Are LLMs a Dead End? : Why AI Industry Betting on Something Bigger!

Every AI is missing one thing a two-year-old already has, a picture of the world. And Wall Street is betting $700 it doesn't matter. Your phone can pass a law exam. It can write poetry. So, why does it still tell you to turn left into a lake? A two-year-old would not. She has never read a single book, but she knows a cup falls down. She has one thing every AI on Earth is missing, a picture of the world running quietly in her head. Earlier this year, in February 2026, Nvidia lost $260 [music] billion in a single day. It happened immediately after a record quarter. The experts don't even agree. Some call it a bubble, some call it a boom. But, watch what the researchers who actually build this AI are doing. They're quietly moving on to something new. This is not doomscrolling. In June, India's [music] central bank warned that stretched AI valuations could spill into the wider market. Apollo argued that the AI trade [music] now belongs at the center of a rebuilt portfolio. One finance paper gave the honest version, [music] "This can be a real technological revolution and still form local bubbles. That is the danger, not that nothing is real, that too much of the system is now leaning on one thing being real for long enough." So, is it a boom or a bubble? The answer is why this moment is dangerous. Both stories [music] can be true. The build-out is real. Chips, power, and data centers are becoming actual infrastructure. But, the return on that build-out depends on assumptions, how long the hardware lasts, how much [music] power costs, and whether customers keep paying. That is the split. One side sees earnings beginning to [music] justify the spending. Other sees a chain of expectations that has to keep holding. >> [music] >> Same machines, same money, opposite verdicts. When the numbers disagree, watch the builders. [music] Yann LeCun left the company where he helped shape modern AI to build something else. Fei-Fei [music] Li's world labs crossed the billion-dollar line on the same bet. That does not prove world models will win. It [music] tells you something more useful. Some of the people closest to this technology thinks scaling systems that predict language [music] is not the whole story. They are putting serious money behind systems that have to model action, [music] space, and consequence. That is the question underneath the market argument. A system can predict what [music] comes next and still have no picture of why it happens. If you want to understand the bet on world models, you have to go back to the first person who asked whether a mind carries a small model of reality 83 years back to 1943. [music] 1943. A British scientist named Kenneth Craik writes one line that barely anyone noticed. >> [music] >> The mind, he says, carries a small-scale model of reality inside your head. A little simulation [music] you use to try things out before they happen. Think about reaching for a cup. Your senses and memory have taught your brain >> [music] >> what solid objects do. Before your hand moves, your mind has already simulated how the cup and table [music] will behave. That idea sat quietly for decades. Then, in 2018, two researchers did something that sounds like science fiction. They trained an AI to drive a race car entirely [music] inside its own hallucinated dream. If you have seen Inception, picture a world building itself from nothing around someone moving through it. That is almost exactly [music] the idea. To be clear, the AI is not dreaming the way you and I do. [music] It is running a simulation, but dream is the word the researchers chose because >> [music] >> that is how it behaves. Forget everything you think you know about chatbots. To understand this, we start with a baby and a cup of coffee. Okay, let's strip away every bit of jargon. >> [music] >> You know a coin flip is 50/50, not because you read it in a book, because you've lived it. You've dropped things, [music] spilled things, bumped into things your whole life. That's a world model. Now think of your AI chatbot. [music] It read every book ever written about swimming and never once touched water. It knows the words. [music] It has never felt the splash. A world model learns the other way, the way a child does. Knock over the coffee, watch it run across the floor [music] and learn cause, effect, consequence. That's the difference between reading about the world and living in [music] it. This is the part nobody says out loud. This, right here, this is mechanically why your AI makes things up. Okay, for the builders watching, here's the exact difference in one line. [music] A language model predicts the next word. Give it the cat sat on the and it guesses mat. That's it. That's the whole trick, a very, very good auto complete. >> [music] >> A world model predicts the next state of the world. Given where everything is and an action [music] you take, what happens next? One finishes your sentence, the other runs a simulation of reality. And those are not the same kind [music] of intelligence at all. Go one level deeper and here is the twist. The hard part was never predicting. It is predicting [music] what matters. To know where Jupiter will be in a hundred years, you do not simulate every atom in the solar system. You need six numbers. That is the whole idea behind Yann LeCun's approach. He calls it Jeppa. [music] Instead of predicting every single pixel of the future, it predicts in a stripped-down space where you have thrown away the noise and kept only the part that is actually [music] predictable. It is not a bigger brain, it is a smarter one. Now the experiment that should keep every AI engineer up at night. Researchers trained a model on the streets of New York, nothing but taxi trips, millions of them, and it got good. Ask it the best route across the city, and it found it 97% of the time. It looked like it knew New York cold. So, the researchers got clever. They opened it up and rebuilt the map the AI had in its head, and the map was impossible. Streets folding through [music] each other, roads flying over other roads, a city that could not exist. It never had a map at all, just a brilliant pile of guesses. And the second they added one small [music] detour, a single closed road, the whole thing fell apart. Think of a tourist who memorized every turn to the station, confident, fast, perfect, until one road is shut and now he's completely lost. He never understood the city. He memorized it. This, right here, is mechanically why your AI sounds so sure of itself and still gets it wrong. It looks like it understands the world. Open it up, and there's no world underneath. So, is any of this actually real yet, or is it all still a dream? It is real, but not evenly real. DeepMind's Genie 3 showed interactive worlds [music] you can move through in real time. World Labs shipped Marble, turning prompts and images into navigable 3D spaces. And in June 2026, [music] Nvidia pushed the field toward physical AI with Cosmos 3, a world model built across language, >> [music] >> image, video, audio, and action. That is the board now, not one lab, not one demo, not one hype cycle. The products are arriving, the research curve is rising, and the next question [music] is no longer whether world models exist. It is whether they are reliable enough [music] to trust. But, don't get fooled twice. World models could be the next bubble, too. Let me show you how to tell the difference. >> [music] >> But, here is where I lose half the hype [music] channels on YouTube. World models could become the next bubble, too. Look closely. The worlds they dream still get physics wrong. [music] A robot trained inside a broken world learns broken lessons. LeCun's new [music] lab raised more than a billion dollars, and it still has to prove the product. And a public 2026 [music] finance paper does not give AI a free pass. It says the boom has real fundamentals, but also localized [music] bubble dynamics. So, I am not here to sell you world models as the answer. I am here to ask the harder question. What would make any of this safe to trust? >> [music] >> Here is the true line the 700 billion missed. It was never really about which [music] design wins. A chatbot, a world model, pick whichever you like. The thing that decides whether it is safe to put in your car, your hospital, or your bank is whether [music] you can trust it. And trust is not a model you can buy. It is a discipline you practice. Remember the [music] top from Inception? You spin it to ask one question. Is this real, or am I still dreaming? That is the whole job. [music] You test what the thing actually knows. You put hard limits on what it is allowed to do. You verify before [music] you believe a word of it. Spin the totem. If it wobbles and falls, it is real. If it just keeps spinning, you are still in the dream, [music] and you cannot yet tell. We call that AI reliability engineering. [music] Because the bubble might pop, but that 83-year-old question Grek asked, that one never goes away. [music] The bubble may pop. The question underneath won't. >> [music] >> Don't trust, verify. I'll see you in the next one. >> [music] [music]

Transcript source: youtube-auto-caption. Use the film as the primary record.