A startup called General Intuition has raised $320 million to develop AI agents—software that learns to think and act—by training them on video game data instead of the usual approach of feeding them internet text. The company is now valued at $2.3 billion, according to GamesBeat and TechBuzz.ai, both reporting on 25 June 2026.

The idea is this: video games provide a better training ground for AI agents than traditional sources. In a game, an agent sees what is happening on screen, makes a decision (move left, attack, build a structure), and gets immediate feedback on whether that choice worked. This mirrors how humans learn through trial and error. Internet text, by contrast, is passive information—it tells you facts but does not teach you to make decisions under pressure. Games do both: they reward good decisions and punish bad ones in real time.

This is not a completely new idea. Companies like DeepMind used game environments to train AI systems that learned to play Go and StarCraft at world-beating levels. What General Intuition is doing is taking that approach and scaling it up—using millions of hours of gameplay from many different games to train a large, general-purpose AI system that could eventually help solve real-world problems, not just play games.

The engineering work required is substantial. Video game data is messy and varied. Each game has different graphics, different rules, different kinds of decisions an agent must make. A strategy game like StarCraft requires long-term planning; a first-person shooter like Counter-Strike requires quick reflexes and spatial awareness. Mixing all that data together and teaching a single AI agent to learn from it is like teaching a person to become skilled at basketball, chess, and cooking all at once by watching videos of people doing each. It is possible in theory, but the machinery to make it work is complex.

The $2.3 billion valuation is bold. The company has not yet released a working product or published independent proof that agents trained this way actually perform better than agents trained the old way. Investors are making a bet on the team's ability to pull it off.

There is, however, a genuine open question: games are not the real world. In a game, the computer knows exactly what is happening at all times, and the rules never change unexpectedly. In the real world, information is incomplete and surprises are constant. An agent that becomes expert at playing chess or Minecraft has learned a lot, but what it has learned might not transfer well to, say, robotic navigation or medical diagnosis. Whether game-trained agents will actually work in practice is something the research community will want to verify once General Intuition publishes its results.

The funding reflects a broader shift in the AI field. Researchers have been training large language models on ever-larger amounts of internet text, but that text is becoming scarce—there is only so much text on the internet, and it is getting harder to find new, high-quality material. Game data is abundant, constantly generated by millions of players, and it has a built-in signal: you can tell whether a decision was good or bad by the outcome. For these reasons, many AI researchers now think games might be a better source of training material than more internet text.