General Intuition raises $320M at $2.3B to train agents on video games
General Intuition just raised $320M at a $2.3B valuation, led by Khosla Ventures, with General Catalyst, Jeff Bezos, Eric Schmidt and racing driver Nico Rosberg in the round. That takes total funding to $454M, eight months after a $134M launch round last October.
The bet is that video games are the cheapest way to teach an agent to act in the physical world. The company sits on hundreds of millions of hours of gameplay from its parent, Medal, and the crucial detail is that the footage comes with embedded action labels, the actual button presses and their timing, not actions guessed from pixels after the fact. That's the whole edge. Most learn-from-video approaches have to infer what the player did; General Intuition already knows. CEO Pim de Witte's pitch is one model that reacts to a Fortnite screen and to real-world dynamics in a way an LLM never could, spatial-temporal reasoning that's supposed to transfer to robotics.
This is the same idea as Qwen-AgentWorld, train in a simulated world and beat real-environment training, but coming at it from the data side instead of the model side. The bottleneck for embodied agents is action-labeled experience, and gameplay is a firehose of exactly that, the kind of clean labeled data nobody else has at this scale. The unproven leap, and the thing worth watching, is whether skills learned dodging in a game actually transfer to a real robot arm. If they do, the data moat here is enormous. Source: TechCrunch, June 25, 2026.
← Back to all articles
The bet is that video games are the cheapest way to teach an agent to act in the physical world. The company sits on hundreds of millions of hours of gameplay from its parent, Medal, and the crucial detail is that the footage comes with embedded action labels, the actual button presses and their timing, not actions guessed from pixels after the fact. That's the whole edge. Most learn-from-video approaches have to infer what the player did; General Intuition already knows. CEO Pim de Witte's pitch is one model that reacts to a Fortnite screen and to real-world dynamics in a way an LLM never could, spatial-temporal reasoning that's supposed to transfer to robotics.
This is the same idea as Qwen-AgentWorld, train in a simulated world and beat real-environment training, but coming at it from the data side instead of the model side. The bottleneck for embodied agents is action-labeled experience, and gameplay is a firehose of exactly that, the kind of clean labeled data nobody else has at this scale. The unproven leap, and the thing worth watching, is whether skills learned dodging in a game actually transfer to a real robot arm. If they do, the data moat here is enormous. Source: TechCrunch, June 25, 2026.
Comments