Your vantage point at Sourcegraph puts you in a really compelling position to opine on a lot of what's happening in AI across the software development life cycle. That's the crux of your platform. For folks who aren't as familiar, give a quick intro to what Sourcegraph is and some of the products you've built, and then we can poke into some of the different areas of the stack.

We started Sourcegraph because we knew how terribly manual it was to build software. We saw this inside the big banks, and we saw this on our open source projects. We wanted eventually to make it so that everyone could code. That's been our mission from day one. Everyone thought that was crazy. They're not saying it's crazy anymore.

But when we started, we had to start out with Code Search, which was incredibly valuable and solved that biggest problem of understanding the codebase. Then, on that foundation, when large language models started to get really good, we came out with Cody, which is the best chat RAG AI assistant for code. That was about two and a half years ago, before ChatGPT. We're working with Anthropic on that.

But it is crazy how fast things are moving. Tool-calling agents basically obliterate all technology that anyone has built in the past around information retrieval. They can use that as a subtask. So Amp is our next-gen coding agent. We came out with that just a few weeks ago, and we've tried to do as much as we can to future-proof it so that it will get better as the models get better. It is riding that model wave. I call this product-model fit, or model-product fit. You really want to be there when you're building a product, and it feels great to be there now.

You also sit in a unique position where you have a really rich understanding of people's codebases. How do you think about building your own stack, whether that's working with the models the folks at Anthropic and OpenAI have put out, versus ingesting data you get access to because of your platform? What are you leveraging within that stack?

Our customers are every dev at Stripe, at Palo Alto Networks, four of the six top US banks, and seven of the 10 biggest tech companies. So the most important thing, before I even talk about the data we're collecting, is that customer trust and relationship, and knowing what it's actually like to be building the software that powers the world and what the pains are of the devs writing that code day to day. That is the most valuable thing. We got that because we built Code Search, and that understands all your code. It helps you find out why something is broken, or who's done this already.

There have been some areas where we've been able, with customer permission, of course, to use data to train models to make our products better. Our autocomplete and our next edit suggestion models are based on that. But when you look at the real tool-calling agents, like Amp and Claude Code, that use state-of-the-art models, it's not about what proprietary data we can bring, because that's going to be a tiny fraction of all the tokens that go into one of those models. So then it's about how we take a customer codebase and make an agent using those state-of-the-art models work really well.

People are scratching the surface there, especially with Claude 4 just coming out. People are still finding new capabilities in it. That's basically what I mean by being on the model wave. Whenever a model comes out, it exposes new capabilities, and it's our job to go figure out, usually working with early-access model releases, how to make the best use of them.

Let's talk about the existing stack. Everybody has a unique angle, whether it's GitHub's position or VS Code, and several folks have come at it from a new-age IDE. A year, two years, three years into the future, what does the core engine of the software developer look like? Where's the strongest position for companies trying to build AI into the process? What are the most important components, and which ones get abstracted away over time?

The AI coding agent is the thing, in capital letters. Everything else is going to be subservient to the AI coding agent. Another way to look at that is that the AI coding agent is the AI version of the software developer. Every tool today serves the software developer, so it makes sense that if AI is writing the vast majority of code, everything serves that.

So your editor: I've seen people change editor preferences they've had for decades because VS Code has better AI editor extensions. People change their programming language. They're changing how their code is structured. They're changing the tools they're using, going from Datadog to something simpler, because AI is starting to consume more and more of the logs and humans are not needing to anymore. Everything else is downstream of the AI coding agent.

And it is not just in software development that that's true. When you have a technology that can build other technology, it is a revolution on the scale of the Industrial Revolution. So we think this is the most important thing to be building. Anyone who's trying to leverage a foothold in the editor or a public cloud to have a good offering here, no, they're getting it mixed up. The AI coding agent is the thing.

What about the questions folks have around differentiation? There's almost a chat-window exhaustion. There are so many different tools and so many different ways to build things, and they have strengths and weaknesses across different aspects of building an application. How do you think about differentiation in this world of the AI coding agent?

A little bit meta, but differentiation is a really hard thing right now. Everything is moving so fast. There is no moat. The models are surprising even their creators with their capabilities. So what does differentiation even mean in a world where AI coding tools make it so much faster to build new features? A feature cannot be differentiation anymore. Even a business model: a company that's been around for 150 years could be disrupted more easily than ever before.

So what does differentiation even mean? We think about this, and we just try to be really open about it. We're trying to build Amp to be the very best agentic coding tool. We've done some very different things, and they seem crazy to a lot of people. They seem crazy to a lot of our competitors, who've really dug in on showing you that model dropdown with 17 different models and doing lowest-common-denominator integrations with each of them. Who have really dug into "AI shouldn't cost more than 40 bucks a month; that's evil if you're only making it for the rich people." They've dug into "this is the most private, secure, whatever. It's self-hosted. It can run in Fort Knox." So that's where our competitors have dug in, and we have just used the customer trust to set ourselves up to build the very best product. That's differentiation.

But in theory, our competitors could wake up tomorrow and say, "Oh, I realized what we were doing was dumb," switch, do what we're doing, and go build it faster than ever. They're not doing that. I think the more interesting question is why they're not, not how our product is differentiated and how it can sustainably be differentiated.

Do you think that's a languishing framework people have in company building, where you build a business and a product with a path to differentiation in mind? You say, "We're going to be the privacy-focused one," or "We're going to be the most SOC 2-compliant enterprise option," and that's how you build differentiated products. In a previous world, not too long ago, that was fair. Maybe it gave you some advantage with a particular customer base. To your point, that's not true. Are the competitors putting their foot down on particular strongholds just running a leftover framework that's totally outdated?

That's right. You cannot predict what the world is going to be like in a year, so pursuing some long-term differentiation strategy is folly. There's one thing that really does matter, and that is customer trust. That's something we built up because we've been around for 12 years. We've got all these customers loving our Code Search and Cody and so on. Some of the things I mentioned that we do, like Amp being multi-tenant and Amp not exposing model selection, we're able to do because our customers trust us. They might initially say, "Whoa, hey, I don't think we can do that." But we talk to them and explain how this helps us build the best product.

Just the fact that we can talk to them and they'll listen to us, and not call the police if we knock on their door like a startup they've never heard of, that is valuable. That customer trust is valuable. But everything else, you can't really plan for. You don't know what the world is going to look like.

Do the methods you used to establish customer trust still hold? Are there still ways to get in and build customer trust, or is that a benefit of having been around for a long time? If you were starting Sourcegraph today, would establishing that customer trust be significantly harder given the pace of change?

I think it's a lot harder. One thing we've seen with Cody: a year ago, in June 2024, it was by far state of the art in AI coding assistants. We had the very best chat RAG, and already that is obsolete. It's not that Cody is a bad product in any way. It is the very best chat RAG. It's great for understanding and fixing your codebase. But a lot of companies might look and say, "Why would I trust Sourcegraph with Amp, this new AI thing they're doing, if Cody is obsolete? They messed up. They must be really dumb if they have an obsolete AI product out there."

We're able to not be weighed down by that. What I would say to that is that we move fast and we have a next-generation product, and you shouldn't disqualify somebody just because they're following through on a previous product that a lot of enterprises, some of them slower moving, really want to see commitment on. So we're continuing to make that better.

But I think a lot of smaller companies are going to get dinged if someone says, "Oh, three months ago their product was totally different. They're all over the map." It's really hard to come out of that. There are going to be a few breakouts. But if you don't get that breakout, and then not only take the reason people started loving your product but evolve as fast as you can and not overfit to the one thing that helped you break out, it's really tough.

Going back to a point you made earlier, one of the advantages you have is that you're already plugged into a customer's existing codebase, with a deep understanding of that code. To your point, the proprietary data you have is a drop in the bucket compared to the underlying training data for the quality of the model. Do you still see a competitive advantage in the depth of your understanding of somebody's existing codebase when it comes to putting generated code into production? Or is that access to the codebase offering less and less of an advantage?

Everything is changing. As of a couple of months ago, there's a new kind of data that you actually need to be collecting to make better agentic coding tools, which no one was collecting before: how people are actually interacting with an agent. What are you typing? What are the feedback loops it's getting? And ultimately, does that code pass tests? Does it work? How can you tighten that up? Those were the kinds of things you could not be collecting, because that workflow did not really exist until 3.6 Sonnet, really 3.7 Sonnet. So in a sense, everyone starts from scratch.

And actually, the products that have set the user expectation that they're not collecting that data, like Cursor and Windsurf and Claude Code and all these other things: you have a session, and then it goes poof. You don't actually track that. With Amp, what we've done is you can actually see what everyone on your team is doing, just like if you git push a branch, everyone can see your branch. All that data is stored so that you can learn from others. If you have a pull request you used Amp for, you put the little link in, and people can see what you actually typed in. It's really valuable to users.

But what it means is that for companies that want us, in the future, to train Amp on all the ways in which Amp has worked, or maybe cases where it didn't work as well, we're now gathering that data. Every other tool out there didn't have the foresight to say, "Hey, maybe this new kind of data is going to be really valuable." But everyone's starting from scratch here. I think we've got by far the most of that kind of data of anyone, other than synthetic or paid data. So it's a whole new race.

Final question for you. Where is there maybe some small pocket of advantage or arbitrage? Occasionally, folks are a little more aware of what's happening at the cutting edge. Look at the release of DeepSeek: before it popped off, a few people had some exposure to it. Maybe it's specific developers using existing models in unique ways. Is there anywhere you look to get a sense of where things are going, or how the space is evolving, to make sure your products are set up for success?

There's an edge to whoever can build a product that stays at the model-product frontier, riding that wave for more than a few months. It's very hard. Have not seen any company really do it. Initially, Copilot had this great autocomplete. Everyone said, "Oh, they're crushing it." And then, bing, and now no one talks about that. If you fast-forward a little bit, huge props to Cursor for coming out with some awesome stuff, but now it's getting super complex, and they've really slowed down since they raised that big round and grew so much. Windsurf came out of the gates really fast, and who knows what's going to happen? Will they have a future? I don't know. And they have multiple things in their product called Legacy mode. That's really overfitting to what worked at a given period of time, and not setting yourselves up to evolve with your customers and with the model capabilities.

That's really hard to do. I know it's really hard to do because we learned a lot of lessons from how we built Cody. But with Amp, we're trying to do it, and anyone who can do that is going to be good. And all these people who go and launch some AI product, and it's cool, and they get all these signups and give a bunch of free credits to everyone: that happens so much. Everyone is so sick of launches.

I think a lot of people are also dissing Apple, for example, for being behind. But it's not like on your Samsung phone you're in love with all of their AI features. I think there's something to be said for doing it right and continuing to do it right. Do not look for that launch-day pop. Look for the thing that will get users to keep using and loving it, that will have a huge impact. That's always what has mattered, but it feels like people have forgotten that in AI.