Shawn Wang: belief

Transcript context

…And I saw Jim Fan at a, at a very good tweet about why it's so impressive. And I think when you have somebody leading the embodied research at NVIDIA and he said that something is impressive, you should probably listen. So yeah, there's basically like, I think you, you mentioned like impacting the world, you know, that we live in. I think that's kind of like the key, right? It's like the LLMs don't have, a world model and Jan Lekon. He can come on the podcast and talk all about what he thinks of that. But I think SORA was like the first time where people like, Oh, okay, you're not statically putting pixels of water on the screen, which you can kind of like, you know, project without understanding the physics of it. Now you're like, you have to understand how the water splashes when you have things. And even if you just learned it by watching video and not by actually studying the physics, You still know it, you know, so I, I think that's like a direction that yeah, before you didn't have, but now you can do things that you couldn't before, both in terms of generating, I think it always starts with generating, right? But like the interesting part is like understanding it. You know, it's like if you gave it, you know, there's the video of like the, the ship in the water that they generated with SORA, like if you gave it the video back and now it could tell you why the ship is like too rocky or like it could tell you why the ship is sinking, then that's like, you know, AGI for like all your rig deployments and like all this stuff, you know, so, but there's none, there's none of that yet, so. Hopefully they announce it and talk more about it. Maybe a Dev Day this year, who knows. Yeah who knows, who knows. I'm talking with them about Dev Day as well. So I would say, like, the phrasing that Jim used, which resonated with me, he kind of called it a data driven world model. I somewhat agree with that. I am on more of a Yann LeCun side than I am on Jim's side, in the sense that I think that is the vision or the hope that these things can build world models. But you know, clearly even at the current SORA size, they don't have the idea of, you know, They don't have strong consistency yet. They have very good consistency, but fingers and arms and legs will appear and disappear and chairs will appear and disappear. That definitely breaks physics. And it also makes me think about how we do deep learning versus world models in the sense of You know, in classic machine learning, when you have too many parameters, you will overfit, and actually that fails, that like, does not match reality, and therefore fails to generalize well. And like, what scale of data do we need in order to world, learn world models from video? A lot. Yeah. So, so I, I And cautious about taking this interpretation too literally, obviously, you know, like, I get what he's going for, and he's like, obviously partially right, obviously, like, transformers and, and, you know, these, like, these sort of these, these neural networks are universal function approximators, theoretically could figure out world models, it's just like, how good are they, and how tolerant are we of hallucinations, we're not very tolerant, like, yeah, so It's, it's, it's gonna prior, it's gonna bias us for creating like very convincing things, but then not create like the, the, the useful role models that we want. At the same time, what you just said, I think made me reflect a little bit like we just got done saying how important synthetic data is for Mm-Hmm. for training lms. And so like, if this is a way of, of synthetic, you know, vi video data for improving our video understanding. Then sure, by all means. Which we actually know, like, GPT 4, Vision, and Dolly were trained, kind of, co trained together. And so, like, maybe this is on the critical path, and I just don't fully see the full picture yet. Yeah, I don't know. I think there's a lot of interesting stuff. It's like, imagine you go back, you have Sora, you go back in time, and Newton didn't figure out gravity yet. Would Sora help you figure it out? Because you start saying, okay, a man standing under a tree with, like, Apples falling, and it's like, oh, they're always falling at the same speed in the video. Why is that? I feel like sometimes these engines can like pick up things, like humans have a lot of intuition, but if you ask the average person, like the physics of like a fluid in a boat, they couldn't be able to tell you the physics, but they can like observe it, but humans can only observe this much, you know, versus like now you have these models to observe everything and then They generalize these things and maybe we can learn new things through the generalization that they pick up.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.