This Week in Startups - Next Unicorns: Empowering robots to think for themselves via AI with Covariant’s Peter Chen | E1796

Episode Date: August 23, 2023

This Week in Startups is brought to you by… Lemon.io - Hire pre-vetted remote developers, get 15% off your first 4 weeks of developer time at https://Lemon.io/twist Vanta. Compliance and security shouldn't be a deal-breaker for startups to win new business. Vanta makes it easy for companies to get a SOC 2 report fast. TWiST listeners can get $1,000 off for a limited time at vanta.com/twist LinkedIn Marketing. To redeem a $100 LinkedIn ad credit and launch your first campaign, go to LinkedIn.com/nextunicorn * Today’s show: Covariant CEO Peter Chen joins Jason to discuss the future of AI in robotics (1:33), the key concepts of reinforcement learning (13:50), and much more! * Time stamps: (0:00) Covariant CEO Peter Chen joins Jason (1:33) AI’s role in robotics and its value in e-commerce warehouses (7:38) Lemon.io - Get 15% off your first 4 weeks of developer time at https://Lemon.io/twist (8:59) Reinforcement learning today and the AlphaGo moment (13:50) The 2 key concepts of reinforcement learning (17:35) Approaches to accessing data for AI in robotics (20:35) Robotics hardware (22:50) Vanta - Get $1000 off your SOC 2 at https://vanta.com/twist (23:57) Covariant's hardware and software in use (27:32) The importance of adaptability (32:40) LinkedIn Marketing - Get a $100 LinkedIn ad credit at https://linkedin.com/nextunicorn (34:08) The Tesla Optimus and humanoid robots (39:58) Overcoming hardware constraints (44:38) Boomerang investors and customers * Follow Peter: https://twitter.com/peterxichen Check out Covariant : https://covariant.ai/ * Read LAUNCH Fund 4 Deal Memo: https://www.launch.co/four Apply for Funding: https://www.launch.co/apply Buy ANGEL: https://www.angelthebook.com Great recent interviews: Steve Huffman, Brian Chesky, Aaron Levie, Sophia Amoruso, Reid Hoffman, Frank Slootman, Billy McFarland, PrayingForExits, Jenny Lefcourt Check out Jason’s suite of newsletters: https://substack.com/@calacanis * Follow Jason: Twitter: https://twitter.com/jason Instagram: https://www.instagram.com/jason LinkedIn: https://www.linkedin.com/in/jasoncalacanis * Follow TWiST: Substack: https://twistartups.substack.com Twitter: https://twitter.com/TWiStartups YouTube: https://www.youtube.com/thisweekin * Subscribe to the Founder University Podcast: https://www.founder.university/podcast

When you're building a company like this and you're trying to get product market fit, you have to find the place where your product can provide the most value in the short to midterm. So you have customers. But at the same time, looking at the long term, what's the biggest opportunity? In a way, you're being reinforcement learned as the founder here, the same way the chess robot we're talking about in terms of reinforcement learning. You could play the short game, which is, hey, we got to get into factories and figure out how to move these. these Dorito chips and batteries into the boxes without crushing them and quickly 24 hours a day, 365 days a year.

But also, hey, winning the game could mean maybe losing some customers about building that general purpose robot that you could put a hundred of them into a factory and just say, go find work. This is really an interesting concept. This weekend startups is brought to you by lemon.io. Need to speed up your product development without draining your budget. Hire vetted engineers from Europe at lemon. Go to lemon.io slash twist to get 15% off for the first four weeks.

Vanta. Compliance and security shouldn't be a deal breaker for startups to win new business. Vanta makes it easy for companies to get a sock to report fast. Twist listeners can get $1,000 off for a limited time at vanta.com slash twist. And LinkedIn marketing. To redeem a free $100 LinkedIn ad credit and launch your first campaign, go to LinkedIn.com slash next unicorn.

All right, one of the most interesting angles for artificial intelligence is how they might impact robots in the real world. Now, of course, robotics has been going on forever. You've been seeing Boston Dynamics or

CafeX making coffee or little tiny robots, the Roomba going around your house and vacuuming. Maybe you've seen the automation at an Amazon factory, incredible to watch. And robots have largely lived inside of factories, and they largely have been programmed by developers, and there is no AI going on. Computer Vision is a small area that's an exception. We invested in a great company, Rood AI, that was picking berries and using some of the hands that have been made,

the hand technology, robotic hand technology from MIT, to very carefully use computer vision to find the right strawberry to pick at the right time. We've talked about it on this program over and over and over again. But now that AI is starting to hit a tipping point, as we've seen, a lot of founders are focusing on, hey, can we get a robot to do reinforcement learning? And we're going to hear all about that today from Peter Chen. He's the CEO and co-founder of Co-variant. Peter, welcome to the program.

Thank you for having me. It's great to be here. All right. Now, you worked at OpenAI for a year or two. You went to Berkeley, got your PhD, and you founded Co-Variant back in 2020. 2017, raised a ton of money, just did a $75 million series C, led by our friends over at Index. Let's talk a little bit.

You heard my preamble robots living in factories. Robots not using AI being programmed to do very vertical specific tasks. Yeah. You know, one robot in the Tesla factory is going to do something radically different than the next one. Exactly. And putting AI in front of these things, it's just not going to work in a lot of cases. and it could cause a lot of damage because these robots are big, powerful, fast, and they can break things, including humans, which tragically we see in these factories.

So what is your approach? You got to this early. What is your approach in terms of putting AI and robotics together to try to hit this future where, my gosh, could robots be learning and using AI to do new tasks in the actual real world? Yeah. So that is a really good preamble in terms of a history to robotics. So I would say robotics is not a new technology and not a new field. There are a lot of robots out there in the world, in car manufacturing plants, in electronics assembly lines, they're robots in all of these different places.

And exactly like you said, Jason, those robots are programmed. And typically what they do is they do just the same motion again and again. And then an automation line and assembly line is so costly because you need to perfectly engineer every step of the process so that a robot that is only doing repeatable motion again and again can succeed. But you can imagine there are a lot of things in the world that just cannot be reduced to repeatable motion again in it. And those are really everything that robotics has not been able to crack before,

including the strawberry picking examples that you mentioned, including really all of the manipulation, like things that require your hands in warehouses and logistics, which is what we focus on. When you think about those facilities, you're handling hundreds of thousands, sometimes millions of different kinds of items that exist in a e-com warehouse, there's no way you can reduce the order fulfillment of that many items

to a perfectly repeatable mechanical process. And those are all the places that we have not seen robots play big role yet. And that's really how I think about AI's role in robotics. It's really not making those robots that are doing mechanical movement again and again better. Like, you don't need AI there. You don't need programming to do that. But what AI can really do is take robots out of those perfectly structured environment where you're just doing the same thing again and again to a much bigger role

where you really need to handle dynamic, diverse circumstances that's changing. every second, every day, every season. And that really opened up a couple more borders of magnetos of robotic applications that are possible. Covariant, we are starting from warehouses and logistics, but we really see the broader world as a fair game for this AI apply to robotics. And they are really... Going into the factory is just to... Because you mentioned that twice.

is obviously a great place to go because you have a high frequency of transactions, as you mentioned. Number two, you have a high variability, the different sizes. I ordered a bunch of straws. You ordered a couch. These are very different sizes, a pack of batteries and a computer, let's say, a laptop. And then on top of that, it is a semi-controlled environment. So you're not building a robot that goes down the street and deliver.

is a burrito and it's going to get kicked over. So while there's variability, it's controlled variability. It's variability on this, you know, a conveyor belt in front of you, let's say, right? Exactly. So, like, I think you can think about the evolution of autonomy, robot autonomy, to going from perfectly structured environment to semi-structured environment, which is what we're handling in this type of warehouses, distribution centers, industrial environments, lots of variability, but still semi-structure.

Like, you're not, you're not going to have, for example, people. kicking around or what self-driving cars can run into is like a turkey chasing a toddler on the street, like really out of bounds scenarios. And so it's kind of like somewhere in the middle and not fully to the open world, but you still expose you to a lot of diversity and complexities of the real world. Imagine this. You got an idea for a tech startup. You're going to change the world.

I know it. But you got a problem. You don't have any engineers. Engineers are hard to come by. They're very busy. they got jobs backed up. Well, you need to find great engineers. You need to find them quickly.

And you need to reduce your burn rate, right? Because you can't be spending like a drunken sailor. You have a limited amount of resources as a startup. Now imagine there was a partner out there waiting to help you, who had a thousand on-demand developers, and they were vetted, experienced, results-oriented, and passionate about helping your startup grow. And what if they charge competitive rates,

you know, reasonable rates? Does this sound too good to be true? Well, you need to head to lemon.io right now. Now, startups choose Lemon.io because they only offer handpicked developers with three or more years of experience with strong portfolios. And if anything goes wrong, Lemon.io will replace your developer as soon as possible. A bunch of launch founders have worked with Lemon.I.O. They've had great experiences. So here's the call to action.

Super easy. To learn more, go to Lemon.io slash Twist to find your perfect developer or tech team in 48 hours or less. And Twist listeners get 15% off the first four weeks. What a deal. So stop burning money, hire developer smarter, and visit lemon.io slash twist. Easy peasy. Lemon squeasy at lemon.io slash twist. In terms of where we're at in this process and where you're at with covariant,

if I were to look at, say, playing games. Yeah. And so you have a very finite game, chess. And then you have an almost a finite game, but a much larger base of possible outcomes. Go. And then you have games that have a massive amount of human variability in them, like, say, poker. So we've watched as those things have fallen and then even DeepMind taking the entire Atari 2,600 catalog, we just had Mustafa on the program, one of the co-founders of DeepMind. Where are we at in that timeline? Are you at like chess? Are you at Go? Are you at, you know,

random video games and do reinforcement learning? If you had to baseline 2023, AI in robotics, reinforcement learning, Where is it at? Yeah, it's a really good question. And the way that you're framing it gave me a lot of feelback to my days of doing the reinforcement learning research at OpenAI and at Berkeley. I was training a lot of reinforcement learning agents like exactly also with Atari suites of games.

So it gave me a lot of flashback of memories. But coming back to this question, this is a really interesting way to frame it. And I would say, where AI for robotics is at, it's really, from my technology, from an algorithm, from my models,

from my compute power perspective, we are at that Atari moment and really beyond. Like, I would say we are at even the StarCraft, Dota. Oh, that's a big chunk too, right? I mean,

the difference between a game like StarCraft and, you know, a 2,600 game like Pong is like the difference between checkers and maybe poker or go, right? That's a big leap. Exactly. So from an algorithmic and from a modeling perspective, we are there.

But what is missing is data. So it's very, very difficult to get robotics data, diverse robotics data that can build this type of AI. So let's even use goal as an example, right? If you think about alpha goal, incredible achievement that can be human champion and goal, incredible breakthrough that DeMind has built a couple years ago. And you have mentioned why this was incredible. This was incredible because Go is a really complex game. If you look at all the possible compositions of a Goal board,

like there's something like 10 to the 170 possible configuration of a Go game, and that's more than the number of all particles in the observable universe. And that's like a crazy... I mean, that is just to pause on that. The Game Go, which just has two different stones, It seemingly seems to be, you know, on the surface value, we just look at Go and be like, checkers. Ah, it's a simple peasant game. It means nothing.

Chess is much more sophisticated. It's not because of the size of the playing board. Exactly. You know, and the number of, I think it's the dynamism of what can happen when you have a multiple angle flip and, you know, three or four different rows change at the same time. It's just so many possibilities of, uh, outcomes. Exactly. Wild.

Wild. Like, wildly complex game. And that was why, like, for a very long time, leading artificial intelligence researchers didn't think we would crack goal in any time soon. And so it was a big shock, like, when deep reinforcement learning was able to crack the game of goal. And, but if you look behind the hood, like, one key thing that really power that is the amount of data. Like, if you look behind the goal winning, um, deep neural net,

is trained on more than 100 years of goal playing experience. I mean, we can go into the details of like, it starts from cell play, expert play,

and all of those kind of good stuff about the data. But just, just pause for a moment and think about the sheer amount of data that is trained on. Yeah.

When you're playing against AlphaGo, you're effectively playing against a goal player that has done nothing in her or his life for 100 years,

just playing. Right. Their life is a hundred years of playing a hundred games simultaneously. Who knows? I mean, it depends on how many H-100s, I guess you have. Exactly. But it's playing

such an amazing number of games and figuring out outcomes, and it doesn't even need to be trained. So this is a good place to pause, given your background. If you were going to explain reinforcement learning, in the case of Goh or Starcraft or, you know, playing Pong, to somebody who had never heard of artificial intelligence. I just want to understand how reinforcement

learning works on a very basic level. What are the three or four key concepts and terms of art in reinforcement learning? So the most two most fundamental concepts that you need to understand for reinforcement learning. One is you learn from doing different actions. So if you only do the same action again and again, there's no reinforcement learning because you have no contrast. And then the second concept that you need to understand. is there needs to be a reward function. So once you do action one, it leads to some outcome one.

You do action two, it leads to some outcome two. There needs to be a reward function that can rate which one is better. Once you can have an agent that do different actions and the actions lead to different outcomes that can be rated by a reward function, then you can start doing reinforcement learn. And at a high level, it's really simple. It's about the agent exploring the world by taking different actions and that lead to different outcomes.

And that outcome is rated by reward function. And then the reinforcement learning algorithm, just look at what are the actions that tend to be better. And you start that learning loop on there. There are a lot of technicalities on how to make that work and how to make that run efficiently and how to make it work well together with a big new net. Like, for example, how do you turn GPD4 into chat GPD4?

Like, there's lots of craft and details that's needed in making that happen. But really at a high level, it's as simple a step. It's taking different actions and figuring out which one is better and try to do that more often. That's really the core basis of reinforcement learning.

Okay. So reinforcement learning, to reflect it back to you, the behavior, you have to have an ABC choice, right? So you have to have a behavior choice. In the case of, you know, chess,

it would be moving one of the pieces according to the rules. And there's only a certain number of pieces that can move in the opening move. And then what is winning? What do you want to reinforce? What do you want to tell it good? And good in chess is having a piece taken or not having a piece taken.

Are those the two basic components there? Yeah, those would be like a good incremental reward function. And then your ultimate reward is whether you have won the game or not. Right. You can imagine losing all of the pieces. But if you won the ultimate game, that's still a good outcome. Well, and this is like a great point because if you look at somebody like Magnus or some of the top chess players. I watch clips of them. I don't know if you've ever watched

clips of them on like TikTok or YouTube. They, they now make little short clips of the best endings. One of Magnus's like incredible gifts is he sacrifices massive amounts of material. He'll sacrifice a pawn and then he'll sacrifice a rook and you're like, oh my God, he's dead. But those sacrifices lead to a series of moves that boom, checkmate. And so it could be that that getting a material advantage is the wrong training. It's like that's the short-term thing that's right to do,

take the pawn, take the rook. So the other player playing Magnus thinks they're doing the right thing, but they haven't thought as far ahead as Magnus, who is now, you know, mate in two, or mate in one,

when you make that error to take the rook. Yeah. And you are referring to a technical concept here in reinforcement learning that is called, like how do you optimize for, a very delay reward. You're optimizing for something that has a

long-term dependencies. You make a move now, and maybe you lose a couple steps, but you ultimately win the game. And a big challenge in reinforcement learning is how do you figure out that delay outcome and how do you figure out that long-term dependencies that you have?

I want to take a step back and coming back to the data question on robotics and on your earlier question of where we are in the AI for robotics evolution. And I make this common that from an algorithm and model perspective, we have what we have. But we don't have data. We don't have the equivalent of goals data of 100 plus years of diverse playing. Or chat GPT ingesting Reddit and Twitter and open crawl or Google indexing the web and

putting it into BART. Exactly. We don't have the Reddit equivalent. We don't have the GitHub equivalent in robotics. And that is the key limiting factor of robotics. Hold on a second, though, does not, I got to think, Bezos, who's a genius, would have cameras all over, you know, the conveyor belts in the factories. Would, you know, a couple of cameras watching humans do this be the potential, you know, a data source or is it not trained enough? So a robot and an AI watching a human pack boxes, would that be if you had a million hours of that enough for you to send a robot in there? So it's a really good question. Like this topic has an academic name.

This is called third person imitation learning. So this is like you're seeing from a third person view, someone else doing it. Can you learn from that? And I would say like the best state of the art is you could learn something from it, but it's never as good as if this is coming from your own actions, how you have tried it. And whether you can actually learn from that yourself. And the reality is actually like even for. For Amazon, they actually release a couple of data sets on the items that exist in their warehouses.

And actually, the data set is much smaller than what we have collected, even here at Covariant, from across the diverse set of customers that we have. A big chunk of that is, it's not just about the data. It's also getting about, like, it's getting exactly the right format of data and getting the right type of data. If you think about the modern movements in large language models, a lot of the secret source is in the type of data that you curate. And this is not something that you can just, oh, let me try to crawl more of the internet. The equivalent of that would be try to put more cameras in these warehouses

and just look at conveyors, right? But are you actually capturing the useful moment, the most meaningful data? And to understand what kind of data you need to capture actually require really deep understanding of what you actually need the robots to solve. So can we collect useful data? Yes, for sure. Like, we can already start collecting that today. But what we have found is that really to teach robots to operate very autonomously,

you need extremely high-quality data. And if you need very high-quality data, they need to be collected in a very useful way. I would think the hardware that you use, is distinctly different than a human hand. Now, they might have modeled it after the human hand, but it's going to move in a different way. It can move faster. It can move in ways that would give us carpal tunnel syndrome,

that they would, you know, the HR department would say, you know, do not bend over like this. It's going to cause, you know, carpal tunnel. It's going to cause back problems. A robot does not have those, right? The robot can just move in any direction. You could, you know, you know, hyper-extend your elbow and dislocate your elbow to move a package with a robot.

So the data you collect has to take that into account is the flexibility of the robot, which can twist and turn in any given way. Where's the, let's talk about where that robotic, the robotic hands are robotic arms, because you can buy a robotic arm now. Exactly. That's capable of doing, you know, 24 hours a day, 365 days a year with very little downtime. Yeah. That can lift hundreds of pounds for how much now. What is the entry level robotic arm that you might see in the Cafe X? coffee machine or you might use in a small warehouse to move, you know, a 10 pound package or something.

What do those go for now? Industrial robots are incredibly robust and mature technology. And so they're really good. Like they last for a long time. Like they can work 24-7. And with proper maintenance, like those robots can go for 10 years. So it's actually really incredible technology that's been built up and tool by the automotive industry. These type of robots depends on size and payload.

It typically goes somewhere from 25K to 50K. So which is not a very significant cost. Like if you really think about like the lane for time. We're talking about super industrial. If you were to compare it to a human arm, a human arm in a factory, would cost you $50 an hour in total compensation. And if it was working, you know, 24 hours a day, that's $1,200 a day, 365 days a year. We're talking about a half million dollars per year, over 10 years, $5 million.

One of these arms could do that for 50K. Exactly. So the robot arm itself is an incredibly cost-effective technology. If you're a SaaS or services company that stores customer data in the cloud, then you need to be, uh, SOC2 compliant. You knew that from a third party. And you need that third party to close big deals. And if you want to get compliant easier and faster, you need to use, V-A-N-T-A-N-T-A. Vanta makes it so easy for you to get and renew your SOC2.

On average Vantage customers are SOC2 compliant in just two to four weeks. Prepare that to three to five months without Vanta. And Vanta can save you hundreds of hours of manual work and up to 85% of compliance costs. This is a total no-brainer. And Vanta does more than just SOC2 compliance. They also automate up to 90% compliance for GDPR, HIPAA, and more. You can't afford to lose out our major customers. We all know that.

Listen, it's a hard year. Last year was hard. You can't lose those major customers because you don't have your compliance dialed in. Just work with Vanta. Get your compliance automated and tight and tight is right. Lock down those big deals. Here's the best part. Vanta's going to give you $1,000 off.

That's $1,000 off at vanta.com slash twist. That's Vanta.com slash twist for $1,000 off. Your sock, too. So we have a video, speaking of robotic arms, maybe we have a good point to stop here. You can take a sip of your coffee there, gee, and then show us this video. And then we can continue the discussion maybe because people, you brought some show and tell. If you're listening and you're not watching on YouTube, just go over to YouTube, type this week in Startups Covariant and you'll find it real quick. So what we're seeing here is the robots that's picking up a really diverse set of items in a really chaotic pile of items.

So what the robots need to do is not just repeating the same motion again and again, but it really needs to understand what's in front of it in 3D, what are the different objects, what are the different ways to, approach the item was the best way to pick it up and really manipulate it and transfer it successfully. And for those of you who is watching the video, you can really see a diverse set of items ranging from items from pharmaceuticals to CPG to candies, food, grocery items. And so you can really see how these items, they come in different orientations in the world. And they also, each one can come in in a different position, but also different deformation. Like if it's a bag, like there's not the same bag that would always appear in the same. If it's a bag of Doritos, which I think I see bags of Doritos and bags of gummy bags,

gummy bears there. If you throw 50 bags of gummy bears, 50 bags of Doritos or gummy bags into a tray, they're going to land in all different ways. It's going to look very different to the robot. But after training, like you're saying, it's going to know this is a bag of Doritos. It has a certain texture. You don't want to crush it. If you hit it too hard, it's going to.

crack some Doritos. Whereas the gummy bears, maybe you could hit those a little harder. So what is that arm called? It's not a hand type arm. It's not a pincher. It looks like it's got like two digits that are kind of like suction cups. What are we looking at there in terms of that extension?

These are vacuum-based hands. So like Jason, like you said, like you can very nicely articulate the difference between the arm, like, which is like the wide body that we're seeing here. This is a robot arm that's manufactured by ABV, one of the. the world's top four robot manufacturers. And then there is the hand, which is actually the part that gets in contact with the physical world. So the hands here, it's actually a very simple mechanism. So there's vacuum that gets generated.

And it's just kind of like one of the vacuum at your home. Like you can suck things into it. And then you have two tubes of vacuum, so two cups that the robot is individually actuating there. So you can choose to use one cup or both cups together. depends on what is trying to pick up. Or I suppose a different percentage on each one. Exactly. And so what we have found is that suction or vacuum

has turned out to be a fairly general hand technology. You can use this fairly extensively in the warehouse setting. It's not going to solve everything in the world. It's not as dexterous as a human hand, but it's actually fairly universal. What we have found to be very important, though, is you can not have the same robot hand everywhere. If you're handling much larger items, you can

imagine you need bigger suction tube. You need more cups to pick up a kettlebell. A 50 pound kettlebell is not getting picked up by the suction arm probably. Exactly. But it's easily be picked up by a pincher, you know, or whatever you call, you know, clasping. Like a finger-based

gripper, right? And what this actually points out is something that's very interesting in AI for robotics is the AI for robots needs the ability to adapt to different kinds of physical hardware. It really need the ability to not just handle one kind of physical body, but actually multiple kinds of physical bodies. Because we haven't been able to build a hand or a robot arm and robot body that is as universal as human body.

What that means is that for different use cases, for different customers, you need different hardware. And now your AI needs to have that ability to adapt to different physical embodiment that it has. And this is actually getting at a pretty interesting idea, right? Then how do you actually build an AI that can learn across multiple different physical bodies? And how can you build an AI that can learn across multiple scenarios and different kinds of item sets that you're building? And that really sits at the core of what we're building, which is what we call covariant

brain, a foundation model for robots. And we say it's a foundation model for robots because similar to chat GPT, like that is learning across translation tasks, coding tasks, all of these different language tasks related together. The foundation model that we also learn across multiple different robot tasks with different robot hardware in different customer scenarios, in different verticals together. And we do that because that is necessary to start. solve the data problem for AI, for robotics that we mentioned earlier.

Imagine, like, if every time I need to come up with a new robot hardware platform, I need to collect a bespoke set of datasets for that hardware. And if I need to collect a bespoke set of data set for one customer, you're never going to build up to that alpha-go moment of 100-plus years of experience. So really the only way to bootstrap this foundation model for robotics is to collect all of them together and you have to build one AI that can learn across all these different tasks together. Which begs an interesting question. Someone like Amazon would see this information, this data, this learning as a proprietary advantage over Target and Walmart. Whereas Target, Walmart or some

other vendors who maybe were far behind, let's say Target was way behind Walmart and Amazon in terms of automated and their factor. I'm just making this up. They might very much want to contribute their data in order to get your solution. So how do you think about the go-to market strategy as a founder, where some people might say, I don't want to give you that data, or how do you get that data, and then how do you negotiate that with your customers? Yeah, so we work with customers that, we work with big e-com customers that are innovation forward, like they know automation with AI power robots is the way to go in the future.

And they also look at themselves and say, there's no way I can build. that competencies internally, like in order to compete with Amazon, in order to keep up with the innovation. I have to work with a startup that takes a partnership approach and can really bring that technology to them. So one of this example is one of the recent customer that we have announced in Europe is with a customer called Auto Group, OTTO Group. They're actually the second biggest e-com company in Europe behind Amazon.

And so they exactly compete against Amazon. So they are a big e-commerce conglomble. They also own, I think the American more familiar brand would be Cray and Barrel. So they own Crayton Barrow. And why do they partner with CoVirion? They partner with Covariant exactly because they see AI robotics as inevitable. And they see that using that capability as one foundation model, one platform to transform multiple places of their supply chain network,

as crucial, but at the same time, they cannot build that themselves. So really, the best move is to partner with covariance and really bring that technology to their network. And yes, they're contributing data to the platform, but they're also benefiting significantly from it. Yeah, give to get. And because we have already built up such a broad data sets over robots across multiple continents, the robots that's deployed at their site

can also achieve much better performance they want before even contribute any data to it. You can already start benefiting from it. And this is a large part of what makes the current state ALAM so powerful, right? Because even before you fine-tune, for example, a llama on your own data sets, it's already quite useful, right?

Because it has already learned so much about the world that even for your own business problem hasn't seen all of the proprietary data yet. it can already perform well out of the gate. And by working with us, we give them the ability to ingest their data onto this large data sets that also make the AI more powerful on their specific use cases. Got it. When you're selling to B2B buyers, you really want to get your pitch in front of the decision

maker, the person who gets to sign the check. Because these upper level execs, they're the ones who make you. the purchasing decision. Everybody can have an opinion on the team. Of course, it's 2023. But there's always somebody where the buck stops. And that buck stops on their desk and doesn't get into your bank account. These high level folks are hard to find. They're hard to target on social media platforms. But LinkedIn is the social network for business and they have 930 million members ready to do business with you. And that includes the 180 million senior level decision makers. Plus, don't tell anybody. There's also 10 million C level executives there. That's

A ton. Purchasing power. LinkedIn ads is built specifically for B2B marketers. No other platform in the world can offer these eyeballs and you can target them obviously by their location, the size of their organization, they're vertical and their title. When you think about business, I want you to just think about LinkedIn. LinkedIn equals business. Business equals LinkedIn. It's that simple folks. When you present them with an opportunity, they will, of course, be in the mindset to receive that because they're not posting pictures of their food for mentally on vacation. make B2B marketing everything it can be and get a $100 credit towards your next campaign

by going to LinkedIn.com slash next unicorn to claim your credit. That's LinkedIn.com slash next unicorn. Terms and conditions apply because LinkedIn is so generous so that this week in startup's audience. What do you think of the projects like Tesla's? I think it's called Optimus,

you know, building an actual, you know, full-on robot that looks like a human and walks around like a human. Obviously, these arms, vertical, you know, they're vertically integrated. They've been out there for now, what is it, 30, 40 years of these arms,

you know, being at scale, I guess, and making a huge difference beating humans every day. But, you know, some folks are going to take the approach like Elon's doing. There's figure, I guess, is the other one of, hey, here's a robot that is a human. It's going to reinforcement learn and walk around your,

factory or walk around your house and put dishes in the dishwasher or pick up and clean up after the dog if it, you know, spills its food over or worse. Yeah. So I'm very glad that someone is working on humanoid robot. Like this is going to be a very key enabling platform to really open up to a wide set of robotics use cases. So if you think about like these industrial robot arms, they are really good, but what's the limitation of them?

The limitation of them, they are largely fixed stationary robots, meaning you have to bring work to it and you have to constantly feed work to it. And those type of robotic applications only make sense if you're running a two to three shift operations and the robot constantly can be busy. And that's the type of use cases that we are solving for our customers. These are heavy industrial environments that there's work constantly happening. and that's where you have this super positive business case. It's a very big market, right? We can easily sell billions of dollars of ARR worth of robots to this type of logistics setups that run two, three shifts of operation.

But it's not everything, right? A lot of things that's in a not as intense industrial setting, like maybe you would only do your dishwasher twice a day at most, right? And that's like... Twice a week, who knows? Yeah, or twice a week, right? So does it make sense to have a dedicated robot arm that's fixed around your dishwasher to only do that? Unless you're running a cafeteria.

And I looked at there's a dish bot, I think it's called, which was using magnetics to pick up the things. But you had to use the same dishes. So if you had a cafeteria and used the same size bowls, the same type of cups, and they're made of like plastic, not China, you could actually use it. But yeah, so with this, you know, Tesla Optimus or, you know, the other ones in the market, they can go find work. Exactly. They go to your backyard and start looking for work.

Exactly. They don't need to work in one fixed setup that only high volume industrial environment can afford. It's like commercial kitchen. It's another type of industrial-ish environment. So this is going to be a very important platform. Basically, it's going to open up robotics to even more use cases where it's starting to go into automating things that aren't frequent, that don't happen frequently.

So very, very important technology building that needs to happen. I personally don't really have a forecast of when this would land, because this is a very tall challenge, right? You are trying to do a lot of things that's not very high value, each one in isolation, but you need to be able to do a lot of these. So that puts a lot of burden on the generality of the hardware platform and the cost of it. Because each single one of this is not going to be very high value.

That also means, like, if your humanoid robot costs a million dollars, that's not. You can't use it in your house unless you've got a lot of dispensable income. But you could use it if you were a military application, you know, or it's a newfangled firefighter that can go into a burning, you know, a pet store and take the pets out of the pet store without getting burned. Like, there are going to be some applications where you're willing. willing to pay a million bucks. And actually, we have bomb robots now.

They just don't look like humans. They look like, you know, little RC cars, right? They drive them around. Such an interesting point you make because when you're building a company like this and you're trying to get product market fit, you have to find the place where your product can provide the most value in the short to midterm. So you have customers.

But at the same time, looking at the long term, what's the, the biggest opportunity. In a way, you're being reinforcement learned as the founder here, the same way the chess robot we're talking about in terms of reinforcement learning. You could play the short game, which is, hey, we got to get into factories and figure out how to move these Dorito chips and batteries into the boxes without crushing them and quickly 24 hours a day, 365 days a year. But also, hey, winning the game could mean, you know, maybe losing some customers but building that general purpose robot that you could put a hundred of them into a factory and just say, go find work.

It is really an interesting concept. It is very interesting. And we believe the key way there is keep building a general AI, right? Because that is the thing that is going to transcend whatever use cases that you're looking at today and whatever platforms, hardware platforms that you're building it on top of. Because a generalized understanding of the physical world and how to interact with it is independent of the use cases, independent of the use case frequency and independent of the hardware platform.

So, like, from a covariant perspective, like, I wish the Tesla Autumus, like, hardware platform exists today, because that can allow us to put our AI, our foundation model on it, to solve a lot more problems. What's the biggest problem in hardware? Like, what is the big, is it the actuators? Is it creating the pulley systems?

I know there's many different types of pulley systems. What seem to be the hang-ups there? From a humanoid robot perspective? Or from just generally? Generally. And then we could go to humanoid. But I think generally, what is it that the robots can't do yet? Is there some blocker that everybody's going, oh, you know, like storage used to be or bandwidth

used to be in the internet. If there was an equivalent in robotics, is it those, like, actuators or the pulley systems that, you know, create the strength where the arm can move? Is it the tips of the fingers to know this is a ripe strawberry versus this is a firm piece of corn. Yeah. How do you think about that? So, the answer to that actually is very different.

If you think about it as a general purpose humanoid robot versus like kind of a more classical robotic automations perspective. And I'm going to say a little bit what's the difference? The difference is that from a more classical robotic automations, the key thing is, can you customize very quickly? because for every single physical problem that you want to solve, you typically can come up with some clever mechanisms that make the problem easier because the problem is not every single thing that a human body needs to do.

So the hardware challenge there is not so much of any single individual one of those, but it's for every new problem you need to customize your hardware design a little bit. And it's your speed of customization that sits at the core of this. The humanoid robot is interesting. The humanoid robot, I would say it's as much as a product problem, as much as a hardware problem as a product problem. We cannot build a humanoid that is as good as human today. But then what is the first humanoid product that you should build?

I would say it's as big a problem as the specific hardware challenges that's there. Like you said, maybe we should build a humanoid robot that focus on bomb removal use cases first. But then, like, once you articulate that product problem, you can find a way to engineer for it. Like, what's very difficult is engineer something that is as good as human. That's too general and too vague. True general. Right.

A problem. I'm to tackle. Well, I'd like the approach that you've taken, which is we know the total addressable market for e-commerce and moving packages around is high volume, high transaction, and high value, right? So it's got a lot of the, if you would put the circles, high volume, yeah, high transaction, high variability maybe, or complexity, maybe a better word, and then there's money involved.

So it might be small transaction sizes, maybe it's $40 on average, but there's like a million of them a day coming out of this factory. Exactly. That means you got $40 million worth of product going out a day, you know, and whatever that is per year. So it's extraordinary how much value is it. When you studied the TAM of other markets, obviously factories were one, but they're well, you know, factories are well-oiled machines.

I don't think they apply as much to what you're doing because there's not variability. So this is actually like, let me make two comments. Like one comment is coming back to your multiple circles. And then the complexity and the variability part is actually a super important part for building a general AI for robotics because the AI that we built in warehouses and distribution centers can see really virtually any objects. that exists in the world, that gives us an extremely good training ground to build AI that can work elsewhere.

So, I mean, given this is a technology podcast, like, I like, make that technology point. And then the second point of what you were saying, and if you think about the type of use cases that we are deploying into, like, they are, you can really think of them as starting ground to build the future of robotics. Because the key insight here is that the AI that understands the physical world and interacts with the physical world is almost independent of use cases and independent of hardware platform. And so by finding this high volume use cases in one industry,

it gives us the ability to start this flywheel, to start building the AI and start solving the data problem for robotics that other people just can't have access to. because they don't have that real world data. They don't have that real world robotics interactions. Amazing. This is just extraordinary. You're in your five or six of your journey.

And it looks like the understanding of AI and the importance of it in the world has caught up to your vision. So that's nice. I guess a lot of investors suddenly that said no to you for the first five years are now banging on your door saying, how do I get an allocation?

talk maybe a little bit on a practical basis about being a founder when people think that's too hard, that's going to be a money burning pit to everybody saying, oh my God, that's the future, it's here now, I need to go make up for this mistake and not backing your company five years ago. You must have

a lot of boomerang investors, yeah? Yeah, it's very interesting because when we started the company, the term foundation model didn't even exist. So when we started telling people what we did was we are building one AI that can learn across multiple tasks, learn across multiple robots.

And then people would be like, I don't see why a specific AI isn't better. Why shouldn't you just train AI that is on one specific use case for one specific customer? No one would make that claim anymore today, because people have seen how GPT4 is better at translation than Google Translate. Even though Google Translate is also a deep learning powered AI-based translation system, that is just a more task-specific one. So it turns out that GPD4, by learning to do a lot of other different language tasks that are not translation, gave it better understanding of the world in terms of semantics, history, memes, grammar,

that actually make it much better at translation than its AI that is specifically built for translation. So it definitely has been tremendously helpful in the last half year or so that the success of our friends at OpenAI and Farbeck or Co-Hare, other places that have done, have really... So some of those VCs came back. Some of those VCs emailed you back after turning you down, I take it.

Definitely. Like I would say, that's definitely a huge increase in interest just because... I think that's the greatest, it's not the greatest feeling ever as a founder, is you get rejections all day long. Man, you must have met with 100 investors, you know, before this AI boom and gotten 97 noses.

Am I approximately correct? 90% knows, 95%? We are somewhat fortunate that we have really big supporters internally. And so we actually had not had the need to fundraise very extensively outside. I guess you're open AI. Mike Vopi from Index has been a big believer in this type of general model very early on. And so we got lucky there. But yes, and that's true, not even just with investors.

Like that's true with customers, right? Like, because, yeah, like this boomerang customers. This whole thesis of general AI being a better approach than specific AI. Also, customers didn't use to believe that, right? But now they really cannot refuel that anymore because the whole world is moving into that direction. Like, no one wants to train specific AI model on their own specific data sets

that's going to get still. That's less performance than a more general AI platform anymore. So that movement has been extremely. both validating to the approach that we have taken for the last couple years, but it also helps us tremendously. Customers eye from the capital market and all of these different aspects. Absolutely fantastic to have you on the program. Peter Chen, Co-Varian AI.

You can follow him on Twitter slash X. He's not super active. Peter X-E-N-C-H-E-N. Co-variant AI. What's your domain name? So I can send people because I know you're hiring. Covariant. Covarying.

Yeah. So if you're hiring, what's the, what's the hardest thing to hire for right now? What do you need? I'd like to help fill some positions for you

to thank you for your time here on the program. Yeah, we're always looking for great engineering talent, great AI talent. Like, if you're interested in solving hot problems that have not been solved, that's where,

which is what we're doing every day. Like, we would love to hear from you. Covariant. aI slash careers. The HR department's going to, uh, Thank me for sending up the careers page.

Yes. Great having you on the program. Continued success. And I would love to check in with you in about a year and see how you're doing on this incredible journey. And we will see all of you next time on this week in startups. Bye-bye.