What happens when AI runs out of text to read? Researchers are turning to video games.
Made with Midjourney
Solving the data problem. The AI industry has an elephant in the room. Large Language Models (LLMs) could exhaust the public supply of written text as soon as 2026. This presents a pivotal question: Without fresh training data, how will AI continue to improve? Some experts believe the answer lies in video games.
Using video games to train AI. Video games offer an endless playground for AI to learn how the real world works. Even a simple game helps teach movement, physics, problem-solving, cause-and-effect, and complex spatial reasoning. And, unlike real-world data that’s expensive and limited, video games provide infinite simulated scenarios where AI can safely experiment as it learns.
Enter: World models. World models are AI models that understand the dynamics of the real world and can simulate cause-and-effect scenarios to predict outcomes. Unlike LLMs, which rely on reading text, world models learn by observing and building mental maps of how objects move, interact, and behave.
AI’s biggest names are already making progress. DeepMind launched Sima 2, an agent that plays and reasons in 3D virtual worlds. World Labs released Marble, which creates entire 3D worlds from a simple prompt. An Yann LeCun — often called a “godfather of AI” — is reportedly leaving his Chief AI Scientist role at Meta to pursue a new startup focused on world models.
If successful, world models will take us a step closer to creating AI that operates alongside us in the real world.
via Superhuman