Startup Monetizes Gaming Data to Train AI Models

A player mashing buttons to dodge an enemy in a video game is, whether they know it or not, generating a small packet of physics data: an action, followed by a consequence. A British startup called Worldmodeldata is betting that this kind of raw, unglamorous gameplay, footage of ordinary people fumbling through 3D environments, could become fuel for the next generation of artificial intelligence.

The idea runs against the current playbook. Large language models got their power from text scraped off the internet, essentially an endless library of human writing. But researchers including Fei-Fei Li and Yann LeCun have started arguing that text alone can't teach a machine how to move through physical space, grip an object, or steer a vehicle. That gap has pushed part of the AI field toward "world models," systems trained to understand cause and effect in three dimensions rather than just predict the next word in a sentence.

The problem is supply. No one has assembled a data set of physical action and consequence anywhere near the scale of what fueled the LLM boom. Video games, it turns out, might be one of the few places that data already exists in bulk.