# Yann LeCun's JEPA Paper: Why a Tiny Model Beat Giant LLMs > Source: > Published: 2026-11-05 00:00:00+00:00 # Yann LeCun's JEPA Paper: Why a Tiny Model Beat Giant LLMs Yann LeCun's new JEPA paper shows a small world model outperforming massive LLMs on world understanding. Here's what it actually proves. ## What did Yann LeCun’s new JEPA paper actually show? Yann LeCun, along with Brown University professor Randall Balestriero, published a paper titled “Proving and Improving the Platonic Representation Hypothesis with JEPA” on October 29. It argues that AI models trained with a “joint embedding predictive architecture” (JEPA) learn an internal map of reality that aligns across different data types, like text and images, under specific training conditions. The paper also reports that a small JEPA-based model outperformed much larger language models on a test of world understanding, which LeCun has pointed to as evidence that scaling text-only LLMs isn’t the only path, or even the best path, to more capable AI. ## TL;DR - **Yann LeCun** , a Turing Award winner and co-creator of convolutional neural networks, has spent years arguing that scaling large language models won’t produce human-level AI. - The new paper, co-authored with **Randall Balestriero** , mathematically examines why models trained on different data types (text, images) sometimes converge on similar internal representations of the world, an idea called the**platonic representation hypothesis** . - LeCun’s architecture, called **JEPA (joint embedding predictive architecture)** , predicts in an abstract “representation space” instead of predicting raw pixels or exact words, which lets it skip unpredictable details. - The paper reportedly found that predicting the next word in a sentence, the core training method behind models like ChatGPT, Claude, and Gemini, is a comparatively weak signal for learning how the real world actually works. - LeCun left Meta at the end of last year to start **AMI Labs** in Paris, which raised**$1.03 billion** in its first funding round, backed by Jeff Bezos’s investment firm, Nvidia, Samsung, and Toyota. - A small JEPA-based model, reportedly around **109 million parameters** , beat models as large as**72 billion parameters** on one world-understanding benchmark, a gap of more than 600 times in size. - The result doesn’t prove JEPA will beat LLMs at everything, but it challenges the assumption that bigger models trained the same way are always the better bet. *Without shadow IT.* **Every app, approved** *✓* *✓* *✓* *✓* *✓* ## Who is Yann LeCun and why does his opinion carry weight? LeCun isn’t a commentator on AI, he’s one of the people who built the field. In the late 1980s at Bell Labs, he helped pioneer convolutional neural networks, a type of AI architecture especially good at processing images. That work became the backbone of check-reading systems used widely in the US by the late 1990s, and today the same basic ideas underpin face recognition, medical imaging AI, and self-driving car vision systems. In 2018, LeCun won the Turing Award, often described as the Nobel Prize of computer science, alongside Geoffrey Hinton and Yoshua Bengio. The three are commonly called the godfathers of AI. LeCun founded Meta’s AI research lab in 2013 and spent more than a decade as the company’s chief AI scientist while also teaching at NYU. At the end of last year, LeCun left Meta to found AMI Labs (Advanced Machine Intelligence) in Paris. He’s said publicly that the entire AI industry has become “LLM-pilled,” with every major lab chasing the same scaling approach because none of them can afford to fall behind by trying something different. AMI Labs raised $1.03 billion in its first funding round, backed by Jeff Bezos’s investment firm, Nvidia, Samsung, and Toyota, making it one of the largest first-round raises in European AI history. ## Why does LeCun think scaling language models has a ceiling? LeCun’s argument rests on a comparison of how much information goes into training large language models versus how much information a human child absorbs by age four. A typical large language model is trained on something like 20 trillion tokens. Each token takes roughly three bytes to represent, which works out to around 10^14 bytes total, the entirety of the publicly available text on the internet. LeCun has pointed out it would take a person several hundred thousand years to read through that much material. Compare that to a four-year-old child. In their first four years of life, a child is awake for roughly 16,000 hours, and visual information reaches the brain through the optic nerve at about 2 megabytes per second. Run that calculation and you land at roughly the same figure: about 10^14 bytes. The child has absorbed a comparable amount of raw information to the biggest LLMs, but through direct experience of the physical world rather than text. And that child understands things no LLM reliably does: that objects persist when hidden, that unsupported things fall, that pushing a glass off a table breaks it. LeCun’s conclusion is that language models are excellent in domains where language itself is the substrate of reasoning, like math and code, but struggle with the physical world because it’s continuous, messy, and not made of words. This is also LeCun’s explanation for why today’s models are so enormous: they’re compensating for a lack of real understanding by accumulating vast amounts of declarative knowledge, essentially memorizing facts, which requires huge parameter counts and memory. ## How does JEPA work differently from a standard LLM? *Nothing gets lost.* **312 apps** across 41 teams · all in one place Most generative AI systems, including LLMs and video models, are trained to predict raw output in full detail: the exact next word, or the exact next pixel in a video frame. LeCun argues this is fundamentally wasteful because a lot of that detail is simply unpredictable. If you rotate a camera around a room and ask a model to predict what comes next, it can reasonably guess the room has walls and maybe windows, but it has no way of knowing who’s sitting in which chair. Training a system to nail details like that “kills it,” in LeCun’s words, by forcing it to waste capacity on noise. JEPA (joint embedding predictive architecture) takes a different approach. Instead of predicting raw pixels or exact words, it first encodes what it observes into an abstract representation, a kind of internal summary, and then makes predictions within that representation space. It can discard information that’s unpredictable or irrelevant and focus on what actually matters for understanding the scene. One risk with this approach is something called “collapse,” where the model cheats by producing the same generic summary for everything, making its predictions technically always correct but practically useless. LeCun’s team uses specific training techniques, referred to in the paper by names like SigReg and VCReg, to stop the model from taking that shortcut and force it to learn genuinely distinct, meaningful representations. ## What is the platonic representation hypothesis, and what did this paper prove about it? In 2024, MIT researchers observed something unusual: as AI models got more powerful, models trained on completely different types of data, images versus text, started organizing information internally in increasingly similar ways. It was as if every sufficiently capable model was converging on the same underlying map of reality, despite learning from entirely different sensory inputs. Researchers named this the platonic representation hypothesis, after Plato’s idea that what we perceive are shadows of a single deeper reality. Until this paper, nobody had mathematically shown why this convergence happens or how to engineer it deliberately. LeCun and Balestriero’s paper demonstrates that under specific conditions, JEPA models trained on different kinds of data will learn representations that are mathematically equivalent, essentially the same map, just rotated. The key condition is what they call “good views”: the model needs to learn by comparing two different observations of the same real event, like a photo of a dog catching a Frisbee alongside a sentence describing that same scene. Because both observations point to the same underlying reality, the model is pushed toward learning what they share in common. A “bad view” is a comparison where the two things don’t actually share the same underlying reality. Notably, the paper argues that next-word prediction, the exact training method behind ChatGPT, Claude, and Gemini, counts as a relatively bad view. The next few words in a sentence often reveal little about the actual state of the world, which may explain why large language models align only weakly with how image-based models represent reality, even at massive scale. ## Is JEPA actually better than large language models? On one specific benchmark testing world understanding, a JEPA-based model with roughly 109 million parameters reportedly outperformed language models as large as 72 billion parameters, a size difference of more than 600 times. That’s a striking result, but it’s a single benchmark, not a sweeping victory across every task LLMs are used for today. Language models remain dramatically better at things like writing, coding, and answering open-ended questions, tasks where language itself is the medium of reasoning. *pass the audit.* What the result does challenge is the industry-wide assumption that making text-based models bigger is the most direct route to broader intelligence. If JEPA-style architectures scale the way LeCun’s team expects, smaller, differently-trained systems could rival or exceed much larger LLMs on tasks tied to physical and spatial understanding, the exact areas where current chatbots and even advanced robots and self-driving systems still struggle. ## Frequently Asked Questions ### What does JEPA stand for? JEPA stands for joint embedding predictive architecture. It’s an AI design proposed by Yann LeCun that predicts outcomes in an abstract representation space rather than predicting raw pixels or exact words. ### What is the platonic representation hypothesis? It’s the idea, first proposed by MIT researchers in 2024, that sufficiently powerful AI models trained on different types of data (like text and images) tend to converge on similar internal representations of reality, as if approaching the same underlying “map” of the world from different angles. ### Did a small model really beat much larger language models? According to the paper, a JEPA-based model with about 109 million parameters outperformed language models as large as 72 billion parameters on a benchmark testing world understanding. This applies to that specific test, not to language tasks generally. ### Why does Yann LeCun think language models won’t reach human-level AI? He argues that language models are trained almost entirely on text, which is a limited and selectively curated form of information. He compares the total text used to train large LLMs to the amount of visual information a child absorbs by age four, and argues the child develops a much deeper understanding of physical reality despite processing a comparable amount of raw data. ### Is Yann LeCun’s company working on JEPA directly? LeCun left Meta to found AMI Labs in Paris, a company focused on what he calls advanced machine intelligence, built around world models rather than scaled-up language models. The company raised $1.03 billion in its first funding round from backers including Jeff Bezos’s investment firm, Nvidia, Samsung, and Toyota.