Yann LeCun argues LLMs are fundamentally limited and will be superseded by AI systems that build world models

In a Newsweek AI Impact interview, LeCun stated that LLMs are unable to represent the continuous high-dimensional spaces that characterize nearly all aspects of the world, and will become obsolete in a matter of years despite Meta's significant investment in the Llama LLM family.

LeCun distinguishes between current LLMs, which operate as System 1 (rapid, pattern-based responses), and the System 2 reasoning (deliberate, analytical processing) that humans use for planning and understanding. LLMs appear to reason through chain-of-thought techniques, but this remains statistical sampling of answer possibilities, not genuine understanding—analogous to students learning by rote memorization rather than grasping underlying concepts.

Effective AI systems require world models that can reason and plan across multiple timescales and abstraction levels, as demonstrated by human travel planning from New York to Paris. Current systems, including automated travel agents, compile statistically likely routes rather than modeling the world in a reusable way that could adapt to different tasks and continuously replan based on environmental feedback.