The TL;DR: The next 5 years? The bubble is probably going to pop on LLMs for the physical world, and we are going to go all-in on World Models. It’s the only way we get AI that doesn’t just chat but actually acts without breaking stuff.
Here is the breakdown of why this is the real shift:
LeCun is on a crusade (and he’s right). Yann LeCun basically told the industry to “abandon” generative LLMs if they care about human-level AI, calling them “completely helpless” when it comes to the physical world. He bounced from Meta to start AMI Labs, which just raised $1.03B at a $3.5B valuation to build models using JEPA (Joint Embedding Predictive Architecture) instead of just next-token prediction. The man is betting his legacy that sensory data > text data.
NVIDIA is building the “Physics Engine” for AI. Check out their Cosmos platform. It’s a suite of “World Foundation Models” specifically designed to generate physics-aware videos and simulate environments . They trained these on 20 million hours of real-world driving and robotics data . This isn’t just for generating cool videos; it’s about giving robots “foresight” to predict the consequences of their actions before they happen.
Startups are solving the “Data Pipeline” problem. The bottleneck isn’t compute anymore; it’s real-world sensor data. LLMs scraped the internet, but robots need LiDAR, video, and telemetry.
The Bottom Line:
For the next five years, the biggest impact won’t be a bigger LLM. It will be the shift from models that manipulate just text tokens to models that manipulate physics. If you are building AI that needs to interact with reality, you are going to be looking at NVIDIA’s Cosmos, feeding data through pipelines built by startups like Encord, and watching what LeCun’s AMI Labs cooks up. Sources: