Equity

Pim de Witte on Why Games Are Physical AI's Missing Dataset

Pim de Witte· CEO of General Intuition at General Intuition
·~26 min·English·TechCrunch
MultimodalRoboticsTrainingAI Company
TL;DR

General Intuition's Pim de Witte argues the next leap in AI isn't more text but world models trained on the one dataset that already encodes space, time, and action: hundreds of millions of hours of gameplay.

01Core Mental Model

Text Deletes Space and Time

<strong>Text throws away space and time;</strong> games keep both — which is why General Intuition bet its pre-training on gameplay, not more internet text.

text fundamentally removes a lot of the information that the real world needs particularly information around space and time.

— Pim de Witte, Equity
Key Insight
The claim isn't just 'more data' — it's a different kind. He frames text as 'described reality' filtered through an author's bias, versus game data as 'perceived reality' captured neutrally. The catch he concedes: losing those stated human preferences also removes a handle useful for interpretability and safety.
02The Result

Generalization Is the Product

Pre-trained on games, the model needed just <strong>8 minutes of real-world data to zero-shot a building it had never seen</strong> — so the generalization itself, not the task-specific tuning, is the product.

the fact that it was actually able to zero shot on just the front camera, no sensors, the office, which with dynamic objects being introduced, people walking by was a very big surprise to us.

— Pim de Witte, Equity
Key Insight
This quietly inverts the robotics playbook. Everyone else collects millions of hours per embodiment; if a general base transfers to a novel space in minutes, the durable asset is the pre-trained prior, not the per-robot dataset — which is why he predicts teams will stop hoarding real-world hours. He caveats that navigation transfers cheaply because games are full of it; a robot arm may need far more.
03The Moat

Go Where the Data Is

<strong>World-model data barely exists anywhere — except in their game clips,</strong> so researchers followed the data and investors backed a company, not an acquisition.

you go where the data is. And in world models, that's nowhere because it doesn't exist

— Pim de Witte, Equity
Key Insight
Notice what he says the moat isn't: data alone. 'If it was only one thing, i.e. the data, then we should have just sold.' The defensibility is data times a team that already wrote the field's foundational papers (Diamond, Iris) — the combination is what turns a dataset into a generational company instead of an acquisition target, reportedly turning down a bid to stay independent.
04Data Quality

Ground Truth Beats Inferred

You can't recover a pilot's rudder input from a video's pixels, so <strong>ground-truth action labels beat inferred ones</strong> exactly where customers live: the long-horizon edge cases.

if you're landing a plane and you're moving the rudder that's not going to be in the pixel stream, right? It's not going to be in the frames.

— Pim de Witte, Equity
Key Insight
The tell is why labs claim pixels are enough: their benchmarks test the general case, but customers pay for the edge cases. Without real action labels a model has to learn to separate the agent from the environment on its own — a different scaling law entirely — and a model trained from scratch on ground-truth data will, in his words, blow the inferred one out of the water.
05The Thesis

Intuition Is the Next Leap

If reasoning was the leap that made LLMs matter, <strong>intuition is the leap world models are chasing</strong> — the physical common sense the company is literally named after.

with LLM's the next big leap was reasoning, right? And that completely changed the game. He's saying with world models, the next big leap is intuition

— Rebecca Bellan, Equity (relaying Vinod Khosla)
Key Insight
Naming the company 'General Intuition' is the thesis, not branding. Reasoning made LLMs feel smart in symbolic domains; intuition is the bet that world models deliver the physical, spatial common sense text can't encode. It's a capability that's far harder to benchmark than a math score — part of why it's been under-invested until a dataset like theirs made the bet fundable.
06Ethics & Defense

Never the Escalatory System

General Intuition will do search-and-rescue but not lethal autonomy, drawing the line at <strong>never becoming part of an escalatory system.</strong>

You're not more important than the democracies you serve.

— Pim de Witte, Equity
Key Insight
He's careful not to be anti-defense — he's anti-harm. The nuance is architectural: whether a violent game clip rewards the model positively or negatively is a training-time decision, so the red line is enforced in how the base model is built, not bolted on at deployment. The Anthropic–DoD friction the interviewer raises is the cautionary backdrop — a lab told it was a 'supply chain risk' for holding its line.
07AI & Jobs

Do Something About Jobs

De Witte's answer to AI job loss isn't a forecast — it's Nerve, a gamer marketplace, because <strong>you don't get to theorize about jobs unless you're building them.</strong>

I don't want to hear how you think that models are going to affect jobs unless you're doing something about it.

— Pim de Witte, Equity
Key Insight
The flywheel is doing double duty. A gamer marketplace that pays people to label data and teleoperate robots creates jobs and generates exactly the ground-truth action data the models need — so his ethics and his data moat are the same loop. That alignment is why he can dismiss lab-leader 'doomerism' as a self-own: if you have a front-row seat to the shockwave, the honest move is to build the jobs, not narrate the losses.