General Intuition is a US World-Model AI startup founded in 2025 by Pim De Witte and Kent Rollins as a spinout of Medal, a platform for uploading and sharing video-game clips. It pursues an action-conditioned approach to World Models, training on observation-action-observation tuples scraped from gaming clips, and raised a $133.7M seed round announced in October 2025.
| Field | Value | |
|---|---|---|
| Type | World-Model AI startup (action-conditioned approach) | |
| Founders | [[entities/pim-de-witte | Pim De Witte]], Kent Rollins |
| Founded | 2025 | |
| HQ | US | |
| Capital raised | $133.7M Seed (early 2026) | |
| Training approach | Action-conditioned models trained on observation-action-observation triples scraped from gaming clips |
Background
General Intuition was spun out of Medal, a consumer platform for capturing and sharing video-game clips, giving the new lab access to a large corpus of gameplay footage as training data. It is organized as a public-benefit corporation (General Intuition PBC) with reported bases in New York and Geneva, and describes its focus as spatial-temporal reasoning — training models that can perceive and navigate three-dimensional environments (Source: siliconangle.com). Its October 2025 seed round, one of the larger seed financings of the year, was co-led by General Catalyst and Khosla Ventures (Source: techcrunch.com; citybiz.co). The company has described target applications spanning in-game agents and embodied systems such as robots and search-and-rescue drones.
Snapshot
Funding
| Date | Round | Amount | Notes | Source |
|---|---|---|---|---|
| 2025-10-16 | Seed | $133.7M | One of the larger seed rounds of 2025; co-led by General Catalyst and Khosla Ventures | (Source: techcrunch.com; siliconangle.com) |
Approach
General Intuition represents the action-conditioned approach to World Models, which De Witte distinguishes from World Labs (generative pixel-space) and AMI Labs (latent / JEPA). The position, articulated by De Witte in his March 2026 Not Boring essay *Computing the Uncomputable* (co-authored with Packy McCormick), is that the action — not the pixel and not the latent — is the compression that lets a World Model generalize. De Witte and McCormick argue that training on enough (observation, action, next-observation) tuples lets the model learn the causality of the world directly, rather than learning to render the world's appearance.
De Witte and McCormick argue that action-conditioned World Models are the path to AGI, on the grounds that they directly learn the predictive structure embodied agents need. In their account, gaming clips are an attractive training corpus because the action is observable (controller inputs) and the next state after the action is observable (the next frame).
Related approaches
The company's training-data thesis sits adjacent to other action-conditioned bets in the field. De Witte and McCormick point to Niantic Spatial's use of crowdsourced Pokémon Go imagery to train a visual-positioning model now powering Coco Robotics last-mile delivery (Source: Heaven, *MIT Tech Review*, March 10 2026). The mechanism differs — visual positioning rather than action-conditioned dynamics — but the underlying wager is the same: that a training corpus already generated by large-scale consumer interaction is a route to general embodied-AI capability.
Relationships
- instance-of: World Models — action-conditioned approach.
- related: World Labs, AMI Labs, Google DeepMind (Genie), Nvidia & TSMC — AI Compute Infrastructure (Cosmos).
- founded-by: Pim de Witte.
- anchor source: World Models: Computing the Uncomputable — Packy McCormick & Pim De Witte (Not Boring, March 2026) (essay co-authored by founder De Witte).
Confidence
- High for the action-conditioned positioning and the McCormick / De Witte co-authored essay.
- High for the $133.7M seed figure and the October 2025 close — corroborated across TechCrunch, SiliconAngle, and Citybiz, independent of the founder essay.