Worldmodeldata licenses 1 million hours of game data for AI
A British startup called Worldmodeldata is packaging video game data to train AI world models, which are designed to understand physical physics and actions. Unlike large language models trained on text, world models require visual and action data to operate robots or autonomous vehicles. The company, advised by Yann LeCun, claims to have licensed nearly 1 million hours of controller inputs and…
Key points
- Worldmodeldata has licensed almost 1 million hours of video game data for AI training.
- Nvidia's Ming-Yu Liu argues game physics are too coarse for fine-grained robotic manipulation.
- The startup aims to solve the lack of physical action data for world models.
The startup argues that video games provide the massive, diverse datasets needed to handle real-world "corner cases" that manual data collection misses. CEO Rhea Loucas believes this data could trigger a "GPT moment" for world models. However, experts remain divided on the approach's effectiveness. Ming-Yu Liu at Nvidia argues that video game physics are often simplified and may not support fine-grained motor control, suggesting the data is better suited for generating realistic video than for precise robotic manipulation.
Academic researchers also note that game physics are coarse approximations. While the hypothesis that larger datasets improve model performance is broadly accepted, the specific utility of game data for physical tasks remains unproven. The industry is currently testing multiple paths, including custom physics engines and sensor-based data collection, to determine the most effective training method.
The story so far
2 episodes →- Worldmodeldata licenses 1 million hours of game data for AIthis story
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- OpenAI claims new model solved Navier-Stokes and 100+ math problems · 2 src
- Researchers release benchmark for AI in systematic review screening · 1 src
- SlideLab framework generates scientific presentations from research papers · 1 src
- Cartograph reduces AI agent tool discovery from O(n) to O(k) · 1 src
- Survey reviews 211 fake review detection studies from 2018 to 2026 · 1 src
Comments
via GitHub Discussions