TL;DR
Yann LeCun's startup AMI has open-sourced H-JEPA, a hierarchical world model that jumps from 18% to 73% task success by stacking planning layers, marking the first concrete research deliverable from LeCun's post-Meta venture.
What happened
- AMI (Advanced Machine Intelligence), LeCun's startup, released H-JEPA jointly with NYU, INRIA Paris, and Brown University on October 10, 2026.
- H-JEPA stacks multiple JEPA layers, each predicting states across different time spans, with high-level layers setting sub-goals and low-level layers refining them into executable actions.
- On the Visual AntMaze benchmark, the three-layer model hit a 73% success rate versus 18% for a single-layer baseline, a 4x improvement.
- The Push-T task showed limits: two layers outperformed three or four, confirming hierarchy has diminishing returns depending on task complexity.
- All code and pre-trained weights are fully open-sourced, allowing direct reproduction of planning experiments.
Why it matters
- LeCun's JEPA thesis, first articulated in 2022 at Meta, now has a working multi-layer implementation with published benchmarks, moving the idea from whiteboard to reproducible result.
- The hierarchical approach solves a core robotics bottleneck: high-level goal understanding and low-level motor control require different abstractions, and H-JEPA trains both end-to-end through shared sub-goals.
- Representation collapse, the classic failure mode where encoders compress all inputs to the same vector, is addressed via SIGReg, a regularization method carried over from AMI's earlier LeJEPA and LeWorldModel work.
- Better performance at lower planning compute cost in simulation tasks suggests the architecture could scale to real-time robot deployment without proportional hardware cost increases.
- Full open weights lower the barrier for robotics and embodied-AI labs to build on AMI's stack, potentially accelerating a JEPA-based ecosystem outside the closed-model mainstream.
What to watch next
- Real-robot transfer: H-JEPA results are simulation-only so far. Physical deployment on quadrupeds or manipulation arms would validate whether the hierarchy holds outside controlled environments.
- Scaling the layer count: the Push-T regression above two layers signals an unresolved tuning problem. Watch for follow-up work on when and how to select optimal hierarchy depth automatically.
- Competitive response from Meta and others: LeCun's departure gives AMI independence, but Meta still owns significant JEPA research lineage. Whether Meta accelerates its own V-JEPA roadmap in response is a key signal.
Originally published on Present of AI, a daily source-linked AI news timeline. Read the full timeline or browse the open dataset.