General Intuition Raises $134 Million Seed Round to Train Agents in Spatial Reasoning with Video Game Clips
Medal, a leading platform for uploading and sharing video game clips, has established an advanced AI research laboratory named General Intuition. This new venture leverages Medal’s extensive collection of gaming videos to train foundational models and AI agents capable of spatial-temporal reasoning, which involves understanding how entities move through space and time. General Intuition contends that Medal’s dataset, comprising two billion videos annually from ten million monthly active users across thousands of games, offers superior training data compared to platforms like Twitch or YouTube.
Pim de Witte, CEO of Medal and General Intuition, explained that gameplay involves transferring one’s perception into various environments, often via first-person perspectives. He highlighted that uploaded clips predominantly feature extreme negative or positive scenarios, providing valuable edge cases for model training due to this selection bias, which aligns with desired training data.
This exclusive dataset reportedly attracted interest from OpenAI, which reportedly attempted to acquire Medal for $500 million in late 2023. This asset also facilitated General Intuition in securing $133.7 million in seed funding, led by Khosla Ventures and General Catalyst, with participation from Raine. The company plans to utilize the recent funding to expand its research and engineering teams dedicated to developing a general agent capable of complex interactions, with initial applications targeting gaming and search-and-rescue drone operations.
De Witte noted that General Intuition’s team has achieved significant progress; their model can comprehend novel environments and accurately predict actions within them using solely visual input. Agents perceive only what a human player would see and navigate spaces via controller commands. This methodology is designed for seamless transfer to physical systems such as robotic arms, drones, and autonomous vehicles, which are commonly operated through video game controllers.
General Intuition’s immediate objectives include generating new simulated worlds for agent training and enabling autonomous navigation in unfamiliar real-world settings. This technical direction informs the company’s commercialization strategy and distinguishes it from competitors offering world models as primary products. Whereas organizations like DeepMind and World Labs sell their respective Genie and Marble models for agent training and content creation, General Intuition focuses on alternative use cases, thereby circumventing copyright concerns. De Witte clarified that the goal is not to produce models that compete directly with game developers. Instead, the company aims to develop bots and non-player characters that outperform traditional deterministic bots, which consistently deliver identical outcomes.
Moritz Baier-Lentz, founding member of General Intuition and partner at Lightspeed Ventures, stated that their bots can adapt to any difficulty level. The objective is to provide dynamic opponents that maintain balanced win rates, enhancing user engagement and retention. Additionally, De Witte’s background in humanitarian work influences the startup’s efforts to empower search-and-rescue drones that must operate effectively in unfamiliar environments without relying on GPS.
In summary, De Witte and Baier-Lentz regard General Intuition’s expertise in spatial-temporal reasoning as vital to the pursuit of artificial general intelligence (AGI). While many AI laboratories concentrate on developing more robust large language models, General Intuition asserts that genuine AGI demands capabilities beyond those possessed by LLMs. As De Witte observed, while humans express events through text, substantial information, particularly relating to spatial-temporal reasoning, is invariably lost in translation.