Genie: Generative Interactive Environments with Ashley Edwards - #696

This article summarizes a podcast episode discussing Genie, a system developed by Runway for creating playable video environments. The core focus is on Genie's ability to generate interactive environments for training reinforcement learning agents without explicit action data. The discussion covers the system's architecture, including the latent action model, video tokenizer, and dynamics model, and how these components work together to predict future video frames. The article also touches upon the use of spatiotemporal transformers and MaskGIT techniques, and compares Genie to other video generation models like Sora, highlighting its potential implications and future directions in video generation.

• •Genie is a system for creating playable video environments for training RL agents.
• •It learns world models from videos without explicit action data.
• •The system uses latent action models, video tokenizers, and dynamics models.
• •It utilizes spatiotemporal transformers and MaskGIT techniques.