General Intuition Funding Long-Horizon Agents Shows Action Data Edge

General Intuition raised $320 million at a $2.3 billion valuation. The round brought total disclosed funding to $454 million. Khosla Ventures led the investment. General Catalyst, Jeff Bezos, and Eric Schmidt also participated.

The company stands out because it owns Medal. That platform has collected hundreds of millions of hours of Fortnite gameplay. Every clip carries exact keypress labels and timestamps.

Funding signals data shift

General Intuition plans to use the cash to scale training on CoreWeave. Pretraining of the next model starts soon. An API release is scheduled before summer ends.

The core thesis is simple. Long-horizon agents improve faster when trained on rich action traces than on additional text alone. Text supplies background knowledge. Action data supplies timing, sequencing, and environmental response.

Medal data powers one model

Medal users record sessions across many games. The platform stores precise controller inputs at millisecond resolution. General Intuition turns those sessions into training sequences for a single policy network.

The model already controls characters in Fortnite for stretches of 100 continuous hours. The same weights, after an eight-minute fine-tune on street footage, let a quadruped robot navigate an office. No separate architecture was required for each domain.

Long-horizon agents need feedback loops

Most current agents rely on language model outputs. They plan steps in text and hope execution matches the plan. General Intuition's approach closes the loop. Every action receives immediate environmental feedback that updates the policy directly.

This feedback loop matters once tasks span minutes or hours. Small timing errors compound. Action traces from games already encode recovery behaviors that text descriptions rarely capture.

Investors see path beyond simulation

The valuation reflects conviction that game-derived policies transfer. Khosla Ventures has backed several robotics efforts. The participation of Bezos and Schmidt adds capital and signal for follow-on rounds.

Competitors still scale text corpora or synthetic trajectories. General Intuition bets that real human gameplay at volume supplies denser supervision. The 100-hour game runs and eight-minute robot transfer serve as early proof points.

remio also requires behavioral context

Office agents face a similar gap. Document retrieval supplies facts. Workflow traces supply the sequence of clicks, approvals, and edits that actually complete tasks. remio captures those traces automatically across meetings, files, and browser activity. The result is grounded output that matches an individual team's real process.

https://www.remio.ai

Remaining uncertainties

The eight-minute robot transfer used limited real-world data. Larger skill sets may require more physical runs. API pricing and usage limits remain undisclosed. Regulators have not yet examined whether game-trained policies carry new safety obligations when deployed on hardware.

What to watch next

CoreWeave compute expansion numbers will show whether training runs stay on schedule. The summer API launch will reveal how developers actually use the policy models. Any partnership announcements with robot hardware makers would confirm transfer progress beyond office demonstrations.