Optimus needs real-world self-play to close its data gap

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 1:20:07

   One intuition you could have for when we get this robotic stuff is that it takes quite a few years to go from the compelling demo to actually being able to use in the real world. So 10 years ago, you had really compelling demos of self-driving, but only now we have Robotaxi and and Waymo and other services scaling up. Shouldn't this make one pessimistic on, say, household robots? Because we don't even quite have the compelling demos yet of, say, the really advanced hand.

2. 1:20:39

   Well, we've been working on humanoid robots now for a while. So I guess it's been five or six years or something like that. And a bunch of things that we've done for the car are applicable to the robot. So we'll use the same Tesla AI chips in the robot as the car. We'll use the same basic principles. It's very much the same AI. You've got many more degrees of freedom for a robot than you do for a car.

3. 1:21:11

   But really, if you just think of it as like a bootstream, AI is really mostly compression and correlation of two bootstreams. So for video, you've got to do a tremendous amount of compression, and you've got to do the compression just right. You've got to compress the, like, ignore the things that don't matter. And like, you don't care about the details of the leaves and the tree on the side of the road, but you care a lot about the road signs and the traffic lights and the pedestrians and even whether, you know, somebody in another car is looking at you or not looking at you.

4. 1:21:50

   some of these details matter a lot so if it is essentially it's got to turn that 1.5 gigabytes a second ultimately into 2 kilobytes a second of control outputs so many stages of compression and you've got to get all those stages right and then correlate those to the correct control outputs the robot has to do essentially the same thing and you think about what humans this is what happens with humans we really are photons in controls out So that is the vast majority of your life has been vision, photons in, and then motor controls out.

5. 1:22:27

   Naively, it seems like between humanoid robots and cars, the fundamental actuators in a car are like how you turn, how you accelerate, et cetera. Where in a robot, especially with maneuverable arms, there's dozens and dozens of these degrees of freedom. And then, especially with Tesla, you had this advantage of like you had millions and millions of hours of human demo data collected from just the car being out there where like you can't equivalently just deploy optimists that don't work and then get the data that way.

6. 1:22:54

   So between the increased degrees of freedom and the far sparser data. Yes. That's a good point. How will you use the sort of Tesla engine of intelligence to train the optimist mind? now you're actually you're highlighting an important limitation and difference between cars it's like we do have we'll soon have like 10 million cars on the road and so that's it's hard to duplicate that like massive training flywheel for for the robot what we're going to need to do is build a lot of robots and put them in kind of like an optimist academy so they can do self play in reality.

7. 1:23:40

   So we're actually pulling that out. So we're going to have at least 10,000 Optimus robots, maybe 20 or 30,000 that are doing self play and testing different tasks. And then Tesla has quite a good reality generator. like a physics accurate reality generator that we made this for the cars, we'll do the same thing for the robots, and actually have done that for the robots. So you have a few tens of thousands of humanoid robots doing different tasks, and then you can do millions of simulated robots in the simulated world, and you use the tens of thousands of robots in the real world to close the simulation to reality gap.

8. 1:24:30

   Close the sim to real gap. How do you think about the synergies between XAI and Optimus, given you're highlighting, look, you need this world model, you maybe want to use some really smart intelligence as the control plane, and so maybe Grok is doing the slower planning, and then the motor policy is at the lower level. What will the synergy between these things be? Yeah, so Grok would orchestrate the behavior of the Optimus robots.

9. 1:24:57

   So let's say you wanted to build a factory. Then Optimus, then Grok could organize the Optimus robots, assign them tasks to build the factory to produce whatever you want. Don't you need to merge XAI and Tesla then? Because these things end up so... What were we saying earlier about public company discussions? We're one more Guinness in, Elon. What are you waiting to see before you say, we want to manufacture 100,000 optimists?

10. 1:25:33

    Is it like... Optimize. Since we're defining the proper noun, we can define the plural of the proper noun too. So we're going to proper noun the plural and so it's optimize. Okay. Is there something on the hardware side you want to see? do you want to see better actuators or is it just you want the software to be better what are we waiting for before we get like mass manufacturing of gen 3 no we're moving towards that i would we're going forward with semester actually but using current um current hardware is good enough that you are gonna you should you just want to deploy as many as possible now i mean it's very hard to scale up production i said uh but uh yeah but i think optimus 3 is the right version of the robot to, you know, to produce maybe something on the order of like a million units a year.

11. 1:26:20

    I think you'd want to go to Optimus 4 before you went to 10 million units a year. Okay, but you can do a million a year at Optimus 3. Yeah. I mean, it's very hard to spool up manufacturing. Yes. So like manufacturing, like the output per unit time is always follows an S-curve. So it starts off agonizingly slow, then it has this sort of exponential increase, then a linear, then a logarithmic outcome until you sort of eventually asymptote at some number.

12. 1:26:50

    But Optimus initial production will be, it's going to be a stretched out S-curve because so much of what goes into Optimus is brand new. There's not an existing supply chain. As I mentioned, the actuators, electronics, everything in the Optimus robot is designed for physics first principles. It's not taken from a catalog. These are custom-designed everything, literally everything. I don't think there's a single thing like that.

13. 1:27:17

    How far down does that go? I mean, I guess we're not making custom capacitors yet, maybe, but there's nothing you can pick out of a catalog at any price. So it just means that the Optimus S-Cove, the units per output per unit time, how many Optimus robots you make per day, whatever, is going to initially ramp slower than a product where you have an existing supply chain. But it will get to a million. When you see these Chinese humanoids,