General Intuition, a Bezos-backed startup with a $2.3 billion valuation, argues that video game data is superior to text for training AI because it combines information density with spatial and temporal dynamics. The company's proprietary dataset from hundreds of millions of hours of gameplay allows it to pre-train world models that can transfer learning to real-world robotics with minimal fine-tuning, such as getting a quadruped robot to navigate an office after just eight minutes of real-world data.
Summarized by Podsumo
General Intuition raised a $320 million round from investors including Jeff Bezos and Eric Schmidt, valuing the company at $2.3 billion just months after its founding.
The startup's core thesis is that text-based LLMs lack spatial and temporal understanding, while video game data provides a rich, ground-truth record of actions and their outcomes in 3D environments.
The company demonstrated a quadruped robot navigating a dynamic office environment after only eight minutes of real-world post-training data, leveraging its pre-trained world model.
General Intuition is building a 'general pre-trained base' for embodied AI, arguing that the generalization capability of the model itself is the product, reducing the need for vast amounts of real-world data.
The company launched 'Nerve,' a marketplace aimed at creating jobs for gamers, including data labeling and teleoperations, to proactively address AI-driven job disruption.
"Text is a one-dimensional sequence of outputting multidimensional thoughts and emotions. Our models are trained in perceived reality, forcing them to be unbiased."
— Pim DeWitt, CEO of General Intuition
"Every researcher in the world wants to do their best work. And their best work... is always going to come from where's the data."
— Pim DeWitt, CEO of General Intuition
"I don't want to hear how you think that models are going to affect jobs unless you're doing something about it."
— Pim DeWitt, CEO of General Intuition