Speaker 0:  everyone is obsessed with the next chatty ai, but jan lecun, just raised a billion dollars to build something completely different. he says llms are dead. they're a dead end, too much text, not enough physical intelligence. today, we're looking at ami labs and why the biggest names in hardware are betting everything on radical new architecture called jepa. so jan lokan, one of the three turing award winners who define modern deep learning, has cofounded ami labs and pulled in a 1,030,000,000 in seed funding. their core argument, lms like chatttp or claude are hallucination machines because they only understand text, not the physical world. ami is betting that j e p a, i'm gonna call it jeppa, standing for joint embedding predictive architecture, which creates world models, internal abstract representations of how reality works with backing from nvidia, toyota, and samsung. this isn't just a theory. it's an industrial play for robotics and autonomous systems. the big question, can they scale this math before llms simply get better? what do you think about this, john?

Speaker 1:  i love this. this excites me so much. i am so excited to dive into this. i think this could be a game changer, but let's talk about this fundamental disagreement of prediction versus understanding. so llms today, they're trained around variations of a simple objective. given a sequence of tokens, predict what comes next. and so scale that process across trillions of tokens with enormous neural networks, and the resulting systems develop extraordinary capabilities in language, coding, mathematics, summarization, and it's increasingly in multimodal reasoning. lacun's objection is that next token prediction may be an indirect route to intelligence rather than the architecture intelligence fundamentally requires. the reason being is because language is only a compressed description of reality. right? there's countless physical relationships that humans understand that people rarely bother writing down because to us, they're obvious. right? so we don't publish billions of sentences explaining that a mug remains on the table when we look away, that pushing it toward an edge would make it fall or that grabbing it requires approaching it rather than simply announcing that we grabbed it. and so lacun, therefore, distinguishes between knowing descriptions of the world and having a model of the world. and that distinction, that's what becomes particularly important once ai systems must act rather than merely answer. so for those of you that i kind of touched on it, but the world model, it's an internal kind of simulator. right? so suppose a robot sees a cup on a table and wants to pick it up. a conventional perception system might identify cup, table, and distance. right? the world model tries to learn something deeper. for example, if i move my arm in this direction, the hand should approach the cup. if i move the gripper at the correct point or if i close it at the correct point, the cup should become attached to the hand. if i move upward afterwards, the cup should rise rather than remain on the table. and then, therefore, the model can internally evaluate possible actions before committing to one. and so this creates the pipeline of observe, represent, predict, plan, act rather than what the llms currently do is prompt, generate output. i love this. i think this is where this is how we're gonna get to, agi, but i don't know. what are your thoughts on this, craig? do you think this is gonna challenge the current model the current, learnings? or what are your thoughts?

Speaker 0:  i think all this stuff is exciting because when anybody takes a new approach to these things, it brings new revelations, new ideas, and even the old players will start building some of this into their tools and products so that, the world gets better. whether they use, jepa or something else specifically, because, like, these companies like to make their own acronyms for whatever they have. that is the equivalent thing. i think this is great because we are moving very fast, and ai is creating most of our software and whatnot now. and when they can share this kind of thing and ais can pick this stuff up quicker than ever before. we'll see this kind of, stuff in active robots within a week of them releasing this rather than months. so that's why i'm excited about how fast these things go. will this be the end all and be all? i don't know. but it's like bluetooth one wasn't great, but here we are. everybody uses bluetooth headsets now.

Speaker 1:  yeah. yeah. and the thing is this isn't, it's not just theoretical anymore. because before leaving meta, researchers worked with him and developed v dash jepa two, which is a video based jeppa system. and so the model was pretrained using more than one billion hours of internet video. it then received less than sixty two hours of unlabeled robot interaction video to create an action conditioned version called v jeppa two dash or two dash a c. so, like, this is like, there is no explicit reward function was required for the task. no training data needed to be collected from those particular robots in those environments. so that's essentially where we're getting to where lacun is getting at least with, you know, it's gonna take less and less training to get to, get to the end result at the end of the day because it's more of a contextual than just straight, predictive learning, if that makes sense.

Speaker 0:  absolutely. and so i like to think of this like an llm is a chef who has read every cookbook in existence but has never stepped into a kitchen. it can describe a souffle perfectly, but if you give it a real whisk and eggs, it doesn't know the physics of how a batter or a beater behaves. ami's world model is the chef who spent years actually cooking. it knows how the ingredients interact. it knows how heat changes the state of the food, and it can simulate the outcome before it even turns on the stove. so just to kind of phrase it in the people that we understand, the investor list is a who's who of hardware. when toyota and nvidia write checks to the same start up, they aren't looking for a new chatbot. they're looking for the brain for the next generation of hardware. so the confidence is high in where this is going. and he just left meta, a couple of months ago. right? i'm just scrolling back here. early twenty twenty six. he left us to found ami labs. so it's interesting to see where this is going. jeppa is not trying to generate an image or text. it's trying to learn from the structure of reality. so it's a little bit different than, what we can kinda do. so we have, the beer case situation where the lm wall hasn't fully arrived. o one style reasoning models, are showing that with enough compute transformers may be able to fake reasoning well enough to handle a beer case, you know, how to fill it out and put stuff in the 24 thing. so that's a model that people are really having trouble with. failure mode. ami gets stuck in a permanent research loops. they're spending $1,000,000,000 on brilliant papers that never translate into a product developers can actually work on. these are skeptics, but they're saying, again, this is not necessarily reality. and real world benchmarks. so ken jeppa holds its own in a dynamic chaotic environment like a sidewalk in downtown tokyo compared to a massive hyper, train transformer. so it's high stakes game, a high reward research bet. it's most important anti trend in ai right now. so it's a big bet, and i think it's going to maybe not, you know, version two or ver but maybe version three. i think this will define, how we start moving out of the tech space and into a world, three d space. that was always a problem with llms. like, if you have a mug on a table and glasses and you fold it like this and then you put it down and you pick it up, l they don't understand hinges. they don't understand how objects can change. and as you build these real world models that actually predict, automated driving will become better. smart glasses will, help people do more things. it'll also record and, have these robots getting smarter. so i'm excited about where this is going. and i love you, yeah.

Speaker 1:  go ahead. let's just talk about why lacun believes llms hit fundamental limits. let's just dive a little bit deeper in that. like, language contains less information than physical experience. we talked about that. actions need understanding. llms don't naturally have persistent world state. like, for example, they don't understand object permanence, which is critical. so a truly autonomous agent will need a persistent representation of the environment that evolves through time, which they don't have. and planning requires evaluating futures. what happens if i do a? what happens if i do b? which outcome gets me closer to the goal? world models create that framework for that form of deliberation. so, i thought i was gonna get a little bit more pushback from you on this one as, like, as far as a little bit more argument, which i'm glad i'm glad we're on the same page with this. think, i think that, i think that lacun has something, and i think if done right, he could, he could become the next big anthropic or open ai. that's my prediction.

Speaker 0:  and we will see because we had some similar technologies that aren't really world models but are quantum in nature that control robotic movement at the edge, understanding if an elbow moves is what happens at the and they can do that in real time. so i'm not a 100% that this technology is gonna be the leader or the winner, but i love the fact that we're looking at these problems from different angles. i let to extrapolate that, openclaw was one of the first harnesses out there that took off, and now how many other harnesses are out there? every major company, claude, kodex, everyone has their own now. and it changed the they're not the same, but it changed the way they looked at how people use ai and how we can get more ai used and how we can use it better and harness these things. and the world changed as a result in six months. this kind of thing can do that at the robotics level because we're still at the infancy stage, and we can take this and apply the other technologies that we're using and come up with something far superior in one generation where we would historically have taken ten years to do these things. so i'm really excited. and i laugh when you say we'll be opposed. like, i'm not you know, his philosophy on open source, and geopolitical and stuff, i don't really care about. and moving stuff forward, if we move these things and it helps everyone, i'm excited because i'd like to see things moving, the ball forward. so i'm excited.

Speaker 1:  yeah. yeah. me too.

Speaker 0:  so ami labs is the most significant hedge against the current ai hype cycle, trading token prediction for physical understanding. we spent the last two years teaching computers how to talk. the next ten will be about teaching them how to be in the real world. drop a comment. do you think l the llm error is peaking, or we are just getting started? let us know what you

Speaker 1:  think.