GitLab CEO on why AI isn't helping enterprise ship code faster

Frederic Lardinois (0:03)

After a bit of a winter break, we're back with The New Stack agents. A new season starts today, and our guest today is, I always want to say Paul Metan, instead of Metan Paul, she treats from Ryder, who leads product over there. Metan Paul, thanks for being here.

Matan Paul (0:21)

Well, it's a pleasure to be here. Nice to meet you, Frederick, and I'm excited for our talk.

Frederic Lardinois (0:26)

Absolutely, absolutely. Yeah, I mean, I followed Ryder for a while, and I've met with some of your colleagues over the years, but never like really in a formal way. So I thought maybe a lot of people might be like me, they saw Ryder in the early days, and it's obviously evolved since then, and evolved quite a bit. So maybe talk to me a little bit of what that evolution was like. And as a caveat, you've been there for two years, the company is a little bit older, but I'm sure you're quite aware of where it came from.

Matan Paul (0:58)

Yeah, for sure.

So the story of Ryder is the story of AI, and we have evolved the same way as the underlying technology and offering has evolved. Me and Wasim, our co-founders, basically started Ryder about five years ago, close to six almost. And at the time, they've also built, essentially, our LLMs from the ground up, together with ChachiPT, at the same time ChachiPT 2 and the original Transformer article came about. Since then, we've been building models on an ongoing basis and building capabilities. Unlike some of the, let's call it, other offerings in the market, we did not start as a research lab, turned consumer, dabbled into enterprise from day one. Our goal and focus was to bring AI and its capabilities to the enterprise world. As I'm sure you know, when you build for companies, for enterprises, there are certain things that are outside of your core product and offering that are as important as not more. So things like data retention, a single-tenant versus multi-tenant deployment, security, how you train your model, what data goes into the model, et cetera, et cetera. And these are all the things that we've built from day one as core primitive for the platform and the product that allowed us to expand in the enterprise. You've also correctly identified that we have evolved. I think Ryker historically in the early days was mostly connected to the marketing space.

Frederic Lardinois (2:58)

Hence the name, right? Hence the name at the time.

Matan Paul (3:03)

Also suggest the name writing. I think where I would challenge that even on the name is our specialization, I would say, and where we really spike on is, it's called the written word. Now, the written word could be a lot of things. It could be marketing, it could be in sales, it could be in code. Code is also written. So when we say content in the context of writer, that has evolved from just like, let's say an ad copy to anything and everything that is the written word. And we've expanded our business in the global 2000 and the Fortune 500 to capture more and more of these use cases across engineering, product, marketing, sales, customer support. But all of these use cases are for internal purposes. We are very rarely customer facing. We tackled the gnarliest, hairiest problems within these organizations, which are usually, how do you take existing operational workflows that they have and basically bring them to the age of AI and help these businesses kind of unlock their potential and continue the growth that they've been showing for hundreds of years.

Frederic Lardinois (4:25)

And the age of AI and now also the age of agents, right? Because at least over the last year or so, you've also built out this agent orchestration layer and kind of tooling both for developers and maybe non-developers as well to build these agents, right? And workflows.

Matan Paul (4:42)

Yeah, so I think during this corner, I apologize in advance, and this is an industry-wide problem. Definitions of names is a bit vague. I will probably use the term LLM, agents and AI interchangeably. But similar to how we've evolved, also the technology has evolved. So if you think of LLMs as the entry point, let's say, when the hype started, they were very sensitive to prompting, very rigid. And then the next generation would be what we call custom apps or eventually custom agents when the marketing language started landing, which is prompt chaining, model selection, that can be a little bit more rigid and indeterministic. And yes, we are now, I would say, in the phase three of the industry, which is agentic systems. And these are where we start delivering on the great promise, or more delivering on that great promise of AI, which is more autonomous, skill-based, can invoke and make, quote unquote, decisions and rationalize these decisions to invoke other agents or other tools to take actions in your system, internally or externally. Writer agent, which is our latest, I would say, iteration of our product, or writer as a whole as it is now, is literally that. It's like an agentic system built on our models and our capabilities to support the enterprises.