AI Pioneer Fei-Fei Li Talks AI Safety, Competition with China

SPEAKER_1 (0:02)

Bloomberg Audio Studios, podcasts, radio, news.

SPEAKER_2 (0:13)

The United States also totally rejects any attempt to construct a globalist scheme to control for the artificial intelligence being spoken of so much now, here and after officially called super intelligence.

Ed Ludlow (0:30)

That was President Trump earlier this hour, talking about super intelligence, as AI safety and development are among the top issues at this year's UN General Assembly. And it's expected to be front and center when President Trump and China's President Xi meet later this week, even as leaders in industry form their own approaches. We're joined now by Fei-Fei Li, a pioneer in artificial intelligence, has been called the godmother of AI, someone who has long advocated for a human-centric approach to developing the technology. She is the co-founder and CEO of World Labs and an AI researcher and professor at Stanford University. It's great to have you back on the program.

Fei-Fei Li (1:08)

Great to be here, Ed.

Ed Ludlow (1:10)

Let's start with the idea of a world model, distinct from an LLM. I appreciate you and I have discussed this already, but for our audience, that distinction is important, the basic definition.

Fei-Fei Li (1:19)

Yes, I'm very excited by world models because if you look at language models, what they output are words. Words can be turned into codes, words can be turned into conversations, but world models, the ones that World Labs has been developing, what we generate are not words. The models generate beautiful pixels, very accurate geometric structure of the 3D or 4D world, and which in turn can allow people, creators, designers, robotics developers, to have precise control of the cameras or the construction of worlds.

Ed Ludlow (1:57)

Your latest generation of world model that you've been working on at World Labs is Atlas. Yes. You haven't said very much about Atlas.

Fei-Fei Li (2:04)

We just shared a technical blog about Atlas, and we are working really hard to release Atlas as a product to the world. And it is the latest, the greatest world model that World Labs has developed. It's the first of its kind in the ability to allow camera control, allow that kind of precise pixel consistency with the real world, with the 3D world. It also tries to solve a very hard problem in the field of computer vision that combines the generation of beautiful pixels with the reconstruction of intricate 3D structure.

Ed Ludlow (2:45)

Does it solve for it?

Fei-Fei Li (2:46)

It does.

Ed Ludlow (2:47)

How good is the underlying model?

Fei-Fei Li (2:48)

It's better than any state-of-the-art specialized model for these tasks.

Ed Ludlow (2:53)

There is the underlying model atlas and you mentioned product. What is the killer use case application for deployment in the real world?

Fei-Fei Li (3:02)

As many foundation models, these use cases are pretty horizontal. We are already engaging with users from creative industry, with design industry, with robotics industry, such as healthcare lab development, pharmaceutical laboratories. So it is pretty horizontal, but one thing that really excites me is by allowing this kind of control for humans, creators can use this as an augmentative tool to really express their creativity or develop downstream technology for their use cases.

Ed Ludlow (3:44)

From a commercial perspective, World Labs has been focused on robotics, right? You acquired Scenics. Explain the use of a world model in the development of physical AI, the deployment of robots in real human-centric settings.

Fei-Fei Li (3:59)

Yes, first of all, World Labs is a spatial intelligence company. A world model is very much useful for spatial intelligence applications from design creativity to robotics, and also which includes healthcare education. What is important for robotics is the ability to understand and construct the world that you can in turn train and evaluate your robotics tasks or policies. There's a lot of steel bottleneck in terms of creating the right training environment or evaluation environment. A robust world model can help unlock some of this bottleneck. And for example, for laboratory sciences, when you want to empower a robot to do pharmaceutical manufacturing or lab science, you don't get that much data, right? These are very highly specialized industries. But using world models to simulate and then turn assimilation into a very realistic training ground, as well as evaluation ground, helps to develop robots.

Ed Ludlow (5:11)

We are talking about the physical world in many cases. And the issue that your industry is confronting right now is alignment. In LLMs in particular, does the model behave as humans intended it to?

Is that the same calculus for a world model? Is there a risk that the world model misbehaves in the same way that we have seen LLMs misbehave?