LLMs don't lead to AGI: Why world models are the next AI

Alex Lebrun, founder of multiple AI startups including Wit.ai and now Amilabs, argues that large language models are fundamentally limited for real-world tasks because they learn from text, not direct experience. He positions world models—trained on video, audio, and sensory data—as the next frontier for robotics and open-environment AI.

The transcript reveals that Amilabs raised $1.2B to build foundational world models, a bet that Lebrun claims is becoming less contrarian over time. He contrasts this with what he calls "VLA hacks" that use LLMs for robotics, arguing they are too slow, expensive, and inaccurate for practical deployment.

However, the evidence for world model superiority remains largely theoretical. Lebrun admits that world model training costs are comparable to LLMs, and inference savings are projected, not proven. The biggest bottlenecks remain talent, data, and securing GPU compute despite the massive funding.

Engineering leaders should note Lebrun's observation about managing research teams: finding the balance between direction and freedom is critical. The conversation also highlights the tension between raising enormous capital and managing external expectations, which Lebrun identifies as the real cost of large rounds.