LeCun-backed AMI Labs challenges LLMs as Anthropic probes Claude reasoning

AMI Labs CRIO Pascal Fung used a keynote at the International Conference on Machine Learning (ICML) 2026 in Seoul on July 7 to argue that LLMs understand the world only indirectly through human-written text. He said real-world AI agents need world models that can grasp physical environments, causality, goals, communication and mental states, warning that hallucinations in robotics could cause collisions.

Fung said LLMs are optimized for fluent language rather than physical causality, making them slower and costlier than humans for real-world understanding. He cited DeepPhy results showing humans achieved a 64.7% accuracy rate, while the highest AI model score was 41.2%. AMI Labs, founded late last year by Yann LeCun, is pursuing world models based on JEPA and raised $1 billion (~1.5 trillion won) in seed funding this year alone.

Anthropic separately reported finding a region called J-space inside Claude that appears to coordinate thinking and reasoning. Researchers said the structure emerged during training rather than by design, and that changing concepts inside J-space altered Claude’s answers, such as replacing “spider” with “ant” and shifting the answer from “eight” to “six.”

LeCun reinforced his skepticism on X on July 4, writing that “The ‘G’ in AGI is nonsense.” Since 2023, he has argued that ChatGPT, Claude and Gemini are not the route to human-level intelligence because language data and autoregressive prediction do not capture perception, physical intuition or causal understanding.