Yann LeCun

AI summary of what Yann LeCun said

Yann LeCun argues that open-source AI is necessary because AI systems will mediate all our digital interactions and must reflect diverse languages, cultures and values, which a few companies cannot provide. He says open-source software is historically more secure, and that fears of AI taking over are unfounded because current systems are far from human or even animal intelligence. He notes that major scientific breakthroughs are still needed for machines to understand the world, remember, reason and plan.

• He says AI tools will become infrastructure mediating all digital interactions, and infrastructure software historically ends up open-source because it is more secure and customizable.
• He says we need free and diverse AI platforms for the same reason we need a free and diverse press, so a single opinion does not dominate.
• He says current AI systems are very far from matching human and animal intelligence, and breakthroughs are needed for machines to understand the world, have persistent memory, reason and plan.

AI summary of what Yann LeCun said

In this interview at Sciences Po, Yann LeCun criticizes advocates of an AI pause and existential-risk narratives, arguing they serve commercial and regulatory capture. He explains why he thinks large language models are a dead end for human-level intelligence, and presents his JEPA/world-model approach and open-source Project Tapestry as paths to European and global AI sovereignty. He also discusses regulation of deployments rather than research, robotics, and animal intelligence.

• He says the advocates of an AI pause, like Dario Amodei, are inconsistent and are using safety arguments to capture the market through regulation and to ban open-source AI.
• He argues that LLMs trained on text cannot reach human-level intelligence because the physical world contains far more information than text, and cites that a four-year-old child sees as much visual data as the largest LLM ingests in text.
• He presents JEPA (Joint Embedding Predictive Architecture) and world models as the path to systems that can plan and understand physical reality, which is the direction of his company AMI Labs.

AI summary of what Yann LeCun said

Yann LeCun, interviewed by Lex Fridman, discusses deep learning's history, why it broke textbook assumptions, and his conviction that intelligence is inseparable from learning. He explains his focus on self-supervised learning as the path to machines that learn world models, and argues human intelligence is specialized rather than general. He also discusses reasoning, memory, causality, patents on convolutional networks, autonomous driving, and why autonomous AI would need emotions.

• He argues HAL 9000 is an example of value misalignment, and that designing objective functions for machines parallels how human societies write laws to shape behavior.
• He explains that deep learning's success with huge over-parameterized non-convex networks broke everything pre-deep-learning textbooks said, and that intelligence is inseparable from learning.
• He advocates self-supervised learning as the key to building world models, saying reinforcement learning alone would require absurd amounts of trial and error, and that human-level AI needs grounding in reality and even emotions.

AI summary of what Yann LeCun said

Yann LeCun delivers a video message for a launch event in Tokyo on July 16th. He introduces himself as an NYU professor, co-founder of AMI, and Turing Award recipient, and argues that the next frontier of AI is physical AI built on world models that can perceive, reason, and act in the real world. He says Japan is well positioned to lead this field and calls for international collaboration.

• He says generative AI over the last five years has been disembodied, strong on language and symbols but unable to apprehend the physical world.
• He says the next generation of AI must learn from continuous, high-dimensional sensor data like video and build internal world models to predict and plan.
• He says Japan's expertise in precision manufacturing and industrial robotics gives it the foundation to lead in physical AI, and that international collaboration is needed.

AI summary of what Yann LeCun said

Yann LeCun, guest on the À LA FRENCH podcast, discusses his career from his thesis and Bell Labs to founding FAIR, and explains his current work on world models and JEPA architectures as the path beyond LLMs. He presents his new company Ami Labs, based in Paris with a global ambition, and describes Project Tapestry, a federated effort for sovereign, open AI models. He argues that generative AI cannot handle the unpredictable real world and that intelligence requires predictive world models rather than accumulated knowledge.

• He recounts his early career: his 1987 thesis on backpropagation, building the SN/Lush simulator with Léon Botou, developing convolutional networks at Bell Labs, and later founding FAIR, whose Paris lab he says had a huge effect on the French AI ecosystem.
• He explains that most information in video is unpredictable, so generative models fail as world models, and presents JEPA as an architecture that predicts in abstract representation space rather than pixels or tokens.
• He describes Ami Labs as a global company headquartered in France to offer advanced AI that is neither American nor Chinese, and Project Tapestry as a federation where participants train a shared open model by exchanging parameter vectors while keeping their data sovereign.

AI summary of what Yann LeCun said

Yann LeCun argues that autoregressive LLMs lack essential characteristics of intelligence such as understanding the physical world, persistent memory, reasoning, and planning, and cannot reach human-level AI on their own. He presents joint-embedding predictive architectures (JEPA) as the path forward, defends open-source AI as necessary for diversity and democracy, and dismisses AI doomer scenarios, saying progress toward human-level AI will be gradual rather than a single event. He also discusses robotics, the printing press analogy, and his optimism that AI will amplify human intelligence.

• He says LLMs cannot understand the physical world, have persistent memory, reason, or plan, so they are missing essential components on the path to human-level intelligence.
• He explains that generative models trained to predict pixels have failed for images and video, and that joint-embedding predictive architectures (JEPA) that predict in abstract representation space are the promising alternative.
• He argues that open-source platforms are the only way to get diverse, unbiased AI systems, since no single system can be unbiased in everyone's eyes.

AI summary of what Yann LeCun said

Yann LeCun, chief AI scientist at Meta, discusses self-supervised learning as the "dark matter of intelligence," arguing that machines must learn world models by observing the world the way babies and animals do. He explains his enthusiasm for non-contrastive joint embedding methods like VICReg and Barlow Twins, offers his speculative hypothesis that consciousness stems from the brain having only one configurable world model, and defends Facebook against media criticism. He also shares advice for young people and his hobby of building electronic wind instruments.

• He argues that self-supervised learning, learning world models by observing the world rather than from labeled data or rewards, is the missing paradigm needed to give machines common sense.
• He explains that non-contrastive joint embedding methods like VICReg and Barlow Twins are the developments he is most excited about, calling them the best path toward predictive world models.
• He discusses his speculative hypothesis that consciousness arises from the brain having only a single, configurable world model, making consciousness a consequence of the brain's limitations rather than its power.