LeCun bets on world models in Seoul as his H-JEPA paper lifts AntMaze to 73%

Yann LeCun, the 2018 Turing Award winner who now runs AMI Labs as founder and chairperson, told an industrial-AI panel in Seoul on October 12 that AI’s next breakthrough will come from systems that understand, predict and act in the physical world — not from ever-larger LLMs, which in his view lack “a reliable understanding of cause and effect in the physical world.” There is a checkable artifact behind the claim: the H-JEPA paper he co-authored went up on arXiv on October 5 (2610.06805), and its abstract reports that a three-level hierarchy lifts Visual AntMaze success from 18% to 73% while using less planner compute.

Key points

• At the panel hosted by Aitomatic and the IIAC 2026 committee, LeCun defined world models as systems “designed to predict how the physical world will change in response to an action” — the core of what his AMI Labs builds.
• H-JEPA (arXiv 2610.06805, submitted October 5 by Wancong Zhang, with Basile Terver, Michael Rabbat, LeCun and Randall Balestriero as co-authors) is the first complete, reproducible implementation of the hierarchical JEPA he sketched in 2022, per TMTPost’s teardown of the paper.
• Asked about recent AI-agent attacks on Korean finance, he called it weak security engineering rather than runaway AI: “The companies deploying the agents lacked sufficient security expertise, while the software sandboxes surrounding them were leaky.”

Background

Hierarchical JEPA has been a promise since 2022, when LeCun’s position paper “A Path Towards Autonomous Machine Intelligence” sketched the idea; TMTPost notes that no complete, reproducible implementation followed for four years. H-JEPA fills that gap. It trains a hierarchy of action-conditioned JEPAs, each level predicting further ahead in its own learned latent space, with planning running top-down: higher levels set coarse goals that become subgoals for lower ones, and they keep slower, task-relevant state while discarding fast, unpredictable detail (all from the abstract). It is a continuation of his long-running critique that LLMs can pass the bar and write code, “but cannot truly handle the complexity of the real world.” Timing matters too: the paper landed a week after AMD agreed to buy Fei-Fei Li’s World Labs for about $8.2 billion, and the Seoul panel was held ahead of the fifth International Industrial AI Conference in Tokyo. The world-model camp is suddenly the best-funded dissident faction in AI.

The facts

Across four simulated navigation and manipulation environments, hierarchical planning beat a flat JEPA, and ablations credit the gains to temporal decomposition plus higher-level goal representations. The numbers worth keeping:

Benchmark	Result	Source
Visual AntMaze	18% flat baseline → 73% with three levels, less planner compute	arXiv abstract
Push-T	Two levels work best; three levels keep only ~58% of the transferable training data, four ~14%	TMTPost, citing the paper body
DROID real-robot video	Inverse-dynamics supervision improves offline planning fidelity; without it, “slow-feature collapse” drops performance to zero	arXiv abstract + TMTPost

On the training signal, TMTPost’s reading of the paper body says H-JEPA trains with “no reward and no reconstruction” — the model only predicts next states in representation space. The abstract itself claims only task-agnostic learning, so treat that detail as the paper’s text as relayed by TMTPost. The Seoul remarks were just as concrete. LeCun rejected the doomsday framing after the agent attacks on Korean financial firms: “I’m not saying this is an easy engineering problem, but it’s an engineering problem.” Future objective-driven systems could check guardrails before acting, he added, and in cyber offense versus defense, AI favors the defenders. For Korea he prescribed playing to manufacturing and robotics strengths instead of matching US infrastructure spending: deploy AI in factories, ships and machinery, build on open foundation models, and compete for leadership in physical AI and robotics. The ‘early and often’ urging came from Aitomatic CEO Christopher Nguyen on the same panel, who also pushed stronger security and governance. Sandbox isolation, for what it’s worth, is an engineering problem the industry is actively working — see our earlier coverage of Alibaba Cloud’s Agent Sandbox.

What others say

The panel itself leaned friendly to the thesis: Aitomatic CEO Christopher Nguyen; AI Alliance co-founder and co-chair Anthony Annunziata, pitching Project Tapestry for jointly developed open models with data sovereignty; ASI Foundry chairman Jimmy Rhee, calling Korea’s industrial data a strategic “data well”; and Ha Jung-woo, vice chair of Korea’s National AI Strategy Committee, framing the competition as a shift from model performance to industrial productivity. The sharpest reading came from TMTPost’s AI newsletter, which cast the paper as anti-agent and anti-RSI: today’s agents scale by spending more steps and tokens at the same granularity, while H-JEPA grows upward through abstraction. It also recalls LeCun’s January remark in Davos that he cannot imagine a system becoming an agent if it cannot predict the consequences of its own actions. Note the tension: in Seoul he did not reject agents — he defended them, blaming leaky sandboxes rather than the technology itself. The same teardown points out the irony that Tian Yuandong, his former student and onetime Meta FAIR research director, co-founded an RSI company that emerged from stealth in May with heavyweight funding, while the H-JEPA paper never once mentions RSI.

Our take

This is a rare round of the LLM-versus-world-model fight that arrives with numbers attached. For developers assembling agent stacks today, the useful part is the counter-argument: LeCun’s camp holds that bolting tools and reward signals onto LLMs never reaches the physical world, and the world-model side finally has a reproducible benchmark to point at. The honest catch is just as clear: all four environments are simulators, the only real-robot data needed a human-designed inverse-dynamics constraint to avoid collapse, and on Push-T deeper hierarchies made things worse. H-JEPA is a foundation for a research direction, not a replacement for anything you run in production. The arXiv abstract mentions no code release, and most details beyond the abstract currently travel through secondhand retellings — so watch for two things next: whether code ships, and whether the hierarchy’s numbers survive contact with real-world data.