The prevailing wisdom in robotics goes like this: to teach a machine how to move through the world, you need to feed it mountains of real-world data. Hundreds of thousands of hours. Maybe millions. One startup just raised $320 million on the bet that this belief is wrong.
According to TechCrunch AI, Pim de Witte, CEO of General Intuition, argues that embodied AI is heading for its own ChatGPT moment. He made the case on a recent episode of TechCrunch’s Equity podcast. His company closed that $320 million round last month at a $2.3 billion valuation, with Vinod Khosla leading the raise. What stands out here is the thesis behind the money: the industry has been solving the wrong problem.
The myth de Witte wants to kill
Most robotics companies today build narrow models. One robot, one environment, one task, trained on data collected specifically for that job. It’s expensive, slow, and it doesn’t transfer. Teach a robot to work in a warehouse and you’ve taught it almost nothing about working in a kitchen.
De Witte says most of that specialized work “will become redundant soon.” His argument mirrors what already happened in language AI. Before GPT-3, companies built natural language models from scratch for each task. Now almost everyone starts with a general model like GPT, Claude, or Llama and fine-tunes from there. General Intuition wants to be that base layer for physical machines.
“The generalization of the model itself is the product,” de Witte told TechCrunch AI. “The fact that it has a base level of reasoning about space and time is going to be the reason why people stop collecting hundreds of thousands or millions of hours of real-world data. Because the reality is, you only need a few minutes.”
Where the training data actually came from
Here’s the part that makes the story interesting. General Intuition didn’t build its foundation model on robot footage. It trained on millions of hours of video game data, including which controller buttons a human pressed and exactly when.
That action data is the whole bet. De Witte and Khosla both argue it’s the key to building human-like intuition for spatial and temporal reasoning. A game teaches a model cause and effect through movement: you press a button, something happens in space over time. Scale that up and you get a model that reasons about the physical world without ever touching it.
The proof, and the surprise
The company says its current model can play a video game for hours and power a four-legged robot. The robot part came after fine-tuning on just eight minutes of real-world robotics data.
That’s the claim worth watching. According to TechCrunch AI, the robot ran on a single front camera, no other sensors, and still handled dynamic objects and people walking by in the office.
“The fact that [the robot] was actually able to zero-shot on just the front camera… was a very big surprise to us,” de Witte says. “I think it’s a sign of what’s to come.”
Why this matters
General Intuition doesn’t want to build robots or self-driving cars. It wants to be the foundation model everyone else builds on. As de Witte put it: “We’re not gonna build a self-driving car company. We’re gonna make it 10 times easier for the next person to build a self-driving car company.”
If the thesis holds, the economics of robotics shift hard. The moat stops being who owns the most data and starts being who owns the best base model. That’s a direct threat to the many startups pouring resources into massive real-world data collection right now.
One caveat worth keeping in mind: a quadruped walking around an office is a long way from a self-driving car or a factory arm handling edge cases. Eight minutes of fine-tuning is a striking demo, not yet a proven pattern across hard, safety-critical tasks.
Still, the direction is clear enough to take seriously. If robotics really does get its GPT-3 moment, the companies hoarding data may find they built the wrong asset. Full details are in the original TechCrunch AI report.