Vipul Ved Prakash: evaluation

Transcript context

…Yeah, makes sense. It's a very important field and I think hopefully there's a good third party that emerges from this. So I just want to touch on one more piece, which is I think I'm appreciating from this discussion that fine tuning is a bigger part of your business than I thought. The other big player in fine tuning is Mosaic. Well, Mosaic is more training, but like there's a bunch of other players in the fine tuning space. If I was a prospective fine tuning customer, what do I come to you with? Do I come to you with my custom data and that's it? Do I also have to write the fine tuning code? What level of engagement do you do with your customers? I think across the spectrum, our customers are training models, pre-training models from scratch and many of them will bring their data sets, you know, user infrastructure and training stack to train their models. There are others who have trained smaller models and want to scale up, scale up across infrastructure, scale up across data. So we'll sort of help them do that. We will have customers who are sort of initially started a little bit more consultative. They have a particular task and idea in mind and we will help them get from there to the data set and the right model to achieve that task. So it's a spectrum and, you know, our goal is to, we're trying to productize as much of this as possible. So that the whole process can be fast and scalable. I would say there is a lot more understanding around fine tuning now, like even the last six months, there are, you know, source tools, recipes, literature, podcasts, discord channels where people are figuring out and it really is in many ways, one of the successes of open source is you have small collectives of, you know, engineers who have created, who are now creating the top models on open source leaderboards. And I have tried out all sorts of different sort of, you know, data recipes, creating synthetic data. Merging models. Merging models. So it's, that's really fun to see. And I think that sort of agency that exists now is exciting. And that is, we see a lot of that sort of being applied into products and, you know, more commercial models that people are deploying in their applications. And then just to, I guess, wrap up the together, it's almost becoming like a platform as a service, because now you release together embeddings. How did you get 92.5 accuracy on 32K retrieval? And do you think we're kind of like getting to embeddings or just like, we did everything that we could, you know, we're getting to like the most optimized it's gonna get and then we should just focus on models and inference or do you think there's still room there to improve?…

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