General Intuition raises $134M to train AI with video game clips

A new artificial intelligence research lab has secured a massive $133.7 million in seed funding to develop advanced AI agents using an enormous collection of video game footage. General Intuition, which spun out from the popular game clip platform Medal, leverages a unique dataset of roughly two billion videos annually contributed by its ten million monthly active users. This vast resource provides a powerful training ground for AI to master spatial-temporal reasoning, the ability to understand how objects move and interact across both space and time.

The company’s leadership believes that gaming environments offer a superior training medium compared to other video sources like Twitch or YouTube. According to Pim de Witte, CEO of both Medal and General Intuition, the nature of shared gaming clips creates a valuable selection bias. Players typically upload moments of extreme success or failure, which naturally provides a wealth of challenging edge cases ideal for teaching AI systems. This distinctive data advantage reportedly drew acquisition interest from OpenAI last year, though neither party has commented publicly on those discussions.

The substantial seed round was spearheaded by Khosla Ventures and General Catalyst, with additional investment from Raine. These funds will accelerate the startup’s mission to build a general-purpose AI agent capable of interacting with complex environments. Initial applications are targeted at the gaming industry and search-and-rescue drone operations.

General Intuition’s technical approach sets it apart. Its models learn purely from visual input, observing game worlds exactly as a human player would, and learning to navigate using simulated controller commands. This method has already shown promising results, with the AI demonstrating an ability to understand unfamiliar settings and predict appropriate actions within them. The company asserts this skill can be directly transferred to physical systems, including robotic arms, drones, and autonomous vehicles, which are often operated via gamepad-like interfaces.

Looking forward, the startup aims to achieve two key milestones: generating entirely new simulated worlds to train other AI agents, and enabling autonomous navigation through completely unknown physical spaces. This technical roadmap also defines its commercial strategy, differentiating it from competitors focused on selling world models.

Unlike other labs such as DeepMind, which market their world models for agent training and content creation, General Intuition is avoiding direct competition with game developers to sidestep potential copyright issues. Instead, its gaming efforts are concentrated on developing advanced non-player characters and bots. These AI-driven entities would move beyond simple, pre-scripted behaviors to offer dynamic and scalable challenges, adapting their difficulty to maintain player engagement.

Moritz Baier-Lentz, a founding member and partner at Lightspeed Ventures, emphasized the goal is not to create an unbeatable “god bot,” but rather to design AI opponents that provide a consistently engaging experience, ideally keeping a player’s win rate around fifty percent to maximize retention.

The company’s ambitions extend beyond entertainment. De Witte’s background in humanitarian work influences the focus on search and rescue drones, which often need to operate in GPS-denied and unfamiliar terrain. The core capability of spatial-temporal reasoning is viewed as a fundamental component in the broader pursuit of artificial general intelligence (AGI). While many major AI labs concentrate on scaling large language models, General Intuition contends that true AGI requires an intuitive grasp of the physical world, a dimension that text-based models inherently lack. As de Witte notes, describing the world through text inevitably strips away rich, intuitive understanding of how things move and interact in space and time.

(Source: TechCrunch)