general intuition’s $2.3b bet that video games can train ai agents for the real world

At a glance:\n- General Intuition raised $320 million, valuing the startup at $2.3 billion after a $454 million total funding run.\n- The company trains AI agents using millions of hours of labeled gameplay from its parent platform Medal, enabling transfer from virtual worlds to real‑world robotics.\n- Its first embodied demo is a quadrupedal robot that learned to navigate a New York office from just eight minutes of street‑captured data.\n\n## The gameplay‑to‑robot pipeline\nGeneral Intuition’s core innovation lies in turning recorded video‑game footage into a rich, action‑labelled dataset. The company’s parent, Medal, lets gamers upload and share clips, producing hundreds of millions of hours of gameplay. Unlike competitors that try to infer actions from raw video, de Witte emphasises that the action labels—exact button presses and timings—are the key to teaching an AI that “the world is not just pixels but a set of controllable events.”\nThe labelled data is fed into a world‑model that learns spatial‑temporal reasoning: how objects move, how ladders can be scaled, how shadows lengthen as the sun shifts. In a demo, a virtual agent walked straight into a series of walls and stopped, whereas other agents sometimes passed through. The model appears to have internalised that “walls are walls,” a concept it learned from millions of hours of gameplay.\n\nThe founding team includes:\n- Pim de Witte, co‑founder and CEO\n- Eloi Alonso\n- Adam Jelley\n- Vincent Micheli\n\n## From simulation to the street\nThe real‑world test came in the form of a large quadrupedal robot that de Witte’s team brought into the New York office. “The same brain powering the agent playing the game is powering the robot,” the CEO said. The robot’s default mode was “exploration,” and with just eight minutes of real‑world data—captured on the street outside the office—it was able to navigate around chairs, avoid trash bins, and continue into the office. The data was not collected inside the office, showing that the agent can generalise from one environment to another with minimal additional training.\n\n## Investor confidence and funding\nOn Thursday, General Intuition announced a $320 million round that valued the company at $2.3 billion, bringing its total disclosed funding to $454 million after the $134 million launch round in October. The round was led by Khosla Ventures, with participation from:\n- General Catalyst\n- Jeff Bezos\n- Eric Schmidt\n- Nico Rosberg\n- Researchers at Google DeepMind and MIT\nThe bulk of the new capital will scale compute capacity through a deal with CoreWeave and focus on pre‑training the next version of the model. A slice of the funding will make the API broadly available by the end of summer.\n\n## Vision for a generational ecosystem\nDe Witte sees General Intuition as an ecosystem enabler rather than a single‑product company. “We’re not gonna build a self‑driving car company,” he said. “We’re gonna make it 10 times easier for the next person to build a self‑driving car company.” The company already has customers in gaming, simulation, and robotics, and plans to test its model in a digital twin of a factory floor, on a human‑like bot inside a gaming studio, and on a quadruped navigating hazardous environments. The goal is to create a data flywheel where each new customer provides real‑world data that feeds back into the model, improving it for everyone.\n\n## Nerve and the future of work\nGeneral Intuition recently launched Nerve, a jobs marketplace that lets gamers earn money using their existing setups. Users begin with data‑labeling tasks and can eventually move toward robot teleoperation and other high‑skill roles. De Witte, who once earned $1.5 million from a private RuneScape server, says the platform is a way to give the generation most exposed to AI‑driven displacement a stake in the future.\nThe company maintains a clear ethical stance: no agents will be employed to harm humans. “We don’t want to be an escalatory part of the system,” de Witte said. “Let’s say I were to come out and say, ‘We’re doing lethal autonomy.’ What do you think would happen in other countries?” He adds that the models can be used for search and rescue missions and other benevolent applications.