Shawn Wang: uncertainty

Transcript context

…Briefly saw it yeah, I'm not, I didn't really take a look at it. Basically, the general idea is that most multimodal models, like Lava or Flamingo, which are late fusion, which is you freeze, freeze, and then you join together, versus early fusion where you do it properly, where, like, everything is, you know, All the modalities are present in, in the, in the early training stage, and it seems like things are trending from late fusion to early fusion. Is is the general thesis with GP four Oh being very obviously early fusion, you guys, I I would class it as early fusion. I, I, I don't know if you have commentary on whether this is obvious to you or this is the, this is the way, or they'll just be, they'll coexist. I think whenever possible, like early fusion is better, I think there will still be a lot of work steps. Dual late fusion just because of like it's a GPU, poor No, no, no. GPU. Okay. Par partially, right. I, I see this as like an art, as an artifact of the line between language research researchers and vision researchers, and more of like, okay, like people who are training language models, they put out like LAMA whatever, and then somebody takes it and then. Do Lakefusion on top of it. It's more like a It's Conway's Law. They're shipping the org chart. Yeah, yeah, yeah, I think so. I don't know what law it was. Conway's Law. Okay, I didn't know about that. But it's kind of like an artifact of the organization, don't you think?…

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