Brief
British Startup Aims to Train AI with Video Game Data
Worldmodeldata plans to package the millions of controller inputs generated by gamers into datasets that could teach AI systems to understand cause and consequence in the real world.
By Felo News Desk · Published
A British startup called Worldmodeldata is proposing to use the vast amounts of controller input data produced by video game studios as a training resource for artificial‑intelligence models that need to understand physical interactions. The company, advised by Yann LeCun, says that the data collected by game developers – which includes millions of recorded button presses and camera footage – can be repurposed to teach AI systems how to predict the effects of actions in a visual environment.
Why Video Games Matter for AI
Researchers have long argued that large language models, which learn only from text, will struggle to perform tasks that require real‑world physics. To address this, a group of scientists, including Fei‑Fei Li and Yann LeCun, are turning to “world models” – AI that learns from visual and action data. However, unlike text, there is no large, publicly available corpus of cause‑and‑effect data for training such models.
Worldmodeldata’s solution is to act as a broker that curates and organizes the controller logs and other data that game studios already collect. Rhea Loucas, the company’s CEO, told WIRED that “there are millions of great games, and they are more and more similar to the real world.” By packaging this data into training sets, the startup hopes to give AI researchers a scalable way to improve world‑model performance.
Industry Interest
Several companies are already collecting video‑game data for their own AI projects. General Intuition and Niantic, for example, have begun gathering data from their platforms. Worldmodeldata claims it can save labs the time and cost of negotiating individual agreements with each game studio.
Current State of the Field
While the hypothesis that larger datasets will improve world‑model performance is widely accepted, the approach has yet to be fully tested. Some labs have tried to generate data manually by attaching sensors to humans and robots, but this method produces limited amounts of data and fails to capture the wide range of scenarios a model might face.
Worldmodeldata’s approach could potentially accelerate progress in AI that can navigate the physical world, from autonomous vehicles to robotic manipulation.
Key facts
- Worldmodeldata is a British startup advising by Yann LeCun (wired.com)
- The company plans to package video game controller logs into AI training datasets (wired.com)
- Researchers argue that large language models lack real‑world physics understanding (wired.com)
- Companies such as General Intuition and Niantic are already collecting video‑game data (wired.com)
- Manual data generation with sensors yields limited data and misses fringe scenarios (wired.com)
Sources
- [1] wired.com — originally reported as “The Next Evolution of AI Is Learning From Your Dodgy Gaming Skills”
