Transcript: Warp Dev The AI Terminal Changing Software Development | Zach Lloyd CEO & Co-Founder of Warp Dev

What's interesting that's happening in the world of AI based development is that I think neither actually the IDE nor the terminal actually makes sense as the sort of primary tool for the future of code. What makes the most sense is a some sort of workbench where uh you as a developer just tell the computer what you want to do. For example, I can say something in English, and it could mean different things for the same exact um, you know, in different context. But code is already logical in the way it is out there. So if you are feeding a logical structure to the LLMs, it's more likely that it performs better in on code than on English language. That was my thesis why I was bullish on like we'll see a lot more coding related products. Now my second company, and I can tell you a couple of things that I learned from the first to the second, because I, the first one I, I basically failed out and learned a lot. Uh, and um, really focus on team. So like really, really, really try to hire great people even if it makes go a little bit more slowly and try to hold that high bar getting. Really try to work on as big of a problem as you can, which I think is counter intuitive to a lot of spirit founders. I think a lot of founders approach starting a company with like find a niche, find a few initial customers. And that that's cool that can work to some extent, but um sort of counter intuitively like the bigger swing you're taking, the easier it is to get people to fund you. Easier it is to attract also people to work with you, but want to work on something really meaningful. I think you use the term level of abstraction, which is my favorite term to use whenever talking about technologies is like we used to write HTML code for websites, and then, you know, we've we've come up with WordPress, and then in WordPress just for e-commerce, we went to Shopify. You know, you don't have to use WordPress for e-commerce, you just use Shopify because everything e-commerce related will come with it. And then like we have bubble and all these tools where the abstraction layer where we are working has completely changed. And maybe we can go to a layer where, you know, Amazon has figured out how to run a worldwide scalable e-commerce website, right? Why can't we just take the best practices of that and instantiate a new amazon.com? Hello everyone, my guest today is Zack Lloyd. Uh, Zack is the founder and CEO of Warp, uh, a company that is developing an intelligent terminal aimed at modernizing the command line experience for developers, software engineers. Uh, they recently raised $50 million in series B led by Sekoya Capital, uh, and participation from existing investors like Jeff Wiener, Mark Benioff, Nileen Feel. He was previously principal engineer at Google, worked on Google sheets and Google doc suite, uh, co-founder and CEO of self-made a venture backed startup, uh, and was also an interim CTO at Time Magazine. In this conversation we'll talk about uh future of software development, uh, evolution of the terminal, uh, AI's role and building new products. If this is the first time you're listening to startup project, don't forget to subscribe to us wherever you're listening to the podcast, it helps us reach to a wider audience. Zack, welcome to the show. So glad to be here. Uh, so I was really excited to have you on the show because uh I think after the you know, the chat GPT moment broke out, um the LLM company sort of were everywhere. I think that's sort of like the first line of value that is being captured. And then I was excited about the application, the new types of applications that we will see. Uh and the most bullish use case for me was uh developer productivity. And the reason being anyone who studied compilers will know that um LLMs are actually looking a lot like compilers in terms of like text completion, auto complete, uh in terms of what they generate uh it it looked a lot like uh compilers. And then there's this aspect of code is very deterministic. Like for example, I can say something in English, and it could mean different things for the same exact, you know, person in different context. Like the exact same phrase could mean a different thing, right? Uh, and it also depends on what emotion you're showing. But code is already logical in the way it is out there. So if you are feeding a logical structure to the LLMs, it's more likely that it performs better in on code than on English language. That was my, you know, um thesis why why I was bullish on like we'll see a lot more coding related products. Uh and I think in some format we're seeing like the biggest use cases are around uh you know, developing new products and especially around software developers. Um, so that's why I was really excited, uh to have uh you and talk about Warp and what you're building there. Um, so I I think a good place to start is uh if you can talk about what is Warp, uh, and how did you come up with this idea of intelligent terminal and why it's important to build an intelligent terminal. I think that would be a good start and sort of level setting for the audience. Cool. Um, so yeah, Warp is as you said, it's an intelligent terminal. The terminal, in case folks aren't familiar, is, you know, it's one of the two most important tools that developers use every day. They use the terminal, they use the code editor. The terminal is basically the place where you tell the computer what to do. Uh and that could mean like building your code, it could be running your tests, could be writing internal tools, uh it could be interacting with your production system. So it's a very ubiquitous and um important tool for developers, but it also is a tool that's tired, you know, stuck getting 40 years ago from a sort of usability perspective. It's it's something that really has not uh evolved much uh from an experience point of view. And you know, Warp started our our goal was like let's modernize this interface. Let's make it more usable. Let's make it work more like a water app. You know, even really simple things like let's make the mouse work in the terminal. Uh, but as the LLMs have matured and come out, uh, the the the product has become vastly different. So at this point Warp is just like a place for developers to talk to their computer and tell the computer what to do. And because they're doing this through the terminal, uh, there's this huge array of tools that already exist uh in the form of these command line apps that can take what a developer says, like you said in English and turn it into a series of app calls that do what the developer wants. And that could mean like setting up a new project or debugging something in production or, yeah, increasingly just writing code, which is obviously that's the biggest about activity. And so that's what we're at today. We we think Warp in the terminal is uh, you know, an amazing interface for uh developer to, you know, tell AI what they want to do and essentially have it done to at least some extent. So, uh developers are really, like every developer I know is like unique in some form. Like everyone is picky about their stack of tools. Um, you know, they have their own slightly different version of terminals or command lens they use. Um, so how how does a developer use warp now? Like go to the terminal and install it or uh how does that how does the existing behavior of the terminal change by installing Warp? Okay, great question. So, if you're a developer, you can just go, you can you go to warp.dev, you download Warp. You, it's a, it's a native app. So if you're running on Mac or you're running on Linux, you're running on Windows, you just open it up and you use it instead of whatever terminal you were using. So if you're using like uh I term or the stock terminal app, the VS code terminal, you just use warp. And despite being a sort of AI native experience, Warp is backwards compatible with your existing uh like stack of things. So the the the way this works really big picture is a terminal is the app you run and then within a terminal you run a shell, and the shell think of it is like a text interpreter. So when you type a command, it's the shell that figures out what program to run. Warp works with all these existing shells. And those are things that are like bash, fish, CSH. And so a big product emphasis for us is like let's meet developers where they are, not make them take a step backwards started, you know, get all the extra benefit of being able to do all this stuff with AI. So basically you allow developers to bring in their sort of existing nuances into warp. That all basically works. So like for 98% of stuff that uh developers have set up in I term or wherever, uh you can just open up warp and it it literally should just work the same, but also be better. At least that's our that's our goal. So uh terminal generally, you know, is a little bit restrictive in terms of like usually they're not intelligent in the sense that some terminals you can, you know, easily go to the previous command and, you know, reuse it. But you have to know the exact command. And this becomes really hard when the developer is in early stages of a career because you have to remember all those commands or like you're constantly going to help or uh, you know, that if you're using, you know, GitHub or like Git and trying to commit uh or trying to do different things with Git. All your you're struggling with basically trying to find the right command to do the right thing, right? Um, so with with what intelligence is warp adding to sort of like uh if you can explain through in terms of like an example or a feature that you have. Yeah, sure. So, you're absolutely right. So one thing that's really, really frustrating for beginners and for experts is like you open up the terminal just a blank screen, and if you want to get something done, you better remember what the command is and these commands can become quite complicated. So like, let's say, um, I don't know, let's say you want to set up a brand new Python project. If you want to do that, you have to install the Python tool chain to and then you might have to like close some git repo and go to clone the git repo. You might find that you don't have SSH keys, and then you're like, oh no, like you're going to start Googling, you're going to stack workflow, how do I recreate my SSH keys to authenticate to GitHub? And so that's annoying. That's like not what developers want to do. Developers want to build things. They don't want to deal with like all this incidental complexity that comes with like uh, you know, fits my path variable to reference the current Python binary. And so what Warp does is you uh you don't have to remember the commands at all. You just need to know what you want to accomplish and you tell the computer to do it literally in English. So you would uh instead of typing a command Warp, you would type help me set up a new Python tool chain, uh clone this repo, make a new branch for me. Make sure it all compiles and runs, and that's that's it. You would type enter. And then what the LLM does is is it says, where am I, what do I try to do and it starts issuing commands to figure out its context. So the LLM might run like LS, it might run git status, it will try to run get clone. When it runs get clone and hits that SSH error, it'll say like uh you know, we had an SSH error. Do you want me to generate these SSH keys for you? And you as a user will say yes, and then it will remember what the command is to generate the SSH keys, and it'll basically do this with you until you get to the spot that you want to be at. And that's a way, way better workflow than uh switching context out of the terminal and like looking up on Google every time you hit some error. So it's it really saves a lot of time and like also a lot of frustration for developers who are doing this kind of thing. So uh I mean the terminal as you said, in its basic form hasn't changed much. But at the same time, I think there's also has been this huge integration between IDs and terminals. Um does that change how you think about Warp? Like uh does Warp have to also now work on the IDE? Okay, great. Uh great question. So the a lot of people use uh the terminal in the IDE and like the uh there's definitely benefits to that like you you have one app. What's interesting that's happening in the world of like AI based development is that um I think neither actually the IDE nor the terminal actually makes sense as the sort of primary tool for the future code. What makes the most sense uh is a some sort of workbench where uh you as a developer uh just tell the computer what you want to do. And so the the standard workflow for someone using an IDE today is they'll open up the IDE, you'll let's say you're trying to build a feature a feature and so you'll open up, you find all the files that might be relevant to building that feature. And then you'll start writing a function, you'll start writing a class definition. Um, maybe if you're using something like co-pilot or cursory, you'll get a bunch of completions, and that will help you do it. But you're basically doing what I call uh coding by hand. So it's like a very much like the level of stretching we're getting at today as developers like you hand build all this stuff. Or likewise, uh if you're using the terminal, what we're typically doing is you're running a bunch of commands um and you're doing those things by hand. But the world that um we're moving towards is something where rather than doing anything by hand as a developer, at least doing by hand from the outset, what you're going to do is you're going to work by prompt. And what that means is you're going to describe the feature that you want to build in English and um the AI with some level of autonomy, and I think increasing autonomy, is going to solicit whatever information it needs from you and from your environment and then it's going to go do that task. And so I guess my my hypothesis is that the IDE is not actually like the right place to do that. Um, it's it's much more of a place for like having a bunch of files open and doing hand editing. And what you see in all of the sort of AI based IDEs, like for instance like cursor, Windsor, which are very popular, is that increasingly they are uh guiding users over to a certain chat panel that sits on the right side of the IDE where um the user can through conversation or prompting build their feature. And that chat panel interestingly enough is like starting to look more and more kind of like a terminal in its interactions where it's like you say something that happens, you see the change. And so Warp's approach to this is not to build an IDE uh although we're trying to support coding use cases, but to build something where if uh you as a developer can like ask for anything you want done and build the interface around like showing the work that's being done directly in that kind of linear fashion where uh the computer is executing your prompt. So my my vision here is like these traditional IDE eternal boundaries are going to kind of blend into something that is really oriented around life what the best workflow for development should be in the future. I mean, historically like if you look at development, like I think you use the term level of abstraction, which is my favorite term to use whenever talking about technologies is um like we used to write HTML code for websites, and then um you know, we we've come up with WordPress, and then in WordPress just for e-commerce, we went to Shopify. Now you know, you don't have to use WordPress for e-commerce, just use Shopify because everything e-commerce related will come with it. And similarly like we've divorced into like sub stack will give you a news letter, right? So we've moved into that abstraction layer where we no longer use WordPress or HTML to write the end output is sort of same, but what you're doing to get it actually changed. So and like we have bubble and all these tools where the abstraction layer where we are working has completely changed. And maybe we can go to a layer where, you know, Amazon has figured out how to run a worldwide scalable e-commerce website, right? Why can't we just take the best practices of that and just instantiate a new amazon.com, right? I think that's what I was just Just that's what's happening. I I think another like um analogy that is is pretty good is you know, back in the day developers used to work in assembler language and that was really, really low level. Like you were controlling the operations what's going in registers, and then you you know, you moved up to a language like C, which you still have to know how memory works. Um, but it it enabled fast productivity and improvements. And then you moved up to a language like Python or JavaScript where, you know, you're still coding, but you don't have to worry about um so much of the underlying system architecture. And this in some ways is like another I think it's the bigger step because you can basically do it through English, but you're you're totally lessening the barrier to working with um, you know, with code just by, you know, moving up moving up to like you just tell computer what to do. Now, I I do think um that I don't know how long this will last, but I think for for now, and for the next like year, two years, maybe longer, you're going to need that um programming expertise to uh build things that are of like what, you know, of any kind of high degree of complexity. And actually in some ways it will be way because Warp really know what's going on because what's going to happen a lot of times with this method of developing by prompt is that the AI will do sort of 80% of something, and then it will get to a point where it gets a loop and gets stuck or it might have bugs that it can't resolve. And if you don't know what's going on, you're going to you're going to be stuck with it. Um, but yeah, it's like the the level of abstractions definitely changing for software. And uh, so I mean we we constantly see this AI hype going up and down. And one of the things I noticed is like doesn't matter if we not like find the next AGI model or, you know, even if we don't make cutting edge improvements where we have so much stuff that we've created till now, and we haven't really moved the needle in terms of like creating new types of applications. We've we've only like copilot is only one farm factor, but we haven't seen new farm factors that much. Like everything is pretty much copilot type farm factor. So I'm really bullish on like seeing new farm factors for like different use cases where you'll sort of build for how would like a word document would if you have to rethink completely writing with AI first. Like how would that look? And I I don't think we have seen that fully, you know, play out. So I'm hopeful like Warp and cursor are trying something in coding. I think coding is might be the first place where we actually see that. Yes. Is the first place where that the LLMs have really shown performance. I think. To to to the point that you made at the top of the episode, it's um these models understand code super well. They're trained on a ton of code. Um, code is like has like a real structure. Um, and so it's it's first of all, it's super cool space to be working in because I I do feel like the advances and how these LLMs help programmers, it's like this is the biggest place for some big people right now. Yeah. So how how has the feedback been from, you know, developers is and also like talk to me a little bit about how the adoption is coming from. Are like developers discovering it and using it in their company and forcing the engineering managers to buy your product or like is it coming from top down, hey, we want to make this, you know, organization more efficiency and uh adopting your product. So it's we're mostly building for developers. So our our go to market motion is bottoms up product growth. Um, we uh it's going really well from like user adoption standpoint like we're well into the multiple hundreds of thousands developers actively using Warp and that's growing really fast. Um, the the uh it's interesting like why people use it. So we have some people who are using it because um they want a better terminal US. Um, and we have some people using it because they're sort of AI early adopters. Um, the it does happen. So if we get enough penetration in a company, like our strategy is like get a lot developers using it, spread it wide, get developers paying for it. Uh, and then when we have enough concentration in a company, we end up having conversations with the seniogators and we do have a bunch of uh enterprise contracts with pretty good companies. Uh, but the primary motion is is product growth bottom up. And then what gets people to pay us, and the revenues have been growing really well, which is cool, is getting getting developers to a sort of aha moment in the app where the AI did something that kind of like blew their mind. Like it it often could be something as simple as like fixing all of their dependency issues in their project. Um, but there is a I think a sort of overlooked challenge where it's like developers are like get the most out of this new tool set need to change their behavior. They need to like think to prompt the AI. And if you don't uh make that really easy, then they don't discover it. And so yeah, a big part of like what's what's really helped us grow is uh sort of like inserting ourselves in developer's existing workflows and ways that's like really little friction but surfaces the value of the AI like, you know, with them doing what still work. What are the examples like which workflows have you inserted in your? Yeah, so some prime example is like you try to build your code, you get a compiler error and Warp just like pops out the fix for the compiler error. And that's like, oh as a developer, all I need to do is accept this fix. Whereas that's very different for being like you run your code, you get a compiler error, and then you expect the developer to know to type in, hey, please fix my compiler error. And so it's the step that we can hook into someone's workflow, guess what they're trying to do and surface the AI as like the fix, it's like that's the best way to get a haul. I think that's one of the reasons that the like the first sort of modality that has really caught on is like basically auto complete is that it's just it's there. It's no work, and it's really low cost if it's wrong. And so looking for those types of hooks um to get developers to like, you know, accept the output of the AI and then eventually train them to change the way they interact with this is huge. Are you um, I was assuming you're not creating your own model and leveraging other LLM models. Um, which models today are doing the best job in terms of, you know, for use cases for Warp? So the best model for developers right now is is Claude Sonic. Um, it's sort of a battle if it's 3.5 or 3.7. Um, we we offer both in Warp. 3.7 is our current default, but it has actually some problems where it um it's a little bit overeager in terms of how it gathers context. There's some information around Twitter on this. Those are those are the. Uh, the other thing that we're doing on the model side that's really interesting is we're now offering for more complex tasks users have the option to sort of do a two-step execution where first they use one of the racing models and so that could be like O1 or R3 Mini or uh Deep 31, like US social version of that. And that they use that for with a plan, and then we switch them to uh a more of a just like uh like a, you know, standard essentially execute the plan. So for those we're using I think the primary why is the I mean, how do you think is is this going to evolve in the next like, you know, two, three years? Like in terms of development, and there's also this um sort of new phenomenon that we're happening is we're calling it as agents, but at the same time we're using this high reasoning models along with our traditional LLMs. And then there is this whole framework around, you know, something being an agent. Uh how is it and how are you trying to use it in your own product? This whole, you know, this overlap of these three things. Yeah, so the the way that I look at this is there's three main modalities that are important for developers right now when it comes to using AI get to work. So what is completions? Uh and I mess it that it's like if you're in there and you're doing coding by hand, completions is incredibly. The second thing is chat. Uh and chat I think, you know, think of it is just like you're pairing with an agent. You're going, you need a beta. It's an interactive mode where you ask something, you get an answer, and the agent does one or two things on its own. The third thing is like the true agent with like real autonomy. I think this is coming. Um, and in this world, you have to change the user experience to be based around higher latency interactions. And so like what does that mean? He's like um, you know, if it devel as a developer, let's say I'm asking an agent to build a feature for me, I don't want to really necessarily like sit there and watch it do it. Um, like might ask you to do it, it might take five minutes to get a plan and gather context. It might take 10 minutes to actually like execute and test. And like you're just not going to sit there and watch it. And so that point's a different sort of interaction modality. And so you know, I think what you need for that is essentially some sort of like workflow management software. So like it's not totally new, it's kind of like GitHub actions or something like that. Um, where you started as job, uh, when it uh finishes, it tells you. If it hits an error, it tells you. Um, you could have multiple of them running at once. Um, you can sort of see the history. And so there's it's like think of it as like workflow automation system that developers need for agents. Uh, I think one of the really important property of this system is that when the agent fails, I think it's going to fail a lot at the beginning or it's going to need more context that it's not a pain in the ass for the developer to hop in and work with it to fix the issue in one of those two other modalities chat or just like can, yeah, code by hand. So I think like the great like fallback mechanism is definitely an important aspect. I mean, traditionally like you know, when developers use terminals, there's always like four or five instances with each terminal doing different task, right? You're almost now like have to think about how do this new farm factor do the same thing, but also in a user friendly way or I don't know, maybe you go back to that. No, that's right. That's like that's why we're bullish on the terminal. Uh this very old school UX actually being weirdly a great uh like UX foundation for a world where you're working with agents because agents are just long running programs. they're long running programs that that the where the big difference between an agent and regular computer program is that the agent can apply its intelligence to a task and make decisions. Um, but they're basically long running programs. And the terminal, like again, going back to like what it was designed for, is something that runs long running tasks. It has all of this multitasking UX primitives built in. You can background a task. Um, you know, it has this interactive uh like grapple mode which is essentially chat and so, you know, this is like our kind of crazy thesis is that this is actually an awesome interface for uh like managing developer agents. And so that's yeah, we're psyched about where we are in the to try and uh, you know, build this new fort factor. Oh, can you talk a little bit about cost because um I haven't looked at the pricing that you're offering developers and enterprises, but there's obviously this whole debate about, you know, LLMs are costly, like the per is still you know cheap enough to make a sustainable business. Like how are you seeing that play out? Like are you seeing a path towards creating a sustainable business model and how are the margins going to look like uh? It's a great, great question. Um, so so the way our pricing works, we have um we have a couple plans for like individuals and small teams. And they're like the $15 and the $40 price point. And the big difference is it's really around like AI requests and it's it's really like it's a hard thing to price. I will say that because the underlying price of these models is actually all based on like tokens and models, but um, I feel like tokens is like a a little too close to the metal to price is is a little too far from the value to a developer. Like the ideal pricing would be like something that's very value aligned, but um, you know, the the types of being start we're trying to miss that hard to do. So we price it in terms of uh requests. Uh for all of our paid users we have like a pretty healthy positive margin, like, I don't know, 30 to 60% something like that. Like, however, it's really hard to do this because the underlying models end up changing both their their cost and they also change like um how much context they want to gather. So for instance, when we put out 3.7 support, our costs went way up because uh it's a very like context hungry model. And so, you know, it's like a a kind of game where we're we're trying to guess at how these models are going to evolve, we're trying to to also offer the best user experience. We also give all of our uh free users some AI because we want them to understand the value and so get them to like, you know, a moment where they want to pay us. And so, um, yeah, the the the costs are like I definitely think there's like a path to a sustainable business here. I I don't know exactly what's going to happen with model cost. Look at on the web, it seems like there's like some amount of commoditization which should bring costs down and be helpful if the app player. On the other end, uh it's like there's still a lot, you know, we want to be offering these cutting edge models that are very sort of costly to operate like the racing models and so on. So it's it's a, you know, a bit of an open question, but we're able to grow in a way which is not like we're not like a searating cash as we uh as we scale our user base, which is important. And I think there there is one path I always thought that these this dependency on LLMs could change completely when you adopt to an open source model and host it in your own cloud and then sort of start to fine tune your own models. I think would that sort of create a completely new profit yeah. So if we were to take like, you know, Deep see or La or something host it, um it's like a totally different level of control over the cost so on a sudden. It's like you're not paying the cloud provider. Like I think if you look at like the sort of who's making money on AI, it's the chip makers. Like I think there's like there's like a big margin on the video product. Then it's like the hyperscalers. So I think like, you know, Gago, AWS, Azure, QCL. The model providers. So there's a lot of people taking margin before you get to the app player. So if it's like you can um move away from the model provider and model, I think you can you can definitely get some of that margin back. We don't do that right now because the quality difference of the models is such that we think it makes more sense to you know, expose the Try to get like basically our our number one concern is like we get users to realize the power of this stuff and convert it to one to pay us. Uh, and then like from a unit economic standpoint, it's like we'll kind of like let's just stay break even at least and then wait and see on in terms of like uh what's the right way to optimize the costs. But it's it's I don't know. I think about this all the time. It's fascinating. Yeah, I mean it it's almost like if I'm using Uber, I don't care if it's on AWS or GCP or Azure, right? Like as long as I establish myself as the service provider, then I am good, then I can increase the margin over time by doing other things. Exactly. So we we want to have the relationship with the developer, um, and we actually like so we we let developers choose the model, but what we're moving towards is a world where we sort of automatically try and do it to give them the best experience is because like it feels like a weird implementation detail, and it feels like at the end of the day as a developer, unless I'm like someone who's like testing LLMs for my job, what it feels like what I should actually care about is like what's the quality of interaction. And so, you know, we we want to abstract that and just take the experience and using whatever the best most cost effective LLM is. Is there any metric that you really focus on uh for your product? Like, you know, for example, one of the things I'm interested in AI product is how much time does it take for a newly signed up customer to get to get me to pay it for it. Like that that's a very uh interesting uh number that I try to find out for new products is because AI products are particularly compelling. And if you hit a use case, it will take you five minutes to get them to pay. And I've seen like customers pay a lot faster than a SAS product or, you know, any other product. So I'm curious like do you have any special metrics or like special time or you know, dimensions that you are looking at which tells you a good feedback whether your product or a feature is working. Yeah, so one really interesting uh metric that we've just started looking at is like what percentage of uh things done in Warp are either done by AI or being asked to AI. And like a normal terminal it's zero. right? Normal terminals you can't even talk to an AI. For Warp, we're like, I don't know, 10%ish of things that are happening in the terminal right now are either the user asking in English or the AI doing something. AI engagement is the leading indicator for