Your gaming data could be the secret to AGI, according to this Bezos-backed startup

We have spent the last two years obsessed with text. We have scraped every digital book, every forum post, and every snippet of code to feed the insatiable appetite of Large Language Models. But lately, there is a sense among founders that we are hitting diminishing returns. You can only rearrange the dictionary so many times before you realize that Shakespeare knew how to write, but he didn't know how to navigate a physical workspace or catch a falling glass.

The Spatial Intelligence Gap

The current crop of AI—your Claudes, your GPTs—are essentially world-class mimics. They understand the statistical probability of the next word in a sequence. What they do not understand is physics. They don't have a concept of gravity, momentum, or the way an object occupies space over time. This is the primary reason why AI video generation still produces nightmare fuel where hands morph into table legs and people walk through walls.

For those of us building in the space, this is the wall. If we want Artificial General Intelligence (AGI) to actually do things in the real world—like operate robotics or navigate complex physical environments—it needs to understand how the world moves. Text can't teach that. Video can teach it to a degree, but video is flat. It lacks the underlying math of 3D interaction.

Why Gaming is the New Library of Alexandria

This brings us to a new venture called General Intuition, which recently caught the attention of Jeff Bezos. Their thesis is simple: the secret to reaching AGI isn't more internet comments; it’s gaming data. When you play a modern video game, you aren't just looking at pictures. You are interacting with a physics engine. Every time a character jumps, every time a car drifts around a corner, and every time a building collapses in a shooter, there is a set of hard mathematical rules governing that movement.

Games provide a sandbox where intelligence can be tested against reality—or at least a high-fidelity simulation of it. Unlike the static data used to train LLMs, gaming data is interactive. It’s a loop of action, feedback, and consequence. If an AI agent tries to walk through a wall in a game, the physics engine stops it. That is a data point you simply cannot get from reading a PDF about architecture.

The Founder Perspective: Scaling Realism

As a founder, I look at this and see a shift in the supply chain of intelligence. For years, we’ve been told that "data is the new oil," but we’re starting to realize that some oil is low-grade. Text is low-grade for spatial reasoning. Gaming data is high-octane. It’s dense, it’s structured, and most importantly, it’s labeled by the very nature of the game’s code.

Building an AI that can "dream" in three dimensions requires a level of spatial awareness that LLMs currently lack. If General Intuition is right, they are building a model that understands the world like a toddler does—by bumping into things and seeing how they fall. This is the bridge between a chatbot and a robot that can actually fold your laundry without breaking the table.

• Spatial Logic: Understanding how objects move in 3D space is the next frontier.
• Feedback Loops: Gaming allows for billions of simulated hours of trial and error.
• Data Quality: Physics engines provide ground-truth data that text-based web scraping lacks.

The Skeptic’s Corner

Now, let's keep it real. We’ve seen the "simulation to reality" (Sim2Real) problem before. Just because an AI can drive a car perfectly in Grand Theft Auto doesn't mean it won't panic when it sees a real-world plastic bag blowing across a highway. Simulations are clean; the real world is messy, dirty, and unpredictable. There is also the question of data moats. Most of the best gaming data is owned by giants like Sony, Microsoft, and Epic. A startup trying to scrape or license this data faces a massive uphill battle against entrenched incumbents who might decide to build their own spatial models tomorrow.

Furthermore, Bezos backing a project is a signal of capital, not necessarily a guarantee of technical success. We have seen plenty of well-funded moonshots evaporate when they hit the hard physics of the real world. The hype around AGI often ignores the sheer computational cost of simulating high-fidelity physics at the scale needed to train a foundational model.

What This Means for Builders

If you are building in the AI space right now, the takeaway is clear: stop looking for more text. The low-hanging fruit of language has been picked clean by the giants. The real opportunity lies in multimodal models that can interpret and interact with physical environments. Whether it’s through synthetic data, gaming engines, or specialized sensors, the winner of the next phase of AI will be the one who teaches the machine how to move.

The move from 'predicting words' to 'predicting physics' is the most significant pivot in the industry since the transformer paper.

We are moving away from the era of the "smart librarian" AI and toward the era of the "capable apprentice." An apprentice needs to know how tools work. They need to know that if they drop a hammer, it falls down, not up. Gaming data provides the most cost-effective way to teach those lessons at scale. It’s a pragmatic approach to a very difficult problem.

The Bottom Line

We shouldn't expect an AGI to emerge from a game of Minecraft tomorrow. However, the logic behind using gaming data is sound. It addresses the fundamental weakness of LLMs: their lack of grounding in the physical world. For founders, the move is to watch how these spatial models develop and think about how physical reasoning can be applied to your specific niche. The era of the chatbot is peaking; the era of the spatial agent is just beginning.

Read the original at TechCrunch AI →