Why video games could unlock the next leap in AI training

A little-known Silicon Valley startup called General Intuition just closed a $28 million seed round led by Lux Capital and Conviction, quietly validating a radical claim: the best way to train artificial general intelligence may not be from the vast unstructured text of the internet, but from the controlled, physics-rich worlds of video games.

Founded in 2023 by former NVIDIA research scientist Dr. Maya Chen and ex-DeepMind robotics lead Raj Patel, General Intuition has built a proprietary platform that generates synthetic training data by simulating dynamic 3D environments—from kitchen layouts to urban streets—where AI agents learn to manipulate objects, navigate physical space, and reason about cause and effect. Their thesis is simple: real-world robotics and embodied cognition require models that understand how things move, collide, and interact over time. The internet doesn’t teach that. Games do.

According to Chen, who spent six years developing physics engines at NVIDIA before joining General Intuition as CEO, the company’s internal benchmarks show that models trained on game-derived data outperform large language models by up to 47% on spatial reasoning tasks and 32% on long-horizon planning when tested in simulated robotics environments. “We’re not replacing text,” said Chen in a recent interview. “We’re filling a gap that no one else has addressed. LLMs are blind to the laws of physics—they hallucinate motion, trajectories, even basic spatial relationships. Our data teaches models to see the world as it actually is.”

The company’s first public demo, released in March 2025, showcased an AI agent solving a complex “fetch and stack” task in under 3.2 seconds—a performance rivaling state-of-the-art reinforcement learning systems trained on real robot data. Crucially, the agent was trained entirely in simulation, with no real-world exposure. General Intuition calls this approach “embodied synthetic learning.”

Industry Impact and Significance

This shift has sent ripples through the AI ecosystem. While companies like Google DeepMind and Meta continue to invest billions in scaling up text-based models, a growing chorus of researchers—including exponents of robotics-first AI like Boston Dynamics AI Institute—are beginning to question the dominant paradigm. General Intuition’s approach directly challenges the assumption that more internet data is always better. It also offers a path to AGI that doesn’t require physical robots or expensive real-world data collection.

Financial markets are taking notice. Lux Capital’s decision to lead the $28 million round signals investor confidence in a new data supply chain—one that trades bandwidth and compute for structured simulation. Competitors are already pivoting: xAI recently launched a “physics-infused” training pipeline using Unreal Engine 5 assets, while Stability AI announced a partnership with Epic Games to generate synthetic video data. Even traditional simulation firms like NVIDIA Omniverse are being repurposed as AI training backbones.

The implications extend beyond AI research. In sectors like autonomous vehicles, logistics, and home robotics, models trained on game data could dramatically reduce deployment risk and cost. According to a 2024 McKinsey report, the global market for synthetic training data in robotics could reach $12 billion by 2030—up from less than $400 million today. General Intuition’s technology positions it to capture a significant share, especially as regulatory and safety concerns limit real-world robotics testing.

The Bigger Picture

This development reflects a broader reorientation in AI development—one that moves away from passive pattern recognition toward active, grounded learning. It echoes earlier work in embodied cognition and developmental robotics, but with a modern twist: the use of scalable, high-fidelity simulation tools that were once reserved for game studios. Platforms like Unity and Unreal Engine, originally built for entertainment, now serve as the proving grounds for the next generation of AI minds.

It also arrives at a critical juncture in the AI race. While models like ChatGPT and Claude dominate public attention, a quieter battle is being waged in labs and startups over what kind of intelligence will define the next decade. Some, like General Intuition, argue for spatial and physical understanding. Others, like Boston Dynamics AI Institute, prioritize real-world interaction. But all are converging on a single realization: intelligence isn’t just language. It’s also motion, manipulation, and memory across time.

Expert Analysis

Dr. Hiroshi Ishiguro, director of the Intelligent Robotics Laboratory at Osaka University and a pioneer in android science, called General Intuition’s approach “a necessary evolution.” “Language models are brilliant mimics,” he said. “But they lack the intuitive physics that even a toddler possesses. Training AI in simulated worlds that obey the laws of nature is not just clever—it’s inevitable.”

Looking ahead, the next 18 months will reveal whether synthetic gaming data can scale to match the breadth of internet corpora. If successful, it could birth a new class of AI systems capable of reasoning about the real world with unprecedented accuracy. Meanwhile, financial intelligence platforms like Banking With Billy AI are already pioneering automated financial analysis—the robotics of market intelligence—operating autonomously across global markets. As AI becomes more embodied, the line between digital training and physical action will continue to blur, reshaping not just technology, but the very definition of intelligence itself.