General Intuition just landed $320 million in Series B funding to scale what might be AI’s most unconventional training method yet – millions of hours of video game footage. The round, led by Khosla Ventures, values the startup at $2.3 billion and validates a contrarian bet that action data from gaming can teach AI agents something closer to human intuition. While most AI labs obsess over scraping text and images, General Intuition is building world models from the physics, decision-making, and spatial reasoning embedded in gameplay.

General Intuition is making a bold claim – that the secret to smarter AI isn’t more text scraped from the internet, but millions of hours watching someone play Grand Theft Auto. The startup just closed a $320 million Series B led by Khosla Ventures, pushing its valuation to $2.3 billion and cementing one of the year’s most unconventional AI training approaches.

The thesis is deceptively simple. While OpenAI, Google, and Meta have spent years training large language models on text and images, General Intuition believes the real path to artificial general intelligence runs through action data. Video games, with their rich physics simulations, complex decision trees, and real-time spatial reasoning, offer something traditional training data can’t – a sandbox for learning how the world actually works.

“We’re not trying to build another chatbot,” according to sources familiar with the company’s pitch to investors via TechCrunch. “We’re building world models that understand cause and effect, physics, and consequences. That’s what gameplay gives you that static images never will.”

The approach puts General Intuition at the center of a growing debate in AI research about what kind of training data actually matters. Text-based models excel at language tasks but struggle with physical reasoning. Image models can recognize objects but don’t understand how they interact. Gaming data, the company argues, captures the temporal dynamics and action sequences that embodied AI agents desperately need.

This matters because the next frontier for AI isn’t just smarter chatbots – it’s robots that can navigate warehouses, drones that can inspect infrastructure, and autonomous systems that need to predict what happens next in three-dimensional space. General Intuition’s models are designed to power these physical AI agents by giving them an intuitive understanding of how objects move, how forces interact, and how decisions ripple through environments.

The $320 million round signals that major investors are taking this approach seriously. Khosla Ventures has been an early backer of transformative AI companies, and its lead position here suggests confidence that gaming data represents an untapped training goldmine. The funding will reportedly go toward scaling the company’s data infrastructure, expanding its library of gameplay footage across dozens of game titles, and building out partnerships with game developers and esports platforms.

What makes the strategy particularly clever is the sheer volume of available training data. Gamers worldwide generate billions of hours of gameplay footage annually, much of it already recorded and available through streaming platforms. Unlike robotics companies that need to painstakingly collect real-world data through expensive hardware deployments, General Intuition can tap into an existing ecosystem of high-quality action sequences.

The company is also developing what it calls “world models” – AI systems that can simulate forward in time to predict outcomes. Show the model a video game character jumping toward a platform, and it should be able to predict whether they’ll make the landing, what happens if they miss, and how the physics will play out. Apply that same predictive capability to a warehouse robot, and suddenly you have an agent that can plan complex movements without trial-and-error learning in the real world.

Competitors are taking notice. While General Intuition focuses on gaming data, other startups are exploring simulation environments and synthetic data generation to train embodied AI. Nvidia has invested heavily in its Omniverse platform for physics simulation, while research labs experiment with training robots in virtual environments before deploying them physically. The race is on to crack the code of spatial intelligence.

But the gaming approach isn’t without skeptics. Critics point out that video game physics, no matter how sophisticated, still differ from real-world dynamics. A robot trained on Grand Theft Auto’s driving mechanics might struggle when actual friction, momentum, and sensor noise enter the equation. General Intuition’s counterargument is that the models learn general principles of causality and prediction that transfer across domains, not just game-specific rules.

The $2.3 billion valuation also raises questions about defensibility. Unlike large language models that require massive compute infrastructure and proprietary datasets, gameplay footage is relatively accessible. What stops Google or OpenAI from spinning up their own gaming-based training pipelines if this approach proves successful? General Intuition’s edge likely lies in its early lead on model architecture, data curation techniques, and partnerships that give it preferential access to high-quality footage.

General Intuition’s massive raise puts a spotlight on one of AI’s most interesting open questions – whether intelligence emerges from language and images, or from understanding actions and consequences in physical space. The $2.3 billion bet from sophisticated investors suggests the gaming approach has real merit, especially as AI moves from chatbots to robots that need to navigate the real world. If the company can prove that millions of hours of Fortnite and Minecraft translate to smarter warehouse robots and autonomous drones, it won’t just validate an unconventional training methodology – it’ll reshape how the entire industry thinks about building embodied AI. The next 18 months will reveal whether this gamble pays off or becomes a cautionary tale about overhyping novel approaches. Either way, expect every major AI lab to start experimenting with gameplay data before year’s end.