The smartest minds in AI are abandoning the chatbot race mid-sprint, and the reason isn't what you think.
The Summary
- Brown PhD dropout Louis Castricato quit studying LLMs to start Overworld, a company building AI that understands physical space, not just text — his diagnosis: "We basically have passed the point of doing real fundamental LLM research."
- Fei-Fei Li's World Labs and Yann LeCun's Advanced Machine Intelligence Labs are both betting on "world models" that teach AI systems how objects respond to physics, how light behaves, how spaces look from angles no camera captured.
- While investors commit trillions to chatbot companies, the cutting edge is moving toward AI that can navigate warehouses, manipulate objects, and build things in the real world.
The Signal
Louis Castricato's pivot tells you everything about where the ceiling is for LLMs. Eight years studying the technology, close enough to see the limits, and he walked away from his doctorate to build something different. His read: the fundamental research is done. What's left is application work. That's lucrative but not interesting if you're trying to build something genuinely new.
The shift to world models isn't about incremental improvement. It's about changing the substrate. Language models learn patterns in text. World models learn patterns in physics. How a ball bounces. How shadows move when light sources change. How a robot arm needs to adjust grip pressure based on object weight and surface texture. Fei-Fei Li describes world models as learning "the statistical structure of space and time" — the same way LLMs learned the statistical structure of language.
"Where language models learn the statistical structure of text, world models learn the statistical structure of space and time."
This is the technical unlock that makes embodied AI agents actually useful. A chatbot can tell you how to pack a box. A world model can predict what happens when you stack irregular shapes, account for center of gravity, and adjust in real time when something shifts. The difference between advice and execution. Between a consultant and a coworker.
The talent migration matters more than the technology itself. Yann LeCun left his role as Meta's chief AI scientist to start Advanced Machine Intelligence Labs in Paris, specifically to work on this. Fei-Fei Li, who built ImageNet and basically taught computers how to see, is running World Labs. These aren't people chasing trend cycles. They're people who built the last platform and can see it's plateauing.
Key shifts happening now:
- Top researchers leaving academia and big tech mid-career to build world model startups
- Investment dollars still flowing to chatbots, but cutting-edge talent flowing to embodied AI
- "World model" becoming the new buzzword, which means we're 12-18 months from real products
The economic argument is straightforward. Chatbots are valuable but their ceiling is knowledge work augmentation. World models unlock physical automation. That's manufacturing, logistics, construction, agriculture. The industries that employ most humans and move most atoms. If you can train an AI system to understand physics as fluently as ChatGPT understands syntax, you can build agents that operate in warehouses, coordinate construction sites, manage farms. Web4 isn't just about software agents. It's about agents that can pick things up.
The Implication
Watch where the talent goes, not where the capital goes. Right now capital is still flooding into LLM companies because that's what investors understand. But the researchers who actually built those systems are moving to world models. That's your leading indicator.
For anyone building agent companies: your advantage window on pure language-based agents is narrowing. The next generation will expect agents that can interact with physical space, not just digital interfaces. Start thinking about how your agent handles a video feed, not just an API response. Start thinking about sensors, not just tokens.