Yann LeCun: Deep Learning, ConvNets, and Self-Supervised Learning | Lex Fridman Podcast #36

LeCun frames AI alignment as an extension of what societies already do with laws—designing objective functions (rules, penalties) that shape behavior toward the common good, suggesting AI ethics will fuse computer science and jurisprudence rather than invent something entirely new.

Modern neural nets with huge parameter counts and non-convex objectives train successfully on relatively modest data via stochastic gradient descent, contradicting pre–deep learning textbooks; this empirical success implies our theoretical understanding of generalization and optimization was too narrow.

LeCun argues that human-like reasoning emerges from systems with hippocampus-like memory, recurrent access to that memory, and energy-minimization style planning (model predictive control), not from static feed-forward models alone.

He critiques logic- and graph-based expert systems for their brittleness and manual knowledge acquisition bottleneck, advocating vector-based “symbols” and continuous functions (à la Hinton and Bottou) as a way to make reasoning compatible with gradient-based learning.

LeCun sees self-supervised prediction (e.g., masked word prediction, video/frame prediction) as the primary route to learning rich world models that later make supervised and reinforcement learning vastly more sample-efficient, mirroring how babies learn physics and causality from observation.

He notes that deep RL systems need the equivalent of years or centuries of experience to reach human performance in games, whereas humans learn tasks like driving in tens of hours because they rely on internal predictive models of physics, not just trial-and-error reward signals.

Using arguments about the structure of the visual system and the vast space of possible Boolean functions, LeCun contends that humans operate over a tiny subset of possible tasks and stimuli—our sense of “generality” is confined to what we can even conceptualize.