Walk into General Intuition's New York R&D floor and the first thing you notice is a monitor showing what looks like someone grinding through a Fortnite match. Except no human is at the controls. The AI agent has been running continuously for 100 hours. Then a four-legged robot the size of a large dog rounds the corner, bumps into a trash can, circles a journalist like a curious golden retriever, and keeps walking. According to co-founder and CEO Pim de Witte, both the game agent and the quadruped are being driven by the exact same model.

That's the pitch. And apparently it's a $2.3 billion pitch, because General Intuition just closed a $320 million Series B led by Khosla Ventures, with checks from General Catalyst, Jeff Bezos, Eric Schmidt, former Formula 1 champion Nico Rosberg, and researchers affiliated with Google DeepMind and MIT. Combined with the $134 million seed round it raised at launch in October 2025, the startup has now disclosed $454 million in total funding. The company is 31-year-old de Witte's second act — his first was Medal, a platform where gamers upload and share gameplay clips, which quietly became the data engine behind everything General Intuition is building.

The Data Edge That Actually Matters

Here's where it gets interesting — and where General Intuition diverges from the crowded field of world model startups that have been vacuuming up VC dollars lately. Most approaches to training embodied AI agents from video data try to infer actions by watching what happens on screen. General Intuition's angle is different: Medal's massive archive of gameplay comes bundled with action labels — precise records of which buttons players pressed and exactly when they pressed them.

That distinction is not trivial. Watching a video of someone driving a car tells you the car turned left. The action label tells you the driver turned the steering wheel 15 degrees at 60 mph while braking. One is an observation; the other is a causal record. De Witte argues that this action data is what lets his model develop a meaningful self/environment boundary — a prerequisite for anything resembling spatial reasoning or physical intuition. Without it, you're teaching a model to narrate the world. With it, you're teaching it to act in the world.

"We have a single model that can respond to information on screen and take action, but also to real-world dynamics in a way that an LLM could never," de Witte has said. That's a bold claim, but the demo provides at least a first-order sanity check.

Eight Minutes of Real-World Data

The robot demo is where things get either genuinely impressive or deeply suspicious, depending on your prior on robotics data requirements. According to General Intuition, it took just eight minutes of real-world data collected outdoors to fine-tune the quadruped's model for indoor office navigation. The robot had never seen the office before. It still bumped into chair legs and misjudged doorways — classic sim-to-real transfer problems — but it navigated with enough coherence to be more than a party trick.

Fine-tuning a foundation model on eight minutes of domain-specific data is believable if the base model already has strong priors about physical space. That's the entire bet here: that hundreds of millions of hours of human gameplay constitute a richer spatial-temporal pre-training corpus than anything the robotics field has assembled through careful data collection campaigns. Whether that bet holds when the stakes move beyond office demos to warehouse logistics or surgical assistance is the unresolved question everyone at General Intuition is presumably losing sleep over.

The Gym, Not the Product

General Intuition's world model — a frame-by-frame simulated environment generated without a traditional game engine — is described internally as "the gym." It's not what they're selling. It's where the agent trains. In demos, the world model reportedly maintains physics correctly: walls are solid, ladders are climbable, shadows track the sun's position. These aren't hardcoded rules; they emerged from the training data. That's actually a meaningful signal about what sufficient scale of action-labeled gameplay can produce.

What the company wants to sell is the agentic model itself, delivered via API. A broader API rollout is reportedly planned for before the end of summer, funded by a slice of the new round. The majority of the $320 million will go toward scaling compute through a deal with CoreWeave, with pre-training the next model version as the primary objective.

Vinod Khosla's Framing

Khosla Ventures led the round, and Vinod Khosla offered a characteristically high-voltage take on why. His analogy: just as reasoning emerged as a step-change capability in large language models, he expects "intuition" — some human-like capacity to anticipate physical cause and effect — to emerge from world models trained on sufficiently rich action data. The gameplay corpus, he argues, is the key ingredient for that emergence.

It's a compelling framing, and it's also impossible to verify until it either happens or doesn't. Khosla has made bold bets before that paid off. He's also made bold bets that didn't. The honest answer is that nobody knows whether game-derived action data generalizes to the physical world at the scale needed for commercial deployment — because it hasn't been done yet.

The Honest Skeptic's Ledger

Let's be clear about what General Intuition has demonstrated and what it hasn't. The demos are coherent and the underlying data thesis is intellectually sound. Action-labeled gameplay data is almost certainly more useful than raw video for training embodied agents. The transfer from gameplay to robot navigation — even imperfect transfer — is a real result worth paying attention to.

What hasn't been shown: robust performance in novel physical environments, reliability under distribution shift, or the kind of consistent real-world data that robotics deployments actually require. The "eight minutes of fine-tuning" claim is exciting but needs replication across diverse settings before it becomes an architecture principle rather than a demo condition.

There's also the competitive landscape to consider. General Intuition is not the only team pursuing world models for embodied AI, and the field is moving fast. Several well-funded competitors are working on similar problems from different angles. Being first to a compelling demo and being first to a deployable product are two very different finishing lines.

None of that means the $2.3 billion valuation is indefensible. It means the valuation is priced on a vision, and the vision has to survive contact with the physical world. That's where the real test begins.