This is Yann LeCun’s 2024 Ding Shum Lecture, delivered at Harvard’s Center of Mathematical Sciences and Applications and posted on the Harvard CMSA channel. LeCun, a Turing Award winner and a chief AI scientist at Meta, uses it to lay out his research agenda and his skepticism that scaling current large language models will reach human-level intelligence.
He argues that machines still cannot learn as efficiently as humans and animals, and proposes what he calls objective-driven AI: systems organized around world models that can predict, reason, and plan, trained largely through self-supervised learning rather than text prediction alone. He presents his Joint Embedding Predictive Architecture as a concrete direction and explains why he considers it a more promising road than autoregressive language modeling.
This is a deep, research-oriented talk that rewards viewers who already understand the basics of deep learning. Its value is in hearing one of the field’s founders articulate, firsthand, a well-developed dissent from the prevailing LLM-centric view, which makes it a useful counterweight in any honest survey of where AI might be heading.