LeCun, Malik, Dupoux challenge AI’s learning vacuum

LeCun and company are calling out the industry's biggest open secret: our "intelligent" models are brittle, static artifacts that cannot learn once the training script stops.

• –Systems A and B (Observation and Interaction) provide a dual path for learning that mimics infant development and animal cognition.
• –System M (Meta-Control) is the key breakthrough, acting as an internal signal to autonomously switch between passive observation and active trial-and-error.
• –The critique of "The Data Wall" suggests that scaling text-only models is a dead end, and the only way forward is through world-grounded, embodied agents.
• –Shifting from human-engineered loss functions to internally generated learning signals could finally break the reliance on massive, manually curated datasets.
• –This represents a significant pivot away from pure scaling laws toward architectural shifts that prioritize temporal and spatial reasoning.