General Intuition raised $320M at a $2.3B valuation to train AI agents on video game footage, the New York startup said Thursday. Khosla Ventures led the round, with General Catalyst, Jeff Bezos, Eric Schmidt, former F1 champion Nico Rosberg, and researchers at Google DeepMind and MIT also writing checks. The deal brings total disclosed funding to $454M after a $134M raise last October, and pushes General Intuition into the small cluster of world-model startups arguing that gameplay, not robot logs, is the cheapest path to agents that can operate in physical space.
The pitch rests on a single piece of proprietary data. General Intuition was spun out of Medal, a clip-sharing platform run by 31-year-old co-founder and CEO Pim de Witte, which has accumulated hundreds of millions of hours of uploaded gameplay. The clips come with action labels — exact records of which buttons a player pressed and when — that most competitors lack. De Witte argues rivals trying to infer actions from video alone are missing the signal that lets a model separate intent from environment.
“Our agent has been playing for 100 hours straight.”— Kent Rollins, Chief Product Officer, General Intuition
In a demo, an AI agent had been playing a Fortnite-style game for 100 hours straight on the company's R&D floor. The same model was driving a quadruped robot walking around the office, navigating by a single camera. The bot occasionally clipped chair legs and bumped into bins, but it had been fine-tuned with just 8 minutes of real-world robotics data — and that data was collected on the street, not in the office it was now exploring.
Key facts
- 01General Intuition raised $320M at a $2.3B valuation, led by Khosla Ventures with Jeff Bezos, Eric Schmidt, and General Catalyst participating.
- 02Total disclosed funding now stands at $454M, following a $134M round in October 2025.
- 03The startup fine-tuned a quadruped robot using just 8 minutes of real-world data, after pre-training on gameplay from sister company Medal.
- 04Training data comes from hundreds of millions of hours of gameplay clips uploaded to Medal, paired with action labels capturing every button press.
- 05API access is expected to broaden by the end of summer, with most of the new capital going to compute via a CoreWeave deal.
Generalization from simulation to embodiment is the entire thesis. "The same brain powering the agent playing the game is powering the robot," de Witte said. The company internally calls its world model "the gym" — a frame-by-frame generated simulation environment used to train agents, not a product sold to customers. The product is the agent itself.
De Witte calls gameplay the next stage of pre-training, on par with the text dumps that birthed large language models. Walls behave like walls in the simulation. Ladders scale. Shadows lengthen as the sun moves. The model picked up causal physics from gameplay alone, then transferred enough of it to drive a robot through unfamiliar rooms with minimal additional data.
“We have a single model that can respond to Fortnite information on the screen and take action, but also to real-world dynamics in a way that an LLM could never.”— Pim de Witte, Co-founder and CEO, General Intuition
Most of the new $320M will go to compute. General Intuition has a deal with CoreWeave and plans to pre-train the next version of the model, with a slice earmarked for broader API access by the end of summer. Current customers span gaming, simulation, and robotics. The model, de Witte said, works on anything controllable with a game controller or keyboard and mouse — which has already included drones and driving games.
Khosla's bet is that proprietary action data is what makes this a generational company rather than an acquisition target. The firm passed on multiple acquisition offers from a major lab in Medal's history, according to chief of staff Brianna Martin, and isn't entertaining one now. "At this point, it would be a data acquisition, which is sort of uninteresting," Khosla said.
De Witte has drawn hard lines on military use. He spent three years in the humanitarian sector, including with Doctors Without Borders, and says no agents built on General Intuition's models will be deployed to harm humans — though search and rescue is fine. He brought on Martin in part because she publicly quit Palantir over its US Immigration and Customs Enforcement work. "I don't know why Silicon Valley does what it does," he said. "There's a reason I'm not there."
There is also a labor angle. De Witte, who made $1.5M as a teenager hosting a private RuneScape server, launched a platform called Nerve — a jobs marketplace where Medal's gamer base can earn money on data labeling and, eventually, robot teleoperation. The cohort, he notes, is also the one most exposed to AI-driven displacement.
The open question is whether the simulation-to-real-world transfer holds at scale. Eight minutes of fine-tuning to walk a robot through one office is a demo, not a deployment. Competing approaches in robotics have foundered on exactly this gap: models that look fluent in simulation degrade fast when the camera feed is messier, the lighting is different, or the task horizon stretches past a few minutes. Khosla concedes nobody has fully answered whether the transfer holds. General Intuition's wager is that more gameplay data, not more real-world data, is the cheapest way to find out.
If the bet works, General Intuition becomes the model layer beneath a generation of robotics, simulation, and gaming startups — the Anthropic or OpenAI of embodied agents, as de Witte frames it. If it doesn't, Khosla and Bezos have funded an expensive demo. Either way, the round is a marker that investors are now willing to value pre-LLM-scale gameplay corpora at billions, on the theory that the next quantum leap in AI capability won't come from more text. The next 12 months — when paying customers start putting the API into factories, drones, and self-driving stacks — will decide whether intuition is a real capability or a pitch deck word.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
