01 logo

The Note on the Podium Said “Don’t Say Closet”

Fei-Fei Li’s CoRL 2026 talk introduced Atlas, a world model that rebuilds Stanford from 33 photos, and ended with Lisa Su onstage.

By JinPublished 2 days ago • 6 min read

In the corridor of the Austin Convention Center, two people in hoodies were talking in low voices about AMD’s acquisition of World Labs. One said, “So today she counts as AMD’s person.” The other didn’t answer, just pointed toward the entrance of the hall. The door was open. In the first row, in the aisle seat, Lisa Su was already sitting down.

When Fei-Fei Li walked onstage, she wasn’t holding a clicker. She first looked down at the podium. A handwritten note was stuck there, the writing small. Later, when the camera swept past, only one line was visible: “Don’t say closet.”

Yuke Zhu gave the opening introduction. He was one of the organizers of this year’s CoRL and also Li’s student. He told a detail from graduation: Li had said to him, “When you were in my lab, you were my student. From now on, you are my friend.” There was applause from the audience. Li stood in the wings, didn’t look at the crowd, and smiled down at the floor.

Her first sentence did not mention the acquisition. “It’s a great honor to be invited to CoRL, especially since several of my former students are organizers. You did a great job. I know, because a lot of my papers and workshop proposals were rejected by you.” The room laughed. She waited for the laughter to stop before saying, “Since we’re at CoRL, let’s start from the motivation in robotics.”

Three boxes appeared on the large screen, labeled “State,” “Action,” and “Data.” She said the fundamental problem robotics has to solve is: given the world state, given the goal, given the robot, what should be done next at every moment? Understanding the world state, whether implicit or explicit can be discussed, but “understanding” itself matters; predicting actions with consequences and planning; and data is always the hard problem. Robotics is a classic long-tail problem, successful deployment is limited, so one always hopes to simulate as many environments as possible to supplement real data.

“These three boxes,” she pointed at the screen, “are what world models need to fill.”

She split world models into three dimensions: rendering, planning, and simulation. Understanding the environment corresponds to rendering; predicting the consequences of actions corresponds to planning; simulating world states corresponds to simulation. She did not think these three are completely independent, but Atlas, which she was going to talk about today, mainly addresses rendering and simulation, especially the geometry of simulated worlds and 3D/4D dynamics.

She read the definition of Atlas in a flat tone: a self-regressive foundation model trained from scratch, a single model, receiving multimodal inputs: text, images, video, depth signals, graybox or 3D structure signals. It outputs pixel frames. It not only generates frames that are beautiful in space and time, but also synthesizes novel views, thereby achieving high-precision, high-fidelity 3D reconstruction.

The slide turned to the next page. It had only one line: Viewpoint is all you need.

She paused for two seconds. “Our formulation is, viewpoint is all you need. Predicting the next viewpoint under strong camera conditions, that is our objective function. Architecturally, it is a multi-view autoregressive diffusion Transformer.”

She began to explain the unification. One extreme: given multiple images, Atlas performs reconstruction; when there are enough images, it performs dense reconstruction. The other extreme: zero pixel conditioning, only a text prompt, and Atlas performs pure generation, generating multiple frames of view-consistent images. Connecting the two ends is sparse reconstruction: given few views, in the extreme only one image, Atlas estimates camera poses, then generates as many views as possible, reconstructing and generating a complete world.

“One model, freely switching from dense reconstruction to sparse reconstruction to pure generation.”

Then she showed Stanford’s Main Quad.

The first slide had two input images. The model “hallucinated” a Main Quad. People familiar with Stanford know this is not the real Main Quad; there was a hint of those arcades, but most of it was generated. Someone in the audience laughed because the arcade was crooked. Li didn’t laugh. She said, “It resembles it, but it isn’t.”

The second slide had eight images. The overall look of the quad basically emerged, but the famous Memorial Church was missing. She pointed her laser pointer at the empty spot on the screen: “Memorial Church is not here, because you didn’t show it to the model.”

The third slide had 33 images. They were all taken by a person walking around on the ground with a camera. When the aerial view appeared onscreen, she paused for a moment and did not explain why ground-level photos could generate an aerial view. She only said, “These images are all sparse, but one of them contains Memorial Church.”

She went through the quantitative results quickly. A few days earlier, a third-party organization they had not known in advance, Rapidata, used about 500,000 human judgments to run a blind test of camera-controlled generation across several leading video models. Atlas ranked first. She did not say who was second, only that it was “clearly ahead.” Atlas can generate view-consistent videos longer than 60 seconds.

At this point she suddenly flipped back one slide, to the previous-generation model from a year and a half ago. Its generation range and geometric consistency were both limited. “Someone criticized us, saying, ‘What you’re making isn’t a world model, it’s a closet model.’” She paused and waited for the laughter to rise. “Now we have walked from the closet into a much bigger world.”

She picked up the water bottle on the podium and took a sip. The cap wasn’t screwed on tight; it tilted when she put it down, and she straightened it before continuing.

Reconstruction comparisons: in depth prediction across seven datasets, Atlas even surpassed the most recent specialized 3D reconstruction and depth prediction models. Atlas itself does not have to use Gaussian Splatting as a representation, but if Gaussian Splatting is attached downstream, then across six datasets, when compared with specialized models on Gaussian Splatting reconstruction error, Atlas matched them. They also adjusted the number of input views, smoothly transitioning from sparse reconstruction to dense reconstruction, and performed well on both datasets.

She showed that famous apartment set: eight images, reconstructing a large space, with the ability to sample various viewpoints. Then a “single-image lift”: one input image, and a Gaussian Splatting world appears. In another example, the model not only generates the scene but also generates depth information, allowing artists to precisely control camera direction.

“Bullet time.” When she said these two words, the screen showed a Matrix-style surround shot. “Back then it required hundreds of cameras in an array. Now, we temporarily set up three iPhones in the office.”

The last demo was real-time. A staff member below the stage was “driving” the viewpoint with a keyboard, and the image onscreen followed. She gave a number: 15 fps. “This is the distilled version; both fidelity and view consistency are still there.”

Robot simulation came last. She said Atlas has potential uses in robot simulation, and they are testing it: both for training navigation robots and for connecting the real world into simulation and then back to real robot manipulation. “These robots were trained only in simulated virtual worlds, yet can be deployed zero-shot to the real world.” After saying this, she reached out, picked up an articulated robot model toy on the podium, and put it back in place, gently.

The promotional segment was short. The Stanford BEHAVIOR Challenge was in its second year, with winners announced tomorrow; she had six papers at CoRL; Atlas would soon be open to the whole world. Follow their X account for the premiere. “I promise it will be a great video.”

When the talk ended, Lisa Su stood up from the first row. The two sat down in high stools onstage. Li’s first sentence was: “Is your chip’s throughput enough to run a 60-second view-consistent video?” Lisa Su laughed, did not answer directly, and first turned her head to glance at the audience.

The fireside chat title mentioned compute, robots, and “the biggest mistake of this era.” The details are in the companion transcript. As the room emptied, someone passing the podium saw the handwritten note still stuck there, its corner curled up. Below the small line, another line had been added, as if written later: “But you can say world.”

tech newsthought leadersapps

About the Creator

Jin

Writer of reamstories

https://reamstories.com/jin

Enjoyed the story? Support the Creator.

Subscribe for free to receive all their stories in your feed. You could also become a paid subscriber, letting them know you appreciate their work.

Subscribe For Free

Reader insights

Comments

There are no comments for this story

Be the first to respond and start the conversation.

Sign in to comment
    Written by Jin