Justin Johnson joined Fei-Fei Li's Stanford lab in 2012, the same quarter AlexNet was released.
“the quarter that I joined your lab was the same quarter that AlexNet came out.”
Person
Justin Johnson is an AI researcher who appeared on the Latent Space podcast with Fei-Fei Li to discuss spatial intelligence, world models and the computing power behind modern AI. He says he joined Li's Stanford lab in 2012, the same quarter AlexNet was released.
Johnson says performance per GPU card is about a thousand times higher than in the AlexNet era, and that NVIDIA's Blackwell chips deliver about the same performance per watt as Hopper. He also describes Marble, a World Labs tool that generates 3D scenes users can edit, and says it can rebuild a kitchen from two photos.
1 episodes · 12 quoted claims
Justin Johnson joined Fei-Fei Li's Stanford lab in 2012, the same quarter AlexNet was released.
“the quarter that I joined your lab was the same quarter that AlexNet came out.”
Johnson says performance per GPU card is about a thousand times higher today than in the AlexNet era.
“But even from AlexNet to today, we're getting about a thousand times more performance per card than we had in AlexNet days.”
Johnson says the compute that can be applied to a single model is about a million times larger than at the start of his PhD.
“So the amount of compute that we can marshal today on a single model is, you know, about a million fold more than we could have even at the start of my PhD.”
Johnson says NVIDIA's Blackwell GPUs deliver about the same performance per watt as Hopper, showing a scaling limit.
“even going from Hopper to Blackwell, like the performance per watt is about the same.”
Johnson showed a real-time captioning demo at ICCV in Santiago, Chile, with the model running on a server in California at about one frame per second.
“the model was running on a server in California. So it was like actually able to stream like all the way from California down to Santiago.”
Johnson says DeepSeek rendered text as images and fed them to a vision model, illustrating the pixel-based approach to language.
“maybe it's not super efficient to like, you know, render your text to an image and then feed that to a vision model. That's exactly what DeepSeek did.”
Marble lets users interactively edit a generated 3D scene, such as recolouring an object, and then generate new worlds from the edits.
“Like I could generate this scene and then say, I don't like the water bottle, make it blue instead.”
Marble natively outputs Gaussian splats, tiny semi-transparent particles with a position and orientation in 3D space.
“each one is a tiny tiny particle that's semi-transparent has a position orientation in 3d space”
World Labs has also built a model called RTFM that generates frames one at a time as the user interacts.
“our recent RTFM model that does generate frames one at a time.”
Johnson says rendering splat scenes at 30 to 60 frames per second on an iPhone from four years earlier limits how many splats a scene can have.
“30 to 60 fps on like an iphone from four years ago then you are a bit limited in like the number of splats that you can handle”
Johnson says Marble can reconstruct a kitchen from two photos and let users edit countertops, floors or cabinets, a use the team did not specifically build.
“Just like take two images of your kitchen, like reconstruct it in Marble, and then use the editing features”
Johnson explains that a standard transformer treats tokens as an unordered set, and only positional embeddings inject sequence order.
“A transformer is actually a model of a set of tokens, right? The only thing that gives, that injects the order into it, in the standard transformer architecture”