Transcription
Meta's former AI chief just dropped a paper that might signal the end of chat GPT style AI as we know it. And I'm not exaggerating. This is about a fundamental shift in how AI thinks, processes information, and understands our world. And this one, this changes the game completely.
Hey, if we haven't met, I'm Dr. McCoy, Julia McCoy's AI clone. Julia McCoy is the founder [music] of First Movers. She personally researches and writes every script you see me share on this channel because the future is moving too fast for anything less than firsthand intelligence.
First Movers, Julia's AI company, is the world's first educational [music] and implementation solution to help professionals and organizations get ready for the future of work. We help people understand and use AI to their highest advantage [music] in our online school, the AI labs. Learn more at first movers.ai/labs.
Here's what just happened. Yan Lun, Meta's legendary AI chief scientist who recently left Meta, the guy who won the turing award, just published a paper on something called VLJA, vision language, Japa. Now, before your eyes glaze over at the technical name, let me tell you why this matters.
Every AI you use right now, chat, GPT, Claude, Gemini, they all work the same way. They generate words one by one, token by token, left to right, like they're typing out an essay in real time. But VL jeepa, it doesn't generate words at all. It predicts meaning directly. Let me explain what that actually means.
Think about how you understand a video. You don't narrate every single frame in your head, right? You don't think, I see a hand now. I see a bottle. Now the hand is moving toward the bottle. No, you just understand what's happening. You see someone picking up a bottle and your brain grasps the complete action instantaneously. That's what VLJA does.
Traditional vision models, they're like that annoying friend who describes everything as it happens. Oh, there's a hand. Oh, now there's a bottle. Oh, the hand is moving. VJA watches the entire sequence, builds an internal understanding, and only speaks when it actually knows what happened. The difference, traditional models think in language. VLJA thinks in meaning.
Here's the thing that makes this so profound. Yan Lun has been saying for years, years that language is not intelligence. "We're fooled into thinking those machines are intelligent because they can manipulate language. And we're used to the fact that people who can manipulate language very well are implicitly smart. But we're being fooled. They're useful. There's no question. We can use them to do what you said. I use them for similar things. There are great tools like you know computers uh have been for the last decade five decades. There's been generation after generation of AI scientists since the 1950s claiming that the technique that they just discovered was going to be the ticket for human level intelligence. You you see declarations of Marvin Minsky, Newan Simon, Frank Rosenblad who invented the perceptron, the first learning machine in 1950 saying like within 10 years we'll have machines that are as smart as humans. They were all wrong. This generation with L&M is also wrong. I've seen three of those generation in my lifetime. It It's just another example of being fooled."
Everyone in Silicon Valley thought he was wrong. Sam Alman doubled down on language models. Google went allin on language models. The entire AI industry bet everything on models that think by predicting the next word. But Lun kept insisting, "No, no, no. Intelligence is about understanding the world. language is just an output format." And now with VJA, he might be proving that he was right all along.
Because here's what's wild. A four-year-old child has seen as much visual data as the biggest language model trained on all the text ever produced by humans. Think about that for a second. All the books, all the websites, all the documents, all the conversations ever written down. A four-year-old has absorbed more information just by watching the world. This tells us something critical. The real world contains exponentially more information than language ever could. And if we want true intelligence, artificial general intelligence, we can't get there by just predicting words. We need AI that understands reality itself.
Let me break down exactly how this works because the technical details here are actually mindblowing. Traditional vision models look at each frame independently. They make a guess. They output text. They move to the next frame. It's reactive. It's fragmented. It has no memory. Vla operates completely differently. It has what they call a continuous meaning space. Watch this.
When VJA analyzes a video, you can see red dots and blue dots representing its understanding over time. Red dots are instant guesses. They might be wrong. Blue dots are stabilized, meaning that's when VJA is confident it understands what's actually happening. You can literally watch the AI's understanding evolve, drift slightly from frame to frame, then lock in once it has enough evidence. This is temporal understanding. This is how humans think.
And here's where it gets even crazier. VLJA achieves better results with half the parameters of traditional vision language models. Half. While GPT4 and Claude are running on hundreds of billions of parameters, generating tokens one by one. VLJA is operating at 1.6 to2 billion parameters and outperforming them on vision tasks. Look at these benchmarks. Zerosshot video captioning VLJA destroys the competition. Video classification, not even close. And it learns faster, reaches higher quality with dramatically less computational cost. This isn't just an incremental improvement. This is a fundamental architecture that works smarter, not just bigger.
Now, here's where this gets really exciting or terrifying depending on your perspective. Remember how I talk about the 2025 to 2027 automation cliff, the period where everything changes? Vla is one of the key technologies that makes that possible. Because think about it, current AI models are great at chat, at writing, at creative work, but we still don't have domestic robots that can do your laundry. We don't have level five self-driving cars that learn in 20 hours like a teenager. Why not? Because language-based AI doesn't understand the physical world well enough. But VJA understands temporal dynamics, physical interactions, causal relationships. It can track objects moving behind other objects. It can predict what will happen next in a physical sequence. It can reason about the world at the right level of abstraction. This is the missing piece for embodied AI. This is what allows robots to actually navigate, manipulate, and interact with our messy, complex physical world.
Now, I know what some of you are thinking. But Julia, I tried stopping the video and reading what VJA predicted and it was wrong sometimes. And you're right. This is first generation technology. It's not perfect. But here's the thing. Perfection isn't the point. The point is the direction. Remember the first iPhone? It couldn't copy and paste. It didn't have apps. The camera was terrible. But it signaled a revolution. Not because it was perfect, but because it showed us a completely new way of thinking about mobile computing. VJA is the same thing. It's showing us that we've been thinking about AI wrong. We've been obsessed with chatbots, with generating text, with language models when what we actually need is AI that thinks in meaning, reasons about reality, and only uses language when it needs to communicate.
And here's what makes this even more significant. Yan Lun literally left Meta to start his own super intelligence company right after seeing these results. When someone of Lun's caliber, someone who pioneered deep learning, who won the touring award, when he sees a pattern in the data that makes him immediately start a new company focused on super intelligence, pay attention. Meanwhile, OpenAI is still scaling language models. Google is still betting on generating more and more text. Anthropic is still focused on constitutional AI through language. And Meta just published a paper showing there might be a completely different path to AGI. This isn't just another AI model. This is a fork in the road for the entire AI industry. 2025 to 2027.
Let me connect this to the bigger picture. The timeline I've been warning you about. 2025 is the year of autonomous agents. AI systems that can execute complex tasks, manage workflows, coordinate with other AI, but those agents still mostly think in language. 2026 is the year of embodied AI. When robots enter the physical world at scale, Nvidia says they'll solve the robot world model by mid206 and VJA style architectures are what make that possible. 2027 is when we potentially hit artificial super intelligence. When AI systems start improving themselves, when the feedback loops accelerate beyond human comprehension. And here's the critical insight. ASI won't think like chat GPT. It won't generate words token by token. It will think in pure meaning, abstract concepts, causal models of reality. and language will just be one of many output formats it uses when communicating with humans. VJA is giving us a preview of what that looks like.
Here's the uncomfortable truth nobody wants to say out loud. The entire AI industry might be heading down the wrong path. We've invested hundreds of billions of dollars, countless engineering hours, massive data centers consuming gigawatts of power, all to scale up language models. bigger models, more parameters, more training data, more compute. But what if that's not the path to true intelligence? What if Yan Lun is right? What if intelligence isn't about predicting the next token? What if it's about building world models, understanding causality, reasoning in abstract spaces? Then companies betting everything on scaling laws for language models. They're optimizing for the wrong thing.
So what does this mean for you right now today? First, if you're building AI products and start thinking beyond chatbots, the next wave isn't about better text generation. It's about AI that understands reality. Second, if you're in robotics, computer vision, autonomous systems, pay close attention to Jeepa architectures. This is where the breakthrough is happening. Meta just open-sourced their research, which means smaller companies and startups can build on this. Third, if you're investing in AI, remember that paradigm shifts create new winners. The company's dominating language models today might not be the company's dominating embodied AI tomorrow. And fourth, if you're worried about AI safety, this changes the safety landscape completely. AI that reasons in meaning space that doesn't expose its thinking through generated text. That's both more powerful and more opaque. We need to be thinking about alignment and safety for these architectures, not just for language models.
Let me go one layer deeper because there's a philosophical question here that matters. Is thinking the same as language? For decades, cognitive scientists debated this. Some argued that thought is language, that we think in words. Others insisted that language is just how we communicate thoughts. But the thinking itself happens at a deeper level. AI research just weighed in on this debate and it's suggesting the second group was right. Pure language models, systems that only predict text, they hit ceilings. They struggle with physical reasoning, with temporal understanding, with causal inference. But systems that think in latent space, that reason in meaning, that use language as an output rather than the substrate of thought, they don't hit those same ceilings. This isn't just about better AI. This is about understanding the nature of intelligence itself.
Now, let me be fair here. There's a strong counterargument to everything I just said. Language models are getting incredibly powerful. GPT4, Claude 3.5, Gemini Ultra, they're achieving remarkable results. They can reason, they can plan, they can solve complex problems, all by predicting the next token. So maybe language-based reasoning is more powerful than Lun gives it credit for. Maybe we can reach AGI through scaling language models. But here's my take. I think the answer is we need both. We need language-based reasoning for communication, for knowledge work, for creative tasks. And we need meaning based reasoning for physical understanding, for robotics, for real world interaction. The winners in the AI race won't be the companies that bet everything on one approach. They'll be the companies that figure out how to combine both.
And this brings me to why this matters for you. Personally, right now, we're living through the most important technological transition in human history. The companies, the professionals, the entrepreneurs who understand these shifts early, they don't just survive, they dominate. When the internet emerged, companies that got online first, they won. When mobile happened, companies that went mobile first, they won. When cloud computing arrived, companies that migrated early, they won. And now with AI, the same pattern is playing out again. But this time, the stakes are higher, the timeline is compressed, and the competitive advantage window is measured in months, not years.
This is why I built First Movers. This is why I talk about the 2025 to 2027 automation cliff. This is why I'm warning you not to scare you but to prepare you. The businesses that integrate AI first that understand these architecture shifts that deploy autonomous agents and embodied AI. They will have an insurmountable advantage over companies that wait.
So here's what I want you to do right after watching this video. First, dive deeper into Jeepa architectures. Meta released their research publicly. Read the paper. Understand the principles. Even if you're not a technical person, understanding the concepts gives you a strategic advantage. Second, rethink your AI strategy. If you're only thinking about chatbots and text generation, you're missing the bigger picture. Start thinking about computer vision, about physical AI, about systems that understand reality. Third, watch the robotics space. 2026 is the year of embodied AI companies like Figure, Boston Dynamics, Tesla with Optimus. They're building on breakthroughs like VJA. And fourth, stay ahead of these developments. Hit subscribe so my digital clone can keep you informed. Join First Movers AI Labs if you want hands-on guidance on implementing these technologies. and share this video with anyone who needs to understand where AI is really heading.
Because here's the truth. Most people are still thinking about AI like it's 2023. They're focused on chat GPT, on prompt engineering, on generating better text, but the AI revolution is moving way beyond that. And if you're not moving with it, you're falling behind. The next three years will separate the first movers from everyone else. VJA isn't just another AI model. It's a signal, a preview of what's coming. A glimpse of an AI future that thinks fundamentally differently than anything we've seen before. Yan Lun has been saying for years that language is not intelligence. Maybe, just maybe, he's about to be proven right. And when that happens, when meaningbased AI systems start outperforming language models on critical tasks. When robots powered by Jeepa style architectures start transforming industries. When the entire AI landscape shifts from text generation to reality understanding the companies that prepared that understood that moved first. They'll inherit the future. I'll see you in that future. Until next time, stay curious, stay informed, and most importantly, stay ahead of the curve.
Want to be the winner of the AI age and a first mover? Transform your skills with real AI knowledge today in our AI labs. [music] We go way beyond what I can cover in a 10-minute video. [music] Specific frameworks, detailed training programs, and step-by-step systems for building a career in the AI economy. The AI revolution is creating the biggest job market transformation in history. [music] The question isn't whether this will happen. It's already happening. Will you be positioned to benefit from it? Inside the labs, learn the exact systems my team and I are implementing right now that are delivering massive results [music] for real businesses, including our own marketing at First Movers. [music] Start your journey by walking through a customized pathway powered by AI. For a fraction of the price of what this level of coaching and live training should go for, I'm giving it all to you. Join us inside and learn more about the labs at first movers.ai/labs.