David Luan: evaluation

Transcript context

…But like with OpenAI, like kind of give their requirements, like co-design it or just work of whatever NVIDIA gave them. So we work really closely with them. There's, I'm not sure I can share all the stories, but examples of ones that I've found particularly interesting. So Scott Gray is amazing. I really like working with him. He was on one of my teams, the supercomputing team, which Chris Berner runs and Chris Berner still does a lot of stuff in that. As a result, like we had very close ties to NVIDIA. Actually, one of my co-founders at Adept, Eric Elson, was also one of the early GPGPU people. So he and Scott and Brian Catanzaro at NVIDIA and Jonah and Ian at NVIDIA, I think all were very close. And we're all sort of part of this group of how do we push these chips to the absolute limit? And I think that kind of collaboration helped quite a bit. I think one interesting set of stuff is knowing the A100 generation, that like quad sparsity was going to be a thing. Is that something that we want to go look into, right? And figure out if that's something that we could actually use for model training. Really what it boils down to is that, and I think more and more people realize this, six years ago, people, even three years ago, people refused to accept it. This era of AI is really a story of compute. It's really the story of how do you more efficiently map actual usable model flops to compute, Is there another GPT 2, 3 story that you love to get out there that you think is underappreciated for the amount of work that people put into it?…

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