How can video games train AI agents? General Intuition's $2.3B bet

An AI startup has completed a major funding round to train AI agents using data from video games. According to TechCrunch, the company, called General Intuition, raised $320 million and reached a valuation of $2.3 billion with the round.

The company's central idea is that video games can be a rich source of training data for AI. Games are interactive environments that contain movement, decisions and outcomes. How a character moves, how it deals with obstacles and how it reaches goals creates a continuous stream of "action data" generated throughout play.

The best-known AI models today are largely trained on text. These models are very capable at understanding and producing language, but they do not have the same natural competence at tasks that require moving and acting in the physical world. What is missing here is a practical intuition for interacting with the world.

General Intuition's bet is aimed precisely at filling this gap. The company argues that action data in millions of hours of gameplay recordings can help AI learn how to behave in an environment. The aim, then, is less about learning to play a game than about extracting a general intuition of "how to move" from games.

The appeal of this approach lies in the abundance of data. Games offer controlled but varied scenarios: behaviours such as jumping, avoiding, planning and homing in on a goal are repeated many times over. This means that action data — which is expensive and slow to collect in the real world — can be obtained far more quickly and safely in a digital environment.

The ultimate goal is to carry this intuition into real-world applications. The company's vision is that behaviour patterns learned in games could be transferred to robotics and other physical systems. If an AI can learn to overcome obstacles and reach goals in a virtual environment, the hope is that some of these skills can be generalised to the real world.

That said, the move from games to the real world is not easy. Virtual environments do not fully reflect the complexity and unpredictability of the real world. Whether strategies that work in a game also hold in the physical world is an open research problem that is being worked on intensively.

Investor interest, for its part, is part of the broader expectation around AI agents. AI systems referred to as "agents" are defined as systems that can not only answer questions but independently carry out tasks in an environment. This field is thought to be a candidate for the next big wave of AI.

Experts say such investments are also a sign of how much data diversity the sector is moving toward. As the limits of text data are approached, companies are turning to new sources such as video, game and sensor data. Games stand out in this search as both an abundant and a structured source of data.

The investment reported by TechCrunch is an example of AI's effort to expand from text toward action. Whether General Intuition's bet pays off will be seen over time, but the idea underlines a growing interest in the view that games can be not only entertainment but also a learning ground for AI.