Yann LeCun Discusses AI, Self‑Supervised Learning, and the Future of AGI

Yann LeCun (referred to as Yang Likun) appears on the "WTF人物" program with Indian entrepreneur Nikhil Kamath to discuss a wide range of topics related to artificial intelligence.

He begins by defining AI through the lens of information theory, emphasizing that intelligence encompasses many aspects beyond simple problem‑solving, and critiques the current dominance of transformer‑based large language models (LLMs) as limited in reasoning and physical understanding.

LeCun explains self‑supervised learning: a method where inputs are deliberately corrupted and the model learns to reconstruct the original, a technique that underlies modern vision and language models. He illustrates this with examples of image corruption and language prediction.

The interview highlights his skepticism toward the claim that transformers alone will achieve AGI, noting that they excel in discrete domains (text) but struggle with continuous high‑dimensional data such as video.

LeCun introduces his research on a "joint embedding predictive architecture" (Jeppo), which encodes video into abstract representations before predicting future frames, avoiding pixel‑level prediction and aiming to capture higher‑level structure.

Regarding the timeline for AGI, he revises earlier predictions of decades to a more aggressive 5‑10‑year horizon, aligning with other futurists.

He advises aspiring AI entrepreneurs to pursue a PhD, arguing that doctoral training fosters genuine innovation and makes fundraising easier, while also stressing the importance of open‑source platforms that will dominate future ecosystems much like Linux did for operating systems.

LeCun discusses the future of human intelligence, suggesting that as AI takes over routine tasks, humans will focus on higher‑level decision‑making, creativity, and abstract problem solving, and that zero‑shot learning—solving new problems without additional training—will become a key measure of intelligence.

The conversation concludes with reflections on the societal impact of AI, the continued need for education, and the inevitability of competition and collaboration in an AI‑augmented world.