Yann LeCun Says Humanoid Robot Demos Are Precomputed Lies

Okay so apparently we've all been watching expensive theater and calling it robotics progress.

Yann LeCun—the guy who literally won a Turing Award for his work on deep learning—just went on record saying that every impressive humanoid robot demo you've seen is basically precomputed choreography. Not autonomous intelligence. Not real-world problem solving. Just very expensive kung fu routines.

"There's a lot of companies building humanoid robots and they do those kinds of, you know, they play kung fu and, you know, impressive thing. This is all precomputed," LeCun said in a recent interview that's been making the rounds. "None of those company—absolutely none of them—has any idea how to make those robots smart enough to be useful."

That's... a pretty sweeping statement. Especially coming from someone currently launching his own AI lab to solve exactly this problem.

The Part Nobody Talks About

Here's the thing that gets me: we all know those Unitree demos look incredible. The ones where the robot does backflips or navigates complex terrain or manipulates objects with what looks like genuine understanding. But according to LeCun, these are carefully choreographed sequences—the robot can judge its environment and move around autonomously, sure, but it's not actually reasoning about what it's doing.

And apparently this extends way further than most people realize. Boston Dynamics unveiled their production-ready Atlas robot at CES 2026—one of the biggest tech showcases of the year. That demo? Teleoperated. A human was controlling it remotely.

Now, to be fair, that makes sense from a risk management perspective. You're putting your flagship product on stage in front of thousands of people and live cameras. Of course you're going to reduce variables. But it does underscore LeCun's broader point: the gap between impressive hardware and actual intelligence is much wider than the marketing suggests.

The robotics companies aren't really hiding this, exactly. Boston Dynamics has Atlas slated for basic part-sorting operations in 2028 and more complex tasks in 2030. That's a pretty conservative timeline for a robot that can apparently do parkour. The timeline itself is telling you: we have the mechanics down, but the intelligence part is... ongoing.

The Real Disagreement

So naturally, Elon Musk weighed in—because of course he did. He's got Tesla's Optimus project, and his take was basically: "He thinks if he can't do it, then no one can."

LeCun's response was immediate and, honestly, pretty bold: "That's actually quite the opposite. I know I can do it and I know how to do it... it's not with the current techniques that everyone is betting on."

Okay, so what is LeCun betting on? Something called V-JEPA (Video Joint Embedding Predictive Architecture), which is his proposed solution to the whole "robots don't actually understand reality" problem.

Here's the core difference: current AI approaches train robots by showing them thousands of examples. Pour water into a cup 10,000 times, and eventually the robot learns to pour water into that specific type of cup. Then you need another 10,000 examples for a different cup shape. And another 10,000 for pouring juice instead of water.

V-JEPA tries to teach concepts instead of patterns. Show the AI a video with chunks masked out, and instead of trying to predict every pixel in the missing section, it learns the underlying physics: liquids flow downward, containers fill from bottom to top, momentum carries objects in arcs. The idea is that if you understand principles, you don't need 10,000 examples of every possible variation.

"The problem is that the approaches that have been successful for language do not work for high-dimensional continuous noisy data," LeCun explained. "You have to use something else."

Where This Gets Interesting

The pushback on LeCun has been... intense. One researcher basically said this is the kind of thing that makes people in the field dislike him—the implication that everyone else is doing it wrong while he alone has the answer.

But there's a counterargument that's actually pretty compelling: when a 17-year-old learns to drive, they're not starting from zero. They come preloaded with 17 years of embodied experience—intuitive physics, spatial reasoning, object permanence—plus millions of years of evolutionary optimization for visual processing and motor control. They're not learning to drive in 20 hours; they're fine-tuning on a massive foundation of world understanding.

Which, interestingly, kind of supports LeCun's point? Both sides agree you need a foundational world model. The debate is about how you build it: through massive amounts of demonstration data (the industry's current bet) or through explicit world-modeling architecture (LeCun's approach).

As one critic put it: "You don't need AGI. You need human sample efficiency." Can you get a robot to learn a new task in 10 examples instead of 10,000? That's the actual benchmark.

LeCun's response to his critics was characteristically blunt: "You do realize that I've been advocating for end-to-end self-supervised training of world models and planning for about 10 years... and I've made a lot of progress over the last five... and actually made it work for simple robotics tasks over the last two."

He's just started a new company to scale this approach. Which means we're going to get an actual test case: can explicit world modeling beat the industry's data-scaling strategy?

What We're Actually Watching

Here's what I find fascinating about this whole thing: nobody's really arguing that current humanoid robots have general intelligence. Even the companies building them aren't making that claim. The disagreement is about path—do you get there by scaling up current methods or do you need fundamentally different architecture?

LeCun's position is that most robotics companies are optimizing hardware while ignoring the harder problem. "A few robotics companies are working on world models and planning," he noted, "but the vast majority are using LLM-derived methods like VLA or diffusion policies with RL fine-tuning in simulation. Those are good for narrow tasks, but the companies building humanoid hardware don't tend to be working on innovative robotics AI."

Translation: they're making robots that look cool and can execute predefined sequences, but they're not working on the intelligence layer that would make those robots actually useful in unpredictable real-world scenarios.

Which brings us back to the uncomfortable reality: your housecat has better common sense than any humanoid robot currently in existence. And LeCun's claim is that we can't just scale our way past that limitation—we need a different approach entirely.

Whether he's right is going to become very clear very quickly. He's got funding, he's attracting talent, and he's making falsifiable predictions. In a few years, we'll know if V-JEPA-style world modeling actually delivers on its promise or if the industry's "just add more data" approach was correct all along.

Either way, at least someone finally said the quiet part out loud: those demos you've been watching? They're not showing you the future of robotics. They're showing you very expensive proof that we've nailed the hardware and are still figuring out the intelligence part.

—Zara Chen, Tech & Politics Correspondent