1029: The Workflow of the Future With Zed

Nathan Sobo joins Scott and Wes to explain why Zed was built in Rust, how GPUI works, and what happens to editors once agents write most of the code. They also talk about DeltaDB, Zed’s new Git-compatible version control system, and Delta, the collaborative agentic editor it powers. Show Notes 00:00 Start 00:35 Welcom to Syntax 01:13 The Journey to Zed: Building the Ultimate Tool 03:26 Why Rust was chosen for Zed? 06:29 Brought to you by Sentry! 07:07 Building a UI from scratch in Rust GPUI 15:55 AI’s role in coding and development 18:42 The role of text editors in the age of AI 21:52 Delta DB: The vision for collaborative development DeltaDB 29:06 The Evolution of Collaborative Coding 44:45 The Future of User Interfaces 52:42 Sick Picks + Shameless Plugs Sick Picks Scott: Wes: Nathan Sobo: Keychron Q11 Grant Green - Idle Moments Shameless Plugs Scott: Wes: Nathan Sobo: DeltaDB Hit us up on Socials! Syntax: X Instagram Tiktok LinkedIn Threads Wes: X Instagram Tiktok LinkedIn Threads Scott: X Instagram Tiktok LinkedIn Threads Randy: X Instagram YouTube Threads

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  Our vision is really, it was always conversations in the code linked to the code rather than snapshots of the code. But what's changed in the past two, three years and really acceleratingly like since last fall, at least for people that write Rust, is that now conversations are becoming sort of the source of our code. Like, we were thinking we'd write all the code and have conversations in it. Now conversations are the beginning and code comes out of that. It's a downstream artifact.

• 0:35Scott Tolinski?

  Welcome to Syntax. Today, we have Nathan Sobo from Zed on the show. Nathan, previously from Warp and then was also at GitHub for nine years, is now the co-founder and CEO of Zed, the text editor that I use every single day, one of my favorite applications on my computer. So Nathan, welcome to Syntax.

• 1:01Speaker 00

  Thank you very much.

• 1:03Scott Tolinski?

  Would you like to get into, I know I gave a little bit of a backstory on who you are, but maybe like you could dive in just a little bit more of how you ended up at Zed?

• 1:13Speaker 00

  Yeah, I mean, I have been obsessed with the idea of building the ultimate tool for software development, basically for myself for a very long time. I think it was right when I graduated college, the year after 2006, that, you know, I'd been using TextMate, which was hot and new at the time. I'd used Eclipse, I'd used Emacs, hadn't used Vim at that point, although that did come along later. And I just wanted to build a tool that combined all the things I loved about these various editors into one product. And I completely underestimated how difficult that would be to do.

• 1:58Speaker 00

  And it took kind of until now, I think, to really even begin to get close to, yeah, basically bringing the speed of TextMate with the features of something like Visual Studio Code. And then also bringing this collaborative dimension that came along eventually. And that was based on my time at this company, Pivotal Labs, which was a pair programming shop. You would show up and there'd be nine to six with a one hour lunch break, just pairing all day long.

• 2:34Scott Tolinski?

  Really?

• 2:34Speaker 00

  And back then, we would plug two keyboards into the same machine, two keyboards, two mice, and just like work through problems together, code together. And that taught me so much. I became such a better software developer so rapidly. But I always wanted a tool that could replicate that experience that we had in this shared office, like over the internet and turn the internet into that whole experience. And Zed is as close as I've ever gotten until what comes next.

• 3:06Scott Tolinski?

  Yes, yes. I can't wait to get into what comes next because I got a preview of it. And it definitely like takes, it makes sense to hear you say that about pair programming, given what I have seen in this. So that's really super, super cool. So you picked Zed and was Rust, because Zed is built on Rust, right, from the ground up? Was Rust the choice initially, or did you land on Rust somehow?

• 3:37Speaker 00

  Well, once upon a time, I talked the founders of GitHub into hiring me to build an editor called Atom.

• 3:44Scott Tolinski?

  Yes.

• 3:44Speaker 00

  Some people may remember.

• 3:45Scott Tolinski?

  Oh, yeah.

• 3:46Speaker 00

  And in the process of creating Atom, we also created this framework called Electron, which is still around and probably running on all of our machines in some shape or form, eating way more RAM than the apps that it's running really justify. So yeah, Atom, the big idea with Atom was we wanted to make it hackable and we wanted it to run on every major platform. And so it really seemed like if we could build on web technology, which was already something people were familiar with, was really a flexible technology. You could load CSS and everybody knows JavaScript. If we could just figure out a way to masquerade a web browser as a desktop app,

• 4:32Speaker 00

  we would sort of achieve our goal of this flexible, hackable system that ran on all the three platforms. And we did. But unfortunately, the trade-off that we made, the devil's bargain that we made, was that the overhead of the browser and sort of the abstraction that it presents, which is this kind of lowest common denominator, common standard that needs to run in a backward compatible way on all these different browsers, although that is reducing down. But still, it's like, it really put a limit on the performance and the user experience that we could yield.

• 5:14Speaker 00

  And so after Atom sort of ran its course and Microsoft acquired GitHub and I was sitting there using Visual Studio Code, I just wasn't happy. Like, I just couldn't see a future where I continued writing all my software in this web browser masquerading as a desktop app. Just the performance wasn't there, the aesthetics weren't there for me. This collaborative dimension, even though they did introduce Live Share, like they never really worked as well as I wanted it to.

• 5:50Speaker 00

  And so that led me to one inexorable conclusion, that we had to get closer to the metal. And luckily, around that time, Rust, you know, yeah, it was, there were, we were working on it for quite a while before we started Zed. And as we worked on it, Rust hit 1.0, async await got introduced. Like Rust really matured as a viable option. And, you know, after being covered in blood from learning the language, we were finally traversed the learning curve and figured out how to build a UI framework, which really didn't exist at the time.

• 6:29Scott Tolinski?

  And if you want to see all of the errors in your application, you'll want to check out Sentry at sentry.io/synthesis. You can sign up today and get two months for free. Sentry is just a really incredible tool for not only tracking your performance, making sure your application has no bugs, but even just seeing what goes wrong when something goes wrong. Because things go wrong all the time when we're coding. And you don't want a production application out there that, well, you have no visibility into in case something is blowing up and you might not even know it. So head on over to sentry.io/synthesis. Again, we've been using this tool for a long time and it totally rules. All right.

• 7:07Speaker 00

  I didn't even know how to begin building UI framework. Because if you've tried Rust, you know the ownership rules, the borrow checker. There's quite a few constraints. It was really relearning how to do desktop app development from the beginning and kind of reinventing it.

• 7:25Wes Bos?

  Yeah, like, can we talk about that? The UI thing just blows my mind because you think if you're going to build a text editor, something as simple as like line wrapping is a pain in the butt, right? And then like, like a text editor has like underlines and things that are overlapping. And if you look at VS Code, they throw a lot of that out of the window as well. A lot of it's just absolutely positioned over top of each other in many layers. That seems really hard. So like, why would you think that's a good idea to build it from scratch? Because it's not like there was something already there where it's like, here is a text box, here is a background, here is like a text node that you can use, right?

• 8:09Speaker 00

  Yeah, I mean, we started with, well, I tried a couple of different approaches. Like at first I tried, oh, maybe we could use Rust in the core of this thing and then we'll wrap it in Electron. And so all the, all the really heavy parts will be in native code and then I won't have to solve the UI problem. But it was just like, performance went out the window as soon as we had to serialize JSON coming out of this native core or whatever. So finally, I got to the point where I'm like, okay, it's a metal graphics surface. And, you know, back then, we didn't have agents to just tell us how to do it. It was like digging up tutorials from the guy that wrote shader toy on YouTube and stuff,

• 8:53Speaker 00

  learning how to write, you know, very primitive shaders, learning how to build a render, a renderer for a 2D app. And then, of course, all the CPU bound stuff that was designed to deliver the data to the graphics processor. But like, for me, the key insight was just like all those layers and all that, all that abstraction that the web developed to give people something to program against that, you know, wasn't a shader, comes at a real cost, right? Like, and ultimately, all those things that I felt like were getting in the way. So by bypassing all of that and going straight from our program to the graphics processor,

• 9:38Speaker 00

  we really reduced the problem down from this sort of, I'm staring at the dev tools in Chrome, like looking at the profiler, looking at all these little slices of time that I have no idea what's going on inside of and watching the garbage collector run and all of this nonsense, style recalculation and reflos and just what? You know, having no control to having complete control. And ultimately, I decided that like, even though it was really hard and we had a lot to do, if we had control, we would be able to deliver. And the goal was, you receive a keystroke and you have pixels on the next v-sync of the display, so that there's zero perceptible lag.

• 10:22Wes Bos?

  Wow.

• 10:22Speaker 00

  And I knew if we had the control, we'd be able to achieve that goal.

• 10:28Speaker 00

  And what is it, six years later? Here we are.

• 10:32Scott Tolinski?

  And this is available, folks, for you to use, GpUI dot RS. If you'd like to use the UI framework, is this the exact framework and just made available from Zed?

• 10:48Speaker 00

  Yeah, it is. Right now, it's available under a permissive license. It's still in the Zed repo. It's not exactly like... You know what 37 Signals did with Rails? Where they really made it this big project that everybody could use? Like, it's open, it's available, but we're not exactly bending over backwards to worry about building this big ecosystem around it because Zed just moves too fast. Like, so up until now, we've just sort of kept it as a crate inside of Zed, but it's there, you can use it. People have even spun up forks, people have built all sorts of apps on top of it.

• 11:27Speaker 00

  I saw a like quant trading app, high-performance quant trading app, you know, doing crazy trading graphs and stuff like that built in it, and there's a bunch of things out there.

• 11:38Scott Tolinski?

  Yeah, you know, one thing I've always appreciated about Zed is the speed at which it operates in general, and I know that that seemed to be a big component of building this editor. But Zed has always felt so lightweight while being so full-featured to me. So it does make sense that that would be born out of the death of Adam and the arrival of Electron. Wes, were you an Adam user?

• 12:10Wes Bos?

  Yeah, well, kinda, because like I was the Sublime text guy, like that was my whole thing. And I use Sublime forever, and I dipped into Adam here and there, but I never went whole hog on it until like VS Code came around and it was like, like VS Code started to get that big momentum. That was probably like 11 years ago, if I were to guess.

• 12:37Scott Tolinski?

  Yeah, I really liked Adam because I always liked designed apps. So for me, it was, Adam was the first text editor where I felt like everything was designed. Where when I went to VS Code, it felt like a massive step backwards in terms of the user experience of it all, but then it just became like the thing to use. So it was like, okay, I guess I'm using VS Code, yeah.

• 13:03Speaker 00

  It just, they had so many more resources they were pouring into it, and the reality was they were a more mature team with a more mature codebase. Like, they were doing Visual Studio Online, Eric Gamma had already created Eclipse before that, and we were just like noob web developers making Adam, like we didn't really stand a chance in that matchup.

• 13:28Speaker 00

  And I think also a really critical thing that we just missed was TypeScript, and I think VS Code really rode to success on the back of TypeScript and the Language Server Protocol that they built around it, and just that whole evolution. And we were still trying to build an editor in like dynamically typed. At the beginning, it was CoffeeScript, don't even get me started, that was not my choice. You know, it took, by the time we realized that the Language Server Protocol, which grew up around Visual Studio Code, was gonna be a big thing, it was just kind of too late for Adam, I think. But, no worries, like, just, that all served as the training ground for what we ultimately applied to Zed.

• 14:13Scott Tolinski?

  Yeah. Would you, like, given what you've gone through in terms of having to build a whole UI layer and everything in Rust, if you were starting over today, July, well, I'm not gonna date this, because if you were starting over today, would you still pick Rust?

• 14:34Speaker 00

  Summer of 2026.

• 14:35Scott Tolinski?

  Yeah, summer of 2026.

• 14:38Speaker 00

  Would I use Rust again?

• 14:39Scott Tolinski?

  Yes, yeah.

• 14:40Speaker 00

  100%, absolutely, yeah. I briefly looked at Zig, and, you know, I don't know enough about Zig, and it wouldn't have been a choice at the time when we started Zed, I don't think it wasn't far enough along. But I think for the nature of the complexity of what we're building with Zed, and it being an open-source project where we're merging pull requests and from strangers on the internet, having the memory safety, you know, having really strict, type-safe security around everything we're doing is critical.

• 15:20Speaker 00

  And I don't really know any, a lot of other options that give us that kind of on-the-metal axis. My one gripe about Rust continues to be that the compiler is slower than I would like.

• 15:32Scott Tolinski?

  I'm always saying that, I have said that, and people in the comments are like, "It's not slow." I'm like, "Oh, come on."

• 15:38Speaker 00

  It's slower than I would like, but the language is incredibly well-designed, I love it. So, yeah, I haven't really seen anything emerge that would be a better alternative for the problem we're trying to solve.

• 15:54Scott Tolinski?

  Sick. So you mentioned you're taking in contributions from outside into Zed. I'm curious about, like, we always, in this day and age, need to ask about like AI usage in building these things, like, what's AI usage for the Zed team look like? And what do some of your processes look like in building?

• 16:17Speaker 00

  We use AI a lot, at least I do, and I think it varies across the team, and obviously, most of the, or everybody on the team is an engineer back from the days when you didn't have an agent to write any code for you, you were typing it out, and, you know, you were lucky if you could get an autocomplete. So I'm really glad that I have that pre-training of my neurons in place to really understand software. Like, I think it would be a different story if I were getting my start now. Some things would be easier, and then certain kinds of lessons might be harder.

• 16:54Speaker 00

  But I write a ton of code agentially these days, and I use it in the pull request review process. When I'm reviewing pull requests, it's amazing to get some help interpreting what the heck is going on in this diff. One thing we don't have a ton of is like this dark factory thing, where there's like loops and things going on super autonomously.

• 17:23Speaker 00

  We have a little of that happening, like, you know, just getting a head start on stack traces and things of that nature.

• 17:31Speaker 00

  But for the most part, definitely my usage of AI is extremely engaged and reading the majority of the output coming out of the thing, and, yeah, being very engaged with it. So, but it's a big part of my workflow. It's changed the calculus a bit, like, there's the volume of contributions, we're having to get much more clear about what we're gonna accept and not, and, you know, and the review load is quite a lot to keep up with, especially with a bunch of other priorities going on.

• 18:05Scott Tolinski?

  Yeah.

• 18:06Speaker 00

  But, I don't know, we get like 40-something bugs a week fixed via external contributions. I'm sure a lot of those were facilitated by agents. I don't begrudge people for getting leverage, especially in a codebase as complex as Zed. You do have to insta-close the flop ones that someone's not putting care into what they're doing. But yeah, it's a big part of it, and I don't have anything against it.

• 18:36Wes Bos?

  Can we ask about you guys building a text editor in the age of maybe we don't even need a text editor?

• 18:45Speaker 00

  Right.

• 18:46Wes Bos?

  And I don't wanna say that like snarkily, but like, like,

• 18:49Speaker 00

  No.

• 18:49Wes Bos?

  Like, what do you think about that? What do you guys, are you moving to something, or like, what do you think there?

• 18:55Speaker 00

  Yeah, I used to call what we were doing, building a text editor.

• 19:00Wes Bos?

  Yeah.

• 19:01Speaker 00

  And now I really think about what we're doing more as building a user interface for interacting with software. And text is a big part of every dimension of that, even more so now that agents are involved, right? Like, if you think about all the problems of interfacing to text, like, text has become even more powerful than it has ever been now that language models are this big part of our lives. Language is made out of text, and I think people... Underestimate how difficult it can be to build a good user interface around text.

• 19:43Speaker 00

  And I also still think that reading code and navigating through code with the facilitation of agents is going to remain a pretty important part of building real software that's not just tossed out, vibe coded, or we're just going to run a loop and torch tokens on this until maybe it's approximating correct.

• 20:05Wes Bos?

  Yeah.

• 20:06Speaker 00

  I think the craft of software is going to involve humans, I think multiple humans, engaging with fire hoses of text.

• 20:18Speaker 00

  And those fire hoses of text interleaving through source code.

• 20:24Speaker 00

  So, yeah, I think a lot of the, yeah, I don't know that necessarily the latency of typing a key inside of a source file and getting a response and autocomplete popping up and choosing from your algorithmically determined list of options.

• 20:42Wes Bos?

  Yeah.

• 20:43Speaker 00

  Is as relevant anymore, but a lot of the skills and lessons that we learn solving that problem, I think, apply even more so to this new world that we're entering.

• 20:54Wes Bos?

  Yeah, whatever that tool looks like, I agree as well. Like, and I'm so glad that you guys are building this because like we need wicked, awesome, I would say, I'm not going to say primitive, but like we talked to the PR computer guys and they're building sidebar tree structure of files, like arrows, like dropdown and diffs. And they built like the fastest one ever. And everybody, everybody needs it. And in this age of being able to just type in the box, whatever you want, we need these amazing, like well done, low level, like you said, on the metal primitives to build this type of stuff. So I'm glad that

• 21:36Wes Bos?

  like, not everybody has like thrown it out the window. And it's like, all right, we're just gonna vibe code absolutely everything and it's all going to be in the browser. Like you need people who care about this type of stuff. So I'm stoked that Zed is still here for that.

• 21:50Speaker 00

  Well, and I think it's worth mentioning that we're building something new.

• 21:55Scott Tolinski?

  Yes, I was gonna say this, Wes, almost kind of unknowingly, because Wes hasn't seen Delta and

• 22:01Wes Bos?

  No.

• 22:02Scott Tolinski?

  I want to applaud you, Wes, for that lead in because it's a great lead in to talking about Delta, which as the z.dev describes on the early access, that it's a version control system that records the work as it unfolds and keeps every change connected to the conversation that shaped it. I think that sentence really undersells what I saw from Delta. So Nathan, I'm going to give the floor to you to maybe tell us and describe to Wes a little bit about Delta.

• 22:30Speaker 00

  Yeah. So, I mean, what you're reading on our site as of the recording of this podcast, which I think by the time of publication will have said a lot more, but it's really about Delta DB. And Delta DB was about a year ago where our head was at about the problems we wanted to solve in general and with respect to the emergence of agents, which is really about filling in the gaps between commits and enabling this new kind of collaboration that we've been thinking about since the very beginning of Zed. The plan was that from the get go was build an environment worthy of the world's best developers where,

• 23:13Speaker 00

  you know, the most discerning, high skilled developers would want to spend their time writing software and then give those developers an environment to connect to one another sort of in the process of writing the software itself. Because I always viewed this diff review, I worked at GitHub, I loved my time there, but it never felt, back in the day, like it was called social coding. It never felt that social to me. I mentioned earlier that time of plugging two keyboards into a computer and like working directly with somebody in the code as we were writing software. That always felt like social coding to me.

• 23:54Speaker 00

  And it, to me, felt like a tools problem that we weren't able to have higher fidelity conversations sort of directly connected to the software we were writing. And we were instead just tying our conversations to these comments that were posted on this webpage that rendered kind of slow. And they were tied to these snapshots. And really like, it's all because of Git. It's all because Linus Torvalds built this tool for like helping people email patches to one another on the kernel mailing list. And everything was sort of designed around snapshots and ran at the pace of email. And it always just felt

• 24:36Speaker 00

  like a higher fidelity conversation would be possible if we could drop below that floor of commits. And so Delta DB has been our vision for, yeah, giving basically every version of the software as it evolves, an address that you can recall it, talk about it, and you can drape over that level of tracking, the ability to comment on any part of the code at any time, and thus be able to have a conversation sort of in the code base as you're working on it. Over the past year,

• 25:15Speaker 00

  the vision of Delta DB, you know, as we built it, as we started building it, of course, we weren't going to shove it into our million line plus IDE with hundreds of thousands of users on day one. We built this little reference client for it. And yeah, over the past year, that little reference client has become essentially a brand new product. So we're going to become a two product company. And yeah, so we have Delta DB and we named the new Agentic Agent Focused Editor that we built around it, Delta. And so in some

• 25:58Speaker 00

  ways, Delta looks at first glance like a lot of the other Agentic development environments that you see out there. But when you peek below the surface a bit and start using it, I think you're going to discover a lot of the power that is enabled by it being built around this more sophisticated version control system that drapes on top of Git. It's still designed to interoperate with Git. And you really don't think about Delta DB as you're using Delta, but what it enables is a new way to collaborate. And yeah, our vision is really, it was always conversations in the code, linked to the code, rather than snapshots of the code.

• 26:44Wes Bos?

  Yeah.

• 26:44Speaker 00

  But what's changed in the past two, three years, and really acceleratingly, like since last fall, at least for people that write Rust, because before then, agents couldn't really write Rust, is that now conversations are becoming sort of the source of our code. Like, we were thinking we'd write all the code and have conversations in it. Now conversations are the beginning and code comes out of that. It's a downstream artifact.

• 27:11Wes Bos?

  So you're sharing like the entire Agentic conversation, like that is stored in your history of Delta?

• 27:20Speaker 00

  Yeah, we have a Delta level. So every single token that flows in, every prompt,

• 27:26Wes Bos?

  Yeah.

• 27:27Speaker 00

  Everything that happens is sort of captured continuously as it's flowing in. And what it enables is you can start this conversation, attach clones of your repository to the conversation, share a link with a member of your team, or, you know, if you've set up a GitHub app or whatever we eventually evolve into, you can sort of broadcast it to your team, hey, I have this thread that I started with an agent. And anybody else can pop into that thread with you and they'll have the same work trees cloned down to their disk. And

• 28:06Speaker 00

  we do a conflict free sync of every single edit. So you can have a copy of the exact work, like in Git, like a work tree is your own personal thing.

• 28:18Wes Bos?

  Yeah.

• 28:18Speaker 00

  If you want to collaborate, you got to commit and push. Delta DB backing this whole thing means like every participant in that conversation with the agent or multiple agents, because you can switch between them. Has a clone of the code that they can touch and manipulate and it will all be synced across all the machines in real time.

• 28:39Wes Bos?

  Oh, man, that's cool.

• 28:40Scott Tolinski?

  Wes, it rips. It rips. And some of the things that you might not know from this is that like, every single line of code, you can determine like which agent wrote that code specifically.

• 28:54Wes Bos?

  And blame.

• 28:55Scott Tolinski?

  Which model. It is like a model blame is really what it is.

• 28:58Wes Bos?

  Damn it, Opus 5.

• 29:00Scott Tolinski?

  And you can annotate responses. It's really, really sick.

• 29:06Wes Bos?

  The people listening right now are probably trying to place where this will live in their thing. Like this is, like, does this replace GitHub? Does this replace Claude Code or Open Code? Or like, is this the, this is the chat app, but also the, like, the history?

• 29:26Speaker 00

  We're not aiming necessarily, I mean, long term, I would love to replace GitHub. And I think there are a number of contenders to that prize. But I think, short term, I think it will be a really good complement to whatever Git hosting provider you have in the very short term. Because what it enables is something that a snapshot-based collaboration can't really provide, which is just the ability to share a thread with somebody. And you know how like, if you share a link in Figma, you can just dive right in and see what the designer's doing and comment on it and just interact.

• 30:02Wes Bos?

  And see their history.

• 30:03Scott Tolinski?

  Or follow, yeah.

• 30:05Speaker 00

  You can come along a day later and maybe they're gone. And you can still sort of pick up. And that's this continuous experience that we're imagining, where you, rather than kind of chatting with an agent all by yourself, having it stamp out these snapshots, downstream artifacts, pushing them up to a server that's hopefully online that day, and then having someone try to load those massive diffs in a webpage that can't even render them, because it's using web technology that can't scale, and having a conversation with your team now on this inscrutable diff that like, maybe you didn't even understand as it vomited

• 30:44Speaker 00

  out of this agent at, you know, fire hose speed. All of that seems like, kind of, just not an efficient way to work anymore, especially in this paradigm. Our vision is more and more, maybe you get something started, then you pull your teammates into that conversation where they can interrogate the agent with you, have a conversation with you. You both have access to the agent and the entire context window leading up to that moment, and the code that's, you know, cross-referenced with all those moments.

• 31:17Wes Bos?

  Yeah.

• 31:17Speaker 00

  And it's just a continuous experience where the collaboration takes place, sort of, in the process of writing the code and we're continuously with it. At the end of that, you can still like, open a pull request and have CI run and merge and do all those things. But a lot of the conversation that you might have had in comments on the diff, if you could get it to render, you would have already had. And you'd had with the benefit of the very agent and all the context and planning and that went into that code existing to begin with.

• 31:51Wes Bos?

  Yeah, I often think about how much is lost in the conversation about how you got up to something, right? Like Stack Overflow is trying this, like Stack Overflow for agents right now, where when the agent learns something, or when it cracks something, like, write it down somewhere, because somebody else is gonna have to go through that process in the future. And wouldn't it be nice if, if that history of you getting up to that point would be, would be visible? Or even just like, you're, you're looking through some code, why is this built like this? Well, hopefully they leave a comment, but if not, you can, you can blame and look at the history of how you got there.

• 32:31Scott Tolinski?

  Yeah.

• 32:31Speaker 00

  Yeah, I completely agree. And there are two big pieces of Delta DB that are like, underlying the experience I'm describing. One is conflict-free replicated work trees.

• 32:45Speaker 00

  And that is sort of starting from some particular Git base and building forward from there.

• 32:52Speaker 00

  And because we're tracking that all in such a fine-grained way, we have a very nice, fluid, open comment on anything, anywhere, in this real time, or asynchronously. You don't have to work in real time, but it's like, the system doesn't impose any, can be synchronous or asynchronous. The system doesn't impose it on you. It's whatever makes sense for you and your team. So there's that, conflict-free replicated work trees. And then this other big dimension that we're working on is this delta-level understanding of the repository as a whole. And what that's going to enable is the conversations taking place in each of these threads that have their own work trees on them, to be aggregated together into

• 33:35Speaker 00

  this global index, where we're essentially turning your codebase into a, like a scaffold for all the conversations that you might want to have, or like a canvas. And so it adds this like metadata layer to the code where you can, sort of in any state of the code that you can see, whether it's committed or not, you can attach an annotation, and then be confident that in any other version of the codebase that that character, or range of characters that you marked, is visible, it will be able to be retrieved. I don't know that we're gonna, we're, that, we might ship the work tree stuff slightly before

• 34:16Speaker 00

  that, the repo stuff, but it'll be following soon after. Or if I can really crank between the recording of this podcast and, and when it comes out, it'll be out there.

• 34:33Scott Tolinski?

  Sick. Yeah. It all makes so much sense to me. And when we were at JS Nation in Amsterdam, React Summit, I was a part of like a panel on like, the future of UI and everything like that. And one thing that we landed on in that panel was like, pair programming and pair prompting needs to become more of a thing. There's too many decisions you're making. There's too much decision fatigue, right? Before, we used to make, you know, 10, 20 decisions a day in regards to maybe even much less than that, like large decisions. And now we're making tons of, it's constant barrage of answer this question,

• 35:14Scott Tolinski?

  answer this question. And just the amount of value that you'd have from bringing someone else into that process cannot go understated. It's like something that like, I'm so missing. And I think about so many instances on like, projects that I'm working on or CJ's, where like, it would be so much better to instill that layer of not just pair prompting, but this pair workflow. I'm like, I'm really excited for this to exist in the world.

• 35:41Speaker 00

  Me too. Yeah, I mean, and again, I think sometimes people hear pairing and they freak out because it's like, you mean I'm gonna have to talk to somebody? And, I mean, the answer is maybe, because I think you're right, that like, it used to be that you could look at a diff and the diff of sort of, was a reasonable proxy for the conversation that needed to happen and the alignment on the decisions. But now, like, the deciding is all that we do, like, like you're saying, right? And so, by the time you're looking at the 10,000-line diff, and you could try to break those down into smaller diffs, etc., but like, in the limit of that, the smallest diffs you can have are deltas,

• 36:23Speaker 00

  right? Like, bring somebody in on the conversation itself. Like, that actually is the true source of our software. It's not the source code anymore. And so, how can you really claim to have an effective review if you're not pulling somebody into that? And yes, you can like staple a transcript and, you know, attach it to the pull request or something for someone to look at later. But why not just make that conversation itself collaborative? That's what, that's what we're after.

• 36:56Scott Tolinski?

  Yeah.

• 36:57Wes Bos?

  The work tree stuff is kind of interesting. Like, Delta DB, it says on the website, virtualizes the work tree, so spinning up a new agent branch is effectively free. Before I have my second question, can you talk about how, how that is technically possible?

• 37:12Speaker 00

  Yeah. So we have our own, it's a database, Delta DB.

• 37:17Wes Bos?

  Yeah.

• 37:17Speaker 00

  We have our own B-tree implementations and hash array map trees. And then, you know, those ultimately write to some underlying backing store, depending on, you know, in the client we are currently using SQLite, but that's fungible. But it's sort of our own abstract representation of the entire file system underlying your tree. Like, we're not critically dependent on the file system. We model the file system. And we do that for a couple of different reasons. One is it lets us have, it lets us make it a lot more lightweight. Like, we use copy-on-write persistent data structures. So clones are

• 38:00Speaker 00

  order one operations for us. You can get that on some exotic file systems, but it's not really something that's like universally available and good. So we just decided, kind of like the same decision we made around the web browser. It's like, let's just own this. Let's take total control. And so, and then another reason is, we're able to make it conflict-free. So, you know, we have a surface representation of this tree of files and all their contents. But underneath that surface representation is a bunch of distributed systems foo that lets us, you know, have multiple replicas of this thing and people

• 38:45Speaker 00

  manipulating it on their own schedules concurrently.

• 38:49Wes Bos?

  So my, my Scott sense is tingling here. Did you build a sync engine then? Is that what you did?

• 38:56Speaker 00

  Yeah, that's a big, that's a part of it. Yeah.

• 38:58Wes Bos?

  Scott, they built a sync engine. Get back here. I don't know where he went, but Scott loves sync engines. Okay, so that means that any, any change ever that anyone has ever made is done in order.

• 39:12Scott Tolinski?

  Sorry, Wes, my dogs are going nuts, so I gotta bring them inside. I'll be right back.

• 39:14Wes Bos?

  Okay. We'll talk about SigEd in a second, but, but that means that because every single Delta, I, I understand the name now, you can just replay every single one and then there would be no conflicts. Is that how it works?

• 39:31Speaker 00

  Yeah, no, of course, if you're collaborating asynchronously,

• 39:35Wes Bos?

  Yeah.

• 39:35Speaker 00

  We could, we could, if we wanted to, like, collate together all of the deltas of people collaborating asynchronously and you would get some convergent result. Everybody would see the same mishmash of stuff slammed together. So obviously, like on an, on a semantic level, there's still the notion of a conflict. But if you're working together synchronously, or multiple agents, more importantly, are working together on tighter time scales, the conflict-free nature of it is really helpful. Essentially, it's giving the illusion, and we also do like mirror it down to the file system. But for us, the file system is just kind of like a user interface.

• 40:17Speaker 00

  Like, we have the source of truth, like, in our database where we fully control things and have the efficiency guarantees and the abstraction and the eventual consistent semantics that we want. Then we mirror that down to the file system to give, if you want to run your compiler, or if an agent that doesn't have access to Delta DB wants to interact with it. Ultimately, the illusion that we're creating is that this work tree exists on N machines at once.

• 40:50Wes Bos?

  Okay. And is that like a, like a mesh of computers? Or is there, will there be like a Delta DB server that's like the, the main boy?

• 40:58Speaker 00

  So we're launching delta.dev, which is going to be our collaborative platform that hosts everything. But the algorithms are designed to be, you know, very, I don't know how you describe it, like, they don't depend on a central synchronization coordinator in a critical way.

• 41:19Wes Bos?

  Yeah.

• 41:20Speaker 00

  But it just turns out that it's the best way for us to deliver a really reliable, solid, good, clear user experience. It's also the best way for us to monetize this because it's always been the plan of Zed to build a service that lets people collaborate. So for now, at launch, it's a centralized architecture for both of those reasons. It just make it stupid simple to go to delta.dev. Yeah, so we have a desktop, like, Delta is a desktop app that you can download and, you know, can read from your file system and do all the same things that you can do in Zed with respect to, you know, native privileged

• 42:01Speaker 00

  access and, you know, the fastest that we can possibly give you. And we also have a web component of it, which is basically very, the same codebase compiled the WebAssembly targeting, you know, WebGL, like, graphics surface. And that means you can share a link and give somebody a URL on delta.dev that they could drop in on and, without even downloading anything, you know, off with their GitHub and be collaborating with you.

• 42:33Wes Bos?

  Whoa. Wait, so the entire, hold on. The entire, the, the interface is built in, like Rust in the, the UI that you built?

• 42:43Speaker 00

  GPUI, yeah.

• 42:44Wes Bos?

  GPUI, but then you compiled that to Wasm or, and have a web component that you can drop into the web, but the UI was written in Rust?

• 42:56Speaker 00

  Right. So it won't be terribly slow.

• 43:00Wes Bos?

  That's crazy. So it's back to the browser, just no HTML. Is that right?

• 43:06Speaker 00

  That's right. Yeah. And the browser, this time, yeah, before, we brought the browser to the desktop. That was the Atom Electron era. We're going full circle by bringing the desktop to the browser, kind of like, bringing the fast tech to the distribution channel and just making it convenient. Really, the browser is about convenience. And we are going to want to encourage people, download the desktop app, you know, like every dual, like linear, right? You know, like, they encourage you to download the desktop app and there are benefits to that. But if you're willing to run in our hosted cloud provider,

• 43:47Speaker 00

  you know, and the cool thing about our, like, I was talking about the work trees.

• 43:51Wes Bos?

  Yeah.

• 43:51Speaker 00

  And, you know, so if you're using the web, then you're obviously going to have to use our hosted runners. But if you're using the desktop app, you could use your own computer and we can actually mix and match in the same thread. So you could have somebody running stuff on their own computer, and then they could share a link and someone joins from the browser and they're running their turn when they submit on a hosted runner. And these, and they could go back and forth, actually, like, yeah, I don't quite, you know, it's going to start simple, but like, kind of the sky is the limit in terms of the flexibility we have and of moving the code around to be worked on in the appropriate spot.

• 44:36Wes Bos?

  Mm hmm. That's really cool. Man, I just, I, I've been thinking about this, like, this is not related to this, but I, I've been thinking about like, like, feature of UI. And everybody's talking about, like, like, I did a video on just like generated UI or like, some people are talking about it will just be streamed in. Do you have, do you have any opinions of like, would that ever be a future? I know games kind of already do that, but like, just some UI generated and simply just streamed to a thin client?

• 45:06Speaker 00

  I mean, the first thing I'm going to say about UI is I think we're going to have a lot less UI. Because

• 45:13Wes Bos?

  Yeah.

• 45:14Speaker 00

  The ultimate user interface is basically language.

• 45:17Wes Bos?

  Yeah.

• 45:18Speaker 00

  And, you know, for the longest time, for the last whatever, 50 years, 40 years, like, the only people that could really use language to talk to a computer were people like us, software developers. And even then, it was like, pretty grueling, brutal process of memorizing a menagerie of commands and flags and programming language syntax and all of it.

• 45:42Wes Bos?

  Yeah. I was trying to explain to my nine-year-old the usage of brackets, semicolons. And I was just like, this is really wild. Yeah.

• 45:51Speaker 00

  So I think now that language, natural language is like, opened up as a way of interacting, like, it's such a superior means of expression over, essentially, like, these gestural user interfaces that we'd been confined to use in the previous era. Like, it's really just, yeah, it's sort of like the, like, I, like, if you, if you couldn't talk, I guess you'd have sign language. So I don't know, maybe I'm, but like, you know, I think the most sophisticated user interfaces, the problem with them is they started to approach sign language, which is really hard to learn. And so we were kind of in this

• 46:30Speaker 00

  common denominator of, yeah, fairly crude gestural interfaces. And I think as language becomes cheaper and we get better at building user interfaces utilizing text and language, it's going to take over more and more of the user interface. But then I still think there is a place for graphics. Like, it's why Delta's a graphical app and not a terminal agent. Because I do think that, yeah, graphics and, you know, subtle use of images and icons and progress bars and things that are not confined to a, you know,

• 47:12Wes Bos?

  A terminal line.

• 47:14Speaker 00

  Monospace terminal grid.

• 47:16Wes Bos?

  Yeah.

• 47:16Speaker 00

  Still have tremendous value. And so I think in that context, absolutely, are we going to have, yeah, even like Mermaid is in Delta, like, a very early example of that. And I think we're going to be adding more capability of that sort over time of like, yeah, Mermaid is this markdown-esque, it's like the markdown of diagrams or something, right? Like the syntax the agent can token efficiently generate to build a little flowchart or whatever it is you need in line. But yeah, why wouldn't that extend to where gestural interfaces are appropriate? Like, why wouldn't you let the agent sort of assemble that on

• 47:58Speaker 00

  the fly? But I also just think you can accomplish so much with language alone.

• 48:03Wes Bos?

  Totally. I agree. I agree. That's good. I'm, I'm glad you said that because I've been on this, like, what does the future of UI kick go for in the, in the next little, last like six months? You know, I've been in talks and talking about MCPUI and some people think the browser is going to go away. And it's somewhere in between. So I was like asking people that whenever I get to chat with them.

• 48:26Speaker 00

  It's funny, like, when I, I forget what website I was trying to navigate. It was maybe like from a, a summer camp for my kids or something like that I was trying to figure out. And like, the website was just like awful and really hard to navigate and understand. And I just finally was like, it didn't actually take me long. I just fed it to My agent. So can you read this and tell me, just answer my questions? Like, I don't want to browse. I want to ask.

• 48:56Wes Bos?

  I don't want this interface at all. I just want to do the thing that I want to do.

• 49:01Scott Tolinski?

  Yeah.

• 49:02Speaker 00

  Yeah. So I think a lot of UI is converging. And everybody's kind of going after it from a different angle. But I think we're converging on these. You see the same app over and over again. I mean, I think one unique thing about Delta is we know a lot about how to build efficient interfaces for text. And so like, we don't model the conversation as a chat.

• 49:21Wes Bos?

  Yeah.

• 49:22Speaker 00

  It is a chat, but it's a document. And that means you can move your cursor around inside of it. And when you put your cursor in, anything that came out of the agent, whether it's a command that it ran, or an edit that it made to your code, or a nested markdown list with 17 different options, lettered A through Z with numbers and letters and like, right? I got sick of remembering the letter of the option that the agent wanted me to follow. And so what we let you do is you just move your cursor around inside all that rich text freely. And wherever you want to annotate, you just start typing. And it immediately like lets you annotate right there.

• 50:08Speaker 00

  So that's an important part of it.

• 50:10Scott Tolinski?

  I'm so amazed by, I mean, the fact that when you all first showed me Delta, I was immediately impressed with the fact that it wasn't another Codex slash any of these apps that, like you said, they all look the same. Here's your projects. Here's your chats. Here's your chat window. And to some extent, there is that, but like, it is different. It is fundamentally different. And I think that takes such a, I think that takes such a different brain to like, what does this actually look like? Surely, we can't just be having a chat GPT chat interface for every single code base where you're just chatting at it. And I think this like, in my experience with the tool, was really impressed by the capabilities there.

• 50:55Speaker 00

  Thank you. Yeah, I mean, it was just born out of the experience of spending a ton of time trying to write code with agents and building what we wanted, or trying to reduce the friction we ourselves were feeling.

• 51:11Scott Tolinski?

  Yeah. Yeah. And it feels like, you know, where, you know, you wanted to create Zed, so you had to create GPUI to solve those problems. You know, wanted to solve these problems of working with agents. And you have a whole version control system. All kinds of new collaborative tools and interfaces. So it takes people like that to think outside the box. And I'm really stoked for the audience to give Delta a try. I'm stoked for Wes to give it a try, honestly. Wes, you want to pair program with me on some stuff?

• 51:47Wes Bos?

  I'm in. I'm, yeah. I feel, I remember this now. I tried to sign up for it when you sent me the link and I hit some Google Forms issue. And I did, I forgot to go back to it. But I just put request early access on there and it worked this time. So I'll give you a shot.

• 52:05Speaker 00

  You can get the binary. And the binary, you know, obviously, we're recording this before the launch date. It's getting ever less alpha and ever more beta. And yeah, we're, you know, we'll be launching the beta and starting to let the first people just from the internet in sort of around the publication date. Yeah. So today.

• 52:28Scott Tolinski?

  Sick. So folks, head on over to, you said this will be at a new URL? It will be at?

• 52:35Speaker 00

  Yeah, delta.dev.

• 52:36Scott Tolinski?

  Cool. Head on over to delta.dev. Nathan, is there anything you wanted to hit that we didn't touch upon here?

• 52:46Speaker 00

  Well, you had, you told me that we had the, you wanted me to share about what I loved, or something that I loved unrelated to.

• 52:52Scott Tolinski?

  Yes. Well, that's what I was going to get into next.

• 52:57Speaker 00

  Sure. Okay.

• 52:58Scott Tolinski?

  Good. No, no, not shoot. Perfect lead in. This is the part of the show where we talk about sick picks and shameless plug. So a sick pick is anything you're enjoying in life right now. What is your sick pick?

• 53:09Speaker 00

  Can I have two?

• 53:10Scott Tolinski?

  Yes.

• 53:11Wes Bos?

  As many as you want. Yeah.

• 53:14Speaker 00

  So yeah, the two thoughts that came to mind. One is my keyboard. It's the Keychron Q11 here. I have two of them on my desk because one, when it's not raining, I take out all my debt. But what I love about it is it's a mechanical keyboard.

• 53:33Scott Tolinski?

  Ooh, split.

• 53:35Speaker 00

  Properly. But it's not weird, right? Like, if you were to put this thing together, it looks normal. And it doesn't have like an ortholinear layout or any of that, like, probably really ergonomic stuff that I cannot get my hands to learn. But it lets me separate my arms.

• 53:57Scott Tolinski?

  Yeah.

• 53:57Speaker 00

  And it has mechanical key switches and it has programmable firmware. And so I'm able to map the, the right or the left spacebar to command. So I'm not curling my thumb under, you know, just getting some control. And then like using the, yeah, using agents to program the QMK firmware.

• 54:21Scott Tolinski?

  Yeah.

• 54:22Speaker 00

  To like, make it light up. You know, I light up the F keys in a certain way so that when I'm coding at night, I like know exactly where they are. Yeah, I love it. It's a great product.

• 54:35Scott Tolinski?

  I don't think I've seen a split that is normal. Looking, all those keyboards are all crazy. And they're all really hard to get used to. Yeah.

• 54:44Speaker 00

  Exactly. And my problem is like, you know, inevitably, as a founder, I'm always like on an airplane flying somewhere. And then I'm on my laptop again. And so I find like, as soon as I build the muscle memory for like the ortholinear domed crazy thing, I go back to my laptop and then I couldn't use it there. I don't know. I just don't have, some people are great at that, like switching UI. I'm very bad at it. I, I like, my muscle memory is very one path. Yeah. So that's the technical recommendation. And then the, the non-technical recommendation, something recently in my life that I love is, yeah, growing up, my dad would listen to like the worst jazz.

• 55:27Speaker 00

  Like, it was like, you know, the smooth jazz, like the kind of stuff that you would listen to, like in the elevator, like at the dentist's office or something like that. But a little over a year ago now, I like flew into, you know, I actually was going to Mississippi for Christmas. My girlfriend's mom, visiting my girlfriend's mom for Christmas. But we flew through Louis Armstrong International Airport or whatever. And I, in New Orleans. I'm like, oh, Louis Armstrong, like, I want to learn more about jazz. And I've kind of, with the help of AI, which knows so much about like this catalog of jazz.

• 56:11Speaker 00

  Like, yeah, really honed in on like kind of 1955 through 64, really starting with Blue Note Records, which was, you know, they were one of the first labels to put African Americans like on the album covers and not be racist jerks. And they gave the musicians time to practice, which was like a rare thing at the time. And they really valued artistry in the musicians. They weren't after just like what would sell. They wanted to find kind of the artistic essence of the music. And yeah, so it's, it was, it just served as like a great

• 56:56Speaker 00

  entry point into a new understanding of jazz that I never had. Takes more patience because I just think the music is a lot more sophisticated than like ACDC or whatever. But it's something that I've really come to, to love. So I figured I'd give you two, like a nerd one and a cultured one.

• 57:15Wes Bos?

  That's good. I like that.

• 57:16Scott Tolinski?

  So any album in particular you want us to check out? Because you would, you would not believe that you are the second guest in a row to Sick Pick Jazz.

• 57:26Speaker 00

  Oh, wow.

• 57:27Scott Tolinski?

  No one has ever Sick Pick Jazz. Dan Abramov, who we just had on the show, Sick Picked a Thelonious Monk album. So.

• 57:34Speaker 00

  Nice. I love this. Well, the first track of it especially, Grant Green, "Idol Moments." It's like, slow, bluesy, just like got so much mood and atmosphere. I don't, I'm not a musician, so I don't, I don't know that I'm, you know, well positioned to like give you all of the musical detail of it. But I just love it. It's got vibraphone in it.

• 58:02Scott Tolinski?

  Ooh, I love a vibraphone. Love a vibraphone.

• 58:05Speaker 00

  I remember that from like Mr. Rogers as a kid or whatever. But like, hearing vibraphone played by somebody that's got this, just like, they're so cool, like blues sophistication going on. I don't know. I just feel like so cool when I listen to that track. Yeah. Idol Moments.

• 58:22Scott Tolinski?

  I haven't heard this before. I don't know any Grant Green. So this is going to, this is gonna go on immediately following this. Can't wait. I love vibraphone. Yes. Big fan. Yes. Sick. And shameless plugs. We already talked about Delta. Is there anything else you'd like to plug?

• 58:40Speaker 00

  No, that's really it. Check out delta.dev. It's a new way to collaborate with agents and your team, written on all the same tech stack as Zed. And, and a lot more.

• 58:56Scott Tolinski?

  Sick. And it's awesome. So check it out. Thank you so much, Nathan. This has been totally, totally cool. Yeah.

• 59:03Speaker 00

  Really appreciate that opportunity.

• 59:06Scott Tolinski?

  Yeah.

• 59:06Speaker 00

  Talking over you again.

• 59:08Scott Tolinski?

  No, talk over. Yes. And fellow Colorado folk, we got to get together sometime. So.

• 59:17Speaker 00

  Agreed. That'd be great. Sign me up.

• 59:23Speaker 00

  Okay. Well, until then, it's been good.