# Language models are changing what robots can understand

Markdown copy of https://realalpha.app/s/language-models-are-changing-what-robots-can-understand/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Kyle Vogt](https://realalpha.app/builders/kyle-vogt/) · Kyle Vogt is CEO of the Bot Company. · Co-founded Cruise, a self-driving car company GM acquired, and now leads The Bot Company's household robots.
- Host: [Jack Altman](https://realalpha.app/people/jack-altman/) · General partner, Benchmark · Co-founded Lattice and ran it until 2023, then started his fund, Alt Capital.
- Show: Uncapped
- Episode: [The Breakthrough For Home Robots with Kyle Vogt, CEO of the Bot Company | Ep. 32](https://realalpha.app/v/XQoFbvyWEy8/), Nov 12, 2025
- Moment: 0:32 to 5:37 of the episode (5 min), playing on https://realalpha.app/s/language-models-are-changing-what-robots-can-understand/; the episode from there: https://realalpha.app/v/XQoFbvyWEy8/?t=32

## Quote

> “You can take all the **common sense that's on the internet** and inject it into a robot brain.”
> — [Kyle Vogt](https://realalpha.app/builders/kyle-vogt/) at [2:22](https://realalpha.app/v/XQoFbvyWEy8/?t=142) · Uncapped

## Takeaways

1. Robots can use knowledge learned from internet-scale data.
2. Neural networks can learn coordinated motion from demonstrations.
3. Most robots will be specialized for particular kinds of work.

## About this segment

Kyle Vogt explains how language-model knowledge and learned motion control remove old robotics bottlenecks. He expects a wave of robots designed for different tasks, not one machine for everything.

From “The Breakthrough For Home Robots with Kyle Vogt, CEO of the Bot Company | Ep. 32”, Nov 12, 2025. [Watch the full episode](https://realalpha.app/v/XQoFbvyWEy8/)

## Transcript

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

[0:38](https://realalpha.app/v/XQoFbvyWEy8/?t=38) You know, it's obviously been really important for a long time, but in the last few years, it seems like a lot of really good entrepreneurs, a lot of good investors have started to pour a bunch of time, money, resources, and effort into this. And I guess I'm curious just to start with sort of laying a foundation of like, can you put this in some context and like what used to be the case and what has changed that's like making people so energized right now?

[0:57](https://realalpha.app/v/XQoFbvyWEy8/?t=57) Yeah, it is. It's like the most excited I've ever seen people in robotics. And, you know, I guess as an engineer, there's something like romantic about building machines to do the stuff that we don't want to do. And that's why I've been doing this for so long. First with, you know, a decade on self-driving cars. But for me, even going back to like teenage years doing BattleBots and then going to MIT to basically build more robots.

[1:18](https://realalpha.app/v/XQoFbvyWEy8/?t=78) but you know during that entire spectrum it's also it's always been this niche thing and frankly like robots have never really lived up to their promise there's always something they're always overly fragile like in a factory environment we put them in cages and if things don't line up like within a millimeter the whole thing doesn't work except the battle bots those were good actually battle bots were yeah you made them with like the saws and everything ours had like a hydraulic axe which was which was pretty cool but the calling these robots is a bit of a stretch they're basically glorified rc cars yeah right with a weapon yeah so so with a weapon what's different now is you know for the first time you have robots that are powered by essentially they have all the brains of an llm built into this robot and we're controlling them with neural

[1:59](https://realalpha.app/v/XQoFbvyWEy8/?t=119) networks instead of classically engineered algorithms and so the difference was before if you have a robot that's like in a room like this uh even saying like go to the whiteboard is almost like an impossibly hard computer science problem it's like okay i have to build an exact 3D map of the world, have a detector that can figure out what a whiteboard is, train it on millions of examples of what whiteboards look like just to be able to do this. And even then, the failure rate would be high if you put it in a different room and it doesn't have a map.

[2:26](https://realalpha.app/v/XQoFbvyWEy8/?t=146) But now it's almost like cheating. You can take all the common sense that's on the internet and inject it into a robot brain. And so if you're like, where's the whiteboard? It knows instantly. You open the chat GPT. If you open video, you can show it anything and it knows what it is. Yeah. And so imagine robots before started with zero knowledge of the world, and now like suddenly have this kind of knowledge of the world.

[2:45](https://realalpha.app/v/XQoFbvyWEy8/?t=165) Like better than us. Like they can look around the room and see stuff better than we can. Yeah. And then on the motion side, you used to have to have a PhD to compute these complex trajectories. You have like 12 joints on a motor or on a robot. How do you get all 12 joints to move in tight coordination to like move an arm to a place? And this is a very difficult and computationally intensive problem.

[3:04](https://realalpha.app/v/XQoFbvyWEy8/?t=184) And now we just kind of jump over that whole thing. And now if you have a way to teleoperate a robot or to put it in a simulation, you can just learn how to move all those joints to mimic the human operator or to accomplish, you know, some reward function or to maximize it. And so you can skip that whole computational challenge. And so those two things together basically mean that everything we thought we knew about robotics or like what kind of businesses were good businesses or bad businesses, all like that slate has been wiped clean. And so I think you're going to see this Cambrian explosion of different robots for different applications that now suddenly just work. Whereas before they were like really struggled to do the most basic things and when you say for different applications is that are you saying it won't be terribly

[3:43](https://realalpha.app/v/XQoFbvyWEy8/?t=223) generalized will it be medium generalized like what made you say for different applications or like different environments maybe yeah i mean classically like a lot of robot businesses like try to really get really really narrow the successful ones we're going to focus on this one problem like factory automation for 3pls for putting things in box at the box. and putting them on a conveyor belt, like very specific.

[4:02](https://realalpha.app/v/XQoFbvyWEy8/?t=242) And that's just so you could narrow the problem enough to be good at it. Now, I think you're going to see people broaden the horizons a little bit because it's much, much easier to go from, you know, a piece of dumb hardware to something that's performing a useful task. You know, I say multiple applications too, because it's my view that there's going to be a whole bunch of different shapes and sizes of robots, each optimized for different types of work.

[4:21](https://realalpha.app/v/XQoFbvyWEy8/?t=261) You know, as opposed to maybe like a humanoid robot, which, you know, is very, very expensive, but in theory can do everything. I think we're probably going to see some of those, But the vast majority of robots will be more special purpose in nature. I feel like there was a moment in AI where the researchers who were sort of closest to the work were like very sure it was going to work before the rest of the world sort of knew.

[4:41](https://realalpha.app/v/XQoFbvyWEy8/?t=281) Is there an equivalent thing in robotic? Like have people crossed a similar threshold to like whatever that was, like the pre-Chat GPT moment in robotics where like some people who are at the very front have been working on it for decades are like, this is definitely happening now. Yeah, if you had like secret microphones in like robotics labs across the country right now, you'd just be hearing, holy shit, holy shit, holy shit. It's like constantly happening. And I think finally, like the light bulb moments are happening. And it all like in the early days of this stuff, it all looks very rudimentary and kind of simple. But if you know what you're looking at, you see the signs of life that mean over the next three to five, 10, you know, even less years of development, this will go from an interesting technology in a research organization to like,

[5:25](https://realalpha.app/v/XQoFbvyWEy8/?t=325) you know, broad mainstream appeal. And yeah, those signs of life are happening. Those light moments are happening all over the place right now. So what are the components? Like there's