Why General Intuition thinks video games are better training data than the web

The internet has been the default feedstock for modern AI, but General Intuition is making a different bet: if you want models that understand motion, interaction, and cause and effect in the physical world, text scraped from the web only gets you so far.

That’s the thesis behind the New York startup, which just closed a $320 million round at a $2.3 billion valuation. Backers include Bezos Expeditions, Coatue, Eric Schmidt, plus researchers tied to MIT and Google DeepMind. The company spun out of Medal TV, the gaming clip platform, and CEO Pim de Witte says its edge comes from data the AI industry has mostly treated as secondary: gameplay.

That sounds contrarian until you look at what today’s models still miss.

Text is cheap. Physics is hard.

Large language models are good at the internet because the internet is mostly language. They handle syntax, summarization, code completion, and broad pattern matching across huge corpora. The physical world doesn’t behave like that.

A robot arm, a drone, a warehouse picker, even a game character moving through a 3D environment has to deal with state changes over time. Objects block each other. Motion has inertia. Actions have consequences a few frames later, not just in the next token. If a model is going to plan in that setting, it needs some notion of world state and temporal dynamics, not just token prediction.

That’s where video games get interesting. Games produce dense interaction data: controller inputs, screen state, rewards, failures, retries, partial observability, multi-step planning, and environment rules that can be simulated at scale. Unlike random web video, game data often comes with a built-in action loop. You can see what the agent did, what changed, and whether it worked.

That matters.

If you’re training a model to reason about the physical world, you want data where action and outcome are tightly coupled. The web is full of passive observation. Games are active systems.

Why game data appeals to model builders

The obvious appeal is scale. Games generate huge volumes of interaction traces, and many of them are already structured enough to be useful. A gameplay recording isn’t just pixels. It can include:

• frame-by-frame visual state
• player actions
• timing information
• reward signals or scores
• map layouts and object interactions
• repeated attempts under similar conditions

For training, that gets you closer to a supervised or reinforcement-learning dataset than a random pile of video clips.

There’s also diversity. Games cover navigation, aiming, prediction, sequencing, resource management, team coordination, and long-horizon planning. Those aren’t toy skills if your end goal is a robot, a drone, or an AI agent that needs to act under uncertainty.

The deeper argument General Intuition seems to be making is that world models may train better on environments where the rules stay consistent and action is visible. In a game, a jump means the same thing every time within that physics engine. On the open web, the world is messier, noisier, and much harder to label.

That doesn’t make games a perfect stand-in for reality. It does make them a better laboratory.

The catch: simulation has a bias problem

Synthetic and simulated environments always tempt model teams into overfitting to the simulator. Games have their own physics, visual style, and incentives. A model that learns Minecraft or Counter-Strike dynamics doesn’t automatically know how to manipulate a warehouse pallet or open a kitchen drawer.

That gap is the whole problem with simulation-to-real transfer. Robotics has wrestled with it for years, and it never quite goes away. You can train on games to build stronger priors about motion and planning, but the last mile still needs real-world data, calibration, and careful evaluation.

There’s another limitation startup pitches tend to skip over: game data is abundant, but not always cleanly licensed for machine learning. If a company wants a commercial training pipeline, provenance matters. So does consent. So do platform terms. The AI industry has already picked enough fights over scraped web data. It doesn’t need another one built on gameplay recordings and user-generated content.

Medal TV is a sensible origin story

General Intuition didn’t start as a grand AGI lab that stumbled into gaming. It came out of Medal TV, a platform built around capturing and sharing gaming clips. That origin makes more sense than it first sounds.

If you run a clip platform, you already sit on a firehose of multimodal data. You see gameplay video, user behavior, metadata, and often enough contextual signals to reconstruct what happened in the session. That’s useful raw material if your goal is to build models that understand sequences of actions instead of just frames on a screen.

The spinout suggests General Intuition isn’t just buying the “games are cool” pitch. It’s betting on data plumbing. Gathering the right traces, organizing them, and turning them into training corpora is usually harder than people think. The model architecture gets the headlines. The dataset engineering decides whether the thing is usable.

That part gets overlooked in a lot of AI startups. The best data wins because the model can only learn from what it sees. If the input stream is noisy, repetitive, or weakly grounded, scaling the transformer doesn’t fix it.

The real competition is for embodied AI

General Intuition is talking about AGI, but the nearer market is physical AI. That bucket now covers robotics, autonomous systems, industrial automation, warehouse orchestration, and agentic control systems that need to understand space and time.

For builders in that world, the question isn’t whether language models are useful. They are. The question is whether text-only pretraining gets you far enough. Probably not.

A robot vacuum doesn’t need to quote Kafka. It needs to know that a chair leg is an obstacle, that a dropped object may roll under a table, and that a path that looks clear in one frame can be blocked in the next. Those are world-model problems.

The strongest labs are already converging on multimodal stacks:

• language for instruction following and planning
• vision for scene understanding
• action traces for control
• simulation or video for temporal reasoning
• real-world telemetry for grounding and correction

General Intuition’s pitch sits in the middle of that stack. If it works, the payoff isn’t just better game agents. It’s stronger priors for agents that have to operate in messy environments where mistakes are expensive.

Defense is where the ethical questions get ugly

De Witte also raised defense applications, and that’s not a side note. It’s the uncomfortable part.

Models trained to understand movement, pursuit, and spatial behavior can be adapted for civilian and military systems alike. That dual-use problem is baked into the category. A world model that improves robotic navigation can also support surveillance, targeting, or autonomous coordination. The same capability that helps a warehouse bot avoid collisions can help a system track moving objects in contested environments.

That means the usual AI safety slogans won’t cut it. If General Intuition’s models become valuable, the company will need real controls around deployment, customer screening, data governance, and model access. Not a policy page. Actual controls.

The defense angle also changes how investors and customers will read the company. Some partners will see strategic importance. Others will see export risk, procurement friction, and reputational headache. That’s the trade-off when your core technology sits close to autonomy and real-world action.

What developers should pay attention to

A few practical takeaways matter here.

First, data modality is becoming as important as model size. If your team still assumes “more text” is the main path forward, you’re probably behind.

Second, action-labeled video will keep rising in value. That includes gameplay, robotics telemetry, screen recordings with input traces, and any dataset where observation can be tied to behavior. If you’re building data pipelines, the structure of those logs matters more than ever.

Third, there’s a split forming in the AI stack. Language models will keep owning general reasoning and interface work. World models may become the layer that handles dynamics, simulation, and control. Those layers won’t stay separate forever, but they’re different engineering problems today.

And finally, the best AI companies are often data companies with a model attached. Boring, but true. Compute gets the attention. Data quality decides whether the system learns anything worth keeping.

General Intuition is betting that gaming data gives it a cleaner path to that kind of learning than the internet ever did. That’s a serious claim, and it deserves more skepticism than startup cheerleading usually allows. But it doesn’t sound crazy anymore.