Yann LeCun Left Meta Because LLMs Will Not Get You to Intelligence

Jacob Effron opens by calling Yann LeCun one of the godfathers of AI. LeCun declines the compliment on the spot. He lives in New Jersey, he says, and in New Jersey a godfather means something else. The joke is a useful tell. At 65, after twelve years at Meta building FAIR into one of the most respected industrial labs in the field, he is not interested in mythology. He is interested in an architectural argument he has been making for a long time, now attached to a company, a raise, and a research agenda he no longer thought Meta would let him run.

He starts by taking a misunderstanding off the table. There is nothing wrong with large language models as the basis for useful products. He uses them himself. They are great at what they do. "They're just not a path towards human-level or human-like intelligence or even animal-like intelligence." That is the claim, and it is the reason the new company exists. AMI stands for Advanced Machine Intelligence. The motto is "AI for the real world."

Language, he argues, is a special case that happens to fit GPT-style architectures unusually well. It is discrete and relatively clean: on the order of a hundred thousand tokens, a probability distribution, a sample, feed it back, repeat. Autoregressive prediction is a natural fit. The physical world is the opposite: high-dimensional, continuous, noisy, and messy. Training a system to understand that world is much harder, and it is what he has spent most of his career on, accelerating over the last five or six years and making what he calls significant progress in the last two. By the end of 2025, Meta was no longer the right place to push it. That is why he left and started AMI Labs.

World model is already becoming a buzzword, LeCun says, in research and starting to leak into industry. He splits the field into factions and dismisses one of them quickly. Vision-language-action models, which reuse LLM technology to map vision and language into robot actions, are "pretty much now seen as a failure": not reliable enough, too hungry for training data. The remaining fight is over how you actually model the world. For him a world model is not a video generator. It is the thing that lets an agentic system anticipate the consequences of its own actions.

He cannot imagine building an agent without that ability. When people act without thinking about consequences, others call them idiots, and he notes there is no shortage of examples on the international political scene. If you can predict consequences, you can plan a sequence of actions that satisfies a goal, by search and optimization. You do not do it by predicting one action after another autoregressively. LLMs, in his account, have neither property. Inference is next-token prediction, not search. Those two missing pieces, prediction of action outcomes and planning by optimization, are the characteristics he thinks are essential for intelligent behavior.

He reaches for the bottle on the table. It is open, uncapped. Push near the bottom and it slides. Push near the top and it probably flips. You cannot predict exactly which way it falls, how the water spills, or whether a slight tilt in the table sends the liquid one direction or another. There is no way to predict that at the pixel level. "Our mental model of the world predicts that at an abstract level of representation." Cognitive science is part of the inspiration, especially the System 2 picture of deliberate behavior as imagining consequences and planning, as opposed to System 1’s reactive instincts. Translating that into a neural architecture is a different problem, and the empirical evidence, he says, is that you do not want to generate pixels.

LeCun has been trying to learn models of the world by prediction for a very long time. The turning point was an epiphany about five years ago: all of the architectures that succeeded at learning representations of images and videos were non-generative, and the generative ones were basically failures. Variational autoencoders are the obvious first idea, train a net to reconstruct its input, but a plain autoencoder learns the identity function, and VAEs and sparse autoencoders never produced representations he found satisfying. Masked autoencoders, cousins of denoising autoencoders and of BERT in NLP, were no better. FAIR ran a huge MAE project. It burned a lot of compute. The result, he says, was disappointing.

The methods that worked were joint embedding architectures. Take an image, corrupt it, run both versions through encoders, and train a predictor to recover the representation of the original from the representation of the corrupted one. That is JEPA, the joint embedding predictive architecture. DINO, through v1, v2, and v3, still running at FAIR in Paris; I-JEPA and V-JEPA; earlier SimSiam and MoCo, mostly from Meta, plus related work from other groups. Predicting in representation space beat predicting pixels. "It just clicked in my mind, but not just mine, that this was the way to go."

Effron puts the obvious objection on the table. Robotics demos from the model companies look increasingly like planning and generalization: a robot walks into a room it has not seen and does a version of a task it has not been trained on. LeCun does not deny the progress. He denies the interpretation. Those systems are trained on enormous amounts of data from teleoperation, from handheld gripper-like devices, or from tracking a person’s hands and fingers, then translating that into robot commands. The recipe is mostly imitation learning, with a little reinforcement learning in simulation. You need a lot of data for every task. The systems are expensive and brittle.

A world model changes the generalization story. If the system can predict the outcome of an action, it can plan its way through a new task without being trained to accomplish that task. Humans do this, and many animals do too. The hope is zero-shot problem solving, or something close, with only a little RL-style fine-tuning. His favorite comparison is driving. "How is it that a 17-year-old can learn to drive in like a dozen hours or maybe 20 hours? We have millions of hours of training data of people driving cars. We still don't have level five self-driving cars." Imitation learning, he says, does not even work for that one task.

Synthetic video for simulation does not impress him as a substitute. You do not need millions of hours of demonstration, and you do not need generated data, if you can learn as fast as a teenager. It is a question of data efficiency. Effron mentions Jerry Tworek, formerly of OpenAI, who expressed a similar unease: keep scaling RL and the results will keep getting more impressive, but there has to be a way that is orders of magnitude more efficient. The problem, LeCun says, is incentive. Frontier labs cannot afford to fall behind one another, so they all work on the same thing. "Everybody is digging the same trench." He put AMI Labs’ headquarters in Paris on purpose, with the American office in New York, not Silicon Valley. It is a bet, but he says they already have results.

Asked where AMI’s technology goes first, LeCun answers with the missing products. Where is your domestic robot? Where is your Level 5 car? A household robot is several years away. A huge number of companies are building robots, he says, and none of them actually knows how to make the machines smart enough to be useful, let alone trusted around a baby. Even relatively narrow manufacturing tasks are unreliable except by imitation learning for a small number of jobs. Making those systems useful is the long-term objective.

Shorter term, he sees a huge set of industrial problems that share a shape. You have a complex system, a jet engine, a chemical plant, a power plant, a manufacturing line, a patient, a human cell, and you need to predict what happens if you change a control variable. Those systems are too complicated to model with a small number of equations. The move is to train a deep network on data to capture the dynamics, get a phenomenological model, and, if it is action-conditioned, a world model you can use for optimal control. "The number of applications of this in industry is mind-boggling."

On the path of progress, he reaches for a Linus Torvalds line. Asked about Linux’s goal, Torvalds said total world domination, and to a first approximation every computer in the world now runs Linux, Azure included. LeCun’s five-year joke is the same phrase. Underneath the joke is a blueprint: systems that can think, even if they cannot talk or listen at first, with LLMs remaining as a language interface on top. Within a year or so he wants a general methodology for training hierarchical world models across a wide variety of modalities, plus demonstrations, some robotic, some industrial process control, some in healthcare, with partners who are already investors. He thinks the realization that the paradigm has to change is happening now and will be "completely obvious to people by early 2027." That does not mean there will be a solution by then. He hopes there will. He is not promising one.

Tapestry is orthogonal to AMI Labs, an idea he has been forming for about three years. People are already replacing search with AI assistants. If Meta’s bet on smart glasses lands, a large share of the information diet will be mediated by a voice in a wearable. If you live outside the United States or China, that assistant was probably built in California or in Beijing, Shanghai, or Shenzhen. It may handle your language poorly, misunderstand your culture, miss your values, and almost certainly fail to represent your politics. The farmer in India and the philosopher in France are not well served by a handful of West Coast or Chinese models.

The proposed fix is an open, free, LLM-style foundation model that anyone can fine-tune for a language, a culture, a creed, or a political bias. A lot of countries that are neither the US nor China want sovereignty for their industry and their citizens. They do not want those citizens "brainwashed by a Chinese model or a Canadian model, actually." Tapestry’s trick is that contributors, India, France, Vietnam, Morocco, Switzerland, Korea, Japan, Kazakhstan, would add data and compute without sharing the data. What they exchange are parameter vectors, in a federated-learning style. Local training stays close to a global consensus vector. The limit of that process is an open model as good as if it had been trained on all the data in the world, which each participant can then specialize.

He thinks two forces pull in that direction: the demand for sovereignty, and the tendency of platforms to become open. Linux wiped out proprietary Unix. The software infrastructure of the internet and of wireless networks went the same way. Who were the big internet-infrastructure players in 1996? Sun Microsystems, HP, Dell. Sun sold Solaris on proprietary hardware. HP had HP-UX. They claimed Unix was far more reliable than Windows. Dell ran Windows NT. "Who is running Windows NT now as a web server?" Linux took the entire internet, Azure included. "OpenAI and Anthropic etc. of today are the Sun Microsystem and HP-UX of yesterday." The closed labs have already run out of publicly available text. They are licensing copyrighted data or training on synthetic data. The window for an open catch-up, in his view, is still there.

Effron puts the scoreboard on the table: IMO gold, the METR task-horizon benchmark climbing. LeCun does not contest the results. He contests the domain. Mathematics and code are two areas where language itself is the substrate of reasoning. When you prove a theorem formally, you manipulate language, and LLMs are really good at that. They are much less good at coming up with concepts and definitions. They are problem solvers. "Mathematics is not just problem solving. Most of it is actually a creative act that those things don't do." The same split applies to software. "LLMs are good programmers. They're not software architects. They're not computer scientists." People move one level up the abstraction hierarchy and decide what to build. The model helps with the building.

What would convince him otherwise? A zero-shot agentic system: give it a new problem it has not been trained to solve and has no script for, and see whether it can do the task. Without the ability to predict the consequences of its actions and use that for planning, he does not think it can, and he does not think you get there with an LLM. Current math and code systems already search, because they can run the program or check the proof. That only works where search can be performed in token space. JEPA’s claim is that you plan in an abstract thought space instead. If token-space search works for a large part of the economy, he is fine with using it. It is still not a path to human-level AI. Without explicit world models, he does not believe you can make the systems fully reliable, or stop them from hallucinating, or deploy them across a widening set of use cases without collecting mountains of task-specific data.

LeCun shared the Turing Award with Geoffrey Hinton and Yoshua Bengio. Effron asks when their views diverged. "In 2023." He did not change his mind, he says. They did, at about the same time, and the trigger was GPT-4. Hinton had never been very interested in LLMs. When GPT-4 appeared he had an epiphany: the systems seemed close to human-level intelligence, possibly with subjective experience. The arithmetic was a cortex of about 16 billion neurons, a guess that you might need ten biological neurons to emulate one backpropagation unit, which collapses the cortex to 1.6 billion equivalent neurons, close enough to GPT-4 that human-level intelligence looked nearby. LeCun does not believe the claim at all.

He offers a less charitable reading of his friend’s trajectory. Hinton spent a career searching for the learning algorithm of the cortex, did not find the real thing, noticed that backpropagation is a good enough substitute, and could declare victory, retire, and tour the world talking about the promises and dangers of AI. LeCun says Hinton is much less vocal about those dangers than he was a year or two ago. He has noticed that current LLMs are not that smart, that a few conceptual breakthroughs still sit between here and human-like intelligence, and that the blueprint of those systems will be quite different from LLMs and probably controllable. Bengio’s worry, in LeCun’s telling, is also not the doomer story of AI taking over the world. It is whether political systems can spread the benefits, avoid making a few rich people richer, and prevent catastrophes from bad use. That is a real danger. It is not as apocalyptic as some people have claimed, "certainly not as apocalyptic as what even Anthropic has claimed." He thinks Anthropic genuinely believes its story, and also has commercial reasons to get governments to believe their systems are dangerous. He does not subscribe to it.

Then he says the thing some Meta colleagues did not like. LLMs are intrinsically unsafe. He does not think they can be made reliable, because you cannot stop them from hallucinating. If they are agentic, you cannot guarantee they will not take an action whose outcome they did not predict. Long coding tasks do not refute that. Coding is a domain where you can check that the program meets a spec, and even there, agents wipe hard drives and do expensive stupid things. The task an LLM performs is only as good as the training that mapped a prompt onto that task. There is no hardwired constraint. There is always a gap between training error and test error, and always a prompt that makes the system do something stupid. A joke was circulating a month before the taping: the car wash is a hundred yards from my house, should I walk? He tried it again two weeks before the interview. They all said yes, except Gemini. Then the gag entered the next training set, and answering it stopped being evidence of common sense.

The alternative he has been proposing is objective-driven AI. You give the system a task. It has a world model. It imagines a sequence of actions, predicts the outcome, and optimizes a cost that says whether the task is done. If that is how the system works, "it can do nothing else." Plenty can still go wrong: the cost function may not actually measure success, the world model may be inaccurate, the system can still make mistakes. But it can predict consequences, which he thinks is indispensable for any agent, and you can add safety constraints that it cannot violate by construction. An LLM can always escape. In healthcare the split is practical. An LLM can regurgitate textbook knowledge and, at best, scale a top doctor’s written competence around the world. Designing a course of treatment for a particular patient needs a model of that patient’s physiology. Telling a stem cell to become an insulin-producing pancreatic beta cell, for type 1 diabetes, needs a model of a human cell. You cannot become a doctor by reading books. You have to do residency, listen to the heart, press on the belly.

Asked what he got most right and most wrong running FAIR, LeCun starts with the lab itself. It produced a lot of the basic methods, science, and tools the rest of the industry uses. PyTorch is the load-bearing example: "the entire industry is built on PyTorch basically, except for a few people at Google." It also produced a culture of openness and scientific process that he thinks is necessary for breakthrough innovation. There is a chain. Blue-sky work happens in universities and in a handful of industrial labs you can count on the fingers of one hand, Google among them, FAIR among them. Then someone pushes an idea until it is practical enough to beat a benchmark. Then the host company puts a large engineering effort behind it. That last stage is where a lot of projects, and a lot of companies, fail. Meta was pretty good at it, he says, far from a Xerox PARC story of inventing the GUI and missing the industry, but it missed steps. You need an organization close to research that is not a product team with a three-month deadline. Facebook and Meta had that, and then they lost it. FAIR ended up isolated, with ideas nobody picked up.

In 2023 Meta created GenAI by taking about 60 or 70 scientists and engineers out of FAIR. The new organization was under so much short-term pressure that it did not have time to talk to the lab it came from. Instead of staying at the forefront of LLM innovation, it focused on near-term work and became conservative. Llama 1 had been a small project inside FAIR in 2022. Early 2023 the Llama people moved to GenAI and started Llama 2. Two of the Llama 1 authors left with someone from Google and created Mistral. Llama 3 and Llama 4, in his telling, inherited the short-term pressure. Zuckerberg was disappointed by Llama 4 and rebooted the organization. The company realized it had fallen behind and refocused on catching up, with the sad side effect that exploratory research lost priority. His own JEPA and world-model work still had interest from Zuckerberg and from CTO Andrew Bosworth, but the rest of the company was entirely on LLMs, and most of the applications sat in manufacturing and other industries Meta was not particularly interested in.

Is that kind of lab still possible inside the race, or is the only path to leave? He thinks a few pockets remain inside Google Research and DeepMind. The industry has otherwise become more closed. Google tightened; even FAIR now has more restrictions on publication. People who want to do breakthrough work get fewer resources and are told not to talk about anything with medium-term relevance. The formula he wants is the early-FAIR, Bell Labs, Xerox PARC formula: hire the best people, people with a nose for what to work on, give them the means to succeed, "and you get the hell out of the way." Pardon his French. For PhD students the advice is even sharper. In academia you should work on the next generation of AI systems, not the current one. Studying how and why LLMs work is descriptive science. If you want to show new things with LLMs, you will not have the GPUs. "Don't work on LLM if you're doing a PhD. Like, there's no point. You cannot contribute."

How did he know it was time to leave? First he wants a misconception killed. He joined at the end of 2013, really starting in early 2014, and spent the first four and a half years as director of FAIR: building the organization, setting the culture, hiring the key people. Then he stepped down, around 58, because he did not want to manage. He thinks of himself as a scientific and technical visionary, not a manager. Joelle Pineau and Antoine Bordes took over the directorship. He became chief AI scientist, reporting to the CTO, and put his own research back in motion. FAIR’s ambition had always been to build intelligent systems. In a 2016 NeurIPS keynote, still called NIPS at the time, he said the way forward was world models that predict the consequences of actions and then plan. Reinforcement learning was too inefficient. Supervised learning had shown its limits. The future was self-supervised learning and world models.

Early video-prediction projects did not pan out. Around 2020 he arrived at JEPA: make the predictions in representation space, then make the predictor action-conditioned, then use it for planning. In 2022 he wrote a long vision paper and, in his phrase, spilled all his secrets, hoping to rally people. It worked. Students came to NYU and Paris to work on it. A FAIR team wanted in. Joelle Pineau suggested it become a major mission. The internal name was Advanced Machine Intelligence, now the name of the company. Zuckerberg read the paper and subscribed. So did Bosworth, former CTO Mike Schroepfer, and chief product officer Chris Cox, who was LeCun’s manager. Then the company refocused on LLMs. Despite support from Zuckerberg and Bosworth, the layers below did not see the point. Politically it became difficult. Meta shut down its robotics AI group, led by Jitendra Malik, now at Amazon. FAIR was under pressure to help the LLM effort. The applications of JEPA world models sat in industries Meta had no interest in. Raising money for AMI was straightforward because investors already knew the story from the paper and the talks.

The Scale AI acquisition and Alex Wang’s arrival were, in the popular narrative, the moment a research lab became harder to run. LeCun says he has no inside information. It is possible Zuckerberg sees in Wang a potential successor, a younger version of himself. The bigger misconception is about Llama. "I had zero technical contribution to Llama, like none whatsoever." His one contribution was arguing to open-source Llama 2. Legal was against it. Policy was against it. Comms and the engineering side, including Bosworth, were for it. For months, about 40 people from Zuckerberg down met every week for two hours. LeCun and Bosworth argued that the safety risks were overblown and that open-sourcing Llama 2 would jump-start an industry. In his view, that is exactly what happened. He never tried to stop LLM work inside FAIR. After 2018 he had no direct influence on what people worked on anyway, except by publishing a vision and rallying volunteers. Starting in early 2024, and certainly in 2025, the way FAIR was managed no longer matched what he thought was necessary to preserve innovation and keep good people. A lot of good people had already left.

Effron asks for one thing LeCun has changed his mind about in the last year. The answer is not LLMs versus world models. It is the fate of the program formerly called unsupervised learning and now called self-supervised learning. The idea was to learn a good representation of the input, then fine-tune with a little labeled data. He tried to apply it to video, because the basic ambition was a system that understands how the world works by watching the world go by. Pixel-level video prediction did not work. He kept arguing for training a system not to solve a particular task but to predict, then reuse the representation. That was the second half of the 2016 keynote. What surprised him is that the program became incredibly successful, but not for video. For language. "LLMs basically are a blindingly successful example of self-supervised learning."

He spends the last word on why the same trick is harder in the real world, and on the paper he wants people to read. An LLM works because the symbols are discrete. The net produces a distribution over something like 100,000 tokens, you sample, you shift, you go autoregressive. You cannot do that with the physical world, so you learn a representation and predict in representation space. The trap is collapse: the encoder emits a constant, prediction becomes trivial. He invented contrastive learning in 1993 as a fix. It works and does not scale well with dimension. Distillation methods, DINO, DeepMind’s BYOL, prevent collapse and nobody really knows why. The cost function you think you are minimizing actually goes up during training. V-JEPA and I-JEPA were trained that way. He is moving off it.

The replacement is an explicit regularizer that tries to maximize the information coming out of the encoder, in the same family as Becker and Hinton in 1989 and Jürgen Schmidhuber in 1992. You cannot measure information content, only upper-bound it, so you take an upper bound and hope. VICReg was an earlier version. The latest is SIGReg, isotropic Gaussian regularization, from Randall Balestriero, a former postdoc of LeCun’s who is now an assistant professor at Brown. It forces the distribution of encoder variables toward a joint Gaussian. Variations can produce sparse or anisotropic representations. With Balestriero and a student at Mila, they trained a world model this way. It is still small-scale. He thinks it is promising. If you read one paper, he says, read that one: *Le World Model*. He is not responsible for the name. Randall picked it.