Why this CEO thinks video games make better training data than the internet
Digital Frontier EditorialJuly 8, 20264 min read
Key Takeaways
General Intuition, valued at $2.3B, bets AGI requires physics intuition — not just language fluency
Gaming data captures cause-and-effect in spacetime; internet text does not
The startup spun out of Medal TV, a gaming platform, giving it proprietary access to millions of gameplay hours
Defense applications loom as an ethical red line the founder acknowledges but won't rule out
The internet is a library. Video games are a laboratory. That distinction sits at the center of General Intuition's $2.3 billion wager on what comes after large language models.
Pim de Witte doesn't hide the ambition. His startup, freshly backed by Bezos Expeditions, Coatue, Eric Schmidt, and researchers from MIT and Google DeepMind, argues that chatbots have hit a ceiling. They predict tokens. They don't predict trajectories. A model that has never seen a ball arc through gravity cannot build a robot that catches one.
The training data problem is simpler than the industry admits. Text describes the world. Gameplay simulates it. Every frame in a first-person shooter encodes position, velocity, collision, occlusion — physics rendered at 60 frames per second across millions of player hours. Medal TV, the streaming platform General Intuition spun out of, already owns that corpus. Proprietary. Labeled. Temporal.
De Witte calls it "world modeling." The phrase sounds academic. The implication is commercial. A model trained on Counter-Strike replays understands cover, flanking, line-of-sight. A model trained on Minecraft understands structural integrity, resource chains, tool use. Minecraft alone has generated more structured physics interactions than every robotics lab combined.
Skeptics will point to simulation-to-reality gaps. Physics engines cheat. Game mechanics simplify. But the gap between "perfect physics" and "game physics" is smaller than the gap between "language statistics" and "any physics at all." LLMs hallucinate forces because they've never experienced one. World models hallucinate less because they've watched cause produce effect ten billion times.
The spin-out logic
Medal TV wasn't built for this. It was built for creators — clipping highlights, hosting tournaments, serving the creator economy. But the exhaust pipe of a platform is often its most valuable asset. YouTube's exhaust became Google's video understanding. Twitch's exhaust became Amazon's computer vision. Medal's exhaust — millions of hours of synchronized gameplay, chat, and metadata — became General Intuition's moat.
The $320 million round buys compute, yes. But it buys time to prove the transfer works. Can a model trained on Elden Ring dodge rolls navigate a warehouse robot around pallets? Can Factorio logistics planning optimize a real supply chain? The investors didn't write checks for maybes. They wrote them for a data advantage no one else can replicate.
The defense question
De Witte doesn't dodge the military angle. He acknowledges it. World models that understand movement, concealment, and coordination will attract defense buyers. The same architecture that helps a drone navigate a forest helps a loitering munition navigate a trench. The founder draws a line — he says — but won't say where. "Ethical red lines" is the phrase he uses. Red lines move when valuations hit ten figures.
This isn't unique to General Intuition. Every physical AI company faces the dual-use trap. The difference: most pretend the trap doesn't exist. De Witte at least names it. Whether he holds the line when a sovereign wealth fund offers nine figures for "border surveillance" is a question the next funding round will answer.
AGI via the arcade
The AGI conversation has stalled on benchmarks. MMLU. GSM8K. HumanEval. These measure linguistic reasoning. They don't measure whether a model can stack dishes without breaking them. General Intuition's bet: the path to general intelligence runs through spatial reasoning, not verbal reasoning. The infant learns physics before language. The model should too.
If they're right, the internet was the wrong training set. Not because it's noisy. Because it's static. A webpage describes a thrown ball. A game engine throws it. The difference is the difference between knowing the word "gravity" and knowing how to catch.
De Witte has the data. He has the capital. He has the pedigree. What he lacks is proof that game physics transfers to atomic physics. The warehouse robot is the test. The battlefield is the temptation. The editorial verdict: watch the warehouse. Ignore the pitch deck. The first deployment tells you everything.