Edwin Chen: evaluation

Transcript context

…That is incredibly interesting. I want to ask you about something I've been very curious about over the past couple years. If you look at Claude, it's been so much better at coding and at writing than any other model for so long. And it's really surprising just how long it took other companies to catch up. Considering just how much economic value there is there, just like every AI coding product sat on top of Claude because it was so good Claude code and writing also. What is it that made it so much better? Is it just the quality of the data they trained on or is there something else? I think there are multiple parts to it. So a big part of it certainly is the data. I think people don't realize that there's almost like this infinite amount of choices that all the frontier labs are deciding between when they're choosing what data goes into their models. It's like, okay, are you purely using human data? Are you gathering the human data in X, Y, Z way? When you are gathering the human data, what exactly are you asking the people who are creating it to create for you? For example, in the coding realm, maybe you care more about front end coding versus back end coding. Maybe when you're doing front end coding, you care a lot about the visual design of the front end applications that you're creating, or maybe you don't care about it so much and you care more about, I don't know, the deficiency of it or the pure correctness over that visual design. And then other questions like, okay, are you carrying [inaudible 00:14:49] how much synthetic data are we throwing into the mix? How much do you care about these 20 different benchmarks? " Some companies, they see these benchmarks and they're like, "Okay, for PR purposes, even though we don't think that these academic benchmarks matter all that much, maybe we just need to optimize for them anyways because our marketing team needs to show certain progress on certain standard evaluations that every other company talks about, and if we don't show good performance here, it's going to be bad for us even if ignoring these academic benchmarks makes us better at the real tasks. " Other companies are going to be principled and be like, "Okay, yeah, no, I don't care about marketing. I just care about how my model performs on these real world tasks at the end of the day, and so I'm going to optimize for that instead." And it's almost like there's a trade-off between all of these different things, and there's like a... One of the things I often think about is that there's a... It's almost like there's an art to post training. It's not purely a science. When you are deciding what kind of model you're trying to create and what it's good at, there's this notion of taste and sophistication, like, "Okay, do I think that these..." So going back to the example of how good the model is at visual design. I'm like, "Okay, maybe you have a different notion of visual design than what I do. Maybe you care more about minimalism, and you care more about, I don't know, 3D animations than I do. And maybe this other person prefers things that look a little bit more broke." And there's all these notions of taste sophistication that you have to decide between when you're designing your post training mix, and so that matters as well. So long story short, I think there's all these different factors, and certainly the data is a big part of it, but it's also like what is the objective function that you're trying to optimize your model towards? That is so interesting. The taste of the person leading this work will inform what data they ask for, what data they feed it. But it's wild it shows the value of great data. Anthropic got so much growth and win from essentially better data.…

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