The Paradigm Shift in AI: From Specialized to General Models
Before the advent of OpenAI’s GPT-3, companies invested significant resources into developing specialized natural language processing models. These models were meticulously crafted from scratch, necessitating extensive amounts of task-specific data for training. However, the landscape has significantly shifted with the introduction of general-purpose models like OpenAI’s GPT series, Claude, and Llama. Today, many organizations begin with these versatile models, subsequently refining or tailoring them to address particular requirements.
Embodied AI’s Future: A New Approach
Pim de Witte, CEO of General Intuition, envisions a similar transformation for embodied AI. Instead of gathering vast real-world datasets to create specialized robotic models, de Witte advocates for focusing on high-quality datasets. These datasets should be capable of developing foundational models that can transfer knowledge about movement and interaction across diverse environments.
“Many companies are currently doing a lot of specialized work focused on individual embodiments, individual environments, and individual robots,” de Witte shared in a recent episode of TechCrunch’s Equity. This approach, he argues, may soon become obsolete with the advent of general models like those developed by General Intuition.
The Potential of Generalized Models
According to de Witte, “The generalization of the model itself is the product.” He asserts that the inherent reasoning about space and time in these models will reduce the need for collecting extensive real-world data. “The reality is you only need a few minutes,” he adds, emphasizing the efficiency of these models.
General Intuition has crafted its base model by training on millions of hours of video game data, incorporating insights such as human interactions with controllers. De Witte and lead investor Vinod Khosla believe that action data is pivotal for developing human-like intuition for space-time reasoning.
General Intuition’s Groundbreaking Achievements
The startup recently secured $320 million in funding at a $2.3 billion valuation, supporting its thesis. Notably, General Intuition’s model has demonstrated the capability to both play video games for extended periods and operate a quadruped robot. The latter was achieved after fine-tuning the model with just eight minutes of real-world robotics data.
“The fact that [the robot] was able to do a zero-shot only on the front camera, without other sensors, in the office with introduced dynamic objects and passers-by, which was a very big surprise to us,” de Witte remarked. “I think it’s a sign of what’s to come.”
A Vision for the Future of Robotics
General Intuition’s ultimate aim is not to manufacture robots but to establish itself as the foundational model for physical AI. This approach allows other robotics companies to build upon its models for their machines. As de Witte succinctly puts it: “We’re not going to create a self-driving car company. We’re going to make it 10 times easier for the next person to create a self-driving car company.”
When you purchase through links in our articles, we may earn a small commission. This does not affect our editorial independence.
Here
“`
