General Intuition

Learning Objectives

• Understand what General Intuition is building and why it trains agents on gameplay video
• Distinguish General Intuition's agent-first approach from world-model makers that sell the model itself
• Identify where spatial-temporal reasoning agents could be deployed beyond gaming

What Is General Intuition?

General Intuition is a frontier AI research lab building foundation models and general-purpose agents for environments that require deep spatial and temporal reasoning — the ability to perceive a scene, anticipate how it will change, and improvise the right action. Founded in 2025 as a spin-out of Medal, the world's largest game-clip platform, the New York-based company is led by CEO Pim de Witte, Medal's co-founder, alongside researchers Eloi Alonso, Adam Jelley, and Vincent Micheli.

The lab's core bet is that the path to capable real-world agents runs through video — and specifically through gameplay, where the world is rich, interactive, and densely sampled. Its models train on Medal's library of roughly 2 billion gaming videos a year from more than 10 million monthly active users across thousands of games. The result, the company says, is an agent that learns purely from visual input: it can make sense of environments it was never trained on, predicting the right actions while seeing only what a human player would see and moving through space by following controller inputs.

How It Differs From Other World-Model Labs

A wave of labs — World Labs, Decart, Runway, and others — are building world models, neural systems that learn how an environment evolves. Most of them aim to sell the world model itself, as training infrastructure or a content-creation engine.

General Intuition's framing is different: it treats world models as internal training infrastructure for its own agents, not as the product. The company says this agent-first focus also helps it sidestep the content-copyright questions that come with selling generated environments. The deliverable, in other words, is the agent's behavior — not a simulator other teams license.

Target Applications

General Intuition has named two initial domains:

• Gaming — agents that can play, test, or populate game worlds, drawing directly on the gameplay data the models are trained on.
• Search-and-rescue drones — autonomous navigation and decision-making in unstructured, never-before-seen physical environments, where an agent must reason about space and act from visual input alone.

The throughline is environments where an agent has to look, predict, and act in real time without a hand-coded map — the same skill whether the environment is a game level or a collapsed building.

Funding and Backing

General Intuition raised a $134 million seed round in October 2025, led by Khosla Ventures and General Catalyst (with participation from Raine). By mid-2026 it was in talks to raise a further $300 million at a valuation just above $2 billion, drawing new backers Jeff Bezos and Eric Schmidt alongside its existing investors. The pedigree of the dataset is part of the draw: OpenAI had previously tried to acquire Medal for the same gameplay video the lab now trains on.

Strengths

• A rare proprietary data moat: exclusive access to Medal's stream of roughly 2 billion gameplay videos a year, each paired with player actions — exactly the action-labeled visual data spatial-temporal agents need
• Agent-first thesis: by shipping agent behavior rather than a licensable world model, it targets end-use value and avoids content-copyright friction
• Top-tier backing: Khosla, General Catalyst, Jeff Bezos, and Eric Schmidt, at a valuation just above $2 billion
• Cross-domain ambition: the same visual-reasoning core targets both gaming and physical-world tasks like search-and-rescue drones

Limitations and Considerations

• No shipped product yet: the first release is expected in late summer 2026, so real-world performance, latency, and reliability are unproven
• Unbenchmarked claims: the "learns from visual input, generalizes to unseen environments" capability is the company's stated direction, not an independently verified result
• Narrow public detail: model architecture, sizes, and evaluation methodology have not been published
• Sim-to-real gap: translating gameplay-trained reasoning to physical drones is a hard, open research problem that many embodied-AI efforts have struggled with
• Data dependence: the moat rests on Medal's dataset and its continued licensing and rights position

Strategic Context

General Intuition sits at the intersection of two fast-moving fields: world models (World Labs, Decart, Odyssey, DeepMind's Genie) and embodied AI / agents (NVIDIA GR00T, robotics foundation models). What sets it apart is the wager that gameplay video is the cheapest, richest source of action-labeled visual experience on Earth — and that an agent trained on enough of it will generalize to the physical world.

If that wager pays off, the search-and-rescue-drone use case is the tell: it would mean reasoning learned in games transfers to environments no one scripted. If it doesn't, the company still has a defensible gaming business on top of a dataset its rivals cannot easily replicate. Either way, the combination of a unique data moat and marquee backing makes it one of the more closely watched bets in the world-model race — which is why it is worth tracking even before its first product ships.

Key Takeaways

• General Intuition is a New York frontier lab, spun out of game-clip platform Medal in 2025, building foundation models and agents for spatial and temporal reasoning
• It trains on Medal's library of roughly 2 billion gameplay videos a year, aiming for agents that learn from visual input and generalize to environments they were never trained on
• Unlike World Labs, Decart, and Runway, it treats world models as internal training infrastructure for its own agents rather than a product to sell
• Initial target applications are gaming and search-and-rescue drones
• It raised a $134 million seed in October 2025 (Khosla, General Catalyst) and by mid-2026 was raising $300 million more at a valuation just above $2 billion, with Jeff Bezos and Eric Schmidt joining; its first product is expected late summer 2026