Meta AI Chief Says True Human-Level AI Requires World Models, Not Just Language
Meta's Yann LeCun pushes back against claims of near-term artificial general intelligence, arguing that current language models lack true understanding. He believes developing three-dimensional "world models" is the real path to human-level AI, though it remains years to decades away.
Meta's chief AI scientist Yann LeCun pushes back against claims that human-level artificial intelligence is just around the corner, directly contradicting optimistic predictions from industry figures like Elon Musk. He states that despite features like "memory" and "thinking" modes in modern chatbots, current AI systems do not actually reason, plan, or understand the world like humans do.
The fundamental limitation of today's large language models lies in their architecture, as they merely predict the next token in a one-dimensional sequence or the next pixel in a two-dimensional image. Because they lack a grasp of the three-dimensional physical world, these systems fail at simple real-world tasks that a human child learns in a matter of hours, such as clearing a dinner table.
To achieve true artificial general intelligence, LeCun advocates for the development of "world models" that can perceive and understand the physical environment in three dimensions. He estimates that this necessary shift in AI development puts human-level AI anywhere from a few years to a full decade or more away from becoming a reality.