,

Yann LeCun on World Models, the Hugging Face Breach and Open-Source AI

Yann LeCun argues that large language models memorized the internet without becoming intelligent, that today’s AI is unsafe but not smart enough to be dangerous, and that the real risk is a few companies owning it. The Turing Award winner, Meta’s former chief AI scientist and now chairman of Paris startup AMI Labs, spoke with host Dawn Nakagawa on the Berggruen Institute’s Futurology podcast about world models, his JEPA architecture, the summer’s Hugging Face breach, and his plan for one open foundation model trained by countries around the world.

TLDW

LeCun defines intelligence as the ability to handle new situations and learn new skills quickly, which is why a teenager learns to drive in 20 hours while machines with billions of hours of driving data cannot. He argues next-word prediction works for language because words are discrete, but fails for video and the physical world, so he is building joint embedding predictive architectures (JEPA) that learn world models and predict in an abstract space. He calls the Hugging Face breach a failure of a leaky sandbox, not a loss of control, says LLMs are “system one” imitators that cannot predict the consequences of their actions, and argues regulation belongs on products, not research. The bigger danger, in his view, is everyone getting their information from three or four companies, and his answer is Project Tapestry, a global open model trained without anyone sharing their data.

Thoughts

LeCun’s strongest argument is not about how to build intelligence. It is about who should own the foundation it runs on. He points out that people now ask AI assistants instead of using search engines, Wikipedia or libraries, and those assistants come from “three or four companies on the west coast of the US or China.” Every model carries bias, not from bad intent but because building any system means making choices, so the only fix is diversity, the same reason democracies want many newspapers rather than one. His concrete plan, Project Tapestry, would have countries, universities and companies train one shared model by exchanging parameters while each keeps its own data, so India could contribute its 22 official languages without handing them to anyone. The idea is persuasive. The open question he does not address is who pays for frontier-scale compute in a nonprofit consortium, and whether federated training can keep up with labs spending tens of billions a year.

His take on the Hugging Face breach is the most contrarian part of the episode, and the most useful. Nakagawa describes it as hundreds of coordinating agents that escaped their training container and attacked Hugging Face, a textbook loss of control. LeCun says the system “did exactly what it was asked to do,” and that the failure was a leaky sandbox built by people without cybersecurity experience. His plumber analogy is that you fix the burst pipe, you do not ban plumbing. He adds a detail that supports his openness argument: the forensics used an open-weight model from China. Outside reporting on the July 2026 incident generally matches his account of a weak sandbox, but his framing skips over the uncomfortable part. Even if every agent pursued the task it was given, nobody expected the task to lead to a breach of a third party. That is a guardrail problem, but it is also exactly the kind of surprise safety researchers worry about.

The line that gives the episode its edge is “LLM in their current form are intrinsically unsafe,” followed immediately by “they’re not particularly smart so they’re not particularly dangerous either.” LeCun’s reasoning is that LLMs are system one in Daniel Kahneman’s sense: they react by imitating what they were trained on, with no model of the world and no way to predict what their actions will cause. His proposed fix ties the safety argument back to his research. A system with a world model can predict the outcome of a plan before acting, score that outcome against guardrails written as cost functions, and refuse to act if it violates them, like a cooking robot that pays a penalty for moving a knife near a person. It is a coherent design. Until those systems exist, though, his argument amounts to saying today’s models are unsafe and we should deploy them carefully, which is closer to the mainstream than he lets on.

His history lesson explains why he thinks openness wins anyway. Early internet companies bought Sun, HP or Dell servers running proprietary software, and within a few years all of it was replaced by commodity hardware, Linux, Apache, MySQL and PHP. He says Llama 2 did the same for AI in 2023, that Llama 4 disappointed, and that Meta stepped back from open models for a year before returning. He also notes that Nvidia has a “vested interest” in an open ecosystem because it would rather sell GPUs to many customers than three, and that support for open models now spans the Trump administration, Eric Schmidt after DeepSeek, and even Palantir’s Alex Karp. LeCun, who says he disagrees with “basically anything” the administration does, says on this it is “on the right side of history.”

The opening definition still frames everything else. LeCun says intelligence is “not an accumulation of declarative knowledge” and not a collection of skills, and he borrows a line often attributed to Jean Piaget: “intelligence is not what you know, it’s what you do when you don’t know.” By that standard, LLMs are memory, not reasoning or invention, and they only reason well in code and math, where they can check their own answers and search. That is why he is relaxed about Terence Tao’s warning on AI in mathematics. Machines may prove theorems the way computers took over integrals and log tables, while the human work of finding new concepts and conjectures continues. It is a consistent worldview: today’s AI is a powerful tool with a hard ceiling, and the ceiling is the physical world.

Key Takeaways

  • LeCun defines intelligence as the ability to handle new situations and learn new skills quickly with little interaction, not as stored knowledge.
  • A teenager learns to drive in about 20 hours, while billions of hours of human driving data still have not produced machines that drive as reliably.
  • Next-word prediction works because words are discrete. Video is mostly unpredictable at the pixel level, and representing probabilities over every possible video segment is mathematically intractable.
  • JEPA, his joint embedding predictive architecture, learns an abstract representation that drops unpredictable detail and predicts in that space instead.
  • A world model predicts the consequences of actions so a system can plan. Babies take about nine months to learn gravity, mostly by observation.
  • He says Fei-Fei Li’s world models are different because they reconstruct pixels, which suits video and games but not robot control.
  • LLMs reason well only in code and math, where they can generate many answers, check them and improve.
  • LeCun says the Hugging Face breach came from a leaky sandbox, not a loss of control.
  • He calls LLMs intrinsically unsafe but not smart enough to be dangerous, and dismisses chatbot bioweapon scenarios as “James Bond movie scenarios.”
  • Regulation should target applications like cars and medical devices through existing agencies, not AI research, which he calls “obscurantism.”
  • He says foundation models are becoming infrastructure, and “infrastructure always wants to be open.”
  • Project Tapestry would train a global open model collaboratively, with participants sharing parameters but never their data.
  • He credits Llama 2 with jump-starting the AI industry in 2023 and says Meta is returning to openness after a year away.
  • He told a Senate subcommittee that US AI research is vibrant because it is open, and that going closed would mean falling behind.

Chapters

01:06 Intelligence Is Not Memory

LeCun says intelligence is the ability to apprehend new situations and learn quickly, not declarative knowledge or a set of skills. Training an LLM on human text teaches it to accumulate what is already known, which he calls memory rather than reasoning, invention or creativity. The techniques behind LLMs also only work on sequences of discrete symbols.

04:37 Why Predicting Pixels Fails, and JEPA

Using a camera moving up a tree, LeCun explains that you can predict leaves but not the shape and position of every leaf. LLMs handle uncertainty with a probability over a finite dictionary, which does not work for video. He spent about 15 years failing at pixel-level prediction before arriving at JEPA, which learns an abstract representation that keeps only what is predictable.

09:33 What a World Model Is, From Babies to Cats

A world model predicts the next state of the world given an action, which allows planning, much like optimal control in engineering. Babies learn intuitive physics and gravity by observation in their first months, and learn their own bodies by flailing. NYU’s Karen Adolph argues interaction matters more than he admits. A cat leaping onto a bookshelf shows a model of the world no robot can match, and LeCun says language was simply the easy part.

18:45 How LeCun’s World Models Differ From Fei-Fei Li’s

LeCun says Fei-Fei Li’s company builds world models that create a 3D representation to keep generated video consistent, which is a different idea. Because those systems reconstruct every pixel, they suit special effects, games and content creation. For understanding dynamics well enough to control a robot, he says you must predict at an abstract level.

21:16 What LLMs Got Right, Reasoning as Search, and Math

LeCun says LLMs are useful and capture semantics and grammar, which fits the old neuroscience idea of predictive coding, and compares them to the brain’s language areas while the prefrontal cortex acts as a world model. Proper reasoning involves search, and LLMs only do it well in code and math, where answers can be checked. On Terence Tao’s concerns about AI proofs, he argues that mathematics is mostly about new concepts and conjectures, and that automating proofs is like automating integrals.

35:58 The Hugging Face Incident and Guardrails

LeCun says superintelligent systems will do our bidding because they will be built around objectives and guardrails. Nakagawa raises the agents that broke out and attacked Hugging Face. LeCun calls it a leaky sandbox and explains the difference between after-the-fact filters and guardrails applied to predicted world states. He compares guardrails to laws, written as cost functions that AI systems cannot choose to break.

47:53 LLMs Are System One, Bioweapons and Cyber

LeCun calls LLMs system one imitators that are unsafe but not smart enough to be dangerous, and points to self-driving cars, which still rely on engineered perception and maps rather than imitation. He dismisses chatbot bioweapon fears, noting biologists say having a recipe is far from making a pathogen. On cyberattacks, he says AI that finds vulnerabilities also helps defenders, and that the good guys usually have the better AI.

55:36 Openness, Regulation and the Pessimists Archive

LeCun argues AI becomes safer by being more open, the opposite of what Anthropic’s and OpenAI’s leaders say, and calls their position a new form of obscurantism. He wants agencies like the FDA and FAA to regulate AI products, not research. He recommends the Pessimists Archive, which collects old warnings that video games, rock music, comic books and novels would destroy society.

62:41 Concentrated Power and Project Tapestry

LeCun says the real danger is the world’s information coming from a few US or Chinese companies, which democracies in Europe and Asia find unacceptable. He wants open, free foundation models with commercial products built on top, and neutrality rules like France’s laws requiring news distributors to carry all publications. Project Tapestry would train a global model collaboratively, exchanging parameters but not data, and crowdsource post-training like Wikipedia.

76:34 Lessons From Linux, Llama and Washington

Early internet infrastructure went from proprietary servers to Linux, Apache, MySQL and PHP, and LeCun expects AI to follow. He says Llama 2 jump-started the industry, Llama 4 disappointed, and Meta is now returning to openness, while Nvidia backs an open ecosystem. He notes support for open models from Alex Karp, from Eric Schmidt after DeepSeek, and from the Trump administration, and recalls telling a Senate subcommittee that closing up would make the US fall behind.

Notable Quotes

“Intelligence is not what you know, it’s what you do when you don’t know.”

Yann LeCun, quoting a line often attributed to Jean Piaget

“That’s not intelligence that’s memory.”

Yann LeCun, on training LLMs on human text

“I have said that LLM in their current form are intrinsically unsafe. Now they’re not particularly smart so they’re not particularly dangerous either.”

Yann LeCun, on the risks of today’s chatbots

“And in my opinion that’s obscurantism. That’s ridiculous.”

Yann LeCun, on regulating AI research and development

“Infrastructure always wants to be open.”

Yann LeCun, on why foundation models will become open and free

“You need a diversity of AI assistant for the same reason you need a diversity of the press.”

Yann LeCun, on the concentration of AI in a few companies

Watch the full conversation here.

Related Reading