Shawn Wang: belief

Transcript context

…Yeah. Maybe I'll even take one more step back and say it's not even entirely worked out. I think the same way it is in text. And so it's an evolving field. If you take a giant step back, I think audio has been lagging images and text for a while. So I think very roughly you can think audio is like one to two years behind images and text. But you kind of have to think today like text was in 2022 or something like this. And you know, the transformer was invented. It looks like it works, but it's, it's, it's far, far less established. And so you know, I'll give you the way we think about the world now, but just with the big caveat that, that I'm probably wrong if we look back in a couple of years from now. And I think the biggest thing is you see both transformer based and diffusion based models for audio in, and in ways that that is not true in text. I know people will do some diffusion for text, but I think nobody's like really doing that for real. And so we, we prefer transformers for a variety of reasons. And so you can think it's very similar to text. You have some abstract notion of a token and you train a model to predict the probability over all of the next token. So it's a language model. You can think in anything, language model is just something that assigns likelihoods to sequences of tokens. Sometimes those tokens correspond to text. In our case, they correspond to music or audio in general. And I think we've learned a lot from our friends in the text domain, from the pioneers doing this of how well these transformer models work, where do they work, where do they not work? But at its core, the way we like to do things with transformers is exactly like it works in text. Let me predict the next tiny little bit of audio, and I can just keep doing that and doing that and generating audio as long as I want. Yeah. I think the, the temptation here is to always try to bake in some specialized knowledge about music or audio. And so, and obviously you will get an improvement in, in your output. If you try to just say like, okay, like here's a set of notes for, you know, here's a set of tokens that only do jazz or only do, you know, like voices. How general do you make it versus how specific do you make it? We've always tried to do things, you know, quote unquote the right way, which means that at the beginning things are going to be hard and worse than other ways. But that is to say, bake in as little kind of implicit knowledge as possible. And so, the same way you don't program into GPT, you don't say this is a noun and this is a verb, but it has implicitly learned all of those things. I've never seen GPT accidentally, you know, put a, put a noun where it meant to put an article in English. We try not to impose anything about music or audio in general into the model, and we kind of let the models learn things by themselves. And I think things are beginning to pay off, but it's, you know, it's not necessarily obvious from the beginning that that was the right thing to do. So, for example, you know, you could take something like text to speech and people will do all sorts of things where you can program in things like phonemes to be the basis for what you do. And then that kind of limits you to the set of things that are expressible by phonemes. And so, ultimately that works really well in the short term. In the long term, it can be quite limiting. And so, our approach has always been to try to do this in its full generality, as end to end as we can do it. Even if it means that in the short term we were a little bit worse, we have a lot of confidence that in the long term that will be the right way to do it.…

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