General Intuition: Why video games beat internet data

General Intuition, a Bezos-backed New York-based startup valued at $2.3 billion, closed a $320 million funding round with Coatue, Eric Schmidt, and researchers from MIT and Google DeepMind joining its investor list. CEO Pim de Witte argued on the Equity podcast that game-derived world models could fill gaps that text-trained models leave open.

What did General Intuition announce?

General Intuition announced a $320 million financing round and reiterated its focus on training world models using video game data rather than internet text. The company described itself as a spinout of gaming platform Medal TV, and CEO Pim de Witte discussed the bet on games as richer training data on the Equity podcast with Rebecca Bellan.

The round adds Coatue, Eric Schmidt, and researchers at MIT and Google DeepMind to the company’s backers. The startup is described in the announcement as Bezos-backed and based in New York, and the company’s valuation is listed at $2.3 billion.

Why use video game data instead of internet text?

Video game environments provide temporally coherent, physics-rich simulations that General Intuition says let models learn how objects move through space and time, a capability the company argues large language models lack. Pim de Witte framed the problem bluntly: "large language models just don’t have what it takes," calling out models like ChatGPT and Claude as strong at text but less able to learn physical dynamics from web text alone.

In the Equity conversation, de Witte positioned gaming worlds as a source of ground-truth trajectories, interactions, and consequence-driven behaviors that can teach what internet text cannot: how things move and interact across time steps. That, he argues, produces world models better suited to physical reasoning and robotics-style generalization than models trained solely on static text.

How did the company get here?

General Intuition spun out of Medal TV, a gaming platform, and carried forward a research agenda focused on extracting world models from gameplay data. The spinoff structure links the startup’s technical focus directly to its origins in recorded and simulated game environments, which the team now treats as training corpora for physical AI.

The new funding round brought high-profile investors into the cap table: Coatue and Eric Schmidt are named investors, and researchers at MIT and Google DeepMind joined the list of backers. The company’s stated valuation in the announcement is $2.3 billion, and the fresh capital totals $320 million.

Why it matters

If world models trained on games can teach spatial-temporal reasoning that text cannot, the result would shift where AI teams source their training data for robotics and embodied intelligence. The move from web text toward simulated, interaction-rich data could change which startups attract capital and which research agendas lead to deployed physical AI. The funding round signals investor confidence that game-based world models are a commercially relevant path to more general physical reasoning.

The debate is not purely technical. The Equity conversation also covered ethical boundaries, noting that as models learn to predict and control physical systems they could be used for defense applications. That raises governance questions about how game-derived models are developed and deployed.

What to watch

Watch for technical papers or demos showing transfer from game-trained world models to real-world robotics tasks, and for the startup’s next product milestones following the $320 million raise. Also monitor who else joins the investor or research roster and any concrete statements from the company about safeguards for potential defense uses.