🎧 Prefer to listen?

I’ve been following General Intuition since their initial $134M raise last year — the startup that trains AI agents using video game clips from Medal’s 10 million monthly users. The pitch was interesting: use gameplay data to teach AI how to navigate the physical world. But I had the same question everyone else did: does watching someone play Fortnite actually translate to a robot walking across an office? Turns out, it does. And the reason why tells you something important about where AI training is headed.

If you’ve been paying attention to AI agents at all, you know the bottleneck isn’t intelligence — it’s understanding. LLMs can write code and answer questions, but they can’t navigate a room or figure out how a door works. General Intuition thinks gameplay data solves that problem. Here’s how.

The Problem with Training AI on Regular Video

Most companies trying to build spatial AI agents — models that understand how to move through physical space — train on regular video. Watch enough footage of people walking, driving, and interacting with objects, and the AI should figure out how the world works, right?

Not exactly. The issue is that regular video only shows you what happens. It doesn’t show you why. You can watch someone walk through a door a thousand times, but you don’t know what they did with their muscles, their weight shift, their hand on the handle. You see the result, not the action.

This is the gap General Intuition’s CEO Pim de Witte calls the difference between “watching someone drive and actually feeling the steering wheel.” Video alone can teach an AI to recognize patterns, but it can’t teach causality — the relationship between an action and its consequence.

That’s why most spatial AI models struggle with novel environments. They’ve learned to recognize patterns in training data, but they haven’t learned the underlying logic of how actions create outcomes. When they encounter something new — a different room layout, an unfamiliar object — they fall apart.

What Makes Gameplay Data Different

Here’s where video games change the equation. When someone records a gameplay clip on Medal, the recording doesn’t just capture the video — it captures the action labels. Every button press, every joystick movement, every click is timestamped and logged alongside the footage.

So instead of just seeing “the character walked through a doorway,” the AI can see “the player pressed W for 1.2 seconds, moved the mouse 15 degrees right, then pressed E to interact with the door handle.” That’s a fundamentally different training signal. It’s the difference between watching a cooking show and actually following a recipe with exact measurements.

General Intuition’s model trains on hundreds of millions of hours of these annotated gameplay clips. The action labels let the AI learn the relationship between decisions and outcomes — what de Witte calls understanding “self” versus “environment.” The model doesn’t just see a virtual world; it understands how its own actions change that world.

This is why, during a TechCrunch visit to their R&D floor, the same model that played Fortnite for 100 hours straight could also power a quadrupedal robot navigating an office. The spatial reasoning transfers because the model learned causality, not just pattern recognition.

From Fortnite to Walking Robots

The demo that convinced me this isn’t just hype: General Intuition showed TechCrunch reporters an AI agent that played Fortnite continuously for 100 hours. Then they loaded the same model onto a quadrupedal robot with a single camera. The robot walked around the office, navigated around chairs, and fine-tuned itself to the new environment in just eight minutes of real-world data.

Eight minutes. That’s the part that got my attention. Traditional robotics requires thousands of hours of real-world data collection — expensive, slow, and dangerous (you can’t exactly let a robot trial-and-error its way through a hospital). General Intuition’s approach uses gameplay as a pre-training environment, then fine-tunes with a tiny amount of real-world data.

The robot wasn’t perfect — it bumped into chairs like a toddler learning to walk — but it was navigating a space it had never seen before using a single camera. No LIDAR, no depth sensors, no pre-built map. Just the spatial understanding it picked up from millions of hours of virtual navigation.

This is what spatial AI agents look like in practice. Not chatbots that answer emails, but models that understand space, time, and cause-and-effect well enough to control a physical body.

The Nerve Platform: Gamers as AI Trainers

There’s a business angle here that most coverage misses. General Intuition didn’t just build a model — they built a marketplace. Nerve is a jobs platform where gamers earn money using their existing hardware. Users start with data labeling tasks (annotating gameplay, validating AI outputs) and can eventually move into robot teleoperation and higher-skilled work.

The logic is elegant: Medal’s user base is the generation most exposed to AI-driven job displacement. Instead of extracting their data and building models that might replace them, Nerve creates a pathway for gamers to participate in the AI economy. As the models get more capable, they’ll need human oversight for edge cases, safety validation, and real-world data collection.

It’s a smart business move too. Having a ready-made workforce that already understands spatial environments — because they’ve been navigating virtual ones for years — is a competitive advantage that’s hard to replicate. These aren’t random crowdworkers; they’re people with thousands of hours of spatial reasoning experience, even if it was built in-game.

Who’s Betting on This Approach

General Intuition’s latest round — $320 million at a $2.3 billion valuation — was led by Khosla Ventures, with participation from General Catalyst, Jeff Bezos, Eric Schmidt, Nico Rosberg, and researchers from Google DeepMind and MIT. Total disclosed funding: $454 million.

That’s a lot of smart money for a company less than a year old. What caught my attention isn’t the dollar amount — it’s who’s writing the checks. When the same people who backed OpenAI and Anthropic are investing in a company that learns from Fortnite, it signals something about where the industry thinks the next breakthrough lives. And it’s not in bigger language models.

The vast majority of the new capital goes toward compute — they have a deal with CoreWeave and plan to focus on pre-training the next version of their model. A slice is earmarked for making their API available by the end of summer, which means developers outside General Intuition will soon be able to build on top of their spatial reasoning model.

What This Means for Solo Builders

If you’re a solo builder or small business, you’re probably wondering why this matters to you. Fair question. Here’s why I think it does:

The API is coming. When General Intuition opens their API this summer, developers will be able to build spatial reasoning into their own applications. Imagine a robot vacuum that actually understands your furniture layout, or a drone that can navigate a warehouse without pre-mapped routes. The applications for small businesses — especially in logistics, delivery, and automation — are significant.

The training data model is replicable. You don’t need General Intuition’s specific dataset to benefit from this approach. The insight — that action-labeled data beats raw video for training spatial AI — applies to any domain where you can capture both observations and actions. Game developers, simulation platforms, and even VR applications could generate similar training data.

It changes the timeline. If eight minutes of real-world data is enough to fine-tune a model for a new environment, the barrier to building physical AI applications drops dramatically. You don’t need a massive data collection operation. You need a good pre-training dataset (which gameplay provides) and a small amount of domain-specific data.

The Competition

General Intuition isn’t alone. Cambridge-based Worldmodeldata recently raised £7 million to turn licensed video-game data into training sets for AI world models. San Francisco’s Origin Lab raised $8 million in May for a similar approach — licensing gameplay data from studios to sell to AI labs.

But General Intuition has an advantage none of them do: Medal. With 10 million monthly users uploading gameplay clips, they have a built-in data pipeline that competitors can’t easily replicate. It’s the same kind of network effect that made platforms like YouTube dominant — the more users contribute, the better the training data, the better the models, the more users want to participate.

The bottom line

General Intuition’s bet is that video game data — specifically gameplay with action labels — is the fastest path to AI agents that can navigate the real world. The $2.3B valuation says smart money agrees. The robot demo says the technology works. And the Nerve platform says they’re thinking about the human side too. If you’re building anything that involves spatial reasoning, physical automation, or robotics, keep an eye on their API launch this summer. For more on how AI agents are changing the landscape, check out the tools I actually use every day or visit /start-here/ to see what’s possible.