4 wild AI predictions from a $39B tech founder

I don't care about that at all. Give like zero shits about that.

I don't care about that at all. Give like zero shits about that.

I think any way I would characterize it is like we're just now, like these companies of mine are just now hitting an inflection point and they're really early, like they can be like really big. So if it works this will like 100x, 1,000x from here. So most of my energy is like how do I make sure that works? There is no flatline here. It's either like it goes down or it goes up, right? Either like it's binary. Either the robots go out of scale or they don't go out of scale. So in like 5 years time, it's either going to be a very big thing or very bad. And so all my energy is going into making this like a thousand or a million x from where we're at here. And so it's like the pressure is on to like really just deliver.

I think at a high level, I think the AI work that we're seeing here now is going to be so much, it's going to be like a hundred times bigger than the internet. It's just like everything is just so, it's just working so well. Like the system is working well, like deep learning works and everything's happening faster than I would have thought. And my, like you know, having done like 15 years of like software and internet, like it was just like nothing was happening faster on a trend line. Here it's happening like that in AI.

We started at, so Hark, I have a new AI lab called Hark about a year ago. I was like very interested in this idea of like kind of building this AI to human symbiosis digitally. It's like, um, Figure is going to be like, I think Figure is going to be like the max ceiling of AGI of like being able to put that out and then there's going to be a version of this in the digital world. It's going to be like a human is going to have this like AI pairing. It's going to have like ultimately maybe your own AI weights, your own memories, maybe your own hardware. It's going like really close. And fundamental to that thesis was like you got to figure out how to get AI to use computers general purpose. You would never hire an assistant that couldn't use a computer. So you got to be able to like give things out to it that can like do everything you can do. Financial models, book flights, like order DoorDash, whatever you need to do, you need to be able to do it all autonomously, but only one in a thousand websites have APIs. So, in most, you know, like most computer use globally is on the internet and browser. My main inclination within two or three years, you'd have a system that you'd be able to talk to and say, 'Go do this or do that.' And to be able to like go online and like maybe use the internet really well like almost like a robot would where you can like move the mouse and use the keyboard. That's what you have to do to solve like general purposeness around a computer is you can't rely on API or MCP. You have to figure out how to like navigate like a human can. Now at Hark, we've just released our first kind of model and research preview last week. It's really hard for us to find now something that we tell it to go do on the internet. It can't do.

Okay, there's a couple things we did a little differently. First is like everybody's tackling this from like using APIs and MCPs. Like the reason OpenClaw got so great, it was like it couldn't use the browser. It couldn't like go on and use DoorDash end to end because DoorDash has no consumer API. So we tried to figure out how to use like a, how to look at a screen. And one is we spin up a virtual computer for every agent. So they don't need like a MacBook or anything. So you can just spin up as many of these environments as you want in the sandboxes and then you need to give it ability to like look at a screen and use like move the cursor and use the keyboard.

Yeah. I don't know if I, yeah, it's got to work well. I mean, that's the whole point. But like if it goes in their funnels, it's like the whole point is like...

It was in our post-training, like it was in our, we have a reinforcement learning process that we think is maybe nobody else in the world has done.

Well, I mean that's kind of all we had at the time we started. So, okay, what is...

I strongly believe like AI will head in two directions like, and then at some point maybe even like maybe like head together. Like the first is we'll have AI out in the physical world that will like do everything in the environment for you like laundry, dishes, cooking, like run the supply chain and be in healthcare. The vessel for that is a humanoid robot. It's just a human form and it will just go out and do like, you like want one piece of hardware that can like, you know, the hardware is capable of doing everything and you put like smart AI into it and it'll go off and do everything in the world. That's what Figure is working on. Separately than that, there's going to be this like really close like digital like AI to human symbiosis that forms. You're going to have like this very special thing that you can like talk to that's with you everywhere you go that will know all your stuff, have access to all your memories, have access to all your accounts and systems and be able to actually go do things for like a superhuman assistant. It'll be like maybe the closest thing is like Jarvis from Iron Man and it will be able to do like, it'll be like superhuman in almost every way. It'll know everything about your life. You'll be able to access it at any moment whenever you need it. It'll be in the background helping you out at all times. If you're on a flight with like a long layover or a short layover and you miss it, it'll like already have backup plans already help you like figure that out. Like it'll just be something with you everywhere you go. We don't have that. We have like really good coding agents. We have really good chat bots, but we don't have like something that can go off and like be my Jarvis. In order to get there, we need to work on the model side. It's got to be just better than text chat. It's got to be able to use computers, have like basically near perfect memory, be able to talk to you like just like a human would back and forth. And we have to have vision in the system. You have to be like look at the world and understand what you're seeing with it. And I think secondly, you need to have, you need to fix the interface to AI. You have like AI over here and a human and you have like an old hardware system in between like a MacBook or iPhone. They were designed 20 years ago. They're complete rubbish for AI. They're not the right interface. So, we went out and we are out there designing what we think comes like after the iPhone for AI. And it's like an upgrade cycle. We see this all the time in startups. You guys see it, right? Like we're in an upgrade cycle with the computers and phones. They're just going to go away. They're going to be new ones. They're going to be all AI computers and phones and systems and they're going to be great. They're going to be all real time. You can always access them and they'll always be like understanding what's happening. They'll always be able to reference things what's going on. You'll be able to abstract away most apps. You'll probably not have an app store, or you'll probably have an AI operating system. It'll be perfect for you. You'll ultimately have your own weights on your own devices that you'll own and have with you everywhere you go. It'll be like a really great pairing. And we hired an incredible team. Teams like, you know, maybe like 80 or 90 now. The guy that leads hardware design previously designed for the last several generations of iPhone, MacBook, MacBook Pro. Like he's just like the, he's a stud. He's great. So, we're designing what we think are the next generation of AI devices that will kill the phone and computer. And then, we're designing the next generation of AI models. The models need to get a lot more multimodal. They need to get a lot more expressive. Like the text encoding is just not enough for us to like really have a real AGI feeling with AI. So, we're working on that. We have our first AI, we did our first research preview for our computer using agent that we came out last week. I think we were like top on some of like the leading like, you know, browser computer use benchmarks in the world. And it'll keep getting better. This will keep getting better and better. Like every month we'll just like it'll be better and smarter using a computer and faster. We're working on a couple other different types of technologies internally on the AI side. And then we'll launch the ability to use Hark on like traditional browser and iPhone and Android in about a month. So you'll be able to start using it and then we'll have hardware coming. We're working on now. We actually have hardware in the lab now we're using testing and it's crazy. Like the stuff is like a sci-fi movie hardware.

I think I've like really changed my mind on this a lot the last like year or so. But we have a really strong opinion here internally. Our opinion is that what sits in the middle is devices that could possibly reach a billion units a year in the world. The only kind of things that we have like that in the world right now are computers and phones that kind of meet that like mega devices. And then you have things on the ancillary around it like orbiting this like big thing that are like AirPods and you know like a watch or things like this that are like they don't sell a billion units a year. They're like 3% of like Apple's revenue and they're like they help the ecosystem as a platform. What we care about at Hark is trying to solve what's in the big middle piece. To solve that, you got to take down the computer and the phone. There's no way around that. So, you have to rebuild a new computer or new phone that's better and replaces your existing systems end to end. And then what's around there is things that like you will have, we will even have at Hark that are like a family of devices that are not a billion units a year but important for the ecosystem.

No, I'm not saying that at all. You're going to want to like really radically rethink everything. The first version hardware we have now in our lab is like unlike anything I've ever seen in my whole life.

What lives outside of here on the edge are like glasses and pendants and wearables and things. They're not the main show. In fact, like the Meta glasses are probably one of the worst products I've ever bought. They're just horrible. They're horrible. Like I can't even like figure out how to use it. It doesn't have its own network. It piggybacks on the iPhone network. It means your app needs to be open on your phone. The pairing's long. Like it doesn't work well. Like I can't think of any reason why I would need this thing strapped to my head for 14 hours a day. Like it's just like the wrong device. It's not like the end state is BCI in the brain and we're going to have like AI language devices for the next 10 years before that. And that like that's the path and it's not glasses. Glasses I think I don't even know if glasses will make our top like 10 list of devices.

Let's get down to like the substrate level here. Like first order, what has changed? What's changed is we have like a new type of computer which is like I think of AI as a new type of computer, new type of automation that's here. That automation can do a few things that are like, when we're designing this we want to design around like key principles that could be like 10x better. If it's like one or two times better your phone or computer you're not going to use it. It's like literally 10x better. What are things now that deep learning brings that are like 10x better? There's a few of them. Like one is AI can like basically now like think and use computers and systems for you just like a human can. It can like talk to you. It can like see, has like visual understanding. It has real-time speech to speech. It can use computers and systems for you as close to as fast or around as fast as a human can. Over time it'll be just as good as a human and faster in terms of success rate. So you have a system that's like almost like humanlike in capabilities. It also can like have memory, meaning you can put memory into it. It won't forget anything. I mean, near perfect over time. So, you have a system that's almost like a human in a box that has all the same affordances a human has. And it's almost like the ability of like, you almost like if you could bring a little human around with a computer on your shoulder everywhere you went, that'd be insane. Like it was only for Sam though. Only Sam could see it. Only Sam could talk to you and only was like there to help with Sam. And that was like your whole life. And it's going to be smarter and better along the way and had perfect memory and could use computers and talk to you and see you'd be like damn that thing would be like, it would be able to do anything you do on a computer.

Okay. And then from there like we got to design around that system. The competitive advantages here are that it is humanlike in capabilities and it has almost near perfect memory can go back and reference over time. My phone doesn't have that. Like I put a contact on my phone like last week and I was like busy when I was like putting the phone number in and like a day later like somebody's like hey did you call that person? I'm like I don't even know the name. I forgot. I can't even ask my phone. Like it's just like it's so stupid. Like the whole system is and then I go in there like order DoorDash like a monkey. Like every day now I'm pushing things like I don't do any of that now with Hark. It does it end to end for me on my drive to work I just like say order me coffee and it's just done. It does it all for me in the background. I don't have to touch anything. It's all abstracted away and it's like if you had that little human with you everywhere you go you would just say like you would even predict probably Brett you want coffee today and be like ah yeah I do like let's order but you know what make it a double shot today and you know like route it to the Hark office instead of Figure. Like I can like I would just and done I got it let me take care of it. I was think like a monkey on my phone for the next like three minutes like trying to do checkout DoorDash.

Start there. And then from there, you got to rapidly prototype. So when you come over, like we have, we've designed everything you could possibly think of. We 3D printed it.

What were designs that didn't work that were kind of cool? The thing is we're building like many different devices now that cover like a pretty wide area of this. We have some pretty crazy stuff that we were like designing. So like it's not like you look at that you're like that looks like this and it well that does well over here. So it's like it's not as easy as drawing those parallels. It's like pretty quite radical. We rapidly prototype all this. We have like a fabrication facility that does this stuff. We have a whole design studio where we work on this. I like use this stuff like over the coming like weeks and months. I'll like either carry it around with me, wear it, whatever. We end up doing it and we'll like kind of down selection. We had like one of the biggest telecom CEOs in the world here that actually helped with the work with like Steve Jobs on iPhone 1 and he was here two weeks ago and he just come from meeting Tim Cook. You know Tim Cook's on his way out as Apple but he was over there at Apple and came over here and he saw our stuff and he's just like holy shit man. This is the first time I've ever seen anybody that could possibly take out the big guys.

The hard problem is not that. We think, we believe now the most important constraint to really solve is like building a really intelligent robot system to the world. Like there's a bunch of robots you can go buy now. You can buy some from China and you get them and they're complete crap. They can't do anything. They like you can joystick them around. That's all you can do. And you like hit a button and it waves. It's got no hands. It's got nubs. And you're like, 'What do I do with this thing? It's a toy.' It's like early when I bought a DJI drone like years ago and I was like playing around with it. Then like a day later, I was like, 'What do I do with this thing?' And it was like hard to set up. It didn't really work well. Like you know, whatever. It's just like I flew into a bunch of trees. It just didn't work. I was like, 'This is what am I doing with this thing?' Robots are like that now. Like where you can, we can go manufacture a ton of them, but like if they're not really smart, like it's not really going to be that helpful. We're trying to crack like the true human level intelligence of Figure. Like we really want to tackle like how do we make it so I can put it into any home. It can do every job I'd want it to do. That's what we're working on. We think that's the largest gap in the schedule of what we need to go solve for. Like it's that then beyond that like you know people generally sometimes confuse like consumer electronics manufacturing with car manufacturing. There's no big company in the world that would look like say like I'm scared of manufacturing this consumer electronics at high rate if there's so much demand like this is just possible to go do. I mean you can make them, we make a billion phones almost like you know pseudo by hand in the world and with some automation but cars is a different story. Cars like you will die trying to manufacture cars. There's like a lot of like you just like it's so and having seen like you know BMW is a commercial customer of us. I haven't been in BMW and a few other groups like it's gnarly early. The reason why cars are so hard is that you can't hold the part in your hand. Phones you can just like always hold in your hand and go change or whatever, move and hold. Like cars you can't, you physically can't. So you need robots that like literally pass it to other robots that put things on the chassis. And if any of those break across like thousands or 800 robots, you're dead. It is, the whole line's done. And so it's just like this huge giant robot you're building that's building the car. And with Figure you can hold any part in your hand. So I think we're like if we're like between cars and like consumer electronics, we're like over here closer to like you know we're like the 40% level over here by like cell phones. Like we just made our 1,000th EVT robot for Figure 3 last week or week before that.

We have like two large customers. We have us as an engineering and AI research org that needs like robots here. Like every engineer needs a robot. We need like every lab needs robots. Like we need like do tons of testing. There's just a lot of work we need to go do internally. We call like maybe like engineering fleet we need to go to and the second one is go to customers. So we have going to both right now. We've actually shipped out robots to our third customer this week.

We do a lot of like logistics stuff right now and packages. We have other stuff we've done in manufacturing mostly just manufacturing logistics just we've done in the past. But like we're also talking to folks about other industries.

They just did a live YouTube video and they had hundreds of thousands maybe millions of views of people watching this robot sort packages off of a conveyor belt.

That's an example of like a very close, one of the works we do.

No, no, no. It's like the pitch is like they come to us and they're saying like we're dying with labor. It's like we have like really high turnover. Some areas have over 100% turnover per year. It's really expensive to find talent. We have like just a large talent shortfall. The talent's really expensive. Like wages are going up and we don't have a solve for this. We can't figure out how to automate all this work and we need you to come in and help us. We have an ability to make a lot of good money in our contracts and the customers make like really good ROI on this. Like you got to think like a robot can do like multiple shifts per day, work seven days a week. Like we can like have a lot of uptime. The task you saw like on the logistics line that we should live stream was actually a real use case for one of our customers. That needs to be done at 3 seconds a package and it initially needs to be done five hours a day. Like I think it's like 5 days a week. We did that 200 hours straight at 2.9 seconds a package. So, we're already at human speeds. We're already doing this here now. They're already having ROI and we're now in the early stages of like getting these out to these customers and scaling it up. Over time, it will just put billions out to these groups.

Yeah, 100%. So I think a few things. One is the amount of like noise in the market for a signal is just like, like you mentioned, it's out of control. Like there's just so much shit out there in the market. It's like really hard to tell what the hell's going on. So, let me summarize what I think is like the most important and work backwards.

What I think the most important thing to do is to be able to ship robots autonomously at scale in useful work environments like they can like you know cook your dinner, like clean your dishes, like make your bed, like run the supply chain end to end, work in healthcare, build a building, like do logistics, like that sort of stuff. That stuff requires fundamentally onboard AI you can run so you can do like autonomous work. You can't solve it with code. You need to do it autonomously. You need to do over long periods of time and you probably need to move around and use like something in your hands and move stuff through the world. You know what I mean? It's like really you got to like do stuff economically like move got to move like electrons around. So I think at a high level like what we care about is not like the best robot that's doing back flips and running the fastest mile or dancing or in a parade or running outside in the woods. Like you know we don't care about that stuff.

Yeah. So like for us like we really want to show and demonstrate the ability to do like real things over a long period of time autonomously.

Man, there's like so many groups out there. I don't want to name names. They're all teleoperating the robots in all their videos. And it's like...

We haven't put out a single thing ever that's ever teleoperated. I only have ever done it. Like we used some teleop app internally for like data collection and for like it's good for like you can do a lot of testing on some of the robots sometimes but like it's got to be like this bad thing now. It's like the equivalent of a self-driving of like some dude in like Kentucky is driving the car and it's driving around the passenger. Everybody's taking a video of it. They're posting on the internet like look at this and it's just like dude that's like completely faked.

I think there's like this like in somehow in the robotic space become okay. And it's like for me I'm just like this is the worst thing I've ever seen in any industry I've been in as an entrepreneur my whole life.

They'll say it's not a like the ones that are like running fully autonomously will like write that in the headline because it's so hard. I remember like Tesla put a video out one time of a guy like the robot folding and then like in the bottom right was like a teleoperation hand glove you can see in the video. So there's like people get caught like doing that. You just know in the industry people will comment at the company later like hey this is like know wasn't autonomous. So like if you're in the industry, you know, like you can follow it. It's also really tough. I don't know if there's any other company in the world that we've seen that you can watch do autonomous work with AI on board on useful work with a humanoid. We haven't really seen it. Maybe there is. I haven't seen it.

I mean, most of all these are online. I mean, the Figure master plan is on the internet on the site. Archer's was up for a long time. I posted about it. Like I think like deep down I really want to find folks that really care and are obsessed and I'm like I think most of my time is not, I know you want to know about the pitch. Most of my time is trying to find those folks. I found that even in the Bay Area where it's probably like the richest AI and engineering folks in the world. 90% of everybody out here is not good at their jobs.

I technically assess them. All of them. Yeah.

I need to know like a certain guiding principles. Like for instance, I need to know like if you did the work or if you like watch somebody do the work. If you've done the work, it's like a scar you carry with you. It's like dug into you. Like you know all the details. You can talk about it freely. You don't need to think. You'll understand how to like reverse engineer everything you've done and discuss it. The folks that haven't done it can't do that. They just like they can't even go like they get one layer and they just like instantly blow up. They can't talk about it. They don't know why.

I'll give you an example. We have like a really challenging process to go through to be a mechanical engineer here at Figure. You have to be able to build like actuators from scratch. There's bearings and motors and, you know, we have a gearbox. We have like other sensors inside the...

system. It's a really, it's very compact, you know. It's just a very difficult thing to do, and like really hard requirements. We've been doing 10 case studies a week for 6 months and have not hired anybody.

It's insane. But when you do get someone qualified and their competing offers are companies that are larger or more liquid than you, and the offers are, I think they're like tens of millions of dollars a year, right?

The AI side is certainly like that. Like the AI side has gotten, and it's mostly all driven from Meta. Like at Hark, like I've never seen, I thought maybe like Meta was like paying these people for like a year ago and it was like would go away. They've not stopped.

I think we gave an offer to somebody that was really senior that was like they were coming from X8i, like XDI completely blew up, like everybody just left. And about 6 months ago, just like every, it was just like macro hard, like got fully disbanded, like there was basically a bunch of stuff that happened. We interviewed a pretty senior guy on the AI infra side. It was great. I think I gave him like a really good package, like of Series A stock at Hark, and it was, I don't know, 15, 20 million of stock.

Over, we do five, like for my companies in the early days, and we transition to four a little bit later, so we're still at five. And you know, I was like, I think we're like a, I think we can like 10x Hark here pretty quick. And so I was like, okay, you have like, you know, 15, 20 million, I think 10x, you have a few hundred million dollars. We 10x one more time, you have a few billion dollars. And I think we can do, I think we like, we have to like, obviously it's going to be hard, but I think we can do it. And he got an offer to go to Meta for $36 million over four years of RSUs, and he's just like, it's kind of guaranteed cash, you know. I go there and I have to like weigh this like maybe like $200 million at Hark or $20 million or maybe like 36 for sure at Meta, and he left and went to Meta. And they've been doing that like every candidate we speak to is like making some absurd, absurd thing. They just haven't stopped. They've just been at it since like for like a year, a year and a half. They've been buying talent. They've been buying their way into the AI race.

I really like it. I think like the AI space, what I found is the folks that really understand how to do like language pre-training and mid-training and post-training, especially pre-training and the infra around supercomputing and data and evals and all the right stuff you need to get put in place to do that right. And the amount of folks that really understand the right kind of like recipes that transformers do well in, you know, around or whatever you're going to look at, I think is really hard to find. It's actually really hard to find the actual folks that know what they're doing. I think there's probably, my rough calculus now is probably like a rough back of the envelope is probably like 20 to 30 people in California know how to build really good AI models.

No, the like the junior guys, like the guys in like their 20s and like the late 20s or something, are getting like, they're making like a few million total. So they're making like 200, 250 in base. They're making like another million or whatever like in a year in like RSUs every year. And so they're going to pay like a million to you anywhere like 750 to like 2 million or so range per year, and that's been driven up by Meta, and but then all the other labs have like followed comp.

I think it was really smart, and I would have done the same thing if I was Mark. I would have bought my way into the race, and I think he's like, he's doing that now. I don't think I would have done that. I want to understand it, and I want to like first order like find the right folks that really care deeply about this and not hire like mercenaries.

And so he hired a bunch of mercenaries. They're just purely money driven. They came over there. Nobody wants to go to Meta. They just, they're going there because they're getting paid a guaranteed RSU package by sitting around. And what's happening is like you don't need like a thousand people or 500 or 300 to design AI models. You make a really good team of 20 or 30 or 40 people, and that you can get there without doing this. And those people probably would care more deeply about the mission and where you're at and be more committed than just if you purely throw money at the problem. But I think if I was like, I think it was a really good strategy and it's working. I think hats off, like really good execution, their recruiting efforts and how they're structuring this stuff, and it's like, I think it's like paying off for them. Jury's still out if they can like actually ship real products. I think like the problem I have with those groups is they've just traditionally have not been able to do things new.

Well, I mean, I think Facebook is probably, Meta is going to go down in like one of the greatest acquirers in all time with like, you know, Instagram and WhatsApp and different ways. They've like bought their way into those spaces. But like, you know, if you look at like the Ray-Bans and everything they're doing, it's just like, it's not great work. And so I think the question really is how do you really do great work here? I think like we're even talking like we're using like the Hark system right now and it's so good. It's so much better than anything I use today.

Yeah.