The robotics AI startup, spun out of gaming clip platform Medal eight months ago, is in talks to raise $300M — and its thesis rests entirely on whether first person gameplay footage transfers to real world spatial reasoning.
General Intuition raised $134 million eight months ago to spin out of Medal, a platform for uploading and sharing video game clips that was generating approximately 2 billion videos per year. Now the New York-based robotics-AI startup is in talks to raise around $300 million at a valuation just over $2 billion — backed by Jeff Bezos, Eric Schmidt, Khosla Ventures, and General Catalyst, sources familiar with the matter TechCrunch. The backer list is a category bet: the same names are spread across the world-model and embodied-AI landscape. But the question General Intuition hasn't answered publicly is whether its core claim — that Medal's gameplay corpus is the right data to train real-world AI agents — actually holds.
General Intuition's pitch is built on a single, unproven assertion: that first-person, interactive gameplay footage is the right training signal for teaching AI agents to navigate and reason about the real world. Medal's pipeline generates roughly 2 billion videos per year from approximately 10 million monthly active users TechCrunch. The startup characterizes this as unique precisely because it captures interactive, first-person behavior at scale — not the passive video that underpins most foundation model training. "Such a dataset — unique because it allows AI to learn from interactive, first-person gameplay — is the perfect base to teach machines deep spatial-temporal reasoning, allowing them to perceive, anticipate, and interact in real time in simulation," according to sources familiar with General Intuition's positioning TechCrunch.
That pitch has attracted attention. Sources tell TechCrunch that OpenAI previously attempted to acquire Medal, and that other large AI labs have come knocking TechCrunch. Whether those conversations reflect genuine competitive tension or exploratory interest is not clear.
The world-model space is crowded. Runway, Decart, and World Labs have all released world models recently, and Google's Genie 3 recently began integrating Google Maps data for real-world simulation TechCrunch. The near-term commercial use cases across the category — gaming and robotics training — are converging.
General Intuition's differentiation is categorical, not incremental. The company builds world models to train agents; it does not sell the model itself. "General Intuition takes a different approach: it builds world models to train agents, not to sell them. The agents are the product," according to TechCrunch's reporting TechCrunch. Whether owning the agent rather than the model is a more defensible business position — or just a later-arriving, harder commercial bet — is exactly the open question the valuation embeds.
The company was founded by Pim de Witte, who co-founded Medal, alongside researchers Eloi Alonso, Adam Jelley, and Vincent Micheli, who bring expertise in world modeling and simulation TechCrunch. The four-person founding team has world-model research credibility; whether the commercial execution matches the research pedigree is a question the company has not yet answered publicly, since it has no public product and no benchmark demo.
General Intuition plans to use the fresh capital to scale compute capacity with the goal of releasing a new product by the end of summer or early fall, according to a source familiar with the matter TechCrunch. The timeline is aggressive for a company with no public benchmark demonstrating that its training thesis is correct. The $2 billion valuation is a bet on the hypothesis — not a confirmation of it.
General Intuition and representatives for Bezos, Schmidt, Khosla Ventures, and General Catalyst did not respond to requests for comment prior to publication. All figures — the $300 million raise amount, the $2 billion valuation, and the backer list — are attributed to sources familiar with the matter and have not been confirmed by the named parties.