Yann LeCun Calls Takeover Fears 'Science Fiction', Says Language Models Hit a Dead End
Yann LeCun Calls Takeover Fears 'Science Fiction', Says Language Models Hit a Dead End
By Daniel Rao, TechPulse USA — October 13, 2026
While the rest of the AI industry is pouring hundreds of billions of dollars into making chatbots bigger, the man who taught machines to see stood in a Seoul hotel on Monday and called the whole strategy a dead end. Yann LeCun — 2018 Turing Award winner, the researcher whose convolutional neural networks made modern computer vision possible — told reporters that artificial intelligence's next breakthrough will come not from ever-larger language models, but from "world models": systems that understand, predict and act in the physical world. And the industry's favourite nightmare, the AI that suddenly spirals out of control into superintelligence? A science-fiction narrative, he said, not an engineering forecast.
It was the kind of blunt, contrarian performance LeCun has been giving for the better part of a decade. What made this one matter is the timing: days earlier, his collaborators had published a new paper showing their world-model architecture learning to plan in simulated environments at dramatically higher success rates — and the next day, the industrial AI conference his panel was fronting for opened in Tokyo. The world's most famous AI sceptic of scaling isn't just arguing anymore. He's publishing.
The next breakthrough is not a bigger chatbot
LeCun's argument, delivered at the Global AI Leaders Seoul Dialogue at the Westin Josun in Jung District, was characteristically direct. Today's large language models are extremely good at capturing knowledge contained in text, images and video, he said — but they lack a reliable understanding of cause and effect in the physical world. They predict the next word; they don't model what happens when you act.
His alternative, which he has championed since his years running Meta's fundamental AI research lab, is the joint-embedding predictive architecture — JEPA for short. Instead of reconstructing every pixel or predicting every word, a JEPA system learns a compressed internal representation of the world and predicts how that representation changes in response to an action. Such a system could plan a sequence of actions needed to achieve a goal rather than imitate patterns in human-generated data, LeCun said.
The contrast with where the money is going could hardly be sharper. The same week LeCun was dismissing bigger models, Anthropic's data-centre ambitions were making headlines — we covered its secret bitcoin-miner deal capping a $530 billion compute sprint. LeCun's message to Seoul was, in effect, that this entire capital-expenditure arms race is aimed at the wrong target. The labs scaling up transformers are digging deeper into the same trench; the breakthrough, he insists, requires climbing out of it.
The scale of his own bet suggests he means it. After leaving Meta in late 2025, LeCun launched AMI Labs, which announced a $1.03 billion seed round in March — reportedly one of the largest seed rounds ever raised — with backing from investors including Bezos Expeditions, Nvidia and Samsung. The company is building its research program around JEPA, aimed at industrial uses like robotics and anomaly detection.
The receipts landed a week before he spoke
LeCun didn't walk into Seoul with only rhetoric. On October 5, a paper co-authored by LeCun and colleagues appeared on arXiv: "H-JEPA: End-to-End Learning of Hierarchical World Models for Visual Planning." We pulled up the arXiv listing ourselves this morning to check the headline numbers rather than trusting the secondary write-ups — and they hold up as stated in the abstract.
The idea is elegant in the way good engineering usually is. Long-horizon planning requires reasoning across different timescales, so H-JEPA trains a hierarchy of action-conditioned JEPA models where each layer predicts further into the future in its own learned latent space. Planning proceeds top-down: the top layer optimizes progress toward the goal, and each layer's predictions become subgoals for the planner below. Higher layers discard fast, unpredictable detail and retain slower, task-relevant state.
Across four simulated navigation and manipulation environments, the hierarchical approach beat flat JEPA. The eye-catching result: on Visual AntMaze, a three-level hierarchy raised planning success from 18% to 73% while using less planner compute. Under inverse-dynamics supervision, the authors say the approach extends to real-robot videos from the DROID dataset, improving offline planning fidelity at lower compute.
Honesty requires the caveat: these are simulated environments and offline evaluations, not a robot doing your dishes. The gap between "plans better in a navigation sim" and "reliable physical intelligence" is exactly the gap LeCun's critics keep pointing at. But it is also, notably, the kind of concrete, falsifiable progress that the world-model program has been accused of lacking. The trend line in 2026 — from V-JEPA 2's zero-shot robot planning last year to this — is bending his way.
"It's an engineering problem": the safety argument that splits the field
The most deliberately provocative part of LeCun's briefing was on AI safety, where he is nearly alone among the field's elder statesmen. Asked about recent AI-related attacks on South Korea's financial sector, he rejected the idea that they showed AI becoming uncontrollable. They exposed inadequate cybersecurity safeguards, he said — "The companies deploying the agents lacked sufficient security expertise, while the software sandboxes surrounding them were leaky," per the Korea Herald's account of the briefing.
His broader position: while today's LLMs can be jailbroken and are hard to constrain through fine-tuning alone, future objective-driven systems could evaluate guardrails before acting. "I'm not saying this is an easy engineering problem, but it's an engineering problem," he told reporters in Seoul. He added that AI could ultimately favour cyber defenders, because defenders have greater resources and collaborate more openly.
That puts him on a direct collision course with the other camp. Microsoft CEO Satya Nadella has spent this year warning companies to operate on the assumption that every AI model is already compromised — we covered his "assume compromised" essay and what it means for enterprise AI. Nadella's posture is defensive: treat the model as the attacker. LeCun's is constructive: build the guardrails into the objective function and the problem is solvable. One of them is describing a security practice; the other is describing a research program. The industry is currently betting its compliance budgets on Nadella's version.
LeCun's framing has an obvious weakness, and his critics never tire of naming it: it rests heavily on trust in the people deploying these systems, and on the assumption that "evaluate the guardrails before acting" survives contact with a sufficiently capable optimizer. But his Seoul appearance was a reminder that the extinction-risk narrative he dismisses as science fiction is, to him, not just wrong but actively harmful — a distraction that diverts attention from the unglamorous engineering of sandboxes, security hygiene and objective design.
Why Seoul: Korea's factory-floor opening
The briefing's real audience wasn't Silicon Valley. It was South Korea. LeCun and the other panelists argued that the shift to physical AI gives Korea an opening in the global AI race — not by outspending US hyperscalers on data centres, but by playing to its strengths in manufacturing, robotics, shipbuilding and engineering.
The prescription was specific. Rather than trying to match American AI infrastructure spending, Korea should move quickly to deploy AI in factories, ships and machinery — the physical environments where it already has a competitive edge — and train its models on factory-floor data. Aitomatic CEO Christopher Nguyen, the chief architect of the AI Alliance's Project Tapestry, urged Korea to deploy industrial AI "early and often" while strengthening security and governance. ASI Foundry chairman Jimmy Rhee described Korea's industrial data as a strategic "data well," warning that waiting for perfect information could let competitors pull ahead.
The institutional pieces are already moving. LeCun serves as chief science adviser to Project Tapestry, the AI Alliance's federated-learning initiative that lets participants jointly develop a large-scale model without transferring their data to a central server — a design aimed squarely at industrial players who will never hand their factory data to a hyperscaler. And LeCun's ties to Korea run deeper than one briefing: he co-directs the Korea-U.S. Global AI Frontier Lab at NYU, and he used the Seoul stage to call on Korea to attract more foreign researchers and strengthen the path from fundamental research to commercialization.
The Korean officials in the room were listening. Ha Jung-woo, vice chairman of South Korea's National AI Strategy Committee, said AI competition is shifting from model performance toward industrial productivity, and that international cooperation will be crucial to turning the country's strengths in semiconductors, automobiles, shipbuilding and batteries into economic value.
It is a compelling pitch — and it exposes the uncomfortable subtext of the scaling era. As we reported last month, startups are now raising millions just to capture real-world human demonstration data to train robots, because the internet's text corpus doesn't teach a machine how to hold a wrench. LeCun made the same point with characteristic bluntness: AI can prove mathematical theorems and pass professional exams, he noted, while robots still struggle with household tasks a 10-year-old can perform without training. The world's data moat, he is saying, is not text — it's the factory floor. And Korea is sitting on a very large floor.
Our read: the hardest — and maybe most honest — bet in AI
So will LeCun's bet land? Here's our honest assessment: his diagnosis is almost certainly right, and his timeline is the gamble.
The diagnosis — that prediction of the next token is not the same as understanding cause and effect, and that physical reasoning won't emerge from scaling text models alone — gets harder to dismiss with every year the robotics benchmarks stay embarrassing. A field that can generate photorealistic video on demand but can't reliably fold a towel has a structural problem, not a scale problem. LeCun identified it before it was fashionable, and the H-JEPA results are the first concrete signal in a while that the alternative program is making real, measurable progress rather than just publishing position papers.
But "world models are the future" is a much easier claim than "world models will work on a commercial timeline," and AMI Labs is still an early-stage company with no shipped product — albeit one with a billion dollars and the field's most decorated contrarian. The scaling labs, meanwhile, aren't standing still; every gigawatt of compute they install makes their approach harder to abandon, whatever its theoretical ceiling. Korea's factory-floor opening is genuine, but seizing it requires exactly the things LeCun named and Korea historically finds hard: open data collaboration across industrial giants, a pipeline from lab to production line, and foreign researchers willing to relocate.
The smartest thing said in Seoul on Monday may have been the least quotable: that controllability is an engineering problem. Not because it's reassuring — it isn't, really — but because it reframes the debate from prophecy to work. LeCun has spent a decade being right about the long game while the industry got rich on the short one. The question now is whether the long game has finally arrived.
Sources
Reporting for this article drew on the Korea Herald's account of the October 12 Seoul briefing, The Investor's and the Korea Times' coverage of the Global AI Leaders Seoul Dialogue, and the arXiv listing for "H-JEPA: End-to-End Learning of Hierarchical World Models for Visual Planning" (submitted October 5, 2026). Background on AMI Labs and LeCun's departure from Meta was verified against March 2026 reporting on the company's funding round.
About the author: Daniel Rao covers AI research, robotics and chip policy for TechPulse USA. He has reported on the scaling debate from both sides of it and still can't fold a towel one-handed.
Comments
Post a Comment