Elon Musk's xAI is working on world models, which are advanced artificial intelligence (AI) systems that learn from videos and robot data to understand and navigate real-world environments, Financial Times reports (open copy).
This puts the company in competition with others like Meta and Google. These models go beyond the large language models (LLMs) used in AI tools like ChatGPT and xAI's Grok. Instead, world models focus on physical interactions.
The company hired two experts from Nvidia, a firm known for its Omniverse platform that builds simulations. These hires help xAI build models for gaming, where they could create interactive 3D spaces, and for humanoid robots that move in the real world. Musk said xAI plans to release a fully AI-made game by the end of next year. Recently, xAI launched a new tool for making images and videos, which is free and improved over older versions.
Current video tools, like OpenAI's Sora, make clips by guessing patterns from data. World models would improve this by having a true grasp of physics, which is the science of how things move and interact, allowing real-time responses in different settings.
Applications and challenges
xAI is hiring more staff for its "omni team," which handles AI for images, videos, and sound, with salaries from $180,000 to $440,000. It also seeks a video games tutor to teach Grok how to make games and help users design them, paying $45 to $100 per hour. Other labs like Google and Meta are also pursuing these systems.
However, building world models is hard. Getting enough data to mimic the real world is tough and expensive. A game developer contacted by Financial Times noted that AI might not fix key issues in gaming, like needing strong ideas and vision, rather than just math-based loops. Overall, world models could open new uses for AI in products beyond computers, with Nvidia saying the market could match the global economy's size. xAI did not comment on the details.