Yann LeCun’s New AI Company and World Models

27:00Los Angeles, California currently, which is um uh >> That's an average March Friday so I don't know. >> Uh we're finally over the rain, I think, apparently, uh which is great. And our last story is is, you know, interesting because it dovetails with a lot of the AI stories we've talked about. And there's both a human personality side to the story and sort of like a an actual science part to the story. And this is related to uh Yann LeCun, who has recently left Facebook or Meta uh in early 2026 and has now raised a billion dollars to build AI that

27:41understands the physical world. And this is related to This was covered by Wired, and there's there's a lot of background to this story. I'll kind of just quickly paint a picture before we kind of get into the physical world AI piece. You know, in terms of if you've listened to any of the frontier model CEOs over the last 12 months, so this is Sardemus Sassabis at Google DeepMind, who runs all of Google's AI products, Gemini, everything that's integrated with Google Workspace. You have uh Dario Amadei at uh CEO of Anthropic, who has the Claude and Claude Code and Co-work. They've been very much

28:23in the zeitgeist and in the mainstream because of their ongoing kerfuffle with the Pentagon over usage of Claude within military applications, specifically autonomous weapons and mass surveillance, uh primarily focused on civilian, but you could argue generally. And particularly with Davos, the Davos speeches that just happened all all of them were there and talking about all the you know all the journalists are like where is this going and when are we going to get to quote AGI artificial general intelligence or ASI artificial super intelligence and can simply scaling the current architectures meaning putting more compute and more

29:04data into it get us there or do are we missing a fundamental architecture or concept that needs to be included in order to actually make that that jump and Jan LeCun has been very vocal about his perspective that the scaling laws are not going to get you to AGI. >> Yeah yeah he's he's been very vocal about that and he's had some real like really nice like little analogies about why he thinks that is he he equates it to you know the models of Ptolemy >> Mhm. >> on the epicycles and how you know before Copernicus everyone was trying to fit

29:44the earth as the center of the universe and so they came up with orbits around earth but obviously Mars does this weird retrograde motion where it goes backwards and so Ptolemy decided okay I'm going to put an epicycle which is Mars has a big circle around the earth but then around that circle there's a little circle that Mars goes around so every once in a while on that little circle it's going to go back but then as you get closer and closer and the data gets harder and harder you just start adding more and more epicycles and it turns out because of something called Fourier transforms in mathematics for infinitely the more the more epicycles you add the closer and closer you'll get to the ground truth and you'll never discover Kepler's ellipses

30:24if you don't have a fundamental change right and and Jan LeCun is saying that basically our scaling hypothesis which is just make the parameters more and more we're just adding epicycles and fitting data but the underlying parameter of Newtonian gravity and the one over r squared dependence were missing that entirely. I like that little >> That's a great That's a great like reference point. >> Yeah. >> And just to briefly kind of touch on So, you know, people talk about chat GPT and cloud and all this and in the fundamental architecture that is undergirding them is this idea of large language models. >> Mhm.

31:05>> And you know, so the idea is you're optimizing language as a as a an abstract or as a construct. And both Demis Hassabis and Yann LeCun in the last 6 to 12 months have discussed the importance of world models. Um, this idea that language models are not actually having this fundamental understanding of the physics of reality, which is what has maybe made it a little bit more difficult to get image generation and video generation that out of the box feels real. >> Mhm. >> What seems to have happened is these LLMs have sort of created their this this like uh abstraction of what they like what

31:46they understand physics to be. >> Yeah. >> But it is not actually grounded to like the real fundamentals. >> Yeah. >> And so this has now become the new hot topic is world models. Actually believe Google's already introduced for public consumption a version of their world model >> Oh, okay. >> where you can go into the interface and it's procedurally generating a 3D environment >> Okay. >> in real time as you move around it. And you can choose the POV of like a cat or you know, a bird flying through the air. And so it's not like a video game where all the levels have been explicitly coded by the developers already. And then when you get to the edge, you can't walk deeper into the forest cuz there's an invisible fence in the Pokémon

32:28forest. Uh, it just continually will procedurally generate generate on the fly as you move around. So, as far as I'm aware, that's the first world model kind of thing at scale that's launched. But, it seems like Jan agrees with this viewpoint and is trying to push forth. And with his new company, Advanced Machine Intelligence, they're trying to focus on this idea of building AI world models to actually understand like physical 3D spaces, which becomes much more valuable in applications like having any kind of humanoid robotics do anything in a physical environment,

33:09whether that's in a manufacturing context, in a residential context. Um Driving's maybe a little bit different uh because it's such a rules-based system. Yeah. Whereas, having a home humanoid robot at home, where you want to be able to have it fold your laundry, do the dishes, pick up stuff off the floor. It's like unstructured tasks. >> Yeah. And I think just in general, too, even for language, right? And like producing text that makes sense, or producing recommendations that make sense. Having a world model is a is a huge deal, right? It'll give you recommendations that actually make sense because it understands how fundamental physics works, and so on and so forth, right? >> There's one other sort of interesting

33:50palace intrigue, corporate drama aspect to this story, which is um Meta ended up acquiring uh so, Mark Zuckerberg ended up acquiring this sort of AI infrastructure company, Scale AI, which was led by this 20-something, you know, super genius uh Alexander Wang. Not that Alexander Wang. Uh different one. And, you know, with no E between the D and the R at the end, which is interesting. And they acquired a Meta acquired their Scale AI for uh $14.3 billion. And the idea was it was they were building like a tagging and training kind of pipeline system that was supposed to accelerate, you know, OLAMA and all of Meta's AI work. Now, he

34:35Alexander Wang was brought in and became Meta's like first chief of AI, right? Or our head of AI, right? And you already have the legend Yann LeCun there, right? And so >> Yeah, and Yann LeCun is not your head of AI. >> Right. And and what seems to have been from the outside looking in part of the drama here is, you know, Yann is more on the researcher uh kind of viewpoint. So, you could probably categorize like Sam Altman and Mark Zuckerberg in the scorched earth philosophy. Just spend money, get to the scale fastest cuz whoever gets there first wins. And then the Yann LeCun and Demis Hassabis are more on the if we have the deepest research bench,

35:15we will be the ones who win. And research is the driver of product, not product being the driver of research or product being the driver of implementation. And Dario at Anthropic is like kind of somewhere in between those two. So, it seems to be the case that he was uh sidelined because he wasn't moving quick enough or Zuckerberg was worried about losing. And one of the reasons that Mark Zuckerberg is so motivated to not lose the AI battle against everybody else and why he was one of the first movers to really expect like do a huge amount of capital expenditure to build out Meta's infrastructure is when we were going through the transition from desktop to mobile in the

35:55early 2010s, Facebook missed the opportunity to build the hardware device for mobile. And what ended up happening is they became victim to Apple and Android as related to getting distribution to their end customers. So, there was a famous battle between Facebook and Apple where they were like, "If you want to keep stealing users' data, uh we're going to kick you out of the app App Store." >> Okay. >> And Zuckerberg never again, he vowed never again to allow someone to be able to control his distribution in that way. That's why he went big on VR and Oculus because he's like if that's the next medium, no one's going to beat me. That's why he went big on AI and all

36:35this because if that's the next medium, no one is going to beat me. However, all Apple has to do to kind of really kill a lot of these and this is kind of the the bare case for a lot of the foundational models. If Apple finally decides to do something good and do a local only relatively high quality model on device