Tesla's INSANE Plan to Make Billions of Humanoid Robots - with James Douma
01:11:33 Share

stuck between a truck and a... Oh, and there's another one over there. Here we are waiting for our robotaxi. Waymo's having a hard time here. But actually, everyone's having a hard time here. Humans are having a hard time. Yeah, yeah. Hey, James, so I wanted to ask about Optimus. So I was thinking, like, one of the things we've talked about the past, I don't know how many years, four or five years, is autonomous driving. And the...

the whole thing of Tesla making the right moves, the right decisions, being on that path, right path. And like as we drive RoboTaxi, it's like, oh my gosh, it seems like it's here, you know, it's like it's rolling out. What do you think about Optimus? I mean, like it seems like Tesla is making all the right moves with

the right focuses, the right strategic decisions, the emphasis on the hands, the right AI software decisions. It just seems like it's following, I wouldn't say it's the same, but similar kind of systematic, pushing it the right way. How are you viewing kind of Optimus development? Are you as excited as you were with autonomous driving with Tesla early on? - This is us.

Yeah, this is right here. I'll go around. All right guys, this is our last ride over the day. Okay. Yours. Yeah, a couple more probably. Yeah, so Optimus, I mean, do you think Tesla is making all the right moves right now with Optimus? Like, are you as excited as you were with autonomous driving back in the day? I feel like Optimus and...

and FSD only have a really superficial resemblance to each other from a strategy standpoint. I can see why, you know, there seems to be a common tech foundation, but... So, like, the starting decision, I think that ends up... I think Tesla has, like, two starting decisions that...

that essentially give them this overwhelming advantage in the space now, years later. And the first one was to bet on neural networks and cameras, right? Which, you know, at the time they did it, it's definitely a bet. I thought it was the right bet. Other people thought it was the right bet.

But most of the people in the industry thought it was a wrong bet, you know, that essentially going sensor heavy and, you know, doing HD mapping and all that kind of stuff was going to be the winning strategy. The other thing was, and this is the thing that's super amazing to me, is that the company invested in putting all the hardware in the cars way ahead of time.

They built the fleet and building the fleet allowed them to gather the data. At the time they made the initial commitment to neural networks, it wasn't clear that the data was going to be as important as it turned out to be. I mean, you could have guessed that it was, but there wasn't a lot of evidence. But, you know, they had a level of conviction that allowed them to like make an enormous capital investment in building out the fleet, which

You know, the combination of making the right bet on the technology and building out the fleet well ahead of time is put them in the position that they're in right now. I feel like, you know, robots are in a super different place. Like when Tesla started on the on building out the technology that became Robotaxi, it wasn't super clear that it was going to work. I mean, most people thought it wasn't going to work.

humanoid robots you know we're at the early stages of humanoid robots and there's like 25 companies doing humanoid robots like the the environment from a business strategy standpoint is completely different the the thing that i think is most interesting about the humanoid robot thing right now and the thing that i think is tesla's advantage is that um

unlike FSD, like arguably the opposite of FSD, the hardware is the thing that matters most in humanoid robots. And with FSD, FSD was basically you're building on cars, which is a super mature industry. Lots of people can make cars.

And the sensor suite that you need in the compute, well, compute is pretty mature. Like you might, you know, you might need to design a new chip or something, but we design chips. You might need a new camera, but we have super sophisticated camera technology at really low cost. Like all the essential hardware ingredients were kind of already available. There were supply chains. They could be built at volume at low cost, right?

But the software was super not ready, right? A lot of work needed to go into the software. A lot of people thought the software wouldn't work at all. And the software is the final piece of the puzzle that comes in and then makes the market work, right? I think robots are going to turn out to be the opposite. I was watching, you know, I don't know, in 2021 or whatnot, when you asked me what else should Tesla do, and I said robotics. The reason was because

I saw the software happening and I didn't see anybody doing the hardware. Like I feel like the software is coming and it's going to be there before the hardware is ready. Now, when I say the hardware is ready, I'm talking about a mass produced, you know, $10,000 robot that's kind of human equivalent in terms of its functionality. Hands are super hard. I mean, there's there are a lot of difficult software problems. But before we have.

you know, my opinion is before we have a mass-producible robot body that is durable, efficient, actually has the, you know, the dexterity and the sensor feedback and, you know, the battery longevity and all of that kind of stuff to be able to do what a human body does, well before that, we're going to have software that makes them super useful, right? So,

you know, in FSD software was the key because the hardware was in a sense easier. Not that... and it was easier because there was already a lot of infrastructure. There was already a supply chain and all that and the core technologies were well developed. In robotics I kind of feel like it's the opposite. Not that the robotics tech software is available, but when you look at the development ramp for the hardware and the development ramp for the software for robotics,

it seems clear to me that the software is going to get there well before the hardware does. So the key to actually succeeding in that space is not being the first person to make a robot that can fold a shirt. The key is being able to get to scale, because scale is what drives hardware costs down, right? Manufacturing scale. Is getting the sensors, getting the actuators, sorting out the design variants and options, and then

getting it to scale. And so, you know, there's a bunch of humanoid robot startups out there that are cranking out robots that can do interesting tricks because they've got the software, right? But, and I don't, you know, I haven't spent a lot of time drilling down, but I'm guessing they're the kinds of tricks where, you know, if you try to get the robot to do that trick every day for a year, the robot body would fail before you got there. And the

reliability of doing the trick is even today probably not super great. That, you know, the hardware's got a significant way to go and that the key right now to kind of becoming a dominant player in humanoid robot space is making the immediate investment in getting to where you can mass produce robots that are reliable and highly functional at low cost. Let's continue this if you don't mind.

All right, so continue. So, I mean, you're talking about how with-- The hardware's the thing. Yeah, with humanoid robots, you think hardware is the harder problem to solve. So I think Tesla's well-positioned because they have hardware capability, right? And they're focusing on hardware.

you know, that's going to be the final piece that drops in and makes everything work, right? Which is going to be the ability to make the bodies, you know, a quality body that's highly performant at very low cost. You can kind of, you know, you can do bodies that are relatively performant and not very...

reliable and not very inexpensive right now. I mean we see various combinations of these things but to succeed in the market you've got to have all three of those things and to have all three of those things you've got to do the hard work of developing the hardware design, refining it, building a bunch of things, putting them in the real world, seeing what works and doesn't, refining the sensors, refining the feedback,

And then building the manufacturing line that can do that at scale so that you can drive the cost down because that is going to be necessary. Like you're not going to be able to hit this price point at like 10,000 robots a year. You know, you're going to need a million a year, 100,000 a year kind of scale to do it. And

This is an area I'm not drilling down super close on all the different players in the space. There's time for other players to build the infrastructure to go to scale. Tesla is the one that seems most clearly to be focusing on the problem of getting a really good design that's really manufacturable and highly performant so that

you know once they've got it together they can go to scale right you know i mean elon's talked about their test fleets being 5 000 robots right and to me like you know a test fleet is 5 000 10 000 20 000 robots um that's that's kind of a minimum number that you need to kind of flesh out you know uh design decisions that you do and those are going to be robots that you throw away because you're going to do another you know revision and then you're going to have to do those tests on them too so do you think i think

Was Elon saying like a couple thousand by the end of this year or was it next year? Didn't he say one Legion this year? Yeah. So that's 5,000. Let's say hypothetically it's more next year. So let's say second half of next year, they have 5,000 robots. So they deploy them internally with their factories, maybe some other companies or what do you think?

It's not a matter of just like, you know, building robots because we know that design they got to build a bunch of robots They got to use them in the world They got to learn what's good and what's not good about the design choices that they made and then they need to build another Generation robot and do that again. I guess they're they're kind of on the gen 3 optimus right now So they've made some progress but they started at such a low level relative to the ultimate goal that I feel like you know, I

you know we're going to be a gen 6 optimus before we get the multi-million unit thing so but they're doing like what a generation every six months or a year or something like that so maybe it's a few years away but it's a few years away from them like having the design nailed down you know for a first pass at a row this it's just like we're so far from what humanoid robots could be i you know

When I think about the humanoid robot thing, I think about, you know, I've got this iPhone 16 in my pocket. And when I compare it to, you know, the iPhone or the iPhone 2 or 3G or 4 or whatever, like all the steps in between there of significant amount of improvement, you know, Optimus, you know, Optimus 3 is probably iPhone 1.

Right and you know and there's gonna be an you know an optimus 12 at some point and it's gonna be better than Optimus 10 was significantly So I see it as an arc right not like this point in time when oh now we have the robot We just mass-produce it we make a billion You know the the billionth robot is gonna be very different than the millionth robot was yeah, so wait let's say they do optimus gen 3

this year produced 5,000 by second half of next year. Let's say Gen 4 end of 2026, Gen 5, 2027, maybe Gen 6, wait, Gen 6 2028 or 2029. You think Tesla can get to a million robots by then, you think? I wouldn't be surprised if you end up seeing an exponential trend on this because so many things tend to fall into that. You know, so once we...

Once they've fielded a thousand robots, we can look at the next, you know, three months, six months, 12 months and start to figure out what the exponent is. And then you might be able to reasonably extrapolate that out to like a million robots to predict what's going to.

Like, I don't see any good reason why they couldn't, you know, be doing a million robots by 2030. You know, have a... And doing a million robots by 2030 is kind of like having a design that's worthy of making a million of them. Right? Having a, you know... What's the advantage of them, let's say...

Getting to a million robots per year faster than any other company. What's I mean? What yeah, how is that advantageous? So first the market has advantages there's network advantages, right? That you get in some places. I don't know if network advantages. I think robots might be more like cell phones right that

It's not like the iPhone makes so many phones that there's no market for Android phones. I feel like that's less of an effect than like telephone networks, which are the classic network effect thing. It might be comparable to taxi networks, probably a little bit less.

depending on whether, I mean, one thing that could kick in is the network effect of if you field a bunch of robots and they're gathering data and the more data you gather lets you make better robots, there could be a positive feedback loop there where getting first to scale lets you, you know, it gives you a big advantage for being first to an even larger scale. I don't think we've seen evidence of that yet, but given the development, like apparently, you know, we hear that they're doing end-to-end networks.

and that Tesla's, from a software standpoint, a lot of their control stuff is neural network focused. So maybe that will turn into a network effect. I think it's not clear. But as I mentioned before, I don't see the software as the critical thing right now. If you have a highly performance software stack and a robot that costs three times as much and lasts half as long as your competition, that's a big disadvantage in the real world, right?

You don't want to think of robot, you probably don't want to think of robots as like what's the cost of owning the robot, but like what's the cost per robot hour? What's the cost per robot year fully invested? And what set of capabilities do you get out of that? So in that sense, like having more functionality and a longer lifetime, you multiply those two together to get the true value of the robot, right? So both of those matter and they leverage each other.

Do you think Tesla will, I mean, develop, like, will they rely on XAI for the talking and LLM features of the robot? It looks that way right now. I mean, Elon already said they're going to put Grok in it. Is there any, like, tension, you think?

Like, for example, like open AI, you know, might be like, hey, we're going to go into robotics and we think, you know. Yeah, we'll see about the robotics thing. It's an interesting assertion that Altman made that he thinks that they could make a robot stack that's better. Once again, I don't think the stack is the critical limiter on the robot side. I think the hardware is a critical limiter. I mean, I could turn out to be wrong, but that's the way I see it right now on balance. The...

So there's a couple of different ways you could see the final robot stack coming together. Right now, the robot stack and language models don't have a huge amount in common. They're kind of synergistic, but the language stack is kind of giving you language interface, and it's not necessarily unique to the robot application. That could change. They could become integrated. The robot, there's a ton of robot-specific applications

components of developing the stack. OpenAI worked on robotics a while back in a very different way than I think Altman has proposed doing it. So maybe they can build on that, maybe they can't, I don't know. Tesla has worked the robotics stack part of things

If they need a language module to drop onto the thing to provide for a shared conceptual space for instructions and planning and that kind of stuff, you know, they could do worse than to get it from XAI and use Grok for that kind of stuff, especially if XAI is willing to

tweak the model or custom train a model that is specific to robotics applications, that can be super useful. Training large language networks at a base level is not super hard. There are a lot of organizations that do it right now. It's hard if you're going to be at the frontier

And it's also hard if you're going to train a really big model if you need a really big model. A lot of applications don't need a big model. Like for instance, a thing that could work is Tesla could get the rights to distill an advanced version of Grok down into a model that then gets integrated with a robotics model to stick in to use for Optimus. And that might all make sense.

It's pretty hard to narrow things down right now because the software stack is pretty flexible. Okay, thinking a little bit more longer term, let's say five to seven years out, let's say Optimus is a great plumber but looks at different plumbing assignments but needs to figure out what the assignment is or what's broken, let's say, to fix. Wouldn't, at that time, the best frontier, let's say Optimus,

whatever they call them, it might not even be LLM's definition, but let's say the best thinking models can be used to help, let's say, problem solve, identify what's going on, what's the problem, how to fix that, let's say plumbing problem. Like that ship has sailed. And five years is on the other side of the singularity right now, so you can't make a prediction five years out. Yeah, but I'm saying wouldn't optimists need

a smarter brain than just, let's say, a so-called physical AI brain. They would need a thinking brain as well. Okay. So I used Grok to solve a plumbing problem I had a while back. I took some pictures of something and I asked it what the fix was and it figured it out. It told me what the pipe threads were, like what parts I had to buy. Like, the plumbing problem...

Like that problem that you're talking about is basically solved. I mean, it may not be solved in a nicely packaged format, but the set of capabilities that you're talking about, about, you know, essentially being able to diagnose and debug like household things that you might want a plumber, electrician, whatever to do, like...

So that capability already exists. And it exists in kind of broadly available models that are open source capable that could easily run on Optimus hardware. That said, Optimus, you know, he's going to have an internet connection. And so like if he needs to upload something to Big Brother in the cloud and get advice, like that is certainly something that is not going to be a challenge to deploying the robots. Yeah, I like...

The compute is not a problem. Batteries are not a problem. Structure is not a problem. I mean, we need to figure out structural elements. We need to refine actuators. We need to return. We need to refine the overall body design.

But compute batteries, those things are solved. And most of the things that we think of as the complicated software part, the high level reasoning and that kind of stuff, that's actually doing reasonably well. Not arguing that robots, that AI does really good high level reasoning right now. But the level of reasoning where like, here's a set of pipes and here's a problem. How do I reconfigure these pipes in order to make the problem go away?

That's a thing, like it's a mix of integrating knowledge that you can gather from different kinds of sources that can be trained into a model ahead of time and a moderate amount of reasoning to explore a couple of alternate paths. Like plumbing problems don't have two to the four possible solutions and a tree you have to search, right? Plumbing problems have like 10 problems and you can brute force that. And language models can do that kind of stuff right now. So, okay, so...

What about Tesla, you think? Do they have some unique advantages that somehow give them, let's say,

some special advantage with humanoid robots. Like, is it, like for example, is their actuarial design going to be that much better than anyone else out there? Or is it more the combination of various things? Like for example, their experience with real world AI and neural nets plus the FSD computer plus, you know,

mechanical, they say actual design, plus a bunch of other things. Is it just kind of like spread across the board that their expertise put together is unique? Or is there one or two things you think that are especially like crucial and important?

Both. Like, they've got a couple of categories that I think they're extremely good at, you know, relative to other players I'm familiar with in the space that are actually really important. But then they also bring the whole stack. Like, both of those things are a strength to them. So I would say a specialty in advanced manufacturing and, you know, scaling rapidly and

having a culture of being very, very solution focused, getting there quickly, you know, just and that the whole like first principles thing, I think that really helps you at the level of development that they're at right now. On top of that, they have, you know, a big cluster for training. They have a bunch of internal AI capability. Those are all important things, too.

As I said before, I don't think the AI is going to be the structural differentiator in getting to... Not that it's not important. It is important. It's just that given the weight of like, is it more important to be a super good manufacturer of robots or is it more important to be the absolute cutting edge...

you know, software, robotic software stack provider. I think, you know, the robot side of the thing is a bigger advantage than the mechanical robot side of things is a bigger advantage at this point in time than the software side is. Both of them need to advance. Both of them will advance. But

You know, I and so that's the thing, you know, how many of the companies out there that are building robots right now are laser focused on building a billion robots, you know, and looking at the arc to get to a billion robots. You know, they're they're like not going to bother introducing a robot into the market until, you know,

you're going to sell 100,000 of them, right? The first 1,000, 10,000, whatever they're going to use internally for testing and that kind of stuff. Like that's a first volume kind of focus that they've got. It's not, you know, let me do some tricks so I can get some more VC funding, you know, so that I can move forward. Let me try to convince some early customers that my robot is worth using, right?

So that, you know, so that I have the stepping stone to get to 10,000 robots and I'm at 100 robots right now. I feel like, you know...

My sense of Tesla's thing is they're not worried about the first thousand, 10,000. They'll use them internally. They'll find a couple of peer industrial companies to take them and help with development and vetting and that kind of stuff. But the first version, certainly the first version they sell to consumers will be designed to sell in the millions. Like it'll be priced at that point. It'll have the durability. It'll have the flexibility to

to address that thing. The first one that'll go into consumer industrial commercial applications might be 100,000 robot kind of scale thing. Like a design where you build 100,000 of them and that's where it hits its sweet spot. Maybe there are other competitors out there who are willing to eat the first 100,000 robots as part of the development process.

So, okay, so do you think that you're saying the hardware might be the biggest differentiator right now? Is it because you think the software is less differentiated, meaning that there'll be similar capabilities in terms of on the software side across, let's say, the top robotic companies/borrowing from other AI models or whatever? Let me put it this way.

just to put some numbers on it. Not that I want to, you know, say these are the right numbers, but say two years from now, the robot stack is done. Like we have all the core robotic software that we need. Two years from now, we are not going to have a body that is worth making a million of or 10 million of or a hundred million of. And there won't be people who have, who are ready to build a million robots at the, remember,

I think it's easy to think of this as a design problem where like you get the design of the robot right and that now we just go build a bunch of them and we're okay. There's this feedback loop where the more you make, the more you can make because the design needs to undergo further refinement to drive the cost down. You know, 10,000 robots at 100,000, you know, unit cost is not a substitute for... It's not a good...

it's not a good stepping stone to you know a hundred thousand robots at a ten thousand point right i think you although you do have to go through that arc right you have to do all of those intermediate steps you have to start with fewer bots that cost more and work your way down and you and that that that whole arc it needs to be continuous like it needs to make sense at every point of the of the arc to move on to the next step but

To succeed long term, you know, you need to be following a path that gets to that endpoint that you're going at. If you focus too much on it making sense right now and you pick a path that won't make sense when you go 10 or 100x larger, you know, this is the Boston Dynamics thing, right? You know, highly performant $1 million robots have no future.

Right. Because there just aren't that many tasks where a humanoid robot is worth spending a million dollars on. They're very, very, very few. You need we need a ten thousand dollar robot that's highly performant. And, you know, if you develop technologies that can't, you know, be incrementally improved to the point where they become the ten thousand dollar robot or they can't easily like that's not their natural path of development, then then you're.

you're sabotaging your long-term success in order to get your near-term success. Like, this is a common problem, right? Because startups need to work, you know, new technologies need to work at all the intermediate points as well. If you have the

If you clearly have a strategy where that long-term goal is clearly what you're looking for, and you're less concerned about the intermediate stuff because you have deep pockets, because you have, you know, Tesla's AI folks. I mean, they have AI folks dedicated to Optimus, but they, you know...

They have AI that they use for other things also. Like the cost of them having an AI group isn't building an AI group just for Optimus. It's ancillary to the AI group that they already need to have. And there's a lot of synergy, right? So the cost and complexity for them to build the AI group is kind of lower, right? Like they have these other...

things in their court already that make it easier for them to get the intermediate points and allow them to focus on the long-term point because they don't have to satisfy like their VCs don't get their need to get their money back in five years right that's not a constraint that they have yeah when I look at it it's we're kind of getting too deep into the weeds you know my my thesis on Tesla being

a good call in robotics. You know, we're looking at different things and I'm defending this point in ways that maybe are not, they won't maybe not stand up super well to detailed criticism. The overall thesis is, you know, the software's coming along fine, the hardware needs an awful lot of focus, and the specific thing that needs a lot of focus on hardware is getting to highly performant, durable, cost-effective bodies

as quickly as reasonably possible. And I think Tesla has both the right strategy, the right resources,

And the motivation to like do that, not just that they want to do it, but like that they're well incentivized to do it. Like, you know, Tesla is a big enough company that making, you know, a hundred million dollars a year on robots is not interesting to them. Right. The Optimus is an interesting market for them at like,

$10 billion, $100 billion, or a trillion dollars kind of scale. Like that's the thing that moves the needle for them. So if you're a small startup, you might be satisfied getting to a billion dollar a year run rate, revenue runway, or $10 billion or something, but that's not gonna satisfy. So Tesla's internal incentives

are to do the thing that is most likely to make them long-term successful, unlike a smaller company that has less resources. So I tend to think they have an advantage in the space, but you know, it's the singularity. Like it is so hard to predict.

The future, three years out, five years is just like impossible right now. It's just like it's over the horizon and we're guessing from the sparks we see on the far side what's going on over there. Interesting. Yeah, I kind of look at it like...

there's a certain economic value a humanoid robot will give at different functionalities. Like a super performant humanoid robot obviously will give higher economic value. And the cost of that robot to the customer needs to be lower than the economic value that it provides that customer. But the problem is, I think, to this point, and then you have to times that by the number of

of those type of customers there are, right? Who can get that type of economic value for that robot. - It's also worth keeping in mind that humanoid robots like robo-taxis, they have a thresholding thing. If you can't deliver robo-taxi service for less than the cost of a human driving a car, you have no market advantage. If you can't deliver humanoid labor for less than what it costs a human to do that job, you have no advantage. So all of this is effort we're putting into these things

in anticipation of crossing that threshold of being able to supply an existing need at a lower cost point or with higher quality or something like that.

The robo taxi has to first outperform the human rideshare driver in order to really have a case. And then once you cross that, all kinds of things become possible. But until you do, nothing is possible. So there's a threshold. Similarly with humanoids. If you can supply a humanoid to somebody that does manual labor but costs $100,000 a year, it can't compete with human labor.

If you can supply a humanoid to somebody that does human labor and costs $5,000 a year, now you've got a product. Like price is everything because there's an existing thing that meets that need that has to be out-competed before the market appears. And I think the hands, I feel like...

are a huge part of the deal. - Hands are important. - I feel like that's the gateway to tools. - It's half the problem. - It's like all tools and you know, it's the avenue. It's like to usefulness in a way, right? - Human hands set an extremely high bar because human hands are like amazing. Human hands are amazing. They're amazing.

capable. I can reach in my pocket and I can count the quarters and nickels and tell how much change I have by feel. I often...

I'm always amazed at the stuff that I can accomplish in the dark. I can tie my shoes in the dark super easily. It's not even hard. I don't even look at my shoes when I tie my shoes. And that's tactile feedback, motor memory, a good understanding of the dynamics of how shoelaces move and how tension works in shoes and all of that kind of stuff.

Some of that is software, but an awful lot of it is having really high resolution feedback and then having all the degrees of motion, right? Having, being able to like, you know, holding a pair of scissors,

and using them is actually an amazingly complicated. There's all these scissors in the world. They're really made well for human hands. I've never seen a robot use a set of scissors the way a human does a robot hand. I think it's possible and it's going to happen. But the robot hand designs I've seen so far that are performant in the sense that they can compete with the strength and power

and range of motion that human hands can do are unlikely

to be something that you can put in a robot and use in a factory for like three or four years, right? Without them like wearing out and falling off. And if, you know, I wouldn't be surprised if the robot hands, like the hand being everything from the elbow to the fingertips, I wouldn't be surprised if that's half the cost of the robot, right? So like, it's not, it's not just, they need to be performant and they need to be durable because building a performant robot hand that only lasts three months is not a good, it's not good enough.

But the hand is going to be really expensive, right? They'll be half the complexity, half the cost of the robot body, maybe, in the short run. What will happen, one of the reasons that we're in this spot is that

the sensors and actuators that the core technologies needed to make really good hands. Nobody's ever worked on that stuff before. There isn't some other market for some other product which is adjacent to building hands, unlike the cameras. The cameras that FSD needed, we use those for security cameras. We use them for cell phone cameras. We use them for

you know we use them in zillions of other products and so the technology was pretty advanced and so by the time that we needed self-driving cars there were really great really cheap cameras out there and you know the the sensors and actuators that we need for hands you know there there isn't a super high volume super capable actuator right um that um that kind of works that way like you know

The human hand, like our muscles are here. We have some muscles here that do some small things, right? But most of the muscles are here and you've got a set of, you know, essentially cables like the tendons, right? That feed through here to enable all of this kind of motion. If you try to mimic this directly, which is like, I think what Tesla's doing right now is they've moved, they started out, the first generation hand had the actuators in the hand.

But it made the palm big and it limited the range of motion. And they also, they didn't have enough actuators that they could get as many degrees of freedom as are useful in the finger. So now they've moved a lot of the actuators into the forearm and they've got cables coming up. Okay, so you've solved some problems, but cables kind of suck. Cables stretch, maintaining tension is there.

cables, they don't pull and push. So like the human hand, it has, this is one actuator and this motion, the opposite motion is a separate actuator. You have a muscle to flex your finger and you have a different muscle to extend your finger. So I've doubled the number of actuators that you need in there. Well, muscle actuators are really small. You know, the muscle in the back of my hand that elevates this finger, it's like,

It's like a piece, it's not a piece of thread. It's like a shoestring running down the back of my arm, right? Like we don't have a technology for making a shoestring size actuator that can do a finger extension right now. So we do these other mechanical shortcuts. Eventually, we're going to get that actuator that will give you a palm, that will give you a forearm and let you have that actuation, right?

you know, in using, say, a cable, right? And we will come up with the mechanical interconnects and the cable materials and whatnot to make that stuff kind of work. The other thing is, you know, humans have like 150-pound grip strength,

Which you don't need 150 pound grip strength to be useful But you know if you don't have a grip strength in the like 50 pound kind of range You can't use a shovel to shovel dirt right at which you know, maybe that's a thing We don't need robots to do but if like my sense of like a robot that can walk your dog and push your lawnmower well now you need the 50 pounds of grips right and

It's got to be able to hold back your St. Bernard and it's got to be able to like lift your push your lawnmower up, you know, the incline in the front yard, even though the grass is too tall.

So all of these things require you to get into the range of strength and dexterity and sensitivity that the hand has if you're going to be a substitute for human labor. Which is not to say there's, you can fall short of that goal and you can have some utility, but the bonanza comes when you get really close, when you can match those things. And it really can drop into a human world and do all the human things. Open, you know.

When it can, you know, carry your 50-pound feed bags from the car to the garage. When it can, you know, reach into a dishwasher and pull out four forks and separate them with its hand. Or, like, you know, when I empty my dishwasher, I frequently pull out two plates or three plates or two cups or whatnot. Those are all...

You know, there are things that don't, those things don't challenge the strength of a robot hand, but they challenge the dexterity of a robot hand very extremely, right? Every single one of those, you cannot have it and the robot just becomes less efficient. But once again, the threshold is the robot, we want it to be as good at the job as a human is.

in order to match the cost efficiency, so now it has a value proposition as opposed for displacing human labor. I mean, it's not that it's not useful, it's just that it really becomes useful when it can do what people can do, right? So that's a useful mark to have, not because it's an absolute, but because it's kind of a reference for where the market really opens up.

Okay. So how does one go about assessing, let's say, the players in the robotic humanoid robot field?

And how do you even like, let's say for a normal person, is it even realistic for them to even attempt to accurately assess, like, for example, where Tesla is at compared to other players? Especially when, for example, we're talking about, let's say the hand. We don't even know how much the hand is costing, right, for different robotics players. We don't even know the durability, right, of the hand.

whether it lasts one week or one year, we don't know any of those things. And then a lot of this stuff is the companies aren't disclosing anything. So it seems like a difficult...

You can't think that you're going to go out and figure out what all the numbers... I mean, there might be hedge funds that are trying to get insider information on all these things and to benchmark who's where in the market. But yeah, as a retail investor, that's all kind of hopeless. The thing is, I...

you know, the reason that I thought it would be good for Tesla to get into the robotics space was because I thought that they had an unusually good set of core capabilities for being able to perform well in this space. And I still see them that way. I don't actually put a lot of time into trying to exhaustively understand the robotics market. There are a ton of players, you know, it's changing

frequently. There are a ton of unknowns. So I reduce it to big picture items that I think are significant, and then I weigh players on how do they do on these big picture items, fully aware that the world is complicated and things change. What are some of the big picture items?

You know, can, you know, do you have a path to building a billion robots, you know, over the next, you know, do you have a path to getting to a billion robots in like a five-year window kind of timeframe, five years, seven years, that kind of stuff? Like,

you know, scaling to that kind of volume, I feel like is only really plausible for entities that already build, you know, big complicated things at very large volume. So of the people who are in the robotic space out there, you know, who has, you know, a track record or significant, you know, expertise in the space of building complicated mechanical things in very high volume at very tight cost points with very high quality standards.

A car manufacturer is kind of an obvious one because it's parallel in a lot of respects. Do you discount or do you just not trust to say a startup that says in seven years we're going to do a billion and in a year we'll have a thousand?

I don't discount it in the sense that I say it's not possible. Like there's two ways you can take discount. One, a discount is like, oh, that's bullshit. And the other discount is I apply a discount factor to the likelihood. So in that latter count, I do discount them. You know, I might discount the likelihood of them achieving that goal by some percentage. And so I, you know, I will weight that statement accordingly. Got it, got it.

Like people who can do it already, you don't have to discount it or you don't discount it much. You know, you're like, well, they can probably get to it. But, you know, somebody who's currently like using a 3D printer to, you know, an injection molding to make components that really obviously need to be, you know...

I don't know, steel or die cast aluminum or something. It's like, okay, where's your aluminum die caster? Where are you gonna get the experience of doing a billion high quality die casts on your way to doing a billion robots? It's not impossible.

It's just the more different things that you don't have right now that are super challenging that you add to the requirement list, the less likely you are to be able to cruise through that to certain success. Interesting. Yeah, I mean, it strikes me like it feels like Elon gets it. It feels like he...

understands what needs to happen. Like he understands the importance of the actuators, the design, the performance of the hands. He understands what needs to happen to ramp.

You know, he seems to be, he seems to check off a lot of the boxes like you're talking about. I mean... Oh, it's not just him, right? I mean, he's got a ton of smart people working for him. So, like I would say, it's less a question of whether Elon gets it. Like, it's important for him to get the outlines of it. But if, you know, if Elon and, you know...

You know the hundred most senior people in the org like if they get it collectively like that's the thing that matters It's like Elon's 99.9% of this work isn't being done by Elon. It's being done by other people. They need to get it Right so that people can be aligned. It's leadership does a lot, but leadership doesn't do the work, right? so But you know the other

there's a lot of synergies with the things Tesla's already done, the things that Tesla is doing today. And I feel like that gives them a leg up over, I don't know, like, I don't want to name names, but like other robotic startups, right? We haven't seen, like, you know, if Apple said they were getting into space, I'd be skeptical because they don't build mechanical stuff, right? If...

if google said they were getting getting into the space i would be skeptical because you know they haven't built really they haven't built complicated things at scale any startup that does it they haven't done anything at scale you know that's that's kind of a challenge so what if gm said that they were going to do it well okay so gm has a bunch of things going for it they don't have the culture they don't have technology they don't have software so that's a really big problem

Who would be another candidate? Airbus, Boeing? No. I mean, the culture's wrong. It's hard to find the software and the hardware, like, manufacturing combo, right? And to have... I mean, one of the... Most large manufacturers aren't, you know, aggressive and, you know, kind of willing to break with tradition, willing to explore radical alternatives. Like, big...

in the United States, this is for sure, I think less so about China, but manufacturers that really push the envelope and bet the farm

you know radical new approaches to doing stuff like that's not a culture thing that you see at GM or Boeing right like who else makes and like electronics is a very different thing like Apple they don't make their printed circuit boards right they they like printed circuit board manufacturing is super tough icing manufacturing super super tough camera manufacturing super super tough Apple doesn't make any of those things they buy all that stuff from other people so

So Foxconn. Could Foxconn make a robot body skill? Oh, that's an interesting question. Could Foxconn do it? So, you know, Foxconn would be, that would be like an interesting kind of situation. They don't make cars. Like, I feel like cars are a much better match against the skill set you need for robots. Right.

especially electric vehicles because like one of the big challenges in robotics is electromagnetic actuators. You need a lot of capability. I think there's an argument to be made that Tesla's electric, like their motor design team is,

I think there's an argument to be made. They're one of the best electromagnetic actuator teams in the world. That's not something where you go out and you hire some people off LinkedIn and you get it. Even if you have the right people, it takes years to build the infrastructure and the internal culture.

you know, the background to be able to do a good job of that. Electromagnetic actuators are very interesting combination of design creativity, mechanical intuition, and, you know, electromagnetic modeling, you know, that doesn't really exist in other spaces, right? So like Tesla's got that, that's a critical missing factor for robots.

So Tesla can very credibly make the actuators themselves from scratch, new actuators designed for humanoid robots that don't exist in the world today, and then take those to very high volume manufacturing, build them, make them reliable, make them low cost. Tesla's super credible there. If you ask me, is Boeing credible there? I would say probably not. Is Mercedes credible there? Absolutely not. They just don't do this stuff.

Right? Like who else? There are probably machine manufacturers out there who can do similar things at low volume.

But who can do similar things at high volume? I mean, they probably just like maybe HP printer manufacturers. Like they might make actuators in volume at super low cost. They're in the wrong performance envelope in order to really compete. But, you know, that's a place you might look. Foxconn, I don't know what they do outside of like Apple cell phones and that kind of stuff. I know they have a big industrial business. Maybe they have...

set of capabilities in the electromagnetic actuator space that would make them credible in that space. I don't know. Chinese EV manufacturers might be another one. Yeah. Are you happy with Tesla's progress with Optimus so far? I got it like I see so little it's hard for me to have an opinion. Like I don't see anything in what I've seen coming, you know that they've talked about with Optimus that makes me unhappy.

Like, it was cool to see them finally adopt the end-to-end thing so they didn't have the old, you know, the Biden shuffle or whatever was going on with the previous version of the robots. Because, like, that's something that you don't need, right? That's something that's been so fallible for a long time. It was nice to see them just, like, make that stuff kind of go away and adopt the current, you know...

not even the current state of the art, the five-year-old state of the art in making that stuff happen. But once again, my focus is how well are they doing on the mechanics? And I don't get any data about-- we got a little bit of data on the torque peak for the motors.

And I think we got a little info on the slew rate on the motors on one of the Optima ID. And they look good. The fact that they're designing actuators that are specific to the needs of humanoid robots, that's a huge win. That's not an easy design space. It's not like you take a motor out of an air conditioner or something and you repurpose it for a robot or a motorcycle or an electric scooter. None of those motors are actually a good fit for the application. You really need purpose-designed actuators

They're doing that. They look like really good actuators to me. As I've mentioned before, I think the hands are a weak point right now. Hands need to get a lot better, not just for Tesla, for everybody. But the reason they're not there is because the core technologies aren't there. People have got to push that stuff forward. I think Tesla's working to push that stuff forward. We just haven't seen the fruits of that yet. Maybe we will in the next version of

What were your thoughts on the dancing videos, the Simterio? Oh, finally. That was my first thought. Like, that has been doable for quite a long... Like, I can see why Tesla didn't bother doing it. It's not...

It's not critical to make a dancing robot just because other people are showing off dancing robots or robots that can walk on uneven terrain or all that kind of stuff. The thing is, from a software standpoint, those are problems. I don't want to call them solved, but they're things we know how to solve. It's the kind of thing the Tesla's engineering team can sit down, decide they want to do it, crank out a solution, and do it.

But they hadn't done it because it wasn't like my interpretation is. They hadn't done it because it wasn't a high priority. But the thing is, it was one of those things where you're looking at, like, why are they using this? Why are they being so clunky when better methods are known? And so it was pretty cool to see a demo where they were doing that and they were showing off the...

I mean, it's not just a matter of having the software able to do that. You have to have...

you know actuators that are quick enough and have low latency and they don't have high lag you know and they have the torque ratio to be able to do that i mean i had no doubt that tesla was designing actuators with the right set of specs to be able to pull that off but we hadn't seen any evidence of it right because we don't see i mean dancing is it's a challenging thing for a robot it's not super super challenging but it's more challenging than walking like you've got a full diaper right so uh

So it was cool to see that and know and see some kind of direct evidence that they were that they were clearly in that space or, you know, so that was cool. And it's fun to just see a dancing robot. Right. Like it's a joyful kind of kind of thing. So, yeah, I really liked it.

Good job on the dancing robot. I can see that it's not necessarily on the direct path to the long-term goal, but it's cool to see. I mean, would you say that their whole, like, I guess they were demonstrating, you know, Sim to Reel. Yeah. Like, is that an important part of the path to the edit? Well, okay, Sim to Reel is pretty important for robots. Yeah.

You want to do a lot of... So this comes down to the hand thing, okay? If you trip a robot and it falls and it catches itself on its hands, do the hands survive the encounter with the floor? I think an awful lot of robot hands right now, they don't survive the encounter with the floor. So that's one of the reasons why we cannot afford to have the robots fall down, because every time they fall down, we lose a set of hands, say.

or maybe we lose a shoulder or something like that. But it's not good for robots to fall down. So when you're initially developing them, when they're still fragile and expensive and difficult to replace, you program them with software that makes it... Or you use them in ways that make it really unlikely that they're going to fall down. Or you put them in harnesses or whatever to make sure that they don't fall down, which constrains other things that you can do. So, you know, taking Optimus off the harness and having him dancing around...

Well, it might show that you have a disposable optimist here and they don't mind if he falls down and breaks his hands. But it also...

It also suggests that they've gotten to a point where they're comfortable with doing that exercise. So that's kind of something. Symptorheal is pretty important because robots, humanoid robots, have way more degrees of freedom. And they have the inverted plenum problem. They have to balance on feet, which is actually harder than it seems to humans because we're used to it. But it's actually a mechanically extremely challenging thing to do as well as humans do it.

Uh...

So yeah, it was cool to see all that stuff come together and that they're clearly using Simtrite Reel, which is an important thing to have. It's a technology that's been around a while. There's a lot of ways to do it acceptably. I wasn't at all surprised to see that they could do it, like none of those components, right? But it's cool that they're pulling all the things together and giving us a little demo. What do you think is going to be possible in terms of demonstration in the next, let's say, six months for Optimus?

We see so little. I don't know. I would guess that to the extent that Tesla gives us more demo. You know, if you think about Tesla's position with Optimus, they don't have a super strong incentive to put a ton of effort into trying to satisfy retail investors or Wall Street or...

or reveal what they're doing to their competition, right? Reveal where they're at to their competition. So it's unsurprising to me that we don't get a ton of information out of them. They don't need to convince investors of anything. They don't need to convince customers of anything because customers are far away and they're not sensitive to investment information.

uh i think you know throwing a little easy you know to uh produce bone to retail investors every little once in a while to keep the story about optimus credible and whatnot like that's not bad they'll probably keep doing that kind of stuff but in that vein i would expect that what they do is going to be things that appeal to retail investors who are not roboticists

So, you know, I mean, the dancing robot kind of feels like it's in that vein. I get a little bit out of it, but like it doesn't surprise me at all that they're able to do it. Like I'd be shocked if they couldn't do it at the point that they're in. So like it's a little bit nice to see them do it. But on the other hand,

Retail investors who've seen lots of videos of other dancing robots, seeing that Optimus can also do that, I think it's probably much more satisfying to robot. And it may be that the kind of demos we get in the future are in the same vein. The thing that I would be most happy to see would probably be a demonstration of really good finger agility on a hand.

Right, which I could see many different ways that that would happen. To me, using a complicated hand tool that is designed for humans, like that would be compelling to me. Especially like Optimus using a set of tin snips to cut, you know, 16th inch sheet stainless steel. Like that would be super impressive. It requires a lot of dexterity and a lot of hand power simultaneously. And to use a tool which is,

like extremely awkward to use even for human hands right like that would be an example but you could come up with other things like you know the whole that trick where you twirl a thing a pencil on your fingers or whatever that's hard i mean it's hard for a human being to do

It's the kind of trick that you can, if you put work into it, you can program a human to do, a robot to do. But if they did it in a context where it was clear that it was a gaming capability that was built on core capabilities of the thing, that would be cool to see. Because it does... One of the tricks with doing that for humans is that...

For a human to do that, it's the pressure on the back. You feel the pencil move. Or like, you know, you can do this. You see people do this with nickels.

It's the feel of the nickel rolling across your finger, which is the feedback mechanism that you have. So that suggests that you've got really sensitive pressure sensors, right? Now, you can duplicate the trick using cameras, but that's a lot less interesting. It's not going to be as good or as convincing, but you can fake it that way, right? But, you know, if Optimus had pressure sensors on the back of his fingers that were good enough to pull that trick off, like, I would be super impressed. Like, that would be amazing. I don't think...

there exists a core technology for doing that right now, for making that kind of

sensor array that is cheap enough. You see a lot of robot hands now that are focusing on adding texture and pressure sensors to fingertips because that's the high value thing for handling things and whatnot. You don't see them put them on the back of the hand like humans have them because it's useful to have if you have a technology for doing it, but it's got to be pretty cheap before you put really good sensors in places where they're not adding a lot of value.

What year, rough guess, do you think Optimus will be able to pick up a cat with two hands, place it on its lap, and pet the cat? Man.

Is that a tough one? Well, assuming the cat's not sedated. Yeah, yeah. Like, you're just... I mean, there's some cats that are super cooperative, but... Yeah. Let's say a cooperative cat, not like a skittish cat. Yeah, no. I mean, but not... I mean, there's some cats that are ridiculously cooperative. But, like, you know, for a median cat where you can pull off that trick, like, if they could do it inside the next couple of years, I'd be really impressed. I think the...

amount of sensitivity to the cat that it needs that that requires is is quite extraordinary and it requires quite delicate handling to avoid like offending the cat such that it will resist you how about peeling an orange by hand is that like a really hard thing

And that really depends on the orange, right? Like mandarins, the skin just falls off. If we, if we, you know, like let's say your average, you know, sun-kissed navel orange, which is ripe, but we'll just be real specific. Uh,

that's it's it's pretty hard but There the thing is there's so many tricks for making it easier like once you break the skin it gets a lot easier That is a really good trick though like it's a that's a good suggestion I mean peeling an orange is not peeling a grape Like that yeah, that would be impressive yeah, it's our

You know, I like, you know, on the one hand, handling a pencil is really hard. On the other hand, if you put a lot of effort into training a robot just to handle a pencil, that's a thing you can do. Humans kind of get pencil handling as a side effect of all the other things we can do with our hands. Not because, I mean, we do spend time training with pencils, of course. But, yeah.

Chopsticks. That'd be a good one. Chopsticks, yeah. Eating food with chopsticks. Yeah. You know, pick up a peanut with chopsticks. Interesting, yeah, yeah. I mean, you know, human-style chopsticks, not some kind of, you know, specialized chopstick. Yeah, that's pretty hard to do, especially if you can pick up a peanut and then also, I'm trying to think of some food where you have to stab the chopsticks in a little bit, you know.

Ramen. Lift ramen noodles out of a bowl in an appropriate fashion to eat. That's pretty tricky. Human hands are so amazing. There's no shortage of things we take completely for granted that are just unbelievably fine dexterous skills. Yeah.

Yeah, tying shoe laces, super hard. Cracking an egg, right? That's another one. Holding the egg in your hand. I can't do this. With one hand? Yeah. The whole, like, crack an egg without making a mess, and then separating the halves so that you pour, you know, the egg into the pan without getting shell in there? Yeah. Hard. That's really hard to do. Or peeling a hard-boiled egg without making a mess of it. That's another thing I can't do very well. Yeah, yeah.

So, yeah. But, like, my vision of a humanoid robot that is a substitute for human labor, it can do all of these things we're talking about. It can pick up the cat. Yeah. It can care for your baby. It can...

help your grandmother out of bed, help her into the shower. It can, you know, help you load hay bales. It can eat with chopsticks or feed you with chopsticks. It can feed your Asian grandfather with chopsticks when he can't use them himself. Like it can do all these things. And

I have no doubt that this is within the realm of possibility. The obstacle is that, you know, unlike building ICs or solar panels or laser printers or

tires we haven't built this enormous industry with a lot of history to you know solve all the core problems that are really foundational level and move forward yeah and i believe that it's going to take time to do the equivalent thing for the technology we don't have in humanoid robots

Yeah, it's interesting. It reminds me of, I mean, not directly, but a little bit about Apple. They had the software experience with Mac OS, with iTunes Music. They had the iPod experience with miniature devices to design that. So when it was time, they were able to pull their, they had the capability, the expertise to design kind of almost a desktop device

quality OS for a mobile experience and a super high quality phone or touchscreen basically interface and hardware that they leveraged what they had, what they were good at. And in a similar way, Tesla can, and it seems like they are, leveraging what they're really good at. The mechanical stuff, the design. It's a good fit for them. Yeah. Like some of that is serendipity.

You know, I don't think Elon made EVs to put himself in a good position to build Optimus, right? He made the EVs and then I was like, hey, you know, we could do humanoid robots. That would be a great market. And we have the ingredients, right? Like I said, that was why...

you know, before all this stuff started, I said, yeah, Tesla, they're the company to do that. I wish they, I think actually what I said at the time was, I wish they would do it. I don't think they have the time, right? Like, what maniac would take on humanoid robots given all the problems they currently have? And now we know the answer, right? Exactly, exactly. Huh, fascinating stuff. I mean, it's so interesting that

you know we get to follow it and talk about it yeah see how this is the best time to be alive yeah this is the best place to watch all this stuff from i'm sure that you know there are other people in other places looking at other things thinking the same thing yeah but this is me here looking at this and thinking wow this is the best thing ever this is so great and it's like it's cool talking about it first week of robo taxi launch like yeah something is this autonomous driving is happening you know like this stuff really does happen i mean it's it's

It's hard to pinpoint the exact timing of when things reach certain points of maturity, but it's like if you can see where it's headed, the direction. This is a Kitty Hawk moment for truly mass market. Once again, not casting shade at Waymo. Super grateful to those guys. Good job.

I love your service. I'm glad I get to use it all the time in San Francisco. I'm really happy with it. But, you know, the promise of robo taxis is the 40 cent a mile unlimited thing that shows up in 30 seconds with just the vehicle that you want. Like, you know, that's that is where we're going. And if we get there 10 years sooner, 20 years sooner, it makes an enormous difference to the world, especially now.

especially now we're kind of at the point like I was really hoping five years ago that just the superiority of EVs was going to help us make the transition away from

You know using motor fuel to power light transport it seems like such low-hanging fruit for me in terms of like making the world a better place and It feels like Tesla did such a great job with that and and for a while there It looked like maybe the world was moving in that direction now. It seems like it

Aside from Tesla, the world has stepped back from the promise of this kind of thing. And so more than I felt was true a few years ago, like succeeding in AVs, because AVs are naturally EVs. And because AVs have that 5X leverage, right? One AV replaces five ICE vehicles worth of...

worth of miles you know that you know it's another promise for you know transitioning the world to sustainable transport I mean to me it's transport that doesn't make me choke when I'm walking around LA and yeah yeah

So I would be happy to see that transition proceed to pace. And AVs are-- you're our only hope. Robotaxi.

All right, James, thank you for hanging out all day. I'm probably going to split this into two videos. Or five. It's a lot of video. Optimus section will be probably the last part. And I'll probably split that. Yeah, it's been fun hanging out. Go ahead and end this. And yeah, we'll see you guys later. Take care, guys.