day in the lab and the warehouse watching these things grow and develop, how far along is Apollo 2 compared to, I don't know, a human laborer in a warehouse setting or in, say, a factory setting? So I think the days of hard coding something to pick up a box, those days are over, right? So we have reached this inflection point, not just in humanoid robotics, but in robotics as a whole, to where it's very clear that learning is going to be the future.

1:09:05So really everyone's, that's why there's so much excitement and or hype, depending on how you think about it, is because there's this paradigm shift in the way that we can train robots all together. And learning is clearly the winner and the new paradigm. I think what we're working on now is as we get the ability to do these tasks, we're starting to climb the ladder towards task mastery out in the wild. So can we actually start to do these tasks at the same rate that a human can do them? And in that, it depends on the task that we're working on. So we're working on a range of different tasks, and we're not as fast as a human today.

1:09:41So I'm not going to make that claim. There's narrow portions of a task or a use case that we can start to approach parity with humans. But if you look at the range of things that are involved in even a single use case, we can't do everything at rate for a human. And so this is the work that we're doing is we're climbing this ladder of performance towards task mastery. Does it matter if a humanoid robot can be at parity with humans for a task at speed? Because my thinking is if you can take a humanoid robot and plop it in where you had a human, it can work not 24 hours a day, but call it 22 hours a day.

1:10:18So it has more capacity to do more total time at work. So does it need to be at the same kind of speed parity? No, it doesn't. That's a great point. I think that's something that some folks get confused about. No, it doesn't. it depends on the application though. So, you know, at the end of the day, it's simple ROI math, right? Is how much is the robot cost relative to what you're paying today to do a task? And then what's the rate of that system? And you're just going to do a calculation to determine is the ROI better and sort of moving to this new technology, or should we stick to the same thing we're doing today?

1:10:52In many cases, the interesting thing is that businesses don't have that decision to make because there's a massive labor shortage. So they don't, you know, know, this idea of, do I stick with what I have today? It's like, well, I have a huge shortage and this is a big challenge and a big pain. And so this shift towards technology to humanoids is very interesting, but some of these tasks run already today, three shifts a day. So they're running 24 hours a day, seven days a week. So there's not that bleed over time that you're talking about where if the robot was half the price, you could put, you know, less robots and let it run more robots and let it run longer.

1:11:28So it depends on the application, but there are plenty of applications that might be two shifts a day where you could let the robots run longer, or you could throw more robots at the application, have twice the robots running, and you could still accomplish the same effective amount of work. So yeah, that's interesting. On the ROI point, I just did some quick Googling and I live near Massachusetts. And apparently, according to the internet, Amazon warehouse workers and messages make about 16 bucks an hour. So when I consider ROI for a humanoid robot that might go in and take on a task where you can't find a human, should I be considering like the value output of that humanoid has to be at least like X dollars per hour and kind of set that rate compared to human wages?

1:12:11I guess my question, Jeff, is that just feels a little outdated to me as a concept. And so I wonder if I'm chasing the rabbit down the wrong hole here? I think that's the simplest way of viewing it, but it all depends on really what, you know, what customers are interested in. So there's a whole component where a lot of times in logistics and other things, you might have a line that's actually doing the work that's kidding or sorting something. Then you might have a quality control line where you have a separate process where you're checking to make sure things are done properly. There's opportunities where you can have a robot do both of those things at the same time.

1:12:45And so then you can look at that equation differently. There's also a sort of fully burdened labor rate. If you look at many of these applications, we're looking at about a$25 an hour fully burdened rate where you have overhead. Sometimes in logistics, you're paying temp agencies for many of these workers, and they're trying to flex up and flex down depending on the total work that's needed. And so you sort of have to take this holistic picture where it's maybe very rudimentary, but sort of coming in, if you're way out of the ballpark, if you're$40 an hour compared to this$16 an hour, you're just not in the game.

1:13:22But there are more sophisticated ways of looking at it. And it depends on the customer, really. Yeah. Fully burdened labor costs. If you've never had an employee, it might not make sense to you. But if you've had an employee, you know exactly what we're talking about. I want to talk about Google DeepMind, though. You guys announced a partnership with them. Google, of course, is deep in the AI game. I was watching F1 this weekend and they had written Gemini on every single thing they could. And so I'm kind of curious what that partnership unlocks for Aptronic. Is it just a lot of like help support?

1:13:49Is it just cheaper model access? Are you guys collaborating directly on feeding in your data into their systems? What does that look like and how does it help? Yeah, well, I think it's interesting if you look at, you know, Aptronic and Google working together, I've characterized this as the space race of our time. And it just so happened you have Apollo and Jim and I kind of teaming up here. We didn't plan that, you know, that working out that way. But, you know, my view is this is something that humans have been thinking about for thousands of years. And if we really want to solve it, we need the best people in the world working together on this.

1:14:23I think there's another story, which is if you look at the East versus the West, if you look at, you know, what's happening in China relative to what's happening in the West, I think that's another thread that we could pull. So, but my view coming into this was, you know, this is the future I've dreamed about since I was a kid. And now that we're in the window to make it happen, we don't want to have an ego about doing everything. We want to work with the best teams in the world that are sort of mission aligned in terms of what we're trying to do. And for us, it was very clear that, you know, Google really had the legacy and the scale to really help us make a dent and really make some progress here.

1:15:00So what we're doing is a range of things. We're working together in a very deep way. Google's a big investor in Aptronic as well. And we're really working together to try to push, you know, the boundaries of what's possible here, really push the industry forward. So them on the model side, us on the platform side, but then really coming together and testing this out in the world and getting it out, you know, from the lab out into the real world, hopefully in a really big way. Does it also imply that Google might become an Aptronic customer at some point in time in the future? Because they have a lot of facilities, data centers, like they have a lot of physical inventory that they have to take care of.

1:15:36I think the sky's the limit. You know, it's still early in terms of our engagement together. We've been working together for about a year. We have a lot of respect for what they're doing, not just the technology they have, but the way that they've done it. I think if you think about things like AI safety and other things, I think it's going to be even more important as you start to put this on a highly capable, dexterous humanoid robot. lot. And I've always respected the way that they've handled this. And so, you know, we'll see the partnership still pretty early on. We still got a lot to prove on our side as well.

1:16:08But yeah, the sky's the limit in terms of what we can do together. Well, just as a small aside, as you approach commercial viability over the next year, I really would love to stay as at best as possible what you guys are working on. So please keep me in the loop because it sounds like this is the kind of rubber meets the road moment of all the work that has come before. And now, you know, You're out in the field actually testing and trying things. But Jeff, you're not the only one. You mentioned China earlier. I've been thinking a lot about competitiveness between our nation and our economic and political rivals over in China.

1:16:40Now, I think everyone watching this is familiar with a number of American human and robotic companies, Eptronic, figure, et cetera. Where does China stack up compared to us in terms of their similar projects? and do you think we're going to be leading at the national level or do we have some catching up to do? I think we're leading today. You know, the U.S. made early bets in the humanoid space. We did something called the DARPA Robotics Challenge in 2013 to 2015. So we injected, you know, call it a hundred million plus into this sector over a decade ago. Google itself actually made big investments in the humanoid space almost a decade ago as well.

1:17:20They bought up a lot of humanoid companies and really injected a lot of capital into the space even then. So we've had a head start in this space. But China is going to be a real contender, and they're moving very quickly. And they have a national strategy, which is something we don't have yet. We're working on this today. But they're several years into their national strategy. They also just announced a one trillion yuan national fund to fund their domestic robotics ecosystem. system. That's about$138 billion for those that don't want to do the math or the conversion. Just divide by six and a half.

1:17:58It's close enough. Yeah. So, I mean, they're going to be a real force. And I think they have the supply chain and they have the manufacturing prowess and they have the will to do it, right? This is society transforming technology. Think about what an economy is. An economy is productivity per person. Change the number of productive units in an economy, you change what an economy is, right? And so this is something that we really need to take seriously and that we need to really compete. And one of the things I think the Chinese are doing well is they're working together. So, so much of what I hear in Western media is this focus on competition.

1:18:32How do we stack up versus Tesla or any of the other groups that we're competing against? And if you look at the Chinese, they're sharing data, they're sharing models, they're really working together with sort of a national strategy. And so I think that's something that we really need to think about as we sort of play this out because the implications are going to be really big as we move ahead. But there are some great companies over there and they're moving quickly. So I don't disagree from a high level about your point about sharing data, working together, moving ahead as a group. But also your last round was$350 million plus another 50 and change that came on.

1:19:10So there's a lot of money behind you expecting you to win. How do we balance the power of private investment and free range capitalism with the idea of sharing and being a little bit more collaborative? Because to me, those seem to be at odds with one another. But if there is a way to move forward faster as a national industry, I would love to hear it, Jeff. Yeah, well, I mean, I think I should say we're playing to win. You know, we've been at this for a very long time. I think, you know, this is we talked about the partnerships that we have. But I do think there are things we can work together on.

1:19:45For example, we worked with groups like Agility and Boston Dynamics pushing a national robotics strategy. That's something where a rising tide lifts all boats. We should be both sort of funding at the R &D level, sort of more investment in research labs and other areas. This is something that China's done a good job of. And we should also be encouraging uptake of robotics nationally. This is a challenge that we've had sort of broadly in the U.S. is sort of uptake of automation overall across a variety of different sectors. And so these are areas that we can work together on that we can help push together.

1:20:22And, you know, we're still going to be competing as fiercely as anybody else. But, you know, where we can find ways to work together, we should. Trying to avoid politics as much as possible. I'm going to phrase this in a way that I think you can respond. It doesn't seem like right now there's a rising appetite at the federal level to invest in basic R &D. Is that something that could slow down aggregate improvement in the American humanoid robotics industry? Or is that actually not a concern in this particular domain? I mean, the simple answer is yes. I think if you look at just competitively us versus China, you know, if they're going to pour$138 billion into their sector, you know, what's our answer to that?

1:21:06Right. And so people talk about the money that us or figure have raised. But if you compare that, you know, it's a drop in the bucket. And, you know, we do have, you know, Tesla that has a significant balance sheet and we have others. But I do think at the fundamental level, we're still early days for humanoids. They're saying batteries, motors, all sorts of areas of fundamental research that we want to advance. And so my view is maybe that's not where this administration wants to play. So then let's look at other areas that are going to be important. Let's focus on the uptake and the adoption of these robots across different sectors.

1:21:44How do we incentivize businesses if they're going to compete globally and we're going to reshore manufacturing? How do we incentivize those businesses to buy American robots? And this is an area that I think that they can and will play. So, you know, it's early days, right? This is thinking this is the 1980s for personal computers or something like that, where it's the beginning of a major industry. I think we're at the front end of a 40 year cycle here in robotics. And so there's a lot of different ways that we can build and grow this sector as we move ahead. If we're going to make fun analogies, I have to make one in return.

1:22:19And so does that mean with us being in the 1980s equivalent of the PC era today, humanoid robotics, does that make the Apollo one like the Commodore 64 of humanoid robotics? Like, where would you rank it compared to the historical progression in PCs? I mean, I hope it's like the Apple II, right? I was going to say the Apollo II is going to be like the Apple II. I was trying to throw you a bone. Yeah, yeah. No, I mean, look, I think, you know, these analogies don't hold perfectly, but I do think the PC analogy is the best analogy. I sort of characterize traditional industrial robots like mainframe computers.

1:22:53And you can think of humanoid robots as effectively the personal computer. My view is that humanoid robots are the best chance for robotics to scale beyond the limited applications that they're in to the broader market as a whole. We need this general purpose platform that's much more versatile than these robots that we had in the past. And if you sort of take that analogy, I think it's important to point out that it's early days. I think 50 years from now, there's these debates in humanoids that I sort of get a kick out of that are like, are humanoid type? Are they reality? You know, is it actually going to be next year?

1:23:28Is it going to be two years? And it's like in the grand scheme of things, nobody's going to care if it's plus or minus two years. My investors certainly care. And I try to be very sort of honest about where I think we're at. But I think we have reached this inflection point. And I think 50 years from now, everyone's going to be telling their grandkids about being alive at the time when humanoid robots came about and all of the ways that the world changed and improved as we sort of rolled these out into broader society. Yeah, they're also coming at a great time because as the world ages, we need more caretakers and so forth.

1:24:00And, you know, that's going to, I think, open up a lot of spaces for this kind of labor over the next 15 to 20 years, which is long enough to get it right, in my view. All right, Jeff, just one more before I let you go. You guys are building in Austin, Texas. We started the show with that with a little bit of a joke. But honestly, there has been a lot of folks that are moving to the Texas area to build hardware companies. Also a lot in Southern California, especially in a defense context. I'm curious why the company picked Austin and has it proved to be net accretive to the business to be in the little blue dot in the middle of a sea of red?

1:24:33Yeah. You know, I mean, we picked Austin because we're from Austin. So we spun out of the University of Texas at Austin. Hook him. Oh, sorry. I had it backwards. Yeah. Hook him. Close enough. But, you know, my thesis was, though, I sort of looked at what I call the new economy in Texas. And, you know, we are traditionally an oil and gas state. One of the ways I've sort of explained is you can think of us like Saudi Arabia, where the Middle East, where it's not just that we want new industries, we're going to need them over some time period. And we have this legacy industry that's been built up.

1:25:07And so my thesis coming out of grad school was that robotics is going to change. AI was going to make a big impact in what robots were going to do. And we were going to need major domestic robotics companies, major domestic OEMs. And the options were really the East Coast in Boston or the West Coast in California. And for a number of reasons, I always believe that Texas was the best choice out of all of those. I think in robotics, geography matters, right? We have the Texas-Mexico corridor, component manufacturing, and things that are going to be very difficult to do here in the U.S. can be done in Mexico.

1:25:40We can build a supply chain up where we can leverage the existing automotive supply chain in Mexico into robotics. and then we can get anywhere in the U.S. in less than 24 hours. And so this is my thesis 10 years ago was that Austin would be the place to build this kind of company. This was prior to Tesla and many of these other groups moving to Austin. And I think it's been a good bet overall. So I think Texas still remains a good bet. It's still early in the Texas sort of story if you think about tech and you think about Austin, where it's going to go. But it's been great for us. It's been awesome to see, you know, all the folks move here.

1:26:17And it's been really interesting to see these bets that you have when you don't know any better as a graduate student actually start to pay off where people are starting to talk about Austin and humanoid robots. And, you know, we were kind of right about some of the bets we made early on. And we'll see how it develops. See, everyone, it's not just podcasters. And Austin, there's a lot more going on there. Jeff, thank you so much. And by the way, when you do ship my Apollo 2, my address, I'll just text it to your team and you can just have it delivered whenever you want. No worries. Appreciate it.

1:26:47Sounds good. All right, Jeff, what is the website? And quickly before you go, a role that you're hiring for, you're having a hard time landing the right person. The website is aptronic.com, A-P-P-T-R-O-N-I-K.com. And we're always hiring technical talent. So all across the engineering stack, if you're a great engineer, you want to build the future, you want to do it in a way that is designed to have the best future for humans, then you should come work for us and come check out what we're doing at Aptronic. All right, thanks, Jeff. Talk to you soon. Thanks for having me. Thanks.