Varun Mohan: evaluation

Transcript context

…Oh, I really appreciate that. I'm so excited to have you here. I feel like just you guys have become this overnight success, which is definitely not an overnight success, but I feel like I've been hearing about Windsurf more and more as people's favorite AI tool. And I just don't think people know the story behind Windsurf, behind Codeium, the company that you built. So I thought it'd be good to maybe just start there and have you just briefly share the history of Codeium and how Windsurf emerged out of Codeium. Yeah. So the company was actually started close to four years ago. As you know, AI coding was not a thing four years ago. ChatGPT was not out four years ago. At the time, we actually started out building GPU virtualization and compiler software. Before this, I worked in autonomous vehicles. My co-founder, who I had known since middle school, worked on AR/VR at Meta. And for us, we believe deep learning would touch many, many industries. It wouldn't just touch autonomous vehicles. It would touch financial services, defense, healthcare. And we believe these applications were hard to build, these deep learning applications. So we made it possible for you to effectively run these complex applications on computers without GPUs, and we would handle all the complexity of being able to actually run the workload on the GPUs for you. And we were able to optimize these workloads a ton. And in the middle of 2022 rolled around and we had a couple million in revenue and we were managing upwards of 10,000 sort of GPUs. We had eight people. At the time, we were free cash flow positive. But I think what we felt was once these generative models started to get very good, we sort of felt a lot of what we built was not as valuable. And this was a very, very hard moment for us at the company. We were only eight people at the time, but we felt, "Hey, would people be training these very bespoke sentiment classifier models anymore that were very, very custom models? Or would they just ask GPT-N, is this a positive or a negative sentiment?" Probably it's going to be the latter, right? And in a world in which everyone was going to run generative AI models, why would an infrastructure company be a differentiator? Because everyone is going to run the same kind of infrastructure down the line. So instead what we decided to do was we believe generative AI was almost going to be like the next internet. And in that case, what we should go out and do is build the next great apps like Google, like Amazon. And we vertically integrated and actually took our infrastructure, our inference infrastructure to go out and build Codeium at the time. And at that time, we were early adopters of GitHub Copilot and we thought the coding space was going to get tremendously disrupted in the next coming years. So we actually took our infrastructure, we ran our own models in massive scale. We even trained our own models. In the very beginning it was very, very simple. It was purely an autocomplete model, which basically means that as the user was typing, we'd complete the next one or two or three or four lines of code. But we provided the product entirely for free in all the IDEs that developers coded. That meant to VSCode, JetBrains, Eclipse, Visual Studio, Vim, Emacs. And the reason why we were able to build it for free was because of our infrastructure background. We were able to optimize these workloads a ton. I guess very quickly after that, some large businesses also wanted to work with us. e to build it for free was because of our infrastructure background. We were able to optimize these workloads a ton. I guess very quickly after that, some large businesses also wanted to work with us. And we built out this enterprise motion to work with these large companies like Dell, JPMorgan Chase. And for them the bigger thing wasn't just, "Hey, could we autocomplete code or could we chat with the code base?" It was, "Could you offer us a secure offering that was also personalized to all the private data inside the company?" So we took our infrastructure and made it so that we invested a ton in making sure that we deeply understood these large companies code bases. And that's what we were working on until six months ago. It's not that we've stopped working on that, but basically what we realized six months ago was we were getting limited by the IDEs that we were already working in. So VSCode, which is a very popular IDE, had a ceiling for the AI capabilities, we could showcase our users. And because of that, we decided to go out and fork VSCode and build our own IDE with some of these new agentic capabilities. And over time in the last couple of years, the model capabilities have also been growing exponentially year over year. And that's sort of where we are right now. I skipped a lot of pieces there, but that's what we're landed.…

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