Fei-Fei Li gives AI a new point of view
The only new Welch Labs upload traces a line from ImageNet's labeled photos to a model that can generate a scene from another angle.
Can Fei-Fei Li Reinvent AI (Again)?
Up next
Why this story? The channel's weekly summary: what went up this week, what is drawing views now, and where the channel stands among similar channels.
"The next view is our next token," Fei-Fei Li says in "Can Fei-Fei Li Reinvent AI (Again)?" It's her shorthand for the difference between a chatbot and Atlas: instead of asking what bit of text comes next, the World Labs model asks what a scene would look like if you moved the camera.
Welch Labs starts with ImageNet, the vast collection of labeled images Li helped build. AlexNet's breakthrough in the 2012 ImageNet challenge showed how much a neural network could learn from those examples. The new video asks where robotics might find a comparable supply of training material. Li and fellow World Labs cofounders Justin Johnson and Ben Mildenhall help trace the route from Marble, which represents scenes as clouds of soft blobs, to Atlas, which predicts new views in sequence. One proposed use is to reconstruct real spaces so robots can practice in simulation. The video includes a sponsored segment.
That gap between a convincing view and a useful simulation is where the comments get interesting. Some viewers ask what counts as a world model in the first place. Others bring up what a robot has to deal with that a camera can't necessarily tell it, like how heavy an object is or how much it will slip.
The interview drew about 154,000 views in its first 30 hours and is the channel's fastest-moving upload. Behind it is "The most cited paper of the century is a brilliant hack," an older explainer about the shortcut connections that make deep neural networks easier to train. Welch Labs added subscribers over the week. Among similarly sized science channels, Reactions had a busier week for views, while Applied Science was quieter.
Published on Trari
Where next?
Things we can learn from the video
Press to jump
- 3:22AlexNet's ImageNet breakthrough showed what deep learning could do with a large collection of labeled images.
- 4:01Li's robotics research exposed a shortage of training data about how objects and spaces behave.
- 10:01Marble represents scenes with soft 3D blobs, while Atlas takes the approach of predicting new views.
- 20:00Video models that look at future as well as past frames are hard to use for live exploration.
- 29:31Atlas generates views in sequence and can work with inputs such as depth maps and camera poses.
- 33:17One proposed robotics application reconstructs real environments so robots can practice in simulation before returning to the real world.
A professional in this field? Want to write about Tech on Trari? Tell us ›
What viewers say
I'm not too worried that a chatbot will take my job. However Fe Fe Li's work can lead to robotics breakthroughs that will change work forever. I'm excited and nervous.
@IkbeneengeitSee it on YouTubeAtlas is a significant step toward this architecture, but it has not yet demonstrated a complete, general-purpose physical world simulator.
@nneislerSee it on YouTubeDo note that LeCun _vehemently_ disagrees that this work is a world model in his sense, so we seem to have two different definitions here.
@coffeeimpliescodeSee it on YouTube
2 more comments
Great video, with great visualizations as always!
One note: the audio balancing in the last segment might've been better if the music were quieter relative to the voiceover.@LiamGerstSee it on YouTubeAlways a pleasure to see a new Welch Labs video!
@wyattbaillifSee it on YouTube
Public comments from the video's page on YouTube
Want to know when somebody answers you?
The videos drawing views on Welch Labs right now
- Channel
- Welch Labs
- Subscribers
- 915,000 ▲ 1,000 this week
- Channel views, last 7 days
- 101,451
- Videos published this week
- 1
- Video published
- November 20, 2026
- Video views so far
- 153,506
- Views on its first day
- 124,053
The numbers
Here's how the Atlas interview and the rest of Welch Labs' channel did this week.
💡 Test your knowledge
Which model made its breakthrough in the 2012 ImageNet challenge?
AlexNet's performance in the 2012 challenge was a landmark for deep learning in image recognition.



