JEPA

Headlines · 4

1. 5d
   BAAI's Wang Pengwei on RoboBrain Orca: world models vs VLA for embodied generalization

   Wang Pengwei, who leads the embodied model research center at China's BAAI, presented the RoboBrain Orca world model at the 2026 BAAI conference. He argues that today's four families of world models — language-, pixel-, 3D-structure- and visual-representation-centric approaches such as LLMs, video generators, World Labs' Marble and JEPA — each capture only a partial world and remain passively task-driven; Orca instead uses Next State Prediction, reasoning forward and backward about state transitions, and separates "unconscious" video learning from "conscious" event-level state-transition learning. He says Orca's pretraining used no action data, drawing on about 125,000 hours of internet video, roughly 130 million expert-model-labeled event-level state-transition samples and causal and general VQA data, and that BAAI is building a benchmark for state-transition world models because none exists; he expects some approach to break the 0% generalization barrier within a year, though 90%-plus would be needed for real usability.

   钛媒体 · 🔥 0
2. 5d
   LeCun: AI's next leap will come from world models, controllable by engineering

   Yann LeCun, Turing Award winner and founder-chairman of AMI Labs, said at an industrial AI discussion in Seoul that AI's next breakthrough will come from "world models" that can understand, predict and act on the physical world, rather than from simply scaling up large language models. He argued that today's LLMs learn effectively from text, images and video but still lack a reliable grasp of cause and effect in the physical world, noting robots still struggle with everyday chores that a 10-year-old can do untrained. LeCun said his AMI Labs is pursuing this route through Joint-Embedding Predictive Architecture (JEPA), though he acknowledged it remains unproven whether such systems can deliver reliable prediction, autonomous planning and large-scale deployment in complex real environments. Responding to recent cybersecurity incidents involving AI agents, including AI-related attacks on South Korea's financial sector, he said the problems stem mainly from firms lacking security expertise and from flaws in sandboxes, and that the risks can be managed through better system architecture and security engineering.

   cnBeta · 🔥 0
3. 5d
   LeCun's H-JEPA hierarchical world model takes aim at agents and RSI

   Yann LeCun's new world model H-JEPA is presented as the first reproducible implementation of the hierarchical JEPA idea from his 2022 paper: multiple JEPAs are stacked, each with its own observation encoder, action encoder and latent predictor, forecasting states at different time horizons, with high levels setting coarse goals that lower levels refine into actions. The article reports that a three-layer H-JEPA raised success on Visual AntMaze to 73% from 18% for single-layer LeWM, but on Push-T two layers worked best and three or four layers degraded, using roughly 58% and 14% of LeWM's transferable training data respectively; training uses neither reward nor reconstruction, and DROID teleoperation experiments needed an inverse-dynamics constraint to avoid slow-feature collapse. It frames H-JEPA as a counterpoint to today's agents, which split tasks into more steps and more tokens, and to recursive self-improvement, which LeCun calls only an amplifier of existing abilities.

   虎嗅 AI · 🔥 0
4. 7d
   LeCun's AMI open-sources H-JEPA, a hierarchical world model for visual planning

   AMI (Advanced Machine Intelligence), the world-model startup chaired by Yann LeCun, released a paper with NYU, INRIA Paris and Brown University describing H-JEPA: multiple JEPA modules stacked so each layer predicts states at its own timescale, with a high level setting a coarse plan and lower levels refining predicted intermediate states into executable actions. The paper reports 73% success for a three-layer model versus 18% for a single-layer model on the Visual AntMaze simulated maze task, with better results at lower planning compute on several tasks; code and pretrained weights were released. The report also says AMI previously raised $1.03B at a $3.5B pre-money valuation, while its website still shows mainly research direction and hiring rather than a public product.

   36氪 人工智能 · 🔥 0

Experience and discussion from the community

Share my JEPA experienceAsk about JEPA