Yann LeCun's case against scaling LLMs as a route to AGI

A post describing Yann LeCun's ETH Zürich talk says he called scaling large language models to reach AGI “impossible.” It says models train on about 30 trillion text tokens, roughly 10^14 bytes, while a four-year-old receives about that much visual data in one year and 10 months. In LeCun's view, scaling adds stored knowledge, not the ability to learn new tasks quickly.