General Intuition raises $320M at a $2.3B valuation to train AI agents using video game action data.

What Happened

General Intuition has secured $320 million in new funding, propelling its valuation to $2.3 billion. The startup is building foundation models for AI agents by training them on millions of hours of video game footage and player telemetry, betting that this data can teach AI systems human-like spatial and physical intuition.

Technical Details

Unlike Large Language Models (LLMs) that rely on static text to predict the next token, General Intuition is building Large Behavior Models (LBMs) focused on action prediction. Video games provide highly structured environments containing continuous state-action-reward loops. By ingesting both the visual rendering of the game and the corresponding input telemetry (keystrokes, mouse movements, controller inputs), the model learns spatial awareness, physics, object permanence, and long-horizon planning. This approach leverages games as massive, diverse imitation learning pipelines, effectively bypassing the slow, expensive process of real-world robotic data collection and physical environment mapping.

Why It Matters

The current bottleneck in agentic AI and robotics is a severe lack of high-quality, real-world action data. While text-based LLMs excel at reasoning, they struggle with physical execution and multi-step planning in dynamic environments. Video games serve as high-fidelity physics engines and logic puzzles that humans navigate intuitively. If General Intuition can successfully execute "sim-to-real" transfer—applying the intuition learned in game engines to real-world tasks—it could dramatically accelerate the deployment of autonomous software agents and physical robots. It marks a critical architectural shift from predicting words to executing actions.

What to Watch Next

Monitor General Intuition's sim-to-real transfer benchmarks. The primary technical hurdle will be generalizing from the constrained, deterministic rules of game engines to the noisy, unpredictable real world. Additionally, watch for early API access or partnerships with robotics firms and enterprise automation platforms, which will serve as the first practical proving grounds for these action-oriented models.