Yann LeCun

Yann André LeCun (born July 8, 1960) is a French-American computer scientist best known as the inventor of the convolutional neural network in its modern trainable form and as the founding director of FAIR, the fundamental AI research laboratory of Meta, where he served as chief AI scientist from 2018 until his departure in 2025. At AT&T Bell Laboratories between 1988 and 1998 he built LeNet, the family of convolutional networks trained by backpropagation that read handwritten digits¹ and, by the late 1990s, a substantial fraction of the checks written in the United States;² the 1998 Proceedings of the IEEE paper with Léon Bottou, Yoshua Bengio, and Patrick Haffner remains the definitive account of that system.³ The deep convolutional architecture he championed was the direct ancestor of AlexNet and of modern computer vision, and for the body of deep learning work he shared the 2018 ACM A.M. Turing Award with Bengio and Geoffrey Hinton, his postdoctoral mentor, an award announced in March 2019.⁴⁵

LeCun is the Jacob T. Schwartz Professor of Computer Science at New York University's Courant Institute of Mathematical Sciences, where he has taught since 2003.⁶ At Facebook, which he joined in December 2013 to create FAIR, he built one of the largest industrial fundamental-research organizations in AI and steered it toward open publication and, later, open-weight model release, most visibly the Llama family of large language models from 2023.⁶⁷ In November 2025 he confirmed that he was leaving Meta after more than a decade there, and in December 2025 he co-founded Advanced Machine Intelligence Labs (AMI Labs), a startup pursuing the world-model research program he had outlined since 2022; he serves as executive chair, with Alex LeBrun as chief executive.⁸⁹

Among researchers he is equally known for a heterodox public position: that the auto-regressive large language models dominating the field are a technological dead end for human-level intelligence, which he argues will require systems that learn world models from sensory input.⁹¹⁰

Career

LeCun was born in 1960 at Soisy-sous-Montmorency, in the suburbs of Paris. He received a Diplôme d'Ingénieur from ESIEE Paris in 1983 and a PhD in computer science from the Université Pierre et Marie Curie in 1987. His doctoral work proposed an early learning algorithm for multi-layer threshold networks, published at the Cognitiva conference in 1985, a variant of backpropagation derived independently of Paul Werbos and of the 1986 paper by David Rumelhart, Hinton, and Ronald Williams that popularized the method.¹¹¹² In 1987 he moved to the University of Toronto for a year of postdoctoral research with Hinton.⁶

In 1988 LeCun joined the Adaptive Systems Research Department of AT&T Bell Laboratories in Holmdel, New Jersey, where the first convolutional networks were trained on handwritten zip codes in 1989.¹ After the 1996 breakup of the Bell System research arm he became head of the Image Processing Research Department at AT&T Labs-Research, where his group completed the LeNet-5 check-reading work, built the MNIST benchmark with Corinna Cortes and Christopher Burges, and developed the DjVu document compression technology with Bottou, Haffner, and colleagues.³⁶¹³

Years	Role	Organization
1987-1988	Postdoctoral researcher	University of Toronto
1988-1996	Researcher, Adaptive Systems Research Department	AT&T Bell Laboratories
1996-2003	Head, Image Processing Research Department	AT&T Labs-Research
2003-	Professor of computer science	New York University
2013-2025	Founding director of FAIR; chief AI scientist from 2018	Facebook / Meta
2025-	Co-founder and executive chair	AMI Labs

LeCun joined New York University in 2003 and was named the inaugural Jacob T. Schwartz Professor at the Courant Institute in 2022; he was also the founding director of NYU's Center for Data Science.⁶ In December 2013 he joined Facebook to found and direct the Facebook AI Research laboratory, keeping his NYU affiliation.⁶ In early 2018 he stepped back from FAIR's day-to-day direction to become the company's chief AI scientist, a role he held through the company's renaming as Meta and through a decade in which FAIR produced, among other work, the wav2vec self-supervised speech models, the Detectron and Segment Anything vision systems, the data-efficient image transformer DeiT, and the Llama language models.⁶¹⁴¹⁵ He is an adviser to Kyutai, the Paris nonprofit open-science AI laboratory founded in late 2023.⁶

On November 19, 2025, following earlier reporting of his plans, LeCun confirmed he would leave Meta to found a startup devoted to world-model architectures.⁸ AMI Labs was established in December 2025; as of July 2026 he is its executive chair and remains on the NYU faculty.⁹⁶

His honors, beyond the 2018 Turing Award, include the Princess of Asturias Award for Technical and Scientific Research in 2022 (with Bengio, Hinton, and Demis Hassabis), election to the U.S. National Academy of Sciences in 2021, appointment as a Chevalier of the French Legion of Honour in 2023, the VinFuture Grand Prize in 2024, and the Queen Elizabeth Prize for Engineering in 2025, the last two shared with groups of contemporaries that included Hinton, Bengio, Fei-Fei Li, Jensen Huang, John Hopfield, and Bill Dally.⁶

Research contributions

Convolutional neural networks

LeCun's defining contribution is the convolutional neural network as a trainable system. Drawing on Fukushima's neocognitron and on the perceptron tradition going back to Frank Rosenblatt, his 1989 Bell Labs work applied backpropagation to a network whose layers used local receptive fields and shared weights, giving translation-tolerant recognition of handwritten zip codes.¹ The architecture matured through the 1990s into LeNet-5, which combined convolutional layers, subsampling, and a graph transformer for segmentation-free recognition of whole check amounts, and was deployed by banks reading, at its peak, an estimated tenth or more of American checks.³⁶ The MNIST dataset assembled for that work became the standard entry-level benchmark of machine learning.³ The 1998 Proceedings of the IEEE survey tied the full pipeline together and remains the canonical reference for the architecture every modern vision network descends from, including the GPU-scaled revival of 2012, when AlexNet won the ImageNet challenge.³⁵

Document compression and infrastructure

With Bottou and colleagues at AT&T, LeCun co-created DjVu, a document image compression technology that became a widely used format for scanned libraries, and co-developed the Lush programming language used for much of the group's research.¹³ His 1990 "Optimal Brain Damage" paper, with John Denker and Sara Solla, founded network pruning, the removal of weights by estimated saliency, a technique that returned to prominence for compressing large models.¹⁶

Energy-based models and self-supervised learning

Across the 2000s LeCun developed energy-based learning as a unifying frame, treating inference as minimization of an energy function over configurations rather than as probabilistic normalization; the 2006 tutorial with Sumit Chopra, Raia Hadsell, Marc'Aurelio Ranzato, and Fu Jie Huang is the standard statement.¹⁷ The framework, intellectually descended from John Hopfield's network and from the Boltzmann machine line, motivated his long advocacy of self-supervised learning: systems that acquire representations by predicting parts of their input rather than by imitating labeled data.¹⁷¹⁰ At NYU and then FAIR this program produced the OverFeat detection network of 2013, a precursor of modern object detectors, and later the joint-embedding predictive architectures below.¹⁸

FAIR, Llama, and open weights

As FAIR's founder and later Meta's chief AI scientist, LeCun shaped an industrial laboratory unusual for its open publication norms, and he was the company's leading internal advocate for releasing model weights publicly. The Llama family, first released in February 2023 under a research license and later openly, made competitive foundation models available outside the largest labs and seeded an ecosystem of derivative open models.⁷¹⁴ FAIR's vision line of the same period, including the Segment Anything model of 2023, likewise shipped with open weights.¹⁵

JEPA and world models

In June 2022 LeCun published "A Path Towards Autonomous Machine Intelligence," a position paper arguing that human-level AI requires agents that learn predictive world models from observation, act by planning against learned objectives, and acquire common sense chiefly through sensory experience rather than text.¹⁰ The joint-embedding predictive architecture (JEPA) is its technical core: rather than reconstructing pixels or tokens, the model predicts representations of missing or future input in an abstract space. FAIR instantiated it for images as I-JEPA in 2023 and for video as V-JEPA in 2024.¹⁹²⁰ AMI Labs, the company he founded on leaving Meta, is the same bet made as a business: world models rather than scaled auto-regression as the route to capable systems.⁸⁹

Selected publications

Ordered chronologically.

1. LeCun, Y., "Une procédure d'apprentissage pour réseau à seuil asymétrique (A Learning Scheme for Asymmetric Threshold Networks)," Proceedings of Cognitiva 85, Paris, 1985. An early independent backpropagation variant.
2. LeCun, Y., Boser, B., Denker, J. S., Henderson, D., Howard, R. E., Hubbard, W., and Jackel, L. D., "Backpropagation Applied to Handwritten Zip Code Recognition," Neural Computation 1(4), 1989. The first trained convolutional network.
3. LeCun, Y., Denker, J. S., and Solla, S., "Optimal Brain Damage," Advances in Neural Information Processing Systems 2, 1990. Founded network pruning.
4. LeCun, Y., Bottou, L., Bengio, Y., and Haffner, P., "Gradient-Based Learning Applied to Document Recognition," Proceedings of the IEEE 86(11), 1998. The LeNet-5 synthesis.
5. Bottou, L., Haffner, P., Howard, P. G., Simard, P., Bengio, Y., and LeCun, Y., "High Quality Document Image Compression with DjVu," Journal of Electronic Imaging 7(3), 1998. The DjVu system.
6. LeCun, Y., Chopra, S., Hadsell, R., Ranzato, M., and Huang, F. J., "A Tutorial on Energy-Based Learning," in Predicting Structured Data, MIT Press, 2006. The standard statement of energy-based learning.
7. Sermanet, P., Eigen, D., Zhang, X., Mathieu, M., Fergus, R., and LeCun, Y., "OverFeat: Integrated Recognition, Localization and Detection Using Convolutional Networks," arXiv:1312.6229, ICLR 2014. Convolutional detection at ImageNet scale.
8. LeCun, Y., Bengio, Y., and Hinton, G., "Deep Learning," Nature 521, 2015. The field's standard survey.
9. LeCun, Y., "A Path Towards Autonomous Machine Intelligence," OpenReview, 2022. The world-model position paper behind JEPA.
10. Assran, M., Duval, Q., Misra, I., Bojanowski, P., Vincent, P., Rabbat, M., LeCun, Y., and Ballas, N., "Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture," arXiv:2301.08243, CVPR 2023. I-JEPA.
11. Bardes, A., Garrido, Q., Ponce, J., Chen, X., Rabbat, M., LeCun, Y., Assran, M., and Ballas, N., "Revisiting Feature Prediction for Learning Visual Representations from Video," arXiv:2404.08471, 2024. V-JEPA.

Views

LeCun is the most prominent technical dissenter from two positions associated with his Turing co-recipients. On capability, he has argued since his 2022 position paper that auto-regressive large language models trained by next-token prediction on text, systems built on the Transformer architecture such as the GPT series, lack the prerequisites of human-level intelligence: persistent memory, planning, reasoning grounded in the physical world, and a learned world model.¹⁰ He has similarly described fine-tuning and reinforcement learning from human feedback as surface-level patches that cannot guarantee controllable behavior.²¹ In a November 2025 Financial Times interview he said that large language models are "basically a dead end when it comes to superintelligence," adding that many at Meta would prefer he not say so publicly; AMI Labs is organized around the alternative.⁹ On risk, he declined to sign the 2023 open letters that Hinton and Bengio endorsed, and he has repeatedly described near-term existential-risk scenarios as premature, arguing that systems can be made safe by architectural design, with objectives and guardrails built in, rather than by pausing or restricting research.⁹²¹ He has been an equally public advocate of open-weight release, crediting openness with distributing power over AI and arguing that the concentrated, closed development pursued by OpenAI and its peers is the greater danger; the Llama program embodied that position inside Meta.⁷¹⁴

Critics, including Bengio, respond that precaution is warranted precisely because capability evidence arrives late, and that open release of the most capable future systems is itself a channel of risk.²² LeCun's stated view, expressed through 2024 and 2025 interviews and in his departure announcement, is that human-level AI remains years to decades away, that today's safety debates underestimate how much architectural progress is still required, and that open world-model research is the fastest safe path.⁹¹⁰

See also

• Geoffrey Hinton
• Yoshua Bengio
• Meta AI
• AlexNet
• Backpropagation

References