The next AI doesn’t predict words.

LeCun spent twelve years as Meta’s chief AI scientist. For most of those years he told anyone who would listen that the models his own company was spending billions on were the wrong kind of machine, and in November 2025 he announced he was leaving. He walked out at the end of the year to build the thing he actually believes in. The company is called AMI Labs, short for Advanced Machine Intelligence, and in March 2026 it raised $1.03 billion at a $3.5 billion valuation to build world models instead of chatbots. If you’ve noticed the AI conversation splitting into two camps, the people scaling up ever-bigger chatbots on one side and a stubborn minority who think that whole approach is a dead end, LeCun is the man standing at the front of the second camp. What follows is his alternative, in plain English, and a look at what you can already go touch today.

Before I explain it, one small correction to the way most people introduce him. LeCun didn’t go to school with Geoffrey Hinton, though it’s an easy thing to assume since the two get mentioned together constantly. Hinton was his boss. LeCun did his PhD in France, Hinton signed on as his external thesis adviser after meeting him at a conference in Paris in 1985, and LeCun then did his postdoctoral work in Hinton’s lab in Toronto. Hinton likes to tell the story of the thesis defense, which was conducted in French, a language he doesn’t speak. The only thing he understood, he says, was the equations, and those were the important part anyway. LeCun, Hinton, and Yoshua Bengio shared the 2018 Turing Award for the neural-network techniques every modern AI system runs on. So we are talking about one of the co-inventors of deep learning leaving the building because he thinks everyone else is building the wrong kind of AI.

The trick every chatbot shares

Every model you have actually touched runs on the same basic trick. ChatGPT, Claude, Gemini, all of them were trained to predict the next piece of text. You show the model a sentence with the ending removed and it guesses what comes next, and the response you get is that guess repeated millions of times at speed. That turns out to be enough to produce something that writes code, drafts emails, and passes law exams, which surprised the people who built it as much as anyone.

LeCun’s objection is not that the trick fails. It’s that the trick teaches a model about words rather than about the world. A next-word predictor has to hedge every guess, because a sentence can continue in a thousand ways, so what it really learns is the statistics of language and a shadow of the world that produced that language. The technology’s own co-inventor is the one saying it.

“Current AI systems are less intelligent, in any real sense, than a house cat.”

What JEPA actually does

JEPA stands for Joint Embedding Predictive Architecture, which is a name that scares people off before the idea gets a chance. The idea itself is that instead of asking a model to predict the next word or the next pixel, you ask it to predict what a scene means, and you judge it on meaning rather than on whether it can render the details.

Think about what actually happens when you watch a ball roll off a table. You don’t mentally reconstruct every pixel of the ball, the table, and the light falling across the kitchen floor. You register something much more abstract: there’s an object, it’s falling, it’s about to hit the floor. Most of the detail never enters your head because most of the detail is irrelevant. A JEPA model is trained the same way. It compresses what it sees into an abstract internal description, in a space researchers call an embedding, and then it practices predicting the abstract description of what happens next from the abstract description of what it has already seen. It never has to produce pixels or words at all.

The training works like a fill-in-the-blank puzzle for meaning. You show the model most of a picture or video, hide a chunk of it, and have it predict what the hidden part looks like at the level of meaning rather than detail. Because the model is never forced to reproduce exact pixels, it is free to discard everything unpredictable, the texture of the grass and the flicker of the light, and keep the structure that matters: where things are, how they move, what causes what. LeCun’s bet is that this is how animals learn. A toddler doesn’t need a billion labeled examples to know a dropped cup falls. The toddler has an internal model of physics built by watching, and JEPA is an attempt to build that model in a machine.

The paper that laid this out is LeCun’s 2022 position paper, A Path Towards Autonomous Machine Intelligence. The architecture is self-supervised, which means it learns from raw video and images without humans labeling anything, the same no-labels trick that made large language models cheap to train, pointed at the physical world instead of at text.

What exists right now, and what you can actually use

So what can a normal person go use today? The honest answer is that you cannot chat with a JEPA model. There is no website where you type a question and get an answer, because the model doesn’t produce words. It produces internal descriptions, and its interface is a robot arm or a computer-vision pipeline. If what you want is to ask it questions, that product does not exist yet. Plenty of this is real and downloadable, though.

The flagship is Meta’s V-JEPA 2, released with open weights in June 2025. Anyone can download the actual model from Hugging Face and run it, which is why researchers outside Meta have kept building on it even after LeCun left. It was trained on over a million hours of internet video, and a fine-tuned version called V-JEPA 2-AC does something genuinely new: show it a starting image and a goal image and it plans the motion in between. In Meta’s labs it drove real robot arms picking up and placing objects they had never seen, with no task-specific training, and it succeeded about 80 percent of the time entirely zero-shot, meaning straight from internet video plus a day or so of arm-camera footage and no human labeling.

People have shipped real things on top of it, too. Nexar’s BADAS 2.0 is a collision-prediction model built on a fine-tuned V-JEPA 2 backbone and running on roughly 350,000 real dashboard cameras. It watches ordinary driving and predicts, before it happens, that a crash is developing. That is a world model doing a job on production roads right now. On the research side there’s a whole family of these models: I-JEPA for images, V-JEPA and V-JEPA 2 for video, VL-JEPA for video plus language, and LeJEPA, LeCun’s attempt to strip the training recipe down to something clean and provable. The code is public on GitHub, and a curated collection called awesome-jepa will take you down the rabbit hole if you let it.

JEPA is not the only world-model game in town, which matters for the can-I-use-it question. Google DeepMind’s Genie 3 generates playable, navigable 3D environments. World Labs, founded by Fei-Fei Li, the researcher behind the dataset that taught computers to see, sells Marble, which turns photos into explorable 3D scenes. NVIDIA’s Cosmos 3 is open-weight and aimed at robotics. Those are mostly generative world models, though, meaning they render pixels and make video. JEPA’s pitch is that rendering is the wrong goal, because reality is predictable at the level of meaning and unpredictable at the level of pixels, and a machine forced to render every detail wastes its capacity on the unpredictable part.

Is this better than the chatbot you use?

For the things you use a chatbot for, writing and searching and summarizing and coding, a JEPA model is not a competitor, and nobody serious claims otherwise. The comparison that matters is the long one: which kind of machine gets to genuine understanding and planning first. The LLM camp is betting that scale plus language gets you there. LeCun is betting that language is a narrow pipe into an intelligence that actually runs on a model of the physical world, and that the pipe will never be wide enough. The argument is genuinely unresolved, and serious people take the other side. Peter Norvig has argued that language models already are world models and need tweaks, not a restart.

Where the evidence currently favors LeCun’s side is embodied work, robots acting in unfamiliar environments and systems that anticipate physical events like a falling object or an oncoming car before they happen. Where it favors the mainstream camp is everything you can buy today. JEPA models are the bet, not the product. LeCun has said AMI Labs needs a few years to show whether the bet pays, and he has also told reporters he thinks the current AI boom is a bubble heading for a big explosion. He means the money, not the ideas.

I’ve been trying to understand this well enough to explain it plainly, and I’ll admit where that runs out: I’m not sure I’m able to explain it so most people fully understand. I’ll keep trying. The one-sentence version is that today’s AI predicts the next word, and the bet behind JEPA is that the next machine worth building predicts what the world will do next instead.