General Intuition raises $320 million for physical AI

General Intuition raised $320 million at a $2.3 billion valuation last month, backed by lead investor Vinod Khosla, to build general-purpose foundation models for embodied AI — systems, such as robots and autonomous vehicles, that perceive and act in the physical world. CEO Pim de Witte argues embodied AI is set to follow the same trajectory language models took after OpenAI’s GPT-3, shifting from specialized, task-specific systems toward general-purpose foundation models that transfer intuition about movement and interaction across many environments.

General Intuition trained its model on millions of hours of video game data, including which controller buttons players pressed and when — action data that de Witte and Khosla argue is key to developing human-like intuition for spatial-temporal reasoning, or reasoning about space and time. The company has shown its foundation model can play a video game for hours and also power a quadrupedal (four-legged) robot, which required just eight minutes of real-world robotics data to fine-tune. In tests, the robot managed to zero-shot — perform an unfamiliar task without additional training — in a dynamic office setting. De Witte said: “The fact that [the robot] was actually able to zero-shot on just the front camera, with no other sensors, in the office with dynamic objects being introduced and people walking by was a very big surprise to us.” De Witte and Khosla contend that as models gain this baseline reasoning, developers will only need a few minutes of real-world data rather than the hundreds of thousands or millions of hours collected today.

De Witte said many companies currently do specialized work built around individual robot embodiments and environments, and he expects much of that work to become redundant soon as general models emerge — with the model’s own generalization serving as General Intuition’s product. Rather than building robots or vehicles itself, the company wants to become the foundation-model layer beneath physical AI: de Witte said General Intuition isn’t going to build a self-driving car company, but instead aims to make it 10 times easier for the next company that does.

Why it matters

By training on video game action data rather than massive real-world datasets, General Intuition is testing whether physical AI development can become significantly faster and cheaper — a shift that could reshape how robotics and autonomous-vehicle companies build going forward.