Why this CEO thinks video games make better training data than the internet

We have spent the last few years feeding every scrap of digital text into massive black boxes. If a human wrote it on a blog, a forum, or a social media platform, it is probably living inside ChatGPT or Claude right now. We have built world-class mimics that can summarize a meeting or write a decent email, but we haven't built anything that actually understands how the world works. Large Language Models (LLMs) are essentially just sophisticated calculators for words. They are great at predicting the next syllable, but they are functionally blind to the physical laws that govern our existence.

The Problem with Text-Only Intelligence

If you want to build Artificial General Intelligence (AGI), you cannot just rely on tokens and scrapings from Wikipedia. There is a massive disconnect between knowing the definition of gravity and understanding how an object falls. Builders in the space are starting to realize that text is a secondary representation of reality. It is a filtered, often messy version of what humans perceive. When we rely solely on text for training data, we are trying to teach a machine to see the world through a keyhole.

This is where General Intuition and other forward-thinking teams are pivoting. They are looking at video games as the primary source of training data. It sounds counterintuitive to some, but to a founder, it makes perfect sense. Games are structured simulations. They have physics engines, cause-and-effect loops, and spatial continuity. Most importantly, they have goals. In a game, an agent has to interact with an environment to achieve a result. It has to understand that if it walks into a wall, it stops. If it drops an item, that item occupies a specific coordinate in space.

Why the Internet is a Bad Teacher

The internet is a warehouse of static information. It is also increasingly becoming a warehouse of AI-generated slop, which creates a feedback loop that dilutes the quality of new models. This is the model collapse problem we have all been hearing about. When an AI learns from AI, it loses the nuance of human experience.

Video games offer a different kind of data: behavioral data. When you watch a player navigate a complex environment in a game, you aren't just seeing pixels. You are seeing a series of decisions based on physical constraints. Games provide a sandbox where the rules are consistent. This consistency is exactly what machine learning models need to develop what we call common sense. A text model might tell you that a glass breaks when dropped, but a model trained on game physics understands the speed, the impact, and the resulting fragments in a three-dimensional plane.

The Move From Chatbots to Agents

For builders, the shift toward gaming data marks the transition from chatbots to functional agents. If we want an AI to do more than just talk—if we want it to actually operate software, manage robotics, or navigate complex workflows—it needs a sense of spatial reasoning. You cannot learn how to drive a car by reading 10,000 manuals; you have to see how the car reacts to the road. Video games are the closest thing we have to a high-fidelity, high-speed laboratory for these interactions.

• Spatial Awareness: Games teach models about depth, occlusion, and navigation in ways a text file never could.
• Temporal Logic: Understanding that action A leads to result B over a specific period of time.
• Safe Failure: You can crash a car in a simulation a billion times until the AI learns the edge cases of safety, something impossible to do in the real world.

The Founder's Perspective on Synthetic Data

There is a lot of hype around synthetic data right now, and for good reason. We are running out of high-quality human text. But not all synthetic data is created equal. Creating fake essays to train a model is a dead end. Creating a rich, 3D simulation where an agent can live, fail, and learn is a different story entirely. This isn't just about entertainment; it is about building a foundation of logic that isn't dependent on human language quirks.

As someone who has looked at dozens of AI startups over the last year, the ones that impress me aren't the ones finding new ways to scrape Reddit. They are the ones building proprietary environments. If you own the simulation, you own the data stream. You aren't at the mercy of a platform's Terms of Service or the fluctuating quality of public discourse. You are building a ground-truth reality for your AI to inhabit.

The gap between predicting words and understanding physics is the final frontier for AGI. If we cannot bridge that, we are just building better search engines.

What This Means for the Future of Development

We are likely headed toward a world where the best AI models aren't trained on the 'open web' but on private, high-fidelity gaming engines. This creates an interesting moat for companies with access to large gaming libraries or the compute power to run massive simulations. It also means that the skills required to build the next generation of AI are shifting. We need fewer linguists and more world-builders and physics engineers.

If you are building in this space, stop obsessing over how to make your LLM sound more human. Start thinking about how to make it understand the cabinet it is trying to organize or the warehouse it is trying to manage. The intelligence we are looking for isn't hiding in a library; it is hiding in the way things move, break, and interact.

The Takeaway

The internet taught AI how to talk, but video games will teach AI how to think. For founders and builders, the opportunity lies in the logic of simulations. We are moving past the era of 'Big Data' and into the era of 'Rich Data.' The companies that can effectively translate the physics of a game world into the logic of a machine will be the ones that actually deliver on the promise of AGI.

Read the original at TechCrunch AI →