This passage comes from *Sources* by Alex Heath, a weekly newsletter devoted to exploring the evolving intersections of artificial intelligence and the broader technology industry. It is a carefully curated publication, available exclusively once a week to subscribers of *The Verge*, that provides insider perspectives on developments shaping the future of digital innovation.
In the early months of the previous year, Pim de Witte, founder of the video game clipping platform *Medal*, embarked on an ambitious outreach initiative. He began contacting several of the world’s most influential artificial intelligence research laboratories to gauge their potential interest in using *Medal*’s extensive repository of player-generated video data to train their learning systems, commonly referred to as agents. His proposal was rooted in the belief that the immense volume of recorded gameplay—capturing complex decision-making, spatial navigation, and reflex-based responses—could provide a uniquely valuable training ground for AI models seeking to replicate human-like adaptation and perception within simulated worlds.
Within mere weeks of initiating those conversations, de Witte discovered that the demand for his company’s data far exceeded his expectations. He soon found himself fielding multiple acquisition proposals from leading AI companies eager to secure exclusive access to *Medal*’s treasure trove of visual and behavioral data. Although de Witte declined to identify specific bidders, industry reports later indicated that OpenAI had extended a substantial $500 million acquisition offer. Reflecting on that period, de Witte admitted that his initial inclination toward these offers stemmed largely from a limited understanding of the strategic potential contained within his own dataset. “At the beginning, we were quite enthused by those proposals,” he recounted, adding that this reaction was primarily driven by an early misjudgment of just how significant his data could be in shaping the next era of AI research.
De Witte’s perspective shifted after he studied a Google DeepMind research paper demonstrating that gaming data can effectively teach artificial agents to navigate three-dimensional virtual environments. This discovery crystallized a striking realization: *Medal*’s vast and continuously expanding archive—comprising roughly two billion user-generated video uploads each year drawn from tens of thousands of distinct games—could form the backbone of a new foundational model designed to extend artificial intelligence beyond the digital sphere and into practical, real-world contexts.
Building upon that insight, de Witte recently announced the creation of a spin-off venture called *General Intuition*, an independent AI research lab emerging from *Medal*’s infrastructure. This new company successfully secured an impressive $133.7 million seed funding round, with the majority of the capital provided by Vinod Khosla, the influential founder of Khosla Ventures and an early backer of OpenAI. Other major investors participating in the round include General Catalyst and the Raine Group, indicating broad institutional confidence in the project’s future. Additionally, Moritz Baier-Lentz, who leads Lightspeed Venture Partners’ gaming investments, has joined the initiative as a part-time founding team member, bringing considerable expertise from the intersection of gaming and venture investment.
Vinod Khosla expressed a profound degree of optimism about *General Intuition*’s prospects, suggesting that its influence on the field of intelligent agents could parallel the transformative impact OpenAI has had on the broader domain of large language models. According to Khosla, their investment represents his firm’s most significant seed-stage commitment since it first supported OpenAI in 2018—a clear indication of the scale of his confidence in de Witte’s approach. “It’s a substantial wager,” Khosla observed, emphasizing that the combination of an unprecedented dataset and a particularly cohesive, capable team gives *General Intuition* a rare strategic advantage.
For those less immersed in the specialized lexicon of AI research, the notion of *world models* may still be relatively obscure. This branch of artificial intelligence investigates how to train systems with an intrinsic understanding of space and physics—capacities that resemble the natural spatial reasoning humans rely upon. In essence, world model research attempts to construct AI capable not merely of analyzing static data, but of predicting dynamic interactions within an environment. For instance, a sufficiently advanced system could foresee the trajectory of a glass of water tipping from a table and act autonomously to catch it before it shatters. On a more practical level, these models help AI agents learn to reason and make predictions within simulated three-dimensional settings—skills critical for robotics, autonomous vehicles, and other physical-world applications.
Among leaders in artificial intelligence, few have championed this area of research as fervently as Demis Hassabis, the CEO of Google DeepMind. Hassabis has consistently argued that world models will prove indispensable to achieving artificial general intelligence (AGI)—the long-sought milestone at which machines exhibit the flexible learning capacity of the human mind. Google recently showcased its latest advancement in this area, a model known as Genie 3, which can dynamically generate interactive, game-like environments from scratch as a user navigates through them. Meanwhile, other ambitious ventures have also entered this race. Fei-Fei Li’s startup, *World Labs*, recently unveiled its own real-time demonstration of a model capable of producing interactive video on the fly, underscoring the growing commercial and research momentum surrounding this emerging field.
For *General Intuition*, de Witte envisions technology capable of seamlessly operating any hardware device that can be mapped to the familiar inputs of a keyboard, mouse, or handheld game controller. In its earliest stages, he expects the company’s first competent model to assist with search-and-rescue drones—machines that can traverse difficult, dangerous terrain using learned intuition rather than rigid programming. Long-term, he sees the methodology scaling well beyond drones, potentially empowering humanoid robots, self-driving vehicles, and other systems requiring real-time environmental reasoning.
Drawing a parallel to how large language models were originally trained on vast collections of textual web data, de Witte believes that gaming environments offer the perfect structured framework to teach AI how to interpret and react correctly within physical and temporal settings. He describes games as the ultimate “verifiable domain” for developing spatial-temporal reasoning because every action within a game can be crisply evaluated as effective or ineffective—good or bad—through immediate feedback mechanisms. This property, he contends, makes gaming-derived data exponentially more instructive for building generalist AI competence than traditional unstructured real-world datasets.
Despite its promise, the path forward for *General Intuition* remains laden with uncertainty. The correct technical strategy for developing robust world models is still actively contested within the AI community, and as Khosla himself acknowledges, no one has yet determined which kinds of data will ultimately prove most advantageous. Members of de Witte’s initial research team have already contributed several noteworthy academic papers in this area, underlining their depth of expertise, yet the startup now finds itself competing against technology titans with far greater financial and computational resources, such as Google and its DeepMind subsidiary. “Someone will ultimately achieve a monumental success in this space,” Khosla remarked, predicting that world modeling could give rise to enterprises valued in the hundreds of billions—or even trillions—of dollars.
De Witte further anticipates that as enthusiasm for world model technology intensifies, traditional gaming companies will increasingly become acquisition targets for these AI research organizations. His decision to found *General Intuition* was largely motivated by the realization that *Medal*’s extensive dataset positioned him not merely as a supplier of valuable information, but as a potential architect of foundational AI systems. Yet he cautions that other gaming executives may struggle to resist the immediate allure of lucrative licensing or buyout offers from major AI firms. “You are at an inherent information disadvantage,” he advised when asked what guidance he might give to peers within the gaming industry. “As these models continue to improve, their hunger for data will diminish; the better they become, the less they need.”
Readers driven to explore this evolving story further are encouraged to follow the topics and authors connected to this coverage, ensuring that similar analyses and updates appear within their personalized homepage feeds and email digests.
— *Alex Heath*, writing for *Sources*, *AI Column*
Sourse: https://www.theverge.com/column/801370/ai-world-models-general-intuition-medal
