Speaker unverified: belief

Transcript context

…cosmetics. Yeah, so I think, you know there's more questions around data you know, I think a lot of people are talking about the interview that Mira Murady did with the Wall Street Journal, where she, like, just basically had no, had no good answer for where they got the data for Sora. I, I think this is where, you know, there's, it's in nobody's interest to be transparent about data, and it's, it's kind of sad for the state of ML and the state of AI research but it is what it is. We, we have to figure this out as a society, just like we did for music and music sharing. You know, in, in sort of the Napster to Spotify transition, and that might take us a decade. Yeah, I do. I, I agree. I think, I think that you're right to identify it, not just as that sort of technical problem, but as one where society has to have a debate with itself. Because I think that there's, if you rationally within it, there's Great kind of points on all side, not to be the sort of, you know, person who sits in the middle constantly, but it's why I think a lot of these legal decisions are going to be really important because, you know, the job of judges is to listen to all this stuff and try to come to things and then have other judges disagree. And, you know, and have the rest of us all debate at the same time. By the way, as a total aside, I feel like the synthetic data right now is like eggs in the 80s and 90s. Like, whether they're good for you or bad for you, like, you know, we, we get one study that's like synthetic data, you know, there's model collapse. And then we have like a hint that llama, you know, to the most high performance version of it, which was one they didn't release was trained on synthetic data. So maybe it's good. It's like, I just feel like every, every other week I'm seeing something sort of different about whether it's a good or bad for, for these models. Yeah. The branding of this is pretty poor. I would kind of tell people to think about it like cholesterol. There's good cholesterol, bad cholesterol. And you can have, you know, good amounts of both. But at this point, it is absolutely without a doubt that most large models from here on out will all be trained as some kind of synthetic data and that is not a bad thing. There are ways in which you can do it poorly. Whether it's commercial, you know, in terms of commercial sourcing or in terms of the model performance. But it's without a doubt that good synthetic data is going to help your model. And this is just a question of like where to obtain it and what kinds of synthetic data are valuable. You know, if even like alpha geometry, you know, was, was a really good example from like earlier this year.…

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