General Intuition chief executive Pim de Witte says embodied AI could follow the same shift that reshaped natural language processing after GPT-3. Instead of training narrow robotics models from large task-specific datasets, he argues companies will rely on broader foundation models that understand movement, space and time across many environments.
De Witte told TechCrunch’s Equity podcast that many robotics companies are still building specialised systems for individual robots, settings and tasks. He said that work could become less important as more general embodied models emerge.
General Intuition is building one of those models. The startup trains on millions of hours of video game data, including player actions such as controller inputs, to teach AI systems how agents move, react and make decisions in dynamic environments.
Video Games Provide Action Data
The company’s thesis is that video games contain valuable action data because they show what humans did, when they acted and how the environment responded. De Witte and lead investor Vinod Khosla argue that this connection between perception and action can help AI develop spatial-temporal reasoning.
General Intuition raised $320 million last month at a $2.3 billion valuation, in a round led by Khosla Ventures with participation from General Catalyst, Jeff Bezos, Eric Schmidt and others, according to TechCrunch. The funding is intended to help the company scale compute and train larger versions of its model.
The company has already shown its model playing a video game for hours. It has also demonstrated the same model powering a quadrupedal robot after fine-tuning on just eight minutes of real-world robotics data.
De Witte said the robot was able to operate using only its front camera, without other sensors, in an office where people and objects were moving around. He described the result as an early sign of what broader physical AI models could enable.
Startup Wants to Power Other Robotics Companies
General Intuition does not plan to build robots itself. Its goal is to provide a base model that robotics companies, autonomous vehicle developers and other hardware builders can adapt for their own systems.
De Witte said the company wants to make it much easier for others to build products such as self-driving cars or robots. The model’s ability to generalise across environments is the product, rather than any single machine.
The approach reflects a broader debate in robotics. Some companies are collecting large real-world datasets from their own robots, while others are using simulation, gaming environments or video to reduce the need for expensive physical data collection.
General Intuition is betting that higher-quality action data can reduce that burden. De Witte said the industry may not need hundreds of thousands or millions of hours of real-world robot data if a foundation model already has a baseline understanding of space, time and interaction.
If that thesis holds, embodied AI could develop more like language AI did. Companies may start with a general model, then fine-tune it for specific robots, vehicles or physical tasks instead of building everything from scratch.
Featured image credits: Magnific.com
For more stories like it, click the +Follow button at the top of this page to follow us.