Convolutional Structure Is the Correct Inductive Bias for Visual Intelligence

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.

Source: Gradient-Based Learning Applied to Document Recognition, LeCun et al., Proceedings of the IEEE, 1998

True Intelligence Requires World Models, Not Just Next-Token Prediction

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.

Source: A Path Towards Autonomous Machine Intelligence, Yann LeCun, OpenReview, June 2022

AI Research Should Be Published Openly; Open Science Accelerates Collective Progress

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.

Source: Yann LeCun interview on open source AI, The Verge, April 2023 / LLaMA: Open and Efficient Foundation Language Models, Meta AI, February 2023

LLMs Cannot Lead to AGI; Autoregressive Text Generation Is a Fundamental Dead End

Large language models are 'stochastic parrots' — they stitch together language through statistical patterns but have no real understanding of the world. LLMs cannot perform reliable reasoning, planning, and causal understanding because they lack foundational perception of the physical world. The path to AGI requires perception-action loops, continual learning, and world models — not larger text predictors.

Source: Yann LeCun Twitter/X posts on LLM limitations, x.com/ylecun, 2023-2024 / A Path Towards Autonomous Machine Intelligence, Yann LeCun, OpenReview, June 2022

Self-Supervised Learning Is the Correct Path Toward Human-Level Perception

Humans and animals learn by observing the world, not by relying on massive human annotation. Self-supervised learning allows models to learn representations from the structure of data itself — this is the scalable path to intelligence. Contrastive learning, masked autoencoders, and JEPA are all important explorations in self-supervised learning.

Source: Self-supervised learning: The dark matter of intelligence, Yann LeCun & Ishan Misra, Meta AI Blog, April 2021

Inductive Bias Design Principle

Encode the true structural priors of data directly into model architecture, rather than expecting the model to learn all structure from scratch.

LeNet encoded visual translational invariance as convolutional weight sharing and local correlations as local receptive fields; these two inductive biases gave it far greater parameter efficiency than fully connected networks on MNIST.

World Model Hierarchy Framework

Intelligent systems need to maintain predictive models of world states in latent space, rather than merely doing pattern matching in observation space.

JEPA (Joint Embedding Predictive Architecture) predicts future states in latent representation space rather than reconstructing images in pixel space, avoiding the 'hallucination' problem of generative models while learning more abstract world models.

Open Research Accelerator Effect

Openly publishing research results and model weights accelerates one's own research progress through external innovation feedback loops while building an ecosystem moat.

After Meta released the LLaMA open-source model series, the global research community produced thousands of derivative studies and applications, which in turn provided Meta with valuable improvement directions and talent attraction.

Energy-Based Model Unified Framework

Use an energy function to unify various learning problems: good predictions correspond to low energy, bad predictions to high energy, and learning is training the shape of the energy function.

Contrastive learning (e.g., SimCLR, MoCo) can be understood as a special case of energy-based models: the energy of positive pairs is pushed down, the energy of negative pairs is pushed up, thereby learning meaningful representations.

Evidence-Backed Contrarianism

Maintain critical distance from mainstream consensus, but every dissenting view must be backed by specific technical arguments, not mere novelty-seeking.

LeCun publicly questioned the feasibility of LLMs leading to AGI at the height of the LLM boom, offering specific technical arguments: LLMs lack physical world models, cannot perform reliable planning, and suffer from hallucination — criticisms increasingly validated by research in 2024-2025.

Bell Labs Invention Phase

1988-2002

Invention of convolutional neural networks and industrial deployment of handwriting recognition

At Bell Labs and AT&T Labs, LeCun invented the LeNet series of convolutional neural networks and deployed them in the US banking system for check recognition, processing over 10% of all American bank checks. This period laid the technical foundation for modern computer vision, but also experienced the marginalization of deep learning during the 1990s 'AI winter'.

NYU Academic Institution-Building Phase

2003-2013

Building NYU machine learning center, advancing deep learning theory and energy-based models

Joined New York University in 2003 and founded the Courant Machine Learning Lab (later renamed CILVR). Persisted in deep learning research during the AI winter, and together with Hinton and Bengio drove the deep learning renaissance of 2006-2012. AlexNet's success in 2012 validated his twenty years of persistence, bringing deep learning from the margins to the mainstream.

Meta AI Industrial Research Phase

2013-2022

Leading Meta AI Research, advancing self-supervised learning and open-source AI

Joined Facebook (later Meta) in 2013 to found and lead FAIR (Facebook AI Research). During this period drove major breakthroughs in self-supervised learning (including precursors to SimCLR), advocated for openly publishing research results, and made FAIR one of the world's top AI research institutions. The release of the LLaMA open-source model series was the concentrated embodiment of this period's open science philosophy.

World Model Advocacy Phase

2022-至今

Critiquing the LLM route, advocating JEPA world models as a new path to AGI

Published the white paper 'A Path Towards Autonomous Machine Intelligence' in 2022, systematically articulating the AGI path based on world models and proposing the JEPA architecture. Against the backdrop of ChatGPT triggering a global LLM boom, he persisted in criticizing the fundamental limitations of LLMs, becoming one of the most controversial voices in AI. Through public debates on Twitter/X, he brought technical arguments into broader public discourse.