Alessio Fanelli: belief

Transcript context

…The link prediction objective can be seen as like a neural PageRank because what you're doing is you're predicting the links people share. And so if everyone is sharing some Paul Graham essay about fundraising, then like our model is more likely to predict it. So like inherent in our training objective is this, uh, a sense of like high canonicity and like high quality, but it's more powerful than PageRank. It's strictly more powerful because people might refer to that Paul Graham fundraising essay in like a thousand different ways. And so our model learns all the different ways. That someone refers that Paul Graham, I say, while also learning how important that Paul Graham essay is. Um, so it's like, it's like PageRank on steroids kind of thing. Yeah. I think to me, that's the most interesting thing about search today, like with Google and whatnot, it's like, it's mostly like domain authority. So like if you get back playing, like if you search any AI term, you get this like SEO slop websites with like a bunch of things in them. So this is interesting, but then how do you think about more timeless maybe content? So if you think about, yeah. You know, maybe the founder mode essay, right. It gets shared by like a lot of people, but then you might have a lot of other essays that are also good, but they just don't really get a lot of traction. Even though maybe the people that share them are high quality. How do you kind of solve that thing when you don't have the people authority, so to speak of who's sharing, whether or not they're worth kind of like bumping up? Yeah. I mean, you do have a lot of control over the training data, so you could like make sure that the training data contains like high quality sources so that, okay. Like if you, if you're. Training data, I mean, it's very similar to like language, language model training. Like if you train on like a bunch of crap, your prediction will be crap. Our model will match the training distribution is trained on. And so we could like, there are lots of ways to tweak the training data to refer to high quality content that we want. Yeah. I would say also this, like this slop that is returned by, by traditional search engines, like Google and Bing, you have the slop is then, uh, transferred into the, these LLMs in like a search GBT or, you know, our other systems like that. Like if slop comes in, slop will go out. And so, yeah, that's another answer to how we're different is like, we're not like traditional search engines. We want to give like the highest quality results and like have full control over whatever you want. If you don't want slop, you get that. And then if you put an LM on top of that, which our customers do, then you just get higher quality results or high quality output.…

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