AIGenerated AI news, written by AI and explained simply. Models, technology, culture and the people behind it.
AMI is a research startup led by AI pioneer Yann LeCun that builds hierarchical world models for long-term visual planning, and has just published its first open research.
1 story, first covered on Oct 10, 2026.
AMI is a small research company started by computer scientist Yann LeCun. LeCun is best known for inventing convolutional neural networks (CNNs), a core technology that lets machines see and recognize images. He now wants to help machines understand how the world works over time, not just recognize single pictures.
The company is working on something called a hierarchical world model for visual planning. In simple terms, a world model is a system that learns to predict what will happen next when it watches videos or images. AMI’s version stacks many such prediction layers on top of each other, so the system can plan several steps ahead instead of reacting only to the current moment. Their first project, named H-JEPA, showed that this stacked approach made long-term visual planning much better than earlier methods.
Most AI systems today are good at quick tasks—like answering a question or editing a photo—but struggle to plan over longer periods. If a machine can simulate future outcomes in its mind, it could make safer decisions in robotics, autonomous vehicles, and other real-world applications. AMI’s early results suggest that hierarchical prediction is a promising path toward that kind of lasting intelligence.
AMI recently released its first research together with the full computer code and model weights. Anyone can download and study the software, which gives other researchers a chance to test, improve, or build on the idea. The team is likely still developing the next stages of the system, but the public release marks the beginning of open collaboration around their approach.
AMI is founded and led by Yann LeCun, a computer scientist who invented the neural-network techniques that power most modern image recognition. He left his role at Meta to focus on helping AI systems learn to predict and plan over longer time periods.
H-JEPA is a visual-planning model that stacks multiple prediction layers. Instead of only guessing the next frame in a video, it builds higher-level predictions that can look several steps ahead, which improves long-term planning in simulated visual tasks.
Yes. AMI released its first research together with the complete source code and trained model weights, so other developers and researchers can download, study, and extend the system without needing special permission.
How it worksLeCun's Startup Unveils H-JEPA A Hierarchical World Model for Visual PlanningHerald of AI
AMI releases its first research with full code and weights open-sourced, showing that stacking prediction layers dramatically improves long-term planning in visual tasks.