Alessio Fanelli: evaluation

Transcript context

…I feel AI companies are really kind of driven by research. And Chris and I had been talking about how to reduce the cost of building models. We felt that there aren't really big data modes around foundation models. They are built from a subset of the web. What is difficult is the cost of capital to build these. And one of the ways in which you can reduce this cost is by making more efficient systems. With that, it was really about finding the right set of co-founders and team. In fact, when Chris introduced me to Ce, and I think within the first five minutes of talking to Ce, I was like, we are starting this company. And our early focus was thinking about this more sort of disparate set of resources, you know, GPUs around the internet. Can we use those to build? And we really have to compress communication for, you know, when we do gradient averaging, there's just a lot of traffic. And if you can reduce that somehow, you sort of open up the possibility of using cheaper compute, you know, across the network. And Ce's research for a decade has been in that subject. You know, and from there, finding, you know, other folks in the network, I think there is generally a lot of excitement and philosophical alignment around what we are doing, which, you know, we publish papers, we publish open source libraries and code, we build open models. And I think the people in academia in, you know, machine learning and NLP, that's really what they want to do. So I think that's been really a kind of kernel for, you know, composition of the company. And we're lucky to have, you know, at this point, attracted some of the best researchers in the field. So I think that's the most important thing. And, you know, the rest of it is sort of driven by us. A couple of these philosophies around independent systems and decentralization and good developer interfaces, you want to make it accessible. That's, you know, just as important. And the rest follows from there, I think. I want to try and fill in some of the blanks in the history of Together. I think people come on your website today and they say, you raised a hundred million dollars Series A. They're like, wow, these guys are like super legit company. But it feels like Red Pajama just came out a year ago. I remember we had Mike Conover in the studio, who had built Dolly at Databricks. And you announced it literally the morning we were recording. So we're like in the studio on our phones, looking at it. And it's like, wow, this is like the first time now there's like a good curated dataset to do open pre-training. So maybe let's start from there. Like, what was the motivation behind it? Why did you decide to do that? It's, datasets are one of the things that most people don't want to work on. They just want to do models, not datasets. Yeah. So, yeah, first one is not the first, right? So I think it's actually built on a whole bunch of amazing effort the community already have. For example, Eleuther have the pile, right? There's a whole bunch of amazing datasets they have, like C4, right, from Google, right? So I think really get inspired by the impact those like datasets have on the community, right? So I think when we did Red Pajama, it was a time that people are really fascinated by Lama, the model, like Lama 1, right? Which I feel like decades ago, right? But it's kind of, people are really excited about the quality, right? So that's really like a big shift in people how to think about open model. People start to see hope, right? So, but the one problem of Lama is the data recipe is being described in a pretty detailed way in the paper, but the data is actually not there. So, and our original thinking is how about we take the recipe and we try to do our best effort reproduction and try to put it out, such that we can learn from our mistakes in the reproduction together, right? So that's essentially the original thinking behind Red Pajama. And we have been pretty happy and excited about what community have been kind of build on it. For example, there's a dataset called Slim Pajama, right? Which do deduplication over our data, right?…

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