Fei-Fei Li’s stealth startup has a new spoiler by herself
The short answer is "Embodied Intelligence."'
The AI startup arena is already very crowded, with startups finding themselves in more and more polarized positions. xAI has secured 50 billion in a record-breaking B round, while several other star companies are seeking a discount sale. However, there are still more companies joining in this race.
Fei-Fei Li’s stealth startup is one of the exciting new players. According to Reuters, Li is building a startup that uses human-like processing of visual data to make AI capable of advanced reasoning. After the news came out, Li shared her TED talk about Spatial Intelligence on X.com, which has been seen as her answer to what the new company is up to.
But the term "spatial intelligence" is actually still a buzzword. What exactly is her startup working on?
It seems that there is a new hint coming out, directly from Fei-Fei Li.
The short answer is "Embodied Intelligence."
What’s going on here
The spoiler came out during her lecture at Stanford CS231n. As a startup founder, Li is officially on a partial leave at Stanford. This CS231n class is one of the very few classes she is still teaching. The lecture, titled Human-centered AI, is basically a Computer Vision history lecture that is open to all. However, this time, after the routine contents, she added and emphasized on one specific topic.
“Computer vision and NLP progress inspires a new north star, embodied AI.” She said. “Embodied AI is closing the loop between perception and action.”
“I really think this is the biggest opportunity,” Li said. “This is the one last thing I must tell you guys.”
The fact that she didn't mention OpenAI throughout the entire lecture, and didn't delve into any new generative models in CV, or talk about the diffusion model, makes this part of the lecture particularly meaningful. It's a hint of her next move.
“Today’s robotics research is skill-level tasks, short-horizon goals, but all are closed world instructions.” She said. “My goal is to bring them into open-world instructions.”
Why this matter
First of all, according to Fei-Fei Li’s past research style, there may be some open-source datasets coming out from her startup that will have a great impact on the entire robot industry. Imagine there’s an ImageNet level datasets for robots.
Second, Li’s approach to embodied intelligence as a CV pioneer could be very different. Today it seems that the mainstream idea is to combine LLM with robot hardware. However, it seems that CV and video generative models are closer to robots than LLMs, there may be some alternative methods emerging.
Last but not least, Li’s answer to the GPU availability challenge is finally here. After claiming that Stanford’s Natural Language computing lab has only 64 GPUs and academia is "falling off a cliff" relative to the industry, she herself chose to join the industry side. Will this become the only path for ambitious researchers to truly make significant contributions in the AI revolution? Her future moves will undoubtedly be closely monitored by her fellow colleagues.