Shawn Wang: belief

Transcript context

…Yeah. Yeah. And so that's the way I, I would rather just lean on human feedback and just continue to make that more and more robust and more and more useful. And, you know, you can say some people are like GPU poor and GPU rich. We're like, we're feedback rich. Like when you've got one and a half million people a day, we get as much feedback from humans as we want. So we're not in a position where we needed to have the evals very much. Yeah. And when we do, we saturate them pretty quick. So a safety one, you know, within a month, we don't need to use it anymore because it's sort of, it's, you know, the issue has been addressed. I think one problem I have, and this is a broader products question maybe, is that the ELOs apply to the whole user population. That's right. Clearly the user behavior, there's segments that have like, I'm a role play person, I'm a therapy person, I'm a not safe for work person. You don't split them? This is why I say like, I think we're in year four of like a 20 year thing where it's like, at the end of the day, I'm a role play person. And I think if we all go on like Spotify or like, imagine if Spotify only had the top five musicians, I think it would retain over 85% of its existing users. Yeah. Right. And I think if YouTube, if YouTube only kept the top five content creators, it would be enough for the vast majority of people. The thing I'm just trying to share here is there's one surprising thing about humans is their preferences are pretty correlated. What you find funny and entertaining, I find funny and entertaining, and he finds funny and entertaining. There might be degrees of variation in it, I might find it super funny, you might find it only slightly funny, but optimizing to a global works very, very well. And for segmentation to be really powerful, segmentation will work amazing if you found a comment super boring, and I found it super fun. If we could segment that, then that would unlock really powerful stuff. But unfortunately, that's not the shape of human behavior, right? It's like, I might rank it 10 out of 10 funny, you might rank it 7 out of 10 funny. And it's like, it doesn't give you... It doesn't give you as much space to play as you would hope. It's an element of the diversity of content that AI can produce right now, which is it's not as diverse as if you consider a platform like YouTube, you can watch a Mr. Beast video, that's totally different to a makeup tutorial. So there's enough diversity there where if you go on my YouTube feed, it is totally different to my sister's one. My sister's one, it's all like women, and if you go on mine, it's all like bald, middle-aged men, either talking about MMA or, right? I think with AI, it's still a bit too early for that degree of segmentation. So I think it all comes, the recommender systems, the personalization. But this is why I like the, don't start with the technology, start with the problem. The problem is UGC. We must give users the tools to build more variety and more engaging content.…

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