Meta's AI chief says world models are key to 'human-level AI' — but it might be 10 years out | TechCrunch

Yann LeCun argues that current AI systems, despite their impressive predictive capabilities in language and pixels, lack true understanding, reasoning, and memory. He contends that existing models are limited to one or two-dimensional pattern recognition and cannot navigate the physical world with human-like intuition or common sense. This fundamental gap means that today’s large language models are far from achieving artificial general intelligence, challenging the optimistic narratives often promoted by other industry leaders. To bridge this divide, LeCun proposes developing “world models,” which serve as internal mental representations of how the physical world behaves. Unlike current systems that merely predict the next token, world models would allow AI to perceive three-dimensional environments and simulate the outcomes of potential actions. This architecture enables machines to plan and reason by understanding cause and effect, similar to how humans imagine cleaning a room before physically doing it, thereby unlocking the ability to perform complex, real-world tasks efficiently. This perspective is critical to open data discussions because it highlights a shift from merely scaling data volume to improving data quality and representation. Achieving world models requires vast amounts of diverse, structured data that reflects physical reality, not just text or isolated images. Consequently, the open data community must prioritize creating comprehensive datasets that support three-dimensional perception and objective-driven learning, ensuring that future AI development is built on a foundation of genuine understanding rather than superficial statistical correlations.

Source: techcrunch.com
Published on 2024-10-18