Every major AI company on Earth is pouring billions into large language modelsA machine learning system trained on vast amounts of text that predicts and generates human language. These systems like GPT and Claude exhibit surprising capabilities but also make confident errors.. Yann LeCun, the Turing Award-winning researcher who helped invent the technology behind modern AI, thinks they are all wrong. In March 2026, he put $1.03 billion behind that conviction, launching AMI Labs to build what he calls “world modelsAn AI system's internal representation of how the physical world works, enabling it to predict the consequences of actions before taking them. AI” – an entirely different approach to machine intelligence.
It is one of the most dramatic bets in the history of artificial intelligence: a founding figure of the field walking away from the dominant paradigm and wagering that the technology powering ChatGPT, Claude, and Gemini is fundamentally incapable of achieving real understanding.
The Case Against Large Language Models
LeCun’s argument is straightforward. Large language models are trained on text. They predict the next word in a sequence based on statistical patterns learned from reading vast swaths of the internet. They do this remarkably well, which is why they can write essays, summarize research, and generate code.
But according to LeCun, that is all they will ever do.
“We are going to have AI systems that have humanlike and human-level intelligence, but they’re not going to be built on LLMs,” he told MIT Technology Review from his Paris apartment. “It’s not going to happen next year or two years from now. There are major conceptual breakthroughs that have to happen.”
The core problem, LeCun argues, is that language models do not understand the world. They can pass the bar exam and write working software, but they cannot predict what happens when you push a glass off a table. They have no concept of gravity, no sense of cause and effect, no model of how physical reality operates.
This is why, despite years of effort and hundreds of billions of dollars in investment, nobody has built a domestic robot as capable as a house cat or a truly autonomous car. The hard part is not language. The hard part is everything else.
World Models AI: Learning from Reality
LeCun’s alternative is called JEPA, or Joint Embedding Predictive Architecture. Instead of training on text, JEPA learns from video, audio, and sensor data. Instead of predicting the next word, it learns to predict how the real world changes over time.
The key insight is in what JEPA does not do. Traditional generative AI tries to reconstruct every detail of what it observes, pixel by pixel. JEPA instead learns abstract representations, focusing on the patterns that matter and ignoring unpredictable details.
“The world is unpredictable,” LeCun explained. “If you try to build a generative model that predicts every detail of the future, it will fail. JEPA learns the underlying rules of the world from observation, like a baby learning about gravity.”
Think of how a child learns. A baby does not need to be told that unsupported objects fall. It watches objects fall, builds an internal model of how gravity works, and then uses that model to predict and plan. LeCun believes machines need to learn the same way.
Why He Left Meta
LeCun spent over a decade at Meta, where he founded and led FAIR (Fundamental AI Research), one of the most influential AI labs in the world. He left in November 2025.
The departure was not amicable in all respects. At the World Economic Forum in Davos in January 2026, LeCun said publicly that Meta’s decision to invest tens of billions in LLM-focused data centers contributed to his exit. “The AI industry is completely LLM-pilled,” he told the audience. “In Silicon Valley, everybody is working on the same thing. They’re all digging the same trench.”
He was diplomatic about Meta in his MIT Technology Review interview, noting that “Mark made some choices that he thought were the best for the company. I may not have agreed with all of them.” But he was blunt about one decision: Meta’s robotics group at FAIR was dissolved, which LeCun called “a strategic mistake.”
The Billion-Dollar War Chest
AMI Labs’ $1.03 billion seed round is the second-largest seed round in tech history, trailing only Thinking Machines Lab’s $2 billion raise. The company is valued at $3.5 billion before the investment.
The investor list reads like a who’s who of global tech capital: Nvidia, Bezos Expeditions (Jeff Bezos’ investment firm), Singapore’s sovereign wealth fundA state-owned investment fund that manages national savings or commodity revenues on behalf of a government, typically for long-term economic benefit. Temasek, Samsung, Toyota Ventures, and Eric Schmidt. The round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions.
AMI Labs is headquartered in Paris, with offices planned in New York, Montreal, and Singapore. Alexandre LeBrun, former CEO of health-tech startup Nabla, is running the company as CEO, with LeCun serving as executive chairman. Laurent Solly, Meta’s former VP for Europe, joined as COO.
LeCun is keeping his professorship at NYU. “I can do management, but I don’t like doing it,” he admitted. “My mission is to make science and technology progress as far as we can.”
Not Everyone Agrees
LeCun’s contrarian stance puts him directly at odds with the leaders of the world’s most valuable AI companies.
At Davos, Anthropic CEO Dario Amodei told the same audience that AI models built on the current architecture would replace the work of all software developers within a year and reach “Nobel-level” scientific research within two. He predicted 50% of white-collar jobs would disappear within five years.
Google DeepMind CEO Demis Hassabis was more measured, saying current systems are “nowhere near” AGI and that “maybe we need one or two more breakthroughs.” He placed 50-50 odds on AGI arriving within the decade, though not through models built exactly like today’s systems.
Critics also question whether LeCun’s specific approach is the right one. Gary Marcus, a prominent AI researcher and longtime LeCun critic, has argued that LeCun “has the right name for a thing we need (‘world model’) but an inadequate implementation.” Marcus points out that the concept of world models goes back to the 1950s, and that Jurgen Schmidhuber proposed adding world models to neural networks as early as 1990. LeCun, Marcus contends, rarely credits these predecessors.
AMI Labs CEO LeBrun is candid about the timeline. “AMI Labs is a very ambitious project, because it starts with fundamental research,” he told TechCrunch. “It’s not your typical applied AI startup that can release a product in three months.” It could take years for world models to produce commercial applications.
What World Models AI Could Actually Do
If LeCun is right, the applications extend far beyond chatbots. AMI Labs is targeting industries that operate complex physical systems: manufacturing, aerospace, biomedical research, and pharmaceuticals. These are domains where errors have real consequences and where understanding physical reality is not optional.
LeCun envisions world models powering autonomous robots that can adapt to unfamiliar environments, smart glasses that anticipate your next action, industrial systems that predict equipment failures before they happen, and truly self-driving cars that understand cause and effect rather than just pattern-matching road scenarios.
“An agentic system that is supposed to take actions in the world cannot work reliably unless it has a world model to predict the consequences of its actions,” LeCun argued. “This is the key to unlocking everything from truly useful domestic robots to Level 5 autonomous driving.”
The Open-Source Angle
LeCun is also positioning AMI Labs as a counterweight to what he sees as a dangerous consolidation of AI power. He advocates forcefully for open-source AI, criticizing both OpenAI’s shift from open to closed and Anthropic’s always-closed approach.
“If there is a future in which all of our information diet is being mediated by AI assistance, and the choice is either English-speaking models produced by proprietary companies always close to the US or Chinese models,” he told MIT Technology Review, “it’s not a very pleasant and engaging future.” AMI Labs plans to publish papers and release code as open source as it progresses.
The Stakes
Whether LeCun is right or wrong, the stakes are enormous. The AI industry has committed trillions of dollars to the LLM paradigm. If world models prove to be the breakthrough needed for genuine machine intelligence, the current gold rush could look like a detour. If LeCun is wrong, he will have burned through a billion dollars chasing a technology that produces little more than interesting research papers.
LeCun has been wrong before, and he has been right before. He was one of three researchers who won the 2018 Turing Award for work on deep learning that the entire AI boom now rests upon. But as a Newsweek analysis observed, nobody knows exactly how current AI systems achieve their results, which makes the debate over their ultimate ceiling far from settled.
What makes LeCun’s bet distinctive is not the critique itself. Many researchers share his skepticism about LLMs reaching human-level intelligence. What is distinctive is the scale of the wager: a billion dollars, a new company, and a public break with the industry consensus, all staked on the conviction that the most successful technology in a generation is a dead end.
Every major AI lab is scaling large language modelsA machine learning system trained on vast amounts of text that predicts and generates human language. These systems like GPT and Claude exhibit surprising capabilities but also make confident errors.. Yann LeCun, co-recipient of the 2018 Turing Award for foundational work on deep learning, believes this is a categorical error. In March 2026, he raised $1.03 billion for AMI Labs to pursue world modelsAn AI system's internal representation of how the physical world works, enabling it to predict the consequences of actions before taking them. AI built on Joint Embedding Predictive Architecture (JEPA), an approach he argues addresses fundamental limitations that no amount of LLM scaling can overcome.
The Technical Argument Against LLMs
LeCun’s critique is grounded in what he sees as an architectural ceiling. LLMs operate in a discrete, low-dimensional space: text. They learn statistical correlations between tokens and generate outputs by sampling from probability distributions over those tokens. This works remarkably well for language tasks precisely because, as LeCun explained in a Newsweek analysis, “human language fits this requirement for a discrete, low-dimensional space.”
The problem is that the physical world is continuous, high-dimensional, and governed by causal relationships that text cannot encode. LeCun frames this through Daniel Kahneman’s System 1/System 2 distinction: current LLMs are System 1 processors. They produce one token after another through a fixed computation graph, reacting to the current context without genuine deliberation.
Chain-of-thoughtA prompting technique where a language model is guided to reason step by step before giving a final answer, improving accuracy on tasks requiring logic or multi-step analysis. reasoning, which appears to add System 2 capabilities, is in LeCun’s view “still just a statistical approach, exploring multiple paths with different probabilities before rendering an answer.” He calls it “System 1.1 at best” and notes the n² explorations these methods require make them fundamentally inefficient compared to actual System 2 reasoning over abstract world models.
“We have systems that can pass the bar exam, they can write code,” LeCun said at Davos, “but they don’t really deal with the real world. Which is the reason we don’t have domestic robots and we don’t have level-five self-driving cars.” This is the Moravec ParadoxThe finding that tasks easy for humans (perception, movement) are hard for AI, while tasks hard for humans (math, chess) are easy for AI. in action: what is easy for humans (perception, navigation, physical reasoning) remains hard for computers, and vice versa.
World Models AI and the JEPA Architecture
JEPA, which LeCun proposed in 2022 before the LLM explosion, takes a fundamentally different approach to representation learning. Where generative models (including LLMs) predict in pixel or token space, JEPA predicts in abstract representation space.
The architecture works through three components:
- Encoders transform raw inputs (video frames, audio, sensor data) into abstract representations that capture essential features while discarding irrelevant detail.
- A predictor module learns to predict the abstract representation of a future state from the current state’s representation. Crucially, predictions happen in this compressed representation space, not in raw sensory space.
- A latent variable represents elements present in the future state but not observable in the current one, allowing the system to handle genuine uncertainty rather than collapsing to a single predicted outcome.
“The key is to learn an abstract representation of the world and make predictions in that abstract space, ignoring the details you can’t predict,” LeCun told MIT Technology Review. “That’s what JEPA does. It learns the underlying rules of the world from observation, like a baby learning about gravity.”
This mirrors how neuroscience models describe human cognition: the brain maintains a phenomenally rich internal model of the world that receives minimal input from the senses to “ground” it at each point in time. The extraneous information not required by the model is eliminated, enabling continuous trajectory computation through mental model space with high efficiency.
The practical distinction matters for downstream applications. An action-conditioned world model, as AMI Labs envisions, can predict the consequences of potential actions before executing them. This enables planning: evaluating multiple action sequences and selecting the one most likely to achieve a goal while satisfying safety constraints. LLM-based agents, by contrast, must rely on external tools and trial-and-error because they lack this predictive capacity.
The Meta Departure and Industry Context
LeCun founded FAIR (Fundamental AI Research) at Meta in 2013 and served as chief AI scientist for over a decade. He left in November 2025 after what he described as a growing divergence between his research agenda and Meta’s strategic direction.
The proximate cause was Meta’s organizational restructuring. Meta struggled to gain traction with its Llama models and saw internal shake-ups as CEO Mark Zuckerberg pivoted away from fundamental research toward applied AI. FAIR’s robotics group was dissolved, a decision LeCun called “a strategic mistake.”
At Davos, LeCun was explicit: Meta’s decision to pour tens of billions into LLM-focused data centers contributed to his exit. His view that LLMs would not lead to human-level intelligence “made him unpopular at the company.”
AMI Labs: Structure and Funding
AMI Labs raised $1.03 billion at a $3.5 billion pre-money valuation, the second-largest seed round in tech history after Thinking Machines Lab’s $2 billion raise. The round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions, with participation from Nvidia, Temasek, Samsung, Toyota Ventures, and individual investors including Tim Berners-Lee, Eric Schmidt, and Mark Cuban.
The leadership team draws heavily from Meta’s AI ecosystem. Alexandre LeBrun (CEO) formerly led FAIR’s Paris engineering division and founded Nabla, an AI healthcare company that is AMI’s first disclosed partner. Saining Xie, formerly of NYU and Google DeepMind, serves as chief science officer. Laurent Solly, Meta’s former VP for Europe, is COO.
Target customers include organizations operating complex physical systems: manufacturers, aerospace firms, pharmaceutical companies, and biomedical researchers. These industries need AI that operates reliably in dynamic environments where errors have real consequences, a requirement LLM-based systems cannot currently meet.
LeBrun is candid about timelines. “AMI Labs is a very ambitious project, because it starts with fundamental research,” he told TechCrunch. “It could take years for world models to go from theory to commercial applications.” The company plans to publish research papers and release code as open source, a deliberate contrast with the closed approach of frontier LLM labs.
Counterarguments and Criticisms
The leaders of the world’s most valuable AI companies disagree with LeCun’s assessment. At Davos, Anthropic CEO Dario Amodei predicted that current AI architectures would replace all software developers within a year and achieve “Nobel-level” scientific research within two. Google DeepMind CEO Demis Hassabis was more cautious, estimating a 50% chance of AGI within the decade but acknowledging “one or two more breakthroughs” might be needed.
There are also pointed criticisms of LeCun’s specific approach. Gary Marcus, a prominent AI researcher, argues that LeCun “has the right name for a thing we need (‘world model’) but an inadequate implementation.” Marcus contends that world models require explicit, structured, directly retrievable knowledge about time, space, causality, and events for formal reasoning, and that JEPA is “just another opaque, uninterpretable neural network that doesn’t lend itself to reasoning.”
Marcus also challenges LeCun’s claims of originality, noting that the concept of world models dates to Herb Simon’s General Problem Solver in the 1950s, and that Jurgen Schmidhuber proposed integrating world models with neural networks in 1990. LeCun, he argues, rarely credits these predecessors.
There is also an empirical challenge: LeCun worked on JEPA at Meta with a substantial team for several years. Results have been promising but incremental. I-JEPA demonstrated strong self-supervised image representation learning, and V-JEPA extended this to video. But neither has produced the transformative capabilities LeCun envisions. Whether $1 billion in independent funding changes this equation remains an open question.
The Structural Bet
What makes LeCun’s position worth analyzing carefully is not the critique itself. The view that LLMs cannot reach human-level intelligence is widely shared among AI researchers and academic scientists. What is unusual is the scale and structure of the bet.
LeCun is wagering that the most commercially successful AI paradigm in history has a hard ceiling, that the correct alternative is an architecture he designed, and that the research community and capital markets are both sufficiently persuaded to fund a multi-year effort with no near-term product. The $1.03 billion says at least some sophisticated investors agree.
The counterargument is equally structural: LLMs are not static. Multimodal models already process images, audio, and video. Reasoning capabilities continue to improve. The question is whether these extensions can close the gap LeCun identifies, or whether the autoregressiveA text generation method where each new token is predicted solely from all preceding tokens in the sequence, processing left-to-right with no ability to revise earlier outputs. token-prediction foundation is, as he claims, a permanent constraint.
If LeCun is right, the trillions flowing into LLM infrastructure represent a historic misallocation. If he is wrong, AMI Labs joins a long list of well-funded research ventures that produced interesting papers but no paradigm shift. Either way, the field is entering a period where the fundamental assumptions behind the dominant AI paradigm are being challenged at unprecedented scale.
