Despite Nvidia's claims of fair use, copyright experts question the legality of using such data for AI training. The Cosmos WFMs are not fully open source, as Nvidia has not disclosed all details about the training data or provided the tools to recreate the models. However, companies like Waabi, Wayve, Foretellix, and Uber are already piloting Cosmos WFMs for various applications, including AI models for self-driving vehicles. Nvidia CEO Jensen Huang expressed hopes that Cosmos will significantly impact robotics and industrial AI, similar to the influence of Llama in enterprise AI.
Key takeaways:
- Nvidia announced the release of Cosmos World Foundation Models (WFMs) at CES 2025, designed for physics-based simulation and synthetic data generation.
- Cosmos WFMs are available in three categories: Nano, Super, and Ultra, with model sizes ranging from 4 billion to 14 billion parameters.
- The models are trained on 9,000 trillion tokens from diverse data sources, but Nvidia has not disclosed specific details about the training data's provenance.
- Companies like Uber and Wayve are piloting Cosmos WFMs for applications in autonomous driving and robotics, although the models are not fully open source.