{
  "id": 10410307,
  "title": "The Next Evolution of AI Is Learning From Your Dodgy Gaming Skills",
  "url": "https://urgent.news/2026/09/28/the-next-evolution-of-ai-is-learning-from-your-dodgy-gaming-skills",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-28T09:00:00.000Z",
  "source": {
    "name": "Wired",
    "slug": "wired",
    "url": "https://www.wired.com/story/the-next-evolution-of-ai-is-learning-from-your-dodgy-gaming-skills/"
  },
  "original_language": "en",
  "account": "The AI world is abuzz with the notion that large language models (LLMs) may struggle to handle the physical realm. Trained solely on text, these models may find it challenging to pilot autonomous vehicles or operate robotic arms. To bridge this gap, researchers like Fei-Fei Li and Yann LeCun are exploring \"world models\" – AI systems that require training on visual and action data to understand real-world physics. However, unlike LLMs, which have a wealth of textual data to learn from, world model researchers are struggling to find suitable training material.\n\nXiatian Zhu, an AI specialist at the University of Surrey, explains that world models need \"cause and consequence\" data, something that is scarce on the internet. Worldmodeldata, a British startup backed by Yann LeCun, attempts to solve this problem by leveraging video game data. This data, comprising controller inputs and other information from video game studios, is typically abundant and diverse, making it a potential treasure trove for training world models.\n\nCompanies like General Intuition and Niantic are already harnessing video game data for their world models, but Worldmodeldata claims to offer a more efficient solution. The startup curates and organizes this data, saving AI labs from having to negotiate individual agreements with countless game studios. Worldmodeldata's CEO, Rhea Loucas, believes that video game environments, with their visual representations and player actions, are both plentiful and varied enough to capture crucial corner cases that ordinary training data might miss.\n\nHowever, not everyone is convinced that video game data is the panacea for world models. Nvidia, which develops world models optimized for its chips, prefers to use a custom engine for replicating real-world physics. The company's chief world model developer, Ming-Yu Liu, argues that video game physics often falls short in terms of fine-grained motor control, making it unsuitable for tasks requiring precise manipulation.\n\nDespite these concerns, world model researchers remain optimistic. They believe that as these models grow larger, they will improve in performance, similar to LLMs. Worldmodeldata aims to make this vision a reality, with plans to license almost 1 million hours of video game data for training. If successful, video game data could become the cornerstone of world model training, potentially ushering in a new era of AI applications.",
  "summary": "A British startup is shaping video game inputs into training data for AI models that can navigate the physical world.",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 2,
    "also_reported_by": [
      {
        "outlet": "Wired Business",
        "title": "The Next Evolution of AI Is Learning From Your Dodgy Gaming Skills",
        "url": "https://urgent.news/2026/09/28/the-next-evolution-of-ai-is-learning-from-your-dodgy-gaming-skills-10410513",
        "published": "2026-09-28T09:00:00.000Z"
      }
    ]
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}