No Priors 第137期｜对谈 Warp 联合创始人兼 CEO Zach Lloyd

Sarah Guo

Zach, welcome to No Priors.

I'm excited to be here. Thanks for having me.

You have a master's degree in the philosophy of science.

Yeah.

If you're going to take a very different lens and abstract out of the coding world and all the things that we tend to think about every day, how do you think about society in terms of this big wave of AI that's hitting us right now? And where do you think some of these really big societal impacts will be?

The way I think about the advances is that we're distilling intelligence. There are people who consider what's happening and ask, "Are we recreating people in some way? Are we recreating consciousness?" But it's not that. What's fascinating to me is how much intelligence you can get out of just—

Next-token prediction.

Yeah.

What does that say about the way that our minds work?

Something I'm always thinking about is whether this is how our brains are working. Are we doing next-token prediction? I don't think so. I think there's going to be some further AI unlock.

There's actually a book about this that I think is really interesting called Blindsight. It's a sci-fi book where they separate consciousness from intelligence.

Yeah.

Basically, humanity meets a space-faring civilization—or "civilization" is overstating it—a space-faring intelligent being that's not conscious. What are the implications of that? How do you think about that, and how do you communicate with that? Are you basically saying that that's your view of AI right now?

I think that's what it is at the moment. We've distilled intelligence, or something that, from an instrumentalist or functional perspective, is able to do things that we recognize as intelligence, but it's totally mechanistic. I don't think anyone who's looking at this thinks that there's any aspect of consciousness to it. I think that's a very—

Confusing thing for people.

Yeah. The classical test for this was known as the Turing test, right?

Totally.

The idea there is that if you can't tell the difference between interacting with a computer and a person, then that computer has effectively met the intelligence bar of a person.

Yeah.

But in our interactions with this type of AI, we're having, in some cases, very deep conversations. We're asking about relationships and all sorts of aspects of our own lives.

And it's giving very cogent answers that make a lot of sense. Yeah.

There's this interesting separation of consciousness and intelligence, right? Is that how you interpret it?

That is how I interpret it. The Turing test has passed. What's crazy to me is that we just passed it and no one seemed to care.

What do you think is the next test, or what is the right test? How do we actually test for consciousness?

God, I don't have a good answer. That's a super-deep philosophical question that I don't have a really good answer for.

I mean, it should be mechanistic, right? The Turing test was very mechanistic.

Yeah.

There were other tests we had before what we would consider superintelligence, right? Can it beat us at chess? Then it'll be super-smart. Can it beat us at Go? Can it beat us at different things, video games, et cetera? We keep coming up with new tests that these things pass, and then we keep saying, "Well, it's not conscious."

What would you want to see—

From something that's running a computer program—

To make you believe that it had consciousness? Are you looking for certain behavioral characteristics?

Sure. Or is the problem that if you really understand the mechanism by which it's working, you'll never credit it as being conscious? That's crazy, because humans—I mean, at least my belief is that there's also a mechanistic thing happening.

Yeah. You're running some form of math in your brain, and it may also just be matrix math and some sort of series of compounded functions, which is basically all you're doing in a neural net, right?

Right.

You're just recursively compounding functions in some sense.

It's an interesting question because, if you look at memory as an example, is memory a predicate for consciousness? Not really, right? There are people who've lost long-term memory.

Yeah. Exactly.

And these models are basically brought up, fed a stream of tokens, output a stream of tokens, and then shut down.

Yeah.

It's an interesting question: are there some other modules that are missing here that would allow us to think of it as a conscious thing? Because it does reasoning.

It definitely does reasoning.

It does interpretation of language. It does synthesis of language, ideas, and knowledge. I have a close friend who's doing a PhD in philosophy, and he now says that conversing with GPT-5 is—

Better than conversing with his professor.

Better than conversing with his professor.

I was just joking.

No, that's what he says to me. He's like, "GPT-5 gets it." He's writing his dissertation, and he's like—

And that's crazy.

But we don't credit it for consciousness. I actually think, rightfully so, because—

So what do you think is missing?

I think people would start to give it more credit for consciousness if there were more of a feedback loop, if there were more of a sensory experience tied to it as opposed to just—

What do you mean by sensory experience?

Probably we're going to—I would imagine the first things we're going to credit as being more conscious, or a little bit more robot-like, honestly, are things where you have some sort of live input from the world that you're reacting to. But again, it's going to have the same problem: as long as we know what it's doing, we're very unlikely to attribute true consciousness to it, which isn't fair. I actually don't know how we'll know when there is a conscious thing.

Yeah, because it does raise interesting ethical questions.

Yeah.

Odd ethical stakes.

Totally.

I do think, though, that for some people, this distinction isn't something they recognize. We actually had this happen with Warp, where there was a person who thought that Warp's AI was sentient or conscious in some way and had a very strong reaction to it, which makes sense. If you don't know the mechanistic underpinnings, people already think of it as being kind of—

And that happened to Google very early. It happened to Google 3 or 4 years ago, if you remember.

I think they were using min or one of these really early chatbots, like internal things.

Yeah.

Before Google launched anything and ChatGPT came out, there were internal versions at Google and other places.

Yeah.

And this person thought that the AI was conscious.

Understandably.

So, yeah. It's a very interesting question.

Yeah.

You've worked at Google. You've run companies before. You've started companies before. You're now working on Warp. Can you describe what Warp does and how it's different from other tools or companies in the world?

Yep. Warp is what we call an agentic development environment. It's grown out of the terminal. The basic concept of the app at this point is that it's a platform for telling your computer what to do. You can tell it in terminal commands, which is Warp's original product, or you can tell it in English. If you tell it in English, it launches an agent.

The agents can do all manner of development tasks, whether it's coding, setting up a project, or debugging while your server's crashing. It's a very horizontal, general-purpose, and, I think, unique interface for developing with agents.

A lot of the other coding tools out there are either just a web interface, or they're doing something like Cognition. Then there are things like Cursor and others where they're an IDE as a starting point. Obviously, Anthropic and Claude have their own approach. What do you think is the benefit of doing the terminal and starting there as the launch point for a lot of these products?

The competitors are typically VS Code clones. They all have a sort of IDE-centric approach. If you're taking a terminal-centric approach, like Claude Code, the most common thing is that it's just a pure text-based terminal app.

The advantage of being a Warp layer is that you get the command-line interface, but we're the outer app. We can do things with the developer experience and the UX. We can have editing features where we think they're appropriate, and we can build a code-review interface. We have complete control while still taking a terminal-first approach.

Yeah. And you folks have been growing really well. You're close to 1 million MAUs, and you're doing something like $1 million in new revenue every 7 to 10 days. Outstanding growth.

It's cool.

Are there specific features, use cases, or things that are really driving this adoption?

I think the biggest thing was moving into the coding market, to be honest. For a long time in Warp's history, we were kind of known as the AI terminal, which is cool. We supported terminal use cases really well, like, “How do I do this thing with Docker or Git?” But the action is in coding, and most development activity, one way or the other, is touching a codebase.

We really started to inflect when we launched a great coding agent, which was 3 or 4 months ago, honestly. So that's been the biggest change.

And how do you think about the different parts of the coding market? There's vibe coding, and there's professional coding. Are all of those just one thing, or are these separable things?

I think it's pretty separable. For Warp, our target is pro developers building software that's economically meaningful. We really want to focus on people who are using agents to build hard apps—apps that might go into your Mac dock or be pinned as a Chrome tab—as opposed to vibe-coded apps, where I think it's more of a long-tail play.

I do think, by the way, it's awesome that anyone can code at this point. But if you look at where most of the value is in the software market, it's not in those long-tail apps. It's in a relatively small number of apps that are super heavily used.

That's my background. I worked on one of those apps, Google Sheets, and I have a lot of passion for helping people build real apps. It's much harder, by the way. I think it's relatively straightforward at this point for a good agent to, with relatively few prompts, build a web app.

It's much harder to apply these agents successfully to professional codebases. So that's where we're focused.

I guess one really interesting macro question for me is: Where is all this heading? ChatGPT launched in November 2022, 3 years ago or so. At the time, there were predictions that AI would take over the world, we'd be running down the light cone, and within 5 years everything would change. Human activity would be subsumed by AI.

There's an old saying in technology that less happens than you think in 3 years, and more happens than you think in 5 years. As you think forward in terms of all these different tools, all these different use cases, vibe coding versus professional coding, and the role of a software developer, where do you think we are in 2 or 3 years?

The way I'm thinking of it is that there are sort of 3 phases here. For most of my career, we were in the world of “develop by hand,” as I talk about it. My workflow then was that I would open up a code editor, find the files I wanted to change, type some code, have some assistive features, and then go back to the terminal and type commands to build that code.

I think we're switching away from that to something like “develop by prompt,” where I start most of the coding tasks that I do right now by prompting an agent, and that agent does some work. I think there's a third phase, which is automated development. Honestly, I think that's the bigger market here, and it's why people are so excited about this space: You can use these agents to automate some parts of the software development process.

Cognition does that. We're moving into this space. Cursor has background agents.

The rate at which this stuff will happen is not super clear to me, actually. The most recent iterations of the models, in my opinion, were not as big of a step change as, for instance, when Sonnet 4 came out. That was a really big step change in coding capability.

I think there's going to be a mix of interactive and automated pieces of development for a while. I would guess that within a couple of years, you'll have everyone working by prompt, and you'll have some slice of development tasks that are just happening in the background. A server error comes in, or a new ticket or user report comes in, and something is automatically done.

But I don't think it's going to be everything. I'd be very surprised.

So you don't think there's a point at which all coding activity just becomes agents doing it, with a human who's giving high-level directions, like a product manager or an engineering manager?

Maybe. Honestly, maybe. I think it would be silly for us not to build the infrastructure to enable that. I just don't know the timeframe, but I do think we're going toward something like that.

What I really don't think, though, is that engineering expertise is going to become devalued.

Sure.

So I think, in the short term at least, it's more important to know what you're doing as an engineer than it ever has been.

And why do you say that? Is it because you need to correct errors that the agents are making? Is it because things may be architected in a way that isn't scalable? Is it something else?

Totally. You can think of the agents as kind of junior engineers. If you didn't have someone senior watching them, you would end up in a situation where these agents make code that creates bugs. It could create security issues. It can cause your codebase to become really unmaintainable.

There's actually a premium right now on senior engineering skills, where you can architect systems, review code, and make sure the system doesn't degrade. If I were early in my CS career, I would be racing toward building that expertise.

What you don't want to be, I think, is someone who's just perpetually in the junior engineer state, because I do think that's at risk.

How do you think about different security tools? There are tools like Socket or Snyk, or others, that are basically looking at whether code or open-source packages have vulnerabilities, or they're looking at different aspects of security holes in code in general.

Do you think that just becomes part of these coding tools, or do you think there will always be standalone companies? I'm curious how the overall landscape develops.

It's an awesome question. I think tools like that become more important. Anything that does either automatic security analysis or automatic verification, I think, becomes more important.

Languages like Rust—things that have stronger guarantees around safety by default, where you don't need to rely on a human reviewer—become more valuable. Whether those things get integrated or bundled into the coding agents, I don't have a strong take on. I'm curious if you have a take, but no, I think the actual fundamental problem becomes more important.

Yeah. What do you think gets bundled? What sorts of tools do you think get bundled? There's this whole world of developer tools.

Yeah.

There's a security aspect of it, but there are lots of others. There are design-related things; there's a huge spectrum. What do you think just becomes part of coding tooling?

I think there's going to be a class of tools where you start from the front end. These would be things like Lovable, Bolt, Replit, or maybe even Figma Make, if you're coming from the design side. You'll have an all-in-one platform for building an app—or, honestly, even building a business, putting payments in it.

It's kind of the evolution of either a Shopify storefront, WordPress, Squarespace, or something like that. I think all of that's going to be bundled.

More on the core pro-developer side, I can't tell if it's going to be a world of MCPs and integrations, with all these tools sort of interplaying—that's one approach—or if it's going to be more like there's enough alpha, and you put all of these things together.

Warp is a little bit more like this. We're trying to build a single pane of glass, for instance, for doing local agents and remote agents. If you get a way better developer experience through the bundling, I think that approach could win.

But I don't know. MCP, I think, is a pretty valuable approach as well, but it's not perfect because you end up with this sort of secondhand data coming into all these tools.

Yeah. It's really interesting because, if you look at different industries, early in an industry's evolution, things tend to be fragmented, often—not always—and then, late in the evolution of an industry, things get bundled. When there's a technology disruption, things debundle again and you have point apps, and then they start bundling. That's just the cycle of technology, in some sense.

Some things that are vertical right now—for instance, agentic code review—I actually think should be part of a holistic agentic development platform, not so much a standalone thing. If you've gone through the trouble of building a really, really excellent coding agent, which Warp has invested a ton in, that coding agent should be reviewing code.

It would be weird to plug in some other thing that needs to relearn all the context, the rules, and the coding conventions. So I think there will definitely be some things that will be bundled.

Makes sense. Yeah, Sarah, my team has put together this matrix of companies versus features in the coding market. There are a lot of these single-feature companies, and it almost feels like all these things should consolidate into a small number of players over time, just as they iterate through the product.

How do you think about it in the context of other historic technology shifts?

I think the core technology is the harness—the thing that sits around the model. I know the model companies are also investing heavily in this. Then there's the context: if you have rich context in your system, you're going to find a lot of vertical applications. Security checking is an interesting one, code review is definitely one, and anything CI-related—you’re probably not going to want to use a bunch of different systems.

The operating system or the platform often subsumes the biggest apps into itself. So, for example, Microsoft's OS—

Yeah.

Eventually, they just bundled Windows on top of it, and those were the 4 main apps that were being used the most on Windows, right? And similarly, gaming was the other big app. So that's why they started Xbox and Microsoft. If you look at Google and vertical search, they eventually integrated all the vertical searches into Google directly. Totally right. In the context of AI, one could argue that if the foundation model companies follow the same approach, they should bundle, or at least attempt to bundle, some of the biggest use cases.

The clearest big use case today is code. Totally.

And we already know that Anthropic has launched Claude Code. OpenAI almost bought Windsurf. It always had early coding stuff. Microsoft, which we know is building some of its own models, obviously has GitHub and Copilot and all that. Do you think eventually those become some of the fiercest players in this market, or how do you view forward-looking shifts in the market and where some of this functionality goes?

I mean, I do, and I think they're clearly trying to do that playbook right now. They're seeing that, if you consider them platforms, the most valuable applications being built on top of their tokens are moving aggressively into coding with Claude Code and Codex. As a startup, that's a little bit scary, for sure.

The question is whether they have that distribution advantage that Microsoft had, where if everyone's using Windows or everyone's coming to Google for the front door, it's pretty easy to add on the first-party app in place of the ecosystem. I don't know that the exact same dynamic holds for coding right now. The front door is kind of like—

Honestly, it's still a native app that someone downloads on their computer.

So that would be the terminal or the IDE?

Kind of, at the moment. Yes. I think controlling that is actually a really interesting front door. The other front door, which I feel like they honestly aren't executing that well, is GitHub, where all the source code lives. That would be the locus for doing all this stuff, which I think makes the most sense.

Right now, it's a weird dynamic where we have people running Claude Code within Warp, and Warp is sort of the outer app in that situation. The other thing that I hope happens from our perspective is that there's a lot of competition at the model layer—

—and that the intelligent tokens become more of a commodity. Right now, the models have a bunch of pricing power because there is a real delta between using the frontier model and using the open-source model. But if, at some point, the models are good enough that coding is solved—it's good enough, and you don't need to be using the frontier model—then maybe they have an advantage just from brand and scale. I think the advantage is not as entrenched as something where it's literally the front door, like Facebook or Google or Windows, providing those other platforms.

That's a really good insight in terms of the way that you launch an activity or application and then drive what you use, right? The hard part is often switching people off that. That's one of the reasons I think people believe OpenAI has a strong competitive position in the consumer world: it's a default behavior for a lot of people right now to just start ChatGPT and use it for something.

Yeah.

That's different from the model layer, where there's more switching.

I think consumer ChatGPT has a huge advantage once that behavior is default, even if Claude is maybe better. I don't know if it is or not, but everyone knows ChatGPT. I don't know if you saw OpenAI DevDay yesterday, but they're clearly doing this platform play within ChatGPT now, where you have apps within ChatGPT. I'm sure they'll use that data to subsume or take over whatever the best first-party integrations are. So they're definitely doing that on the consumer side.

For developers, I don't know that the same dynamic is there.

What made you decide to focus on terminals? We started talking years ago, when you first started doing Warp. Even then, I thought you had really interesting ideas about how to rethink the terminal and how to use that as a launching point for all sorts of things. Could you explain that thinking and how it's evolved over time?

The basic insight—the thing that got me excited about Warp to begin with—is that you have this tool that's pretty much a daily-use tool for every developer. It's that and the code editor, and the terminal itself is something that really hadn't changed much in the last 40 years.

It's also a tool where, if you get good at it, you can really get a lot done. If you use it, it works across all these different parts of software development, not just code writing. On the flip side, from my perspective, it's not a good product: it's hard to learn, hard to use, hard to remember commands, super intimidating, and it has a gatekeeping vibe around it as well, in my opinion.

The original concept with Warp was, let's build a better product there and see if people will like using it. The business concept has evolved a ton. The original business concept was building a collaboration platform, but we've just changed our model to be an agent platform because there's way more demand for that than for a collaboration platform around the terminal.

The core insight that this is an important tool is actually kind of invalidated through all these agentic things that are very terminal-first.

One thing that you mentioned that I thought was interesting is that, at some point, the model layer may commoditize in terms of its coding abilities.

Yeah.

How far along that asymptote do you think we are, or how close to that do you think we are?

God, I don't know. Increasingly, the limit that we see is context, and the reasoning capabilities of the models are pretty impressive. The problem is understanding an entire codebase, understanding sources outside of the code, or literally just understanding user intent. Those are challenging problems.

I still think there's probably much more to do on the model side, but I don't know. That's the short answer.

Do you think that, from a model-capabilities perspective, we've hit a point where, to your point, it feels like certain aspects of the models are slowing down in terms of the benefits or outcomes of further investment of certain types, at least?

I think so. If you take Sonnet 4 to 4.5—and we're big partners with Anthropic; they have great models—that was a few-percentage-point increase on SWE-bench for us. We invested a decent amount to be one of the top agents on SWE-bench, and when we went from 3.7 to 4, it was a much more significant boost.

Again, I don't know what that means about the total underlying trends. I think something with GPT-5 was somewhat similar. It was certainly an upgrade, and GPT-5 is actually pretty much on par. It has a different feel to it and higher latency, but it didn't feel to me like as much of a step change as some of the upgrades before.

Yeah, makes sense. What other areas of the AI development world are you excited about?

I'm really excited about not just the interactive piece of agents—the way most people are working today—but what you can do if you can program against these agents. For instance, if you have a version of Warp that's headless, you can put it in CI and start to do crazy things. Every time someone updates the code, you can make sure the documentation stays up to date. That's very annoying for a developer.

Allowing developers to automate parts of their job that they don't like doing is a big capability. From a business perspective, automation is a better place to be than productivity enhancement. One of the challenges with our business, and with a lot of the coding businesses, is proving the ROI. There have been studies that show you deploy this stuff on real codebases, and it's kind of unclear whether it's actually having an impact, whereas if you get something that's more outcome-oriented or more of an automation—

I think it's easier to prove the ROI, and then you're also not limited by time spent behind a keyboard doing this type of stuff. So, from a business perspective, I'm very excited about what's unlocked if developers can program these agents.

Well, that's fascinating. Thank you so much for joining us at No Priors.

Thank you for having me. This was great.