Yann LeCun

Chief AI Scientist at Meta AI. Turing Award winner (2018, with Geoffrey Hinton and Yoshua Bengio). Pioneer of convolutional neural networks and deep learning. Professor at NYU. One of the most prominent critics of the current LLM paradigm and the AI safety consensus.

Background

Developed convolutional neural networks (CNNs) in the late 1980s and 1990s, foundational to modern computer vision. Joined Facebook (now Meta) in 2013. Coined the term “Joint-Embedding Predictive Architecture” (JEPA) as an alternative to generative models. Argues that autoregressive language models are a dead end for achieving human-level intelligence and that energy-based models and hierarchical planning architectures are the correct path.

Notable public disagreements with Geoffrey Hinton and Yoshua Bengio on AI existential risk, which LeCun largely dismisses as overstated.

Appearances in this wiki

Key positions

• LLMs lack four essential properties of intelligence: world understanding, persistent memory, reasoning, and planning
• Bandwidth argument: language encodes ~2×10^13 bytes total; infant vision processes ~10^15 bytes per year — language is insufficient training data for world models
• JEPA is the correct architectural alternative: predict in representation space, not output space
• Open-source AI is the structural safeguard against dangerous power concentration — a bigger risk than misalignment
• Rejects “doomer” AI risk framing; believes people are fundamentally good and distributed AI is net positive