Matan Grinberg - Factory

It turns out that the killer use case for code generation AI is not in fact building more zero-to-one CRUD apps. Despite all the sexy demos that you've seen on Twitter, unfortunately, enterprise engineering organizations and VPs of engineering actually don't really care about building CRUD apps. If you've ever worked at an organization that has a 30-year-old COBOL code base, You know that the problems in these types of organizations are a lot less sexy, a lot less demoable, but also these are problems that VPs of engineering would open up their checkbook if you can solve. These are the problems that we are solving at Factory. So how do we get here from theoretical physics? Quick background on me. So prior to Factory, for 10 years I was doing string theory, which obviously has nothing to do with AI or software engineering. Took me 10 years to realize that I was only doing it because it was hard and not because I actually, you know, loved the problem space. You know, some people takes a little less time to realize that, but it took me 10 years. And while doing my PhD at Berkeley, I ended up going to a hackathon in SF and having what I call intellectual love at first sight with Eno, who would become my co-founder. In a period of 8 days, I dropped out of my PhD, Eno quit his job at Hugging Face, and We got some money from Sequoia and started Factory. And so what have we been up to since then? Well, our mission is to bring autonomy to software engineering. What that means more concretely, we are building autonomous systems called droids. These droids automate tasks in the software development lifecycle, in particular, tasks that engineers don't want to do. So reviewing code, refactoring code, migrating code, testing, documentation, These are things that enterprise engineers spend 90% of their time on and wish they spent 0% of their time on it. We started in April of last year and we're now a team of 20 people in SF every day. And it's truly the privilege of my life working with some of the smartest people I've ever met. They come from companies such as these and we are about— of the 20, it's about 15 on the engineering side. And here are the droids. In particular, they cover the different steps that we know exist within the software development lifecycle. In this problem of code generation, and in particular autonomous code generation, the main thing that needs to be proven is reliability and value that can actually be delivered to enterprises. And when your main problem is reliability, you want to take as many simplifying assumptions as possible. And so instead of having an open-ended agent, We are building droids that are focused on these particular tasks. How do we actually do this? Well, unfortunately, doing semantic similarity search over a vector database is not enough to make a fully autonomous agent. Who would've thought? And so we've taken a very first principles approach, which is what do we as humans need to be good engineers? We need to have access to GitHub, Slack, Linear, and we need to be able to reason about the code base. And these droids in particular need to have the ability to reason similar to how we as humans do. Which is much more similar to knowledge graphs, static code analysis, and much less like semantic similarity search. We also need to give droids the ability to execute code, which is what these platform environments are. These are some of the things that we're working on in terms of the cutting edge of autonomous systems. So we've had state-of-the-art in agent encoding. We've got breakthroughs in orchestration, planning, and retrieval. This is outside of the model, outside the black box. Because that's where there's the most low-hanging fruit in terms of pushing forward the leading edge. We're hiring across the board on engineering. Thank you all. I'll be in the north corner of the building