LeCun's AMI debuts H-JEPA: hierarchical world model hits 73% in maze planning
量子位 · wechat · 2026-10-10
Advanced Machine Intelligence, Yann LeCun's world-model startup, released its first paper H-JEPA with NYU, INRIA and Brown, open-sourcing code and weights. The model stacks multiple JEPA layers that predict at different time horizons: high layers set coarse plans, lower layers refine them into executable actions, with subgoals bridging levels and SIGReg preventing representation collapse. On the VisualAntMaze task, a three-layer H-JEPA reached 73% success versus 18% for a single layer, while using less planning compute; deeper stacks hurt on short tasks like Push-T due to limited high-level training data. The work extends LeCun's long-held view that prediction in representation space — not LLMs — is the path to human-level AI. AMI raised $1.03B in March at a $3.5B pre-money valuation but has no public product yet, while rival World Labs agreed to an $8.2B all-stock acquisition by AMD.
More from Companies & People
- OpenAI accused of firing employee days after contact with safety evaluator METR — YvesMulkers · 2026-10-10
- Jensen Huang Admits AI Incidents Have Happened: 'If It's Not Ready, Don't Release It' — AlexTensor · 2026-10-10
- ChatGPT anniversary: almost nobody saw the demo that changed everything coming — LucaAmb · 2026-10-10
- Google DeepMind's new Polaris team pays $550k–$1.1M, no degree required — brianryhuang · 2026-10-10
- David Sacks: Treating Claude Like a Human Is Why Alignment Has Failed for 10 Years — AlexTensor · 2026-10-10
- Senior architects: seniority is about tradeoffs and risk judgment, not collecting more tools — _jaydeepkarale · 2026-10-10