This Startup Thinks Robotics Is About to Have Its ChatGPT Moment

General Intuition CEO Pim de Witte argues that robotics is on the verge of the same paradigm shift that transformed natural language processing after GPT-3. His company has built a physical AI foundation model trained on millions of hours of video game data, including detailed action data recording exactly when and how human players pressed controller buttons.

De Witte told TechCrunch that many robotics companies today are doing redundant specialized work — building bespoke models for individual robot embodiments, environments, and tasks — that will soon be made obsolete by general-purpose foundation models. “The generalization of the model itself is the product,” he said. Once a model develops basic reasoning about space and time, only minutes of real-world data are needed to adapt it to a new robot.

The startup raised $320 million at a $2.3 billion valuation last month on the strength of this thesis, led by prominent investor Vinod Khosla. The core insight is that action data — records of what a human did in a specific situation — is the crucial ingredient for developing human-like spatial-temporal intuition in AI systems.

General Intuition has demonstrated its model’s versatility across two radically different domains: playing a video game for hours, and powering a quadrupedal robot that navigated an office environment after fine-tuning on just eight minutes of real-world robotics data. De Witte noted that the robot’s ability to operate zero-shot using only its front camera — with no other sensors, while people walked by — was a surprise even to the team.

The company’s endgame is not to build robots itself, but to become the foundation model layer for physical AI — a base platform that other robotics companies build upon for their own specific machines. “We’re not gonna build a self-driving car company,” de Witte said, emphasizing that the general-purpose model itself is the product they are selling.

If this approach succeeds, it could fundamentally reshape the robotics industry’s development model: shifting from collecting massive real-world datasets for each robot and scenario, to fine-tuning a capable physical foundation model with minimal task-specific data. This mirrors exactly the transition NLP underwent with the rise of large language models.

Key challenges remain. The gap between simulated gaming environments and the physical world is vast, and the safety and precision requirements of real-world robotics far exceed those of game-playing. The coming months will reveal whether General Intuition can attract enough robotics companies to adopt its platform and whether its model delivers the promised performance at industrial-grade precision levels.