The AI Spotlight Follow
A 109-Million-Parameter AI Just Beat a 72-Billion-Parameter Giant on One Key Test. Yann LeCun Says That's the Whole Point
Model Wars

A 109-Million-Parameter AI Just Beat a 72-Billion-Parameter Giant on One Key Test. Yann LeCun Says That's the Whole Point

For years Yann LeCun has said bigger chatbots won't reach human-level intelligence. His new paper with Randall Balestriero shows a tiny JEPA model lining up with how vision models see the world better than 70-billion-parameter language models, using a fraction of the data. Here's what it proves, and what it doesn't.

Advertisement

Contents

The contrarian with a billion dollars

While the biggest AI companies spent hundreds of billions of dollars making large language models bigger, one of the founders of modern AI kept repeating the same warning.

Scaling up chatbots will not get us to human-level intelligence.

Yann LeCun has said it in lectures, on panels and on social media, often bluntly. He has dismissed the near-term idea of a "country of geniuses in a data center" as nonsense.

Now he has published one of his strongest pieces of evidence yet. In a paper posted on October 29, 2026, LeCun and Randall Balestriero showed a text model with 109 million parameters lining up with how vision models see the world better than language models up to 72 billion parameters, on one important benchmark. The small model is roughly 660 times smaller.

Advertisement

That doesn't mean small models now beat ChatGPT. It means something subtler, and possibly more important. Here's the full story.

Who Yann LeCun is

LeCun is not just another opinion online.

  • In the late 1980s at Bell Labs, he helped pioneer convolutional neural networks, a kind of AI that is very good at understanding images. An early use was reading handwritten digits, and systems based on that work went on to read a large share of U.S. bank checks.
  • That idea still runs underneath face recognition, medical imaging and the cameras in self-driving systems.
  • He shared the Turing Award, computing's equivalent of the Nobel Prize, with Geoffrey Hinton and Yoshua Bengio. The three are often called the "godfathers of AI."
  • In 2013 he founded Facebook's AI research lab, and spent more than a decade as Meta's chief AI scientist while teaching at New York University.

Then, in late 2025, he did something unusual. He left Meta to start AMI Labs (Advanced Machine Intelligence), headquartered in Paris.

His explanation has been consistent. Silicon Valley, he says, has become "LLM-pilled." Everyone is digging the same trench and poaching each other's engineers, and no one can afford to try something different without falling behind. He has said Meta's own reshuffling toward LLMs was one reason he left.

In March 2026, AMI Labs raised $1.03 billion at a $3.5 billion pre-money valuation, one of Europe's largest initial rounds. Co-leads included Bezos Expeditions, and backers included Nvidia, Samsung and Toyota Ventures. LeCun is chairman; Alexandre LeBrun is CEO. The lab has teams in Paris, New York, Montreal and Singapore.

So this is not one professor shouting from the sidelines. It is one of the people who built modern AI, with a billion dollars, betting the industry took a wrong turn.

The mistake he says everyone is making

Large language models learn mostly by reading enormous amounts of text and predicting the next word. That turns out to be incredibly powerful. It's why chatbots can write essays, code and answer almost anything.

LeCun's argument is that text has a ceiling. He points to a strange gap:

  • AI can pass the bar exam, but we still don't have a robot that can tidy your house.
  • AI can solve calculus problems, but no self-driving car learns to drive in 20 hours of practice like a teenager.

His explanation: language models are brilliant where language itself is the substrate of reasoning, as in math and code. The physical world isn't made of words. It is messy, continuous and unpredictable, and that's where these systems struggle.

He also argues that LLMs are enormous because they accumulate declarative knowledge, huge piles of facts. That's why they need so many parameters and so much memory. His alternative is to replace accumulation of knowledge with understanding.

The four-year-old versus the internet

LeCun's favorite back-of-the-envelope calculation is one of the clearest ideas in AI.

The text side:

  • A large language model trains on roughly 20 to 30 trillion tokens.
  • Each token takes around three bytes.
  • That's about 10^14 bytes, close to all the public text on the internet. A human would need hundreds of thousands of years to read it.

The child side:

  • A four-year-old has been awake about 16,000 hours.
  • The optic nerve carries roughly 2 megabytes per second into the brain.
  • Multiply it out and you get about 10^14 bytes again.

In other words, a small child has taken in about as much raw data through vision as the biggest language models get from all of the internet's text.

But the child learned things those models never fully grasp: dropped objects fall, a ball rolling behind a sofa still exists, a glass pushed off a table will probably break. All without reading a single word.

His conclusion: we won't reach human-level AI by training on text alone. Machines need to understand the real world.

World models and JEPA, in plain English

LeCun's alternative is a world model: an AI that builds an internal understanding of how the world works so it can predict what happens next.

Think about catching a ball. You don't calculate every gust of wind or every stitch on the ball. Your brain predicts roughly where it will be.

The problem with predicting pixels. Many AI systems try to learn about the world by predicting the next frame of a video, pixel by pixel. LeCun's classic example: film a lecture hall while slowly panning the camera, then ask the system to continue the video. It can guess there are more seats and walls. It cannot possibly know what every person in the room looks like. Forcing a model to predict unpredictable details wastes its effort and, in his words, kills it.

JEPA's fix. His architecture is the Joint Embedding Predictive Architecture (JEPA), which he proposed in 2022:

  • Instead of reconstructing raw details, JEPA first turns each observation into an abstract summary, called an embedding.
  • It then makes predictions in that summary space.
  • Unpredictable details can be thrown away. The model focuses on what matters.

An analogy: two people describe the same football match. One lists every blade of grass and every fan's shirt. The other says, "The striker is breaking toward goal and the defender is about to tackle." The second description is less detailed and far more useful for predicting what happens next. JEPA aims for that second kind of understanding.

The catch: collapse. A JEPA can cheat by giving the same summary for everything, which makes its predictions trivially "correct." Researchers prevent this with anti-collapse training rules, with names like VICReg and SIGReg, that force the model to learn something meaningful.

The Platonic idea behind the paper

The new paper is titled "Proving and Improving the Platonic Representation Hypothesis with JEPAs," by Balestriero (Brown University and AMI Labs) and LeCun (NYU and AMI Labs).

The background comes from a 2024 observation by MIT researchers. As AI models got more powerful, models trained only on images and models trained only on text started organizing information internally in increasingly similar ways, as if they were all converging on the same underlying map of reality.

They named it after Plato's cave: different models see different shadows, but all are learning the same reality behind them.

What nobody had shown was why it happens, or how to make it happen on purpose.

Good views and bad views

Balestriero and LeCun turned the hypothesis into what they call the Platonic Representation Theorem.

In plain terms: under idealized conditions, JEPAs trained separately on completely different kinds of data (text, images, video, audio, even actions) learn the same representation of the world, just rotated. Like the same map turned to a different angle.

The key condition is good views.

  • A good view is when the model learns by comparing two different observations of the same moment in the world. A photo of a dog catching a Frisbee and a sentence describing that exact scene. Two photos of the same room. Because both point to the same piece of reality, the model is pushed to learn what they share: the actual state of the world.
  • A bad view is when the two things being compared are just corruptions or fragments of one observation, so they don't truly share a world state. The model ends up spending capacity on quirks of its own data type that no rotation can line up.

Then the provocative part. The paper argues that next-word prediction, the way chatbots learn, is itself a bad-view JEPA. A prefix and its continuation come from the same document, but the next few words often say little about the actual state of the world. That, the authors suggest, is why language models line up with vision only partly, and mostly through brute-force scale.

The results: 109 million versus 72 billion

The team tested how closely text models' internal "maps" matched those of image models, using the benchmark from the original Platonic hypothesis paper.

The headline numbers:

Model Parameters Alignment score
Good-view text JEPA 109 million 0.184
Qwen2.5-72B 72 billion 0.154
Llama-3-70B 70 billion 0.151
CLIP text encoder — 0.124

The small JEPA beat every language model they tested. The paper notes it read 116 billion text tokens, while those big models trained on 15 to 18 trillion. That is a model about 660 times smaller trained on over 100 times less text.

Efficiency:

  • On the same corpus, a good-view JEPA matched next-token prediction's alignment using 21.6 million samples instead of 900 million, about 42 times fewer samples, or around 14 times fewer text tokens.
  • Even a 33-million-parameter JEPA trained on texts from 50 million web images matched Qwen2.5's 3B and 7B models on the benchmark.

A scaling law. Alignment followed a predictable curve in data and the share of good views, accurate enough to predict 85 new training runs before they were trained.

Plug-and-play captions. Because the text and image models learned the same map, rotated, the researchers connected them with a single rotation. That alone produced a system that can describe images, without the usual heavy paired training. They report it beat every pair-free zero-shot classification result reported with language models up to 8 billion parameters.

LeCun summed it up on X: language models accumulate knowledge, which is why they need so much memory; JEPA world models replace that accumulation with understanding.

What the paper does not prove

The headlines got carried away, so be precise:

  • It won on one measure. The 109M model scored higher on how closely its internal representation aligns with image models. It was not tested on math, coding, writing or general knowledge. It is not a chatbot.
  • It isn't "the end of scale." As commentators pointed out, beating a 72B model on one benchmark is a much narrower claim. Big models can win on breadth while better architectures win on understanding.
  • The theory is idealized. The authors themselves list limits: the theorems describe a finite world with perfect conditions, and real data is lossy.
  • The good views came partly from AI. This is the paper's biggest caveat. On its largest dataset, two of each image's three captions were written by an image-captioning model (LLaVA on Llama 3 8B, using a CLIP vision encoder), which means some image-text knowledge was distilled in. Without generated captions, alignment dropped substantially. The authors note the comparisons between views on identical texts don't depend on this, but the headline comparison with big LLMs does.
  • It hasn't been independently reproduced yet.

Critics have also asked whether this undercuts LeCun's own argument: if language models learn world representations too, why exclude them? The fair reading of the paper is that they do learn something about the world, just far less efficiently.

Others say JEPA still lacks practical use cases. LeCun's reply to one such critic was characteristically blunt: humans mostly don't think in words, animals certainly don't, yet they think.

What happens if it scales

The AI industry's main strategy is scale: more data, more chips, more data centers, more money. If that's the only path, whoever has the most cash wins.

This paper suggests another path. If you can get the same understanding with dozens of times less data and a model hundreds of times smaller, the economics change:

  • AI that understands the world could run on a phone, glasses or a robot, not just in a giant data center.
  • Training costs could fall, letting universities, smaller labs and countries without trillion-dollar budgets build serious systems.

There are already early hints. In March 2026, researchers including LeCun and Balestriero (with collaborators at Mila, NYU and Samsung) released LeWorldModel, a JEPA world model with about 15 million parameters trained from raw pixels on a single GPU in a few hours. Its compact representation let it plan up to 48 times faster than a larger foundation-model-based world model.

Robots are the biggest prize. LeCun often says we don't have robots anywhere near as smart as a house cat. A cat judges a jump onto a counter perfectly and plans routes through a room, with a fraction of a human's neurons and without reading a word. If world models can learn the way animals do, by watching and predicting, that could be the missing piece for robots, self-driving cars and assistants that work in the physical world.

Three more implications:

  1. The data wall. AI companies have used much of the internet's quality text. Video, images and sensor data are nearly unlimited.
  2. The bubble debate. LeCun has said AI is not a bubble in the sense that there's plenty of useful work for LLMs, and data center investment won't be wasted as billions use AI. The bubble, in his view, is the belief that pushing the current approach will reach human-level intelligence.
  3. Modular AI. If separately trained models naturally share one map, future systems might be assembled from parts that fit together, rather than one giant model trained on everything.

LeCun is not alone. Google DeepMind's Genie world models, Fei-Fei Li's World Labs, Nvidia's robotics world models and Meta's continuing JEPA research all point the same way. The world-model race is real and moving fast.

Is this a path to AGI?

LeCun has named four things he thinks matter most for real intelligence:

  1. Understanding the physical world
  2. Persistent memory
  3. Reasoning
  4. Planning

He has also said the ideas he finds most exciting look today like obscure academic papers that people will get excited about years from now. This is one of those papers.

On its own, it is not a path to AGI. It shows how to make AI understand the world more efficiently. It does not solve memory, reasoning or planning. But combine efficient world understanding with those abilities, at scale, and you get something closer to how humans and animals think than today's chatbots.

LeCun also points to a 70-year mistake in AI: assuming that if a machine does something we find hard, like chess or calculus, it must be intelligent. The things we find easy, like walking across a room or picking up a cup, turned out to be the hardest for machines. That's exactly the kind of intelligence he's chasing.

Why it could still break

This is what separates real excitement from hype.

  • The bitter lesson. Rich Sutton's famous essay argues that, again and again, methods that best exploit more compute beat clever designs. LLMs are the ultimate example. JEPA must prove it keeps its advantage at much larger scale.
  • Brittleness. World models have shown sensitivity to small visual changes, like an object's color or background.
  • Research, not product. AMI says its results are research results. Its CEO has said it could take years to reach commercial use. LeCun expects early applications in industry, such as digital twins for power plants, jet engines and chemical plants, long before home robots.
  • Absorption. Big LLM labs could simply borrow the good ideas and fold them into their own models. That happens constantly in AI.

What to watch: whether others reproduce the results, what happens when AMI trains much bigger JEPAs, whether the approach shows up in real robots, and whether the big labs quietly adopt it. If they do, that would be the clearest sign LeCun was right.

Why this matters

For years, the AI debate has been framed as "scale versus skeptics." This paper reframes it.

It suggests how a model learns may matter as much as how big it is, and it offers a mathematical explanation, not just a benchmark win. If that holds, the next leap in AI might not come from the company with the largest data center. It might come from the team with the best idea.

LeCun's lab has been moving fast: the simpler LeJEPA training method in November 2025, AMI's founding, the billion-dollar raise in March, LeWorldModel with academic collaborators, and now this theorem, from a team small enough to fit in a few offices across four cities.

FAQ

Did a 109M model beat a 72B model at everything?

No. It beat Qwen2.5-72B and Llama-3-70B on one benchmark measuring how closely a text model's internal representation aligns with image models. It was not tested as a chatbot.

What is JEPA?

Joint Embedding Predictive Architecture, proposed by Yann LeCun. It predicts in an abstract representation space instead of reconstructing every pixel or word, so it can ignore unpredictable details.

What are "good views"?

Training pairs that are two different observations of the same real-world moment, such as a photo and a caption of that same scene. The paper proves these lead separately trained models to the same representation.

Why does the paper call next-word prediction a "bad view"?

Because a text prefix and its continuation come from one document and often reveal little about the actual state of the world, so the model spends capacity on language quirks rather than shared reality.

What is AMI Labs?

LeCun's Paris-headquartered world-model startup, which raised $1.03 billion in March 2026 and publishes its research openly.

The bottom line

Yann LeCun hasn't proven the entire AI industry wrong overnight. Language models are extremely useful and will keep improving.

But he and Balestriero have put a precise finger on what may be the industry's biggest blind spot: the assumption that the only way to make AI understand the world is to make it bigger and feed it more text.

LeCun was right once before, when almost nobody believed in neural networks. If good views scale, he may be right again.

Sources

Nvidia Just Lost Jensen Huang. The Bigger Fight Is Whether AI Should Be Free, and China Is Betting It Will BeUp next · 12 min readNvidia Just Lost Jensen Huang. The Bigger Fight Is Whether AI Should Be Free, and China Is Betting It Will BeMonths after Jensen Huang became the loudest voice against banning open AI models, a spy scandal ended his run as Nvidia's CEO. Behind the headlines is a bigger fight over whether AI becomes an expensive product or a free utility, and China has picked a side.Keep reading →