This Startup Is Convinced Robotics Is About to Experience Its ChatGPT Moment

Before the introduction of OpenAI’s GPT-3, which initiated the foundation model era, organizations primarily created specialized natural language processing models from scratch, depending heavily on extensive datasets tailored to specific tasks. Today, many companies prefer using general-purpose models like OpenAI’s GPT series, Claude, or Llama, and then fine-tuning or prompting these models to suit their specific needs.

Pim de Witte, CEO of General Intuition, believes that embodied AI will follow a similar evolutionary path. He argues that rather than collecting vast real-world datasets to develop specialized robotic models, the emphasis should be on high-quality datasets that can create foundation models capable of delivering intuitive insights about movement and interaction across various settings.

“At present, many companies are engaged in specialized work focused on specific embodiments, environments, and robots,” de Witte noted during a recent episode of TechCrunch’s Equity.

He suggests that much of this specialized work may soon become outdated with the emergence of comprehensive models like those being developed by General Intuition.

“The generalization of the model itself is the product,” he said. “Its foundational reasoning abilities related to space and time will ultimately prompt people to move away from aggregating vast amounts of real-world data. In fact, a mere few minutes is sufficient.”

General Intuition has crafted its own foundation model by training on millions of hours of video game data, which encompasses details regarding player actions, including button presses on a controller and the timing of those actions. Both de Witte and General Intuition’s lead investor, Vinod Khosla, claim that this action data is crucial for developing a human-like intuition for spatial-temporal reasoning.

Recently, the startup raised $320 million at a valuation of $2.3 billion based on this concept. The company has demonstrated that its current model can not only play video games for prolonged periods but also control a quadrupedal robot—achieving this feat after fine-tuning the model with just eight minutes of real-world robotics data.

“The ability of [the robot] to perform zero-shot tasks using only the front camera, without any additional sensors, in an office setting with dynamic objects and people moving around was quite surprising for us,” de Witte remarked. “I believe this indicates what lies ahead.”

General Intuition’s primary objective isn’t to develop robots, but rather to serve as the foundation model for physical AI, acting as a base that other robotics companies can use to create their own designs. As de Witte puts it: “We’re not going to create a self-driving car company. Our goal is to simplify the process for someone else to establish a self-driving car company by tenfold.”