Startup Predicts Robotics’ ChatGPT Moment
▼ Summary
– AI development has shifted from building specialized models from scratch to fine-tuning general-purpose foundation models like GPT-3.
– General Intuition CEO Pim de Witte argues that embodied AI should follow a similar path, focusing on quality datasets for foundation models rather than collecting huge real-world datasets.
– The company trained its foundation model on millions of hours of video game data, including human controller inputs, to develop spatial-temporal reasoning.
– General Intuition raised $320 million at a $2.3 billion valuation, demonstrating its model can power a quadrupedal robot after fine-tuning on just eight minutes of real-world data.
– The company aims to become the foundation model for physical AI, enabling other robotics companies to build upon it rather than creating robots itself.
Before OpenAI’s GPT-3 redefined the AI landscape, companies painstakingly built specialized natural language models, training each one from scratch on massive, task-specific datasets. Today, the norm is to start with a general-purpose foundation model like GPT, Claude, or Llama, then fine-tune it for particular use cases. Pim de Witte, CEO of General Intuition, believes the world of embodied AI is about to undergo the same transformation.
Rather than amassing enormous real-world datasets to train custom robot models, de Witte argues the industry should prioritize higher-quality datasets that yield foundation models capable of transferring intuitive knowledge about movement and interaction across diverse environments. “A lot of companies right now are doing lots of specialized work focused on individual embodiments, individual environments, and individual robots,” de Witte told TechCrunch on a recent episode of Equity. He contends that much of this specialized effort will soon become redundant, thanks to emerging general models like the one his startup has been developing and deploying.
“The generalization of the model itself is the product,” he said. “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.”
General Intuition built its own foundation model after training on millions of hours of video game data, including precise records of which buttons a human pressed and when. Both de Witte and the company’s lead investor, Vinod Khosla, emphasize that this action data is crucial for developing a human-like intuition for spatial-temporal reasoning.
The startup recently raised $320 million at a $2.3 billion valuation, driven by this thesis. The company has demonstrated that its current model can both play a video game for hours and power a quadrupedal robot after fine-tuning on just eight minutes of real-world robotics data. “The fact that [the robot] was actually able to zero-shot on just the front camera, with no other sensors, in the office with dynamic objects being introduced and people walking by was a very big surprise to us,” de Witte says. “I think it’s a sign of what’s to come.”
General Intuition’s ultimate goal is not to build robots itself, but to become the foundation model of physical AI , a base layer for other robotics companies to build upon for their own machines. 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.”
(Source: TechCrunch)
