Yann LeCun
Father of convolutional neural networks, the contrarian reshaping AI's future through open science and world models
Yann LeCun is one of the three godfathers of deep learning, the inventor of convolutional neural networks (CNNs), and a 2018 Turing Award laureate. His LeNet, invented at Bell Labs, is the cornerstone of modern computer vision, directly enabling today's image recognition, face detection, and autonomous driving vision systems. In the 1990s his handwritten digit recognition system was widely deployed by US banks, processing over 10% of all American checks. In 2003 he joined NYU to found the Courant Machine Learning Lab, then became Meta (Facebook)'s Chief AI Scientist while retaining his NYU professorship. He is a steadfast advocate for open science, arguing that AI research results should be published openly. In recent years he has been a sharp critic of the LLM route, arguing that large language models cannot lead to AGI, and advocates for a new approach based on world models and JEPA (Joint Embedding Predictive Architecture). His outspokenness on Twitter/X has made him one of the most controversial public intellectuals in AI.
Methodologies
- Convolutional Inductive Bias Design Method - When designing neural network architectures, first analyze the invariances and local structures of the data, encoding these priors directly into the architecture rather than expecting the network to learn all structure from data.
- World Model as Alternative to Text Prediction - The criterion for evaluating whether an AI system has genuine intelligence: does it maintain a causal model of the physical world in latent space, or does it only do pattern matching in observation space?
Key decisions and timeline
- 1983-09 Entered Université Pierre et Marie Curie for PhD, Beginning Neural Network Research - Persisting in research in a direction unsupported by the mainstream paradigm requires great confidence in technical intuition; entering a correct but neglected field early yields enormous long-term returns.
- 1988-01 Joined Bell Labs, Beginning CNN Development for Character Recognition - Industrial research institutions can provide resources and application scenarios that academia cannot match; combining theoretical breakthroughs with real-world needs is the most effective way to accelerate innovation.
- 1998-11 Published LeNet-5 Paper, Establishing the Complete Architecture of Modern CNNs - Systematizing engineering practice into theoretical papers is the key step to amplifying research impact; a well-timed, rigorously argued paper can define a field for decades.
Beliefs and mental models
- Belief 1 - The visual world has translational invariance, local correlations, and hierarchical compositionality. Convolutional neural networks encode these physical priors directly into network structure through weight sharing and local receptive fields — this is the fundamental reason for their success. Good architecture should reflect the true structure of data, not rely on brute-force computation.
- Belief 2 - LLMs learn by predicting the next word, a method that cannot enable models to understand the causal structure of the physical world. True intelligence requires world models that can predict world states in latent space — similar to how infants learn about the world through physical interaction. JEPA (Joint Embedding Predictive Architecture) is the right path toward this goal.
- Belief 3 - Closed AI research not only slows overall progress but creates the dangerous situation of a few institutions monopolizing technology. Meta AI's commitment to releasing the LLaMA open-source model series is the practice of this belief. Scientific progress depends on open peer review and knowledge accumulation; commercial competition should not be an excuse for closed research.
- Model 1
- Model 2
- Model 3