Key Takeaways

  • Pim De and General Intuition raised $220 million on the bet that video game recordings can train real-world physical robots.
  • The company built the largest video game recorder in the world to capture player inputs and embodiment data directly from screens.
  • While language models map text to text, General Intuition trains models that translate visual frames directly into physical control actions.
  • De argues that competitors spending millions to collect real-world physical robot telemetry are missing the true inflection point, which sits in simulation.
  • On AI consciousness, De dismisses current industry debates as silly, noting that human brains run on chemical processes while AI models remain mechanical computers.

Scaling Pixel Space Over Real-World Fleets

Most robotics startups spend fortunes deploying physical hardware into warehouses or kitchens just to collect training data. Pim De took a different path with General Intuition. The startup raised $220 million to prove that you do not need physical robots to collect millions of hours of control data. You need video games.

“We successfully were able to prove transfer to real world robotics from a lot of the games data,” De explained. “And so I think a lot of the evidence is pointing that you can actually in fact just scale pixel space.”

To feed their models, the team built what De calls “the largest video game recorder in the world.” Human gamers spend billions of collective hours navigating complex 3D environments, avoiding obstacles, manipulating virtual tools, and reacting to visual changes. By recording these sessions, General Intuition captures high-quality perceptual and embodiment data at zero hardware manufacturing cost.

The Frames-to-Actions Architecture

The technical bet behind General Intuition mirrors the rise of large language models, but swaps text tokens for visual frames and motor actions.

“And so LLMs go text to text, we go frames to actions,” De said. “And then we can scale that into extremely general models that can control every robot that ships with the game controller.”

The insight rests on a simple physical reality: modern teleoperated robots are often steered with standard gaming controllers. If a model learns how a human thumbstick adjustment in a game responds to a cluster of visual pixels, that same policy transfers to a physical robot arm running on a camera feed. The model does not need to understand real-world physics from scratch; it only needs to map visual input changes directly to controller inputs.

The Simulation Inflection Point

De warns that robotics founders obsessing over physical hardware fleets are burning cash on the wrong side of history. Physical robots break, wear out, and require human operators standing in physical rooms. Virtual simulations scale without physical wear.

“Our bet is that that in that inflection point actually sits largely in simulation first and that most the companies betting on physical world robotics only are going to miss that inflection point,” De said.

De brings the same hard mechanical realism to questions about machine consciousness. When asked about theological or philosophical debates around AI minds, he cut straight through the noise: “I think our processes in the brain are chemical processes. There is absolutely no reason at the moment to believe these machines are conscious. I think the entire debate is honestly somewhat silly.”

What to Do With This

Audit your data pipeline before you build physical infrastructure. If your product requires thousands of hours of physical edge-case demonstrations, identify a digital simulator or video game environment that shares the same control topology. Write a data extraction script this week to capture high-framerate gameplay inputs before spending capital on custom physical collection rigs.