Startup teaches AI to understand the real world using video games

The American startup General Intuition has announced that it has raised $320 million in investments. The company is developing a universal AI model that learns to navigate the real world using data from video games.

The funding round was led by Khosla Ventures. Other investors include General Catalyst, Jeff Bezos, Eric Schmidt, Nico Rosberg, as well as researchers from Google DeepMind and MIT. Including the previous round, the total volume of attracted investments has reached $454 million, as reported by TechCrunch.

The idea behind General Intuition is that player behavior in video games can serve as a data source for training intelligent agents and robots.

The company grew out of a project called Medal, a platform for sharing gaming video clips. Over years of operation, the service accumulated hundreds of millions of hours of gameplay. However, the main asset turned out to be not the recordings themselves, but information about the players' actions—which keys they pressed and at what exact moment.

According to the founder and CEO, Pim de Witte, it is this specific data that allows the model to better understand cause-and-effect relationships and distinguish its own actions from the surrounding environment.

During a demonstration, the developers showed that the exact same model was controlling a character in a game resembling Fortnite and a four-legged robot navigating autonomously through an office.

According to the team, adapting the model to the robot required only eight minutes of data collected during a walk outside. After that, the robot was able to navigate a completely new indoor environment.

The company also demonstrated its own "world model"—a virtual environment generated by artificial intelligence frame by frame, rather than created by a traditional game engine. Thanks to training on gaming experience, the model already understands the fundamental properties of the surrounding world: walls cannot be walked through, stairs can be climbed, and the position of shadows changes along with the sun.

Unlike many competitors that rely on the costly collection of data from real-world robots, General Intuition expects that video games will significantly accelerate AI training.

The bulk of the new investments will be directed toward expanding computing power and training the next generation of models. In addition, the company plans to make its API available to a broader circle of clients by the end of the summer.

Currently, the technology is being tested by companies in the gaming industry, robotics, and simulation fields. In the future, the model could be used to control robots, drones, and other devices that can be operated via a gamepad, keyboard, or mouse.

General Intuition is betting on an unusual approach to artificial intelligence training: instead of expensive data collection in the real world, the company uses the experience of millions of video game players. Having raised $454 million in total investments, the startup aims to create a universal model capable of operating with equal efficiency in a virtual environment and controlling real robots. While the technology looks promising, its effectiveness in large-scale, real-world scenarios has yet to be proven.

Follow NEWS.am Tech on Facebook and Twitter