AMI Labs calls for world models beyond LLMs

AMI Labs co-founder and chief research and innovation officer Pascal Fung used a keynote at the 2026 International Machine Learning Society conference in Seoul to argue that LLMs are poorly suited to AI agents operating in the physical world. LLMs learn from human-written text, he said, giving them only an indirect grasp of reality and goals centered on fluent language rather than physical causality.

Fung described world models as systems that can read both physical and mental context, comparing the task to a soccer player understanding space, causality, team goals, communication and emotion in real time. He warned that hallucinations may be relatively harmless in text but could cause collisions in robots. On the DeepPhy physical reasoning benchmark, humans recorded 64.7% correct answers, while the best AI model reached 41.2%.

AMI Labs, founded late last year by New York University professor Jan LeCun, has raised 1 billion dollars (about 1.53 trillion won) in seed funding. Fung said real-world agents need perception, prediction, planning and memory, and said AMI Labs is building a JEPA-based world model that predicts what comes next rather than generating the next pixel.