903: Fork Yeah! Microsoft open sourcing Copilot

Scott and Wes are joined by Erich Gamma, creator of VS Code, and Kai Maetzel, Copilot Lead, to share some big news about the future of VS Code and Copilot. They discuss what it means for developers, how AI is shaping the future of coding, and why staying open to the community is key. Show Notes 00:00 Welcome to Syntax! 01:00 The inception of VS Code. 02:49 VS Code adoption. 04:31 Brought to you by Sentry.io. 04:55 Syntax Denver Meetup! 05:19 The big announcement. 06:25 The current state of Copilot and VS Code. 08:31 The challenges with LLMs running outside of the codebase. 09:31 How to make a business case for AI. 10:47 The maturing of the AI landscape. 13:01 The limitations of extensions. 14:06 Open source vs closed source. 14:49 Copilot’s context is public. 19:23 Is context language-specific? 21:23 How does this affect paid Copilot features? 23:27 Secrets of Copilot’s server-side. 28:36 What will be open and what will not? 29:03 Is Copilot’s UI influenced by VS Code forks? 31:31 Maintaining VS Code identity in forks. 33:07 What does open-sourcing GitHub Copilot mean for Cursor and Windsurf? 38:42 Were you surprised to see VS Code forks? 40:03 Are other extensions able to tap into the AI offerings? 43:20 There’s work to be done. 44:13 The timeline. 45:39 Simulation Tests (S Tests). 48:07 How to test LLMs. 49:10 The future of software development with AI. 52:47 What’s your favorite model? 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.

• 0:00Wes Bos?

  Welcome to Syntax. Today we've got a really exciting episode, a big announcement about VS Code and GitHub Copilot. We got Eric Gamma and Kai Metzl on, who are some of the early on, they've been at, been working on VS Code for, for many, many years. And they're gonna, they're gonna talk to us all about it. So welcome Kai, welcome Eric. Thanks so much for coming on.

• 0:22Speaker 01

  Thanks for being here, Wes.

• 0:23Wes Bos?

  Yeah, real quick, let's give a quick introduction before we hit the announcement. Just gives people some context of who you are within the VS Code project.

• 0:32Speaker 01

  So I'm, I initiated when I jumped Microsoft 15 years ago, I started the VS Code project and was leading it for quite a while. And more recently, you know, I stepped down, became more an IC little contributor and Kai is now running the show.

• 0:50Speaker 02

  And I joined the VS Code project almost 10 years ago now.

• 0:55Wes Bos?

  10 years, wow, yeah. Is VS Code really 15 years old? What was the initial like creation? Like, why, why did you start that project in the early days?

• 1:06Speaker 01

  So that we noticed there are kind of these new kind of developers that we often refer to as spawning the web developers. And we wanted to also give them tools so that we are relevant to them, right? We had a big tool, great, great visual studio, all integrated, one of the best, but we wanted to disrupt this a little bit. And really attract only the web developers that don't like the all integrated one. They like command lines. They like to build their own tools, use many languages, whatever. That was the creation moment and from that went on.

• 1:42Wes Bos?

  Oh, well, it worked. You were just telling me a story about how like, like I'm a web developer. I'm probably the target audience. I moved over from Sublime Text many years ago. Are you, are you able to tell that story really quickly? Cause I thought that was kind of funny.

• 1:56Speaker 01

  Yeah, no, you were one of the sublime evangelists, eventually, right? And you saw that, I guess you wrote a book, had a theme about it. And you really talked about how can we convince the community with VS Code, right? And we saw you and we said, yeah, if you were able to convince this boss. Nice. VS Code, that's a good milestone, right? And then I guess you wrote to see and you gave feedback. And so we somehow accomplished that. I haven't, I still missed a book from you on VS Code.

• 2:30Wes Bos?

  Oh, yeah. Yeah, I've been, I've been considering it. I maybe at some point, there's, there's quite a bit that a lot of tricks and whatnot that I have. I actually recently got off the Sublime Text key mapping, which I was on for like seven years. And now I feel like a baby again. I can't remember all of them.

• 2:49Scott Tolinski?

  I remember when VS Code first came out, it did seem like it caught fire immediately. Because yeah, we had Adam, we had Sublime, we had all these editors. But it does feel like a substantial portion of the community moved right to VS Code. From your perspectives, was that how it felt internally as well?

• 3:11Speaker 01

  We sometimes talk about that it's an overnight success, 10 years in the making, right? So we worked quite a bit before, right? So when we announced it, we worked for five years on it in a monthly rhythm shift every month. And we then gradually have grown, right? It wasn't a big bang, right? So and it was always this growth approach, right? We have continually growth and we are now at, right?

• 3:39Speaker 02

  Forty million.

• 3:41Speaker 01

  Forty million.

• 3:42Wes Bos?

  Fourteen million, like what like users?

• 3:45Speaker 02

  Four zero.

• 3:45Wes Bos?

  Forty?

• 3:46Scott Tolinski?

  Four zero. Yeah. Oh, yeah.

• 3:48Wes Bos?

  That's nuts.

• 3:49Speaker 02

  The growth rate, right? When you compare it to today's growth rates for, for some of other tools, right? Then maybe it was not that big, right? But for the time, it was, it was quite impressive, right? How we actually grew over time. And I remember, so the first million we hit, right? That was a big party. And then the second million we hit, that was a big party. And then, you know, the third million, so the parties got a little bit smaller. And I was like, okay, we doubled it again. And we doubled it. And so it's the, and I was like, oh, yeah, okay. Now we hit 40 million. That's.

• 4:22Wes Bos?

  Yeah.

• 4:24Speaker 01

  I remember the one million cake. For one million, we had a cake, but I don't remember the other cakes.

• 4:32Scott Tolinski?

  And if you want to see all of the errors in your application, you'll want to check out Sentry at sentry.io/syntax. 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/syntax. Again, we've been using this tool for a long time and it totally rules. All right.

• 4:55Wes Bos?

  One more thing, real quick. May 29th, Denver, Colorado. We're doing a syntax meetup. You gotta be there if you're anywhere within a horse ride of Denver, Colorado. We got the folks from Vercel, the folks from Mux, folks from Sentry. We're all coming out, of course. We're gonna have Scott there. CJ is gonna be there. I'm gonna be there. It's gonna be a really great time. If you want to grab some tickets, they're free, but you gotta get tickets. Go to syntax.fm/meetup. VS Code, absolutely huge. In the last, probably, what, four years or so, copilot, GitHub copilot, absolutely huge, sort of changing how we, we write code and it's been a massive one. So the big announcement, lay it on us. What do you, what is it today?

• 5:39Speaker 01

  Well, it's about open sourcing. VS Code is open source since 10 years. And we wanna now open source the AI features from Copilot to Copilot in VS Code. So it's about open source, which we talked about VS Code's success before.

• 6:00Wes Bos?

  Yeah.

• 6:00Speaker 01

  And open source is one of the reasons I think we got all this developer love, right? We try hard, we try hard to be a good open source project. We try to be transparent, make PR submission easy, listen to feedback and so on. And yeah, open sourcing is very important here to our heart. It's a way how we develop and that's why this announcement is so exciting to me.

• 6:25Scott Tolinski?

  Incredible.

• 6:26Wes Bos?

  That's really exciting. So what that means is that like, like, explain to us how Copilot and VS Code works right now, because it's, it's like a, is it a plugin or is it like a, was it a closed part of VS Code?

• 6:41Speaker 01

  Let me explain how it is. And then you explained the whole business story because you initiated the whole thing and you can tell this very well, right? So today, it's an open course, open core pattern we are using, right? VS Code is open source, of course. It has some AI naming features in there which are open source. But it's open core, but the Copilot is a closed source extension. And there was a good read, there is a good reason why we started like that and why there is an evolution.

• 7:16Speaker 02

  And that's where I jump in.

• 7:17Wes Bos?

  Yeah, yeah.

• 7:19Speaker 02

  So, yeah, the, you know, when you think about the, the story as such, right? It started actually with GitHub Copilot just completions, right?

• 7:29Wes Bos?

  Yeah.

• 7:29Speaker 02

  And there at the time, we had really close working relationships, actually, with the folks at GitHub Next, who did this in the beginning, right? Because it was about what is, what should the UI look like, right? And this is when the ghost text in a way came in, ghost text completions, right? Before we really just had suggest widgets, right? So, and, and there it was in a way, like, a lot of the smarts were actually just in the extension. It was a fine tuned three, five model, right? So there was a lot of craft actually in the prompt creation and all of these kinds of things, right? And, and Copilot was, when I say Copilot, in this case, the completions, right?

• 8:14Speaker 02

  They were really runaway success, right? You saw a lot of folks kind of seeing this and saying, oh, this is great, this is magical. There were still some who were kind of upset that it was not correct, right? And then sometimes proposed something that was not quite spot on. But overall, I think it was a runaway success. And then the GPT four, right? So jet GPT announcement and GPT four came, right? And I think at this moment, everyone really, even the last person started thinking about what are all the other things that you actually can do. With with AI, right? And what became relatively uh obvious was that I think the UI pieces of AI need to be really within the flow of what you actually do.

• 8:57Speaker 02

  So it really starts permeating all of this, right? But at the same point in time, it was still extremely hard to to actually get reasonably functioning AI functionality, right? So there was still a lot of work that actually you had to put in the prompts, the pre-processing of the data in order to create the prompt, then actually post-processing, right? The models were really bad in actually answering in a format that you wanted them to answer. So there was a lot of code that you actually had to write in order to make that happen, right? Many different kinds of prompts for all different kinds of use cases and so on, right? So that was innovate the technical side, right? On the on the business side, I think you also have to think about, no one really knew how to make a business out of this site.

• 9:41Speaker 02

  Early on, it was like, oh, this is insanely expensive to run, right? So what do we do? We don't want to give anything away because no one really knows where the market goes and so on, right? So in the end, we ended up, as Eric said, in this kind of open core approach, which is that the UI elements, a whole bunch of the UI elements actually started more and more moving into core. We had to enable some of them through, for example, uh APIs, right? So we created more APIs and so on, and then parts were in the in the Copilot extension. When you then actually think about what really happened over the last few years, right? I mean, you just go, let's say you go to OpenAI and the OpenAI cookbook, right? I don't know, literally tells you, oh, here's the best ways of how you write your prompt

• 10:27Speaker 02

  in order to get, you know, reliable answers. Anthropic, the same story, right? Um, Google just had this sixty-eight-page PDF, right, for for how to prompt and so on, right? Then you have something like Codex that OpenAI open sourced, right? Where you literally can follow along how the whole prompt is created and all of these kinds of things. I think the landscape really massively shifted over the last year, maybe two years.

• 10:54Speaker 02

  Right in the way how mature this, how good the models actually became, even working with small little prompts, right? Prompting overall became much simpler.

• 11:04Wes Bos?

  Yeah.

• 11:05Speaker 02

  So, and and if you look, if you take this all together, right? Then, and now also better understanding of the business models, right? That uh what you really have to have on the servers, right? And that is pretty much your your secret sauce. And what you have on the clients in a way became more and more of the of the way of, how should I say? Well, not too much uh specific pieces in there that really differentiate you. And from that point of view, right, we we think it's it's really a product time, right? To to put all of this in open source. Like you wanna have people, people to be able to to kind of go and say, oh, VS Code, I cloned it, right?

• 11:48Speaker 02

  Now I have cloned it, now I built it, now I launch it, now I have code OSS, right? This is the out of source kind of version that you run, right? And with that, you can pretty much type something in the chat, right? And follow along, right? You can debug all the way how that actually makes it then finally to the server, right? And what the server actually responds, right? And then you can see how these things actually come. You can optimize forward, right? You can find problems, uh you can fix issues, can create PRs if you want, right? So I think that's that's really the story, right? I think overall, the same reasoning really applies to open source. That applied to to VS Code in the first place, right? Which is when you achieve a certain majority of a market, right?

• 12:30Speaker 02

  Then it makes sense to actually have a solution that is really open for for all developers, right? Where people can see what is happening behind the scenes, right? Can participate, can have discussions with you about the features that you actually want to have and so on, right? And then not just rely on us to to actually do something about it.

• 12:51Wes Bos?

  Yeah.

• 12:51Speaker 02

  And there is a discussion, but can can come and contribute, right? So we we strongly believe that this moment is here now, so that's why we made that call.

• 13:00Wes Bos?

  Interesting.

• 13:01Speaker 01

  So maybe what I like to add, now what what I see also, for instance, now we have a competition like Pairswoop. What they really show us and what made me always jealous, right? When you are kind of an extension of VS Code. You are not really a little bit bolted on, right? But what you wanna do with this move? You wanna have the same thing, right? It's an AI editor. It's not just bolted on. It's really a core part. When you do code OSS, right? It's there. It's not that you have to install whatever, right? So it should just be this experience. And maybe also, once you know the whole prompt thing that Kai mentioned, it's in the extension right now, right?

• 13:45Speaker 01

  The whole prompting, how we do it and talk to the model, prompt construction. That's in the, in the extension. And this thing right now, we also think we wanna make this open source, get more eyes on it, right? Get contributions on it. And yeah, a security benefits and so on. And that's a good thing. It's also a very good thing for the developer side. And I like it. It's because for us as a developer, it's a little bit stressful to kind of straddle between open source and closed source. Right? We do both, right? And we notice we talk differently, open source to closed source. Open source, we love to talk about and we get love. Closed source is different, right? Because it's not as friendly.

• 14:30Speaker 01

  People are more demanding because I don't know. It's just different. And for us as developer team, it's always stress to play both hats, right? And for us, it's a relief to have this open source consistently for the build process and for our thinking process.

• 14:48Wes Bos?

  Yeah. So part of Copilot and probably a big part of it is, is figuring out what, what parts to send, send to the model. Correct me if I'm wrong, but like, I'm, I was always curious about like, and we, we've had the folks from Windsurf on and, and whatnot. And we always ask them like, like what, what do you send? Like, are you looking at my clipboard? Are you looking at the open tabs? Like, how do you, how do you crawl for that perfect context? And, and how, how all that works? So with Copilot being open source, we'll be able to see exactly. What is is being sent to the model? And and how those and then also improve upon is that is that one of the big ideas here?

• 15:29Speaker 02

  Certainly one of them, right?

• 15:32Wes Bos?

  Yeah.

• 15:32Speaker 02

  It changed a little bit, right? In the beginning, for example, you were absolutely right, right? The the models didn't support reasonably tool calling, right? And from that point of view, you had to pretty much pack up everything upfront. Right, that you needed and send it to the model, right? But then there was also the needle in the haystack problem that when you send too much, right, the the model kind of got confused in these kinds of things, right? We we still do this for for completions, right? That we pretty much pack all of the context upfront and then let them the model complete that, right? And there's still a lot of thinking going into it. What needs to be sent along like, for example, opening tabs?

• 16:11Speaker 02

  We we compute some additional kind of context depending on if the language server actually supports that. You know what what types, for example, should be sent along in in a stripped out way and so on, right? But then there are also other parts like that are very where that doesn't play so much into it anymore, like agentic flows, right? In agentic flows, you use something that we sometimes internally refer to as as breadcrumb prompts, where you pretty much say, so so there is a problem in the problems panel. There's output in in the terminal. And so you really just say these things and then you give the tool extreme.

• 16:52Speaker 02

  Uh, sorry, the model tools in order to to retrieve this information, then it thinks that it actually needs that information.

• 17:00Wes Bos?

  Okay, okay.

• 17:01Speaker 02

  And and that is a is a much simpler approach, actually, right? And sometimes you you have to help a little bit along with the right prompting approach to say what kind of aspects to prioritize in its considerations and so on. But overall, right, that uh tool calling made that much simpler.

• 17:22Wes Bos?

  Interesting.

• 17:23Speaker 01

  Now, when you say prioritization, kind of, you also have infrastructure, right? We use kind of TSX to describe prompts, it's not just string concatenation. And this gives us flexibility, right? That we can say the nodes in a tree have a priority and we can add more or less depending on on the context window.

• 17:41Speaker 02

  Yeah, the this is how we render prompts, right? What I was saying is more like in the prompt itself, you you tell the tool that or you sell the model, what tools first to look at, right? For example, check, you know, if there is some problem in the problem, do this first, right? Don't don't look, for example, in the output of the terminal or something along those lines, right? So you give it some guidance how to proceed through the context. What Eric is referring to is, of course, right? We still have limited context windows, right? And there's a real trade off here, right? You have context windows that are either technically constrained by the model or you actually want to constrain it so that you get a certain throughput.

• 18:25Speaker 02

  So the rule of thumb clearly is with the simplest possible prompt that you can can send, you want to have the best possible outcome. And so if you pump pump the full context window, right? Then the real question is, and you get a certain outcome, how small can you make it in order to get the same outcome? Assuming that the outcome actually is good, right? And so that is then where pretty much our whole prompt rendering story comes in, right? What are parts in the prompt that you can can drop at which points in time? Right? Once you hit the context windows, once you hit a certain size for a certain portion, like think about history, right? You don't want to just pump pump pump history into into the context window

• 19:09Speaker 02

  until the context window is full, right? So what is a what's a natural way of capping this? Right? At which point did the conversation flip, for example, to to be about something else? And this is not relevant anymore, right? So these kinds of things.

• 19:23Wes Bos?

  And does that include logic for different types of programming languages? Or is that left up to the the model? Like, does the Copilot extension have logic for how to treat TypeScript and JavaScript versus like Java?

• 19:38Speaker 02

  Um. So actually, the prompts are not too specific. So there's a history here.

• 19:44Scott Tolinski?

  Okay.

• 19:45Speaker 02

  And the history is in the early days, absolutely. Right. We had an approach that we called, for example, cookbooks. Right. Where we're saying, oh, different languages, different language tools have, for example, different like different linters have different errors. And different error messages and so on, right. And then we created these kinds of cookbooks by saying, oh, if you're in a Python file and you use PyLint. And you see the particular kind of errors, here are ways of how you can fix that. And you know what this really means and how you can fix it. And that was quite language specific, right? Sometimes you had to give also hints about syntax and so on. Right. Funny that not funny, right. Just what I said before, an unmaturation curve.

• 20:28Speaker 02

  These kinds of things are way less necessary these days. What we still do clearly, language specific is what we talked about completions context, right? Because that is, I mean, that's clearly language specific, right? What, what? If a language supports type, let's say interfaces, right? Yeah. And then you need to write some code in order to find the right interfaces that you want to include. And so on, right?

• 20:51Speaker 01

  I think also what is kind of semantic, but not language specific, is when it comes to workspace knowledge, right? So the whole thing, how you find the relevant snippet to add to a prompt? That has a very sophisticated indexing infrastructure that is based on tree set of base know, AST know how of of ranges and so on, and then computing embeddings. But that's that's done at the level of grammars, right? It's not very, yeah.

• 21:23Scott Tolinski?

  I, this is all just so exciting to me, but I, I often, I'm like wondering about, I know to we're getting into some different things here, but I, I was more curious about the implications of what this would mean for paid copilot features within VS Code.

• 21:41Speaker 02

  So now, first of all, Copilot today has a free option, right? So you you just can sign up and you're on free, right? There are some limits associated with this, right? How many premium requests you can get, right? To premium models, there is a base model that we actually have in in Copilot, right? Where you can continue, uh, right? There are no limits on those calls and so on. With free, you can also bring your own key, right?

• 22:08Wes Bos?

  Yeah.

• 22:09Speaker 02

  You put it in whatever API key you already have and you pay for it, right? You you can actually use. So will Copilot be? So is there a relationship between you paying for a really advanced Copilot uh functionality and open source? No, no, not really, right? Because the the point is, we still someone has to to pay for the compute.

• 22:37Wes Bos?

  Right, yeah.

• 22:38Speaker 02

  To run the L L Ms and all of this, right? And and that is really unrelated from the open source question, right? So open source means that you can follow along everything that works and happens on the client. And you can see what we send to the servers, right? And what you pay for then in the end, assuming that you pay and you're not on free, right? Is pretty much the compute on on the backend.

• 23:02Wes Bos?

  Okay.

• 23:03Speaker 02

  And like for enterprise, some some additional service like indemnification and these kinds of things.

• 23:09Speaker 01

  And the key thing to keep in mind, right? Bring your own key, this is the infrastructure where you can bring your API key. And then we go to your model directly without going through the GitHub backend service, right? To get the completion. And that's, I think, is a nice compliment that you already have for this story, right?

• 23:26Wes Bos?

  Yeah. And is there any like secret sauce that you put on on the server side of Copilot? And and if so, because I I was chatting with Burke on our live stream a couple of weeks ago. And he's like, yeah, he's like, bring your own key. Most of that you can run through. Like, if you want to bring your own key, you can run it straight to the OpenAI model or Anthropic model or Google model, whatever. But he also said there is a like a fine tune model. Is that is that true? That are for some parts of of Copilot?

• 23:58Speaker 02

  Yeah, uh, there's not just one, there, there are several. Think about it this way. Yes, you can have ring your own key. There are additional services that co-pilot, so GitHub, the GitHub co-pilot service gives us. Right, and Eric made a reference to one of them, like finding relevant source code snippets. So what happens, right? Is GitHub, when I say GitHub, right, uh, the team on the GitHub side is actually indexing all of those repositories that are all open source repositories, uh, you can opt in for, for your private repositories to be indexed and so on, right? But what indexing really means is you, I'm pretty sure you've heard that before, is, uh, the, the code pretty much gets chunked.

• 24:42Speaker 02

  And then embeddings are computed for each of those chunks.

• 24:45Wes Bos?

  Hmm.

• 24:46Speaker 02

  Right, and then when we look for relevant code, for example, what happens is we take that query that the user types in. Right, and then we compute the embeddings for that query. And then we pretty much go and find embeddings that are relatively close to this in the repository. Right, and that, for example, functionality that, that GitHub provides us, right, is that secret sauce. Uh, there are some, some clever things that are being done, right, but it's not, I wouldn't call it, uh, I, I wouldn't call it secret sauce, but it's clearly about, you know, scalability of all of this. Think about this, there, you know. I don't know what the latest number is, but something like five hundred million repositories of, on, on, on GitHub and so on, right?

• 25:30Speaker 02

  So there's, there, there's a massive scaling infrastructure behind that.

• 25:34Wes Bos?

  Yeah.

• 25:35Speaker 02

  Then there are some other use cases where, for example, uh, new models are really good in editing files, for example, right? So Sonnet has the replace, uh, string. Right, so you can search for a certain string, right, find it and then replace it with something else. And the model is very reliable in giving you this. Uh, the new four one models that OpenAI is, they give you something like, uh, it's called the Play Patch. Right, so they really give you this kind of diff patch format to some, right? And they're also very good in doing this. So, but there is a, a lot of other cases where models actually are not good in, in doing something like this.

• 26:18Speaker 02

  Right, and there's another technique where you just go and tell the model, hey, pretty much give me a rewrite of the file. And allied all of the code that is pretty much the same uh of by stuff that has not changed, just say here is existing code, right? And then you pretty much give the original file plus that kind of rewritten file with the allieded spaces, you give this to a model and and now fiddle it all together so that a reasonable outcome comes out, right?

• 26:48Wes Bos?

  Mm hmm.

• 26:48Speaker 02

  And that, for example, is a is a custom model that does that, right?

• 26:52Wes Bos?

  Okay.

• 26:53Speaker 02

  And we do that, we do that uh many times a day, right? So that needs to be cheap to run and so on. And when you want to run it cheap, right, then you need to have custom models that you the smallest possible model that does a really good job at it, right? So that and right, that's for example, and then the completions models are are clearly custom. Right? Um then next edit suggestion, right? That is the tap, tap, tap kind of functionality, right? You you had some particular page, you made a change here, all the subsequent changes that you should make in order to make it consistent again, right? That is a is a custom model.

• 27:33Wes Bos?

  Okay.

• 27:33Speaker 02

  I think this is what you will will see more and more, right? What would really happens is you can take one big model, right? And the big model is really smart at doing things, but it's also really expensive at doing things. And it's slow at doing things, right? And what you want is you wanna make it for the user fast. And for you, as you know, you wanna put little load on your service infrastructure. And so you've got it then pretty much distill and distill, distill rate in multiple steps until you have the smallest possible model for given use cases, right?

• 28:06Wes Bos?

  Okay.

• 28:06Speaker 02

  And it will vary over time, how many of those custom models you have. Right? Can you achieve the same, but just as I said before, to talk to a big model? Yep, right? But then again, quite expensive and slow.

• 28:20Wes Bos?

  Yeah, yeah, you're running that. Like, I often think about that. I'm coding for ten hours a day and it's it's running every single, I don't know how often, but like keystrokes, right? Yeah, right. And I often think like, how are they making money on this? It's gotta be expensive to run this, yeah.

• 28:36Speaker 01

  May I wanna say one point, since you talk about all these backend model and services, right? For the open source announcement, right? So the the backend services, they will not be open sourced, right?

• 28:48Speaker 02

  That's right.

• 28:48Wes Bos?

  Okay.

• 28:49Speaker 01

  But of course, the whole VS Code, all that, I know how we render divs that get streamed in from the model, how we do inline chat, how we do kind of the agent UI and so on. All that, all that is open source, right?

• 29:02Wes Bos?

  Okay.

• 29:03Scott Tolinski?

  I do wonder about the UI bit of it because we've seen kind of the rise of the VS Code fork and experimenting in different UIs for these types of things. How much do those design choices modify your thinking in terms of what the VS Code's interface and how it should relate to working with AI? Or do you have your own experiments? I mean, I'm sure you do, but is everything that comes out into VS Code informed by your own experiences or is there influence coming in from others in the space working in these types of UIs?

• 29:39Speaker 02

  Mark, clearly it's informed by others as well, right? I mean, even today, issues for Copilot, for example, are open, right? There's the VS Code Copilot release repository where, you know, you can look at issues. People make feature proposals there. In the VS Code core repo, right? Many issues actually land directly in the open source VS Code repo. Like we're saying, oh, I do this, right? And you have this button here, but I really like to have this button over here because there are two others who put the button over there.

• 30:15Wes Bos?

  Yeah.

• 30:15Speaker 02

  So from that point of view, clearly, right, there, there's influence comes also from others. But I wouldn't say that this isn't the, this is just like tool development always happen, right? You take good inspiration from others, right? You would be stupid not to do so. But at the same point in time, you need good scrutiny in what you actually really do and what you don't do because you need to maintain your identity. You need to think about the users that are already there, the UI patterns they already do understand. Then also a big difference for us is our extensibility story. So like, for example, we always need to think about something like.

• 30:59Speaker 02

  Oh, if we put a button here, who else wants to put the button here? And then in the end, you have fifteen buttons and then, right, that doesn't work. So how do we make UI scalable, keep it still lean and clean? Don't lose our own design language by still picking smart ideas that that other folks had, right? So I think in AI, exactly the same role supply that always applied to to to to a development, right?

• 31:31Speaker 01

  So actually, that's a concern that I have, right? So we see these forks, right? And I sometimes see they missle it with the spirit of VS code, right? They add the blue button for AI everywhere, right? And probably whatever. Of course, you can do that, but we have a platform mindset, right? We wanna be good for many others. You cannot be monopolistic and say now everything is like that. And that also the other thing I sometimes miss when I look at what the folks are doing, kind of the consistency we have, right? In in we in we in in VS code. And that's one of the concerns I have with the forks. Maybe an example, right?

• 32:09Speaker 01

  Cursor, they have changed the UI totally, yeah, unrelated, that they have the activity bar on top, right? Which I think is a great UI idea, right? Why not, right? In the meantime, we have added this option as well. I have a hard time to understand why didn't they contribute it back, right? Because it's it's USM. So that I I don't understand it because in particular, when we did it and what we noticed, the implementation, they give to the user is not complete, right? So we get batches where you see how many changes you have from for the for the Git view, right? And they don't do that. So that's kind of that hurts me, right? This loss of consistency we have kind of established and we care about when we do a feature.

• 32:53Speaker 01

  Because sometimes the forks, they optimize where they wanna go and forget about the platform. And that's to me the whole long term view we have. We wanna be built to last for a long time and not for just one wave.

• 33:07Wes Bos?

  Makes sense. And I want to talk a little bit about like the the business decision. I know. I know you guys are engineers and whatnot, but if you're comfortable talking about that, like, what does open sourcing GitHub Copilot mean for Cursor and Windsurf? Like, they forked the editor, so you're saying have the rest as well? You know, what does it? What happens there?

• 33:33Speaker 01

  What do you mean by have the rest? Sorry.

• 33:35Wes Bos?

  Like, so they forked VS Code and then add their own AI stuff on top of it. And and now you're saying, well, here's our AI stuff as well. You can have that as well. Like, anybody can have it. Anybody can build upon this as well. Is there like a strategic business reason that Microsoft is doing this?

• 33:56Speaker 01

  Maybe one thing is what we heard from some organizations. They they really don't like closed source IDEs. So by making the other pieces open source, I think we address that need, right? And that, but that's more for this bigger organizations. And of course, then they have a choice between closed source and open source. Yeah.

• 34:17Speaker 02

  You put the disclaimer upfront, right? We're engineers on the project.

• 34:21Wes Bos?

  Yeah.

• 34:22Speaker 02

  A pure business person might actually answer that somewhat differently.

• 34:26Wes Bos?

  Yeah.

• 34:27Speaker 02

  But when I think about this, the the part is really, could they take whatever we have? Right? The whole thing. Could someone take the whole thing, put a new batch on it? And say, this is this is now ours and mark it under their name. Sure. Right? But we also know how pretty much the open source community has reacted for for many years when someone tried to do this with VS Code proper. Right? There was always immediately this question, why would you go some? Why would you do this? Why would you not bring the changes back to VS Code? Why would you not work in in favor of the overall developer ecosystem?

• 35:07Speaker 02

  Right? I think the competition, at least that that's my point of view, right? They'll move more and more to the Services that you actually provide around this. And really what's running on the backends and all of these kinds of things, right? Right now, for example, we have, and I'm purely speculating here, of course, right? But right now we have access all to the best frontier models, right? So OpenAI, Anthropic, Google, all of them give the models out, right? But you could also imagine a world where, where that is not the case, right? Where people actually try to create very, very specific models and that and and keep certain kinds of secret model source just to themselves and to their own AI tools and so on, right?

• 35:52Speaker 02

  So I think again, as it was in many other cases, right? The what really matters, right, is on the servers.

• 36:00Speaker 01

  Okay.

• 36:01Speaker 02

  How that is done. This is right, how you scale, how you provide in the end lasting value to your customers. Right? And what happens on the client? Those things you want to trust, you want to see what is happening there. If you're a big enterprise, you want to see and and pretty much audit those things, right? And make sure everything is is good, right? Like we have discussions, right, with with enterprises that come and saying, yes, so what do you put in your prompt? What really is this, right? And like you, V, right, you ask, so what's the context you put in there? And people are like, whoa, really suspicious about what you put in here and and all of this, right? I was like, yeah, I have a look now, right? Go, look, right?

• 36:44Speaker 02

  So i- it's all there, you can do it, right? You can see it. So I I cannot really tell you what what it does to Curse's business. But it's also their business, not mine. So but I I can, right, think about that or what what I just mentioned, right? I think it moves through the server. That's where where real competition happens. Right? And what we do on the clients again, right? It also informed our decision. I I mentioned that earlier, right? Like the the prompt crafting, all of these kinds of things, right? There there are like more more understood, right?

• 37:17Speaker 01

  Uh huh. Me also know there is kind of some conversions, right? On the totally, as Kai said, the secret sauces on on the server, all these custom models. On the UI side, there is kind of more conversions, right? In live chat, how you do that, how you show the div stream, get the get streamed in and. This was actually, I just wanna, when I, when you, when you talk to these forks, why the fork? One of the reasons they said, well, the VS Code API was not powerful enough to enable the AI innovation they wanted, right?

• 37:49Scott Tolinski?

  Yeah.

• 37:49Speaker 01

  And that's correct. We, we don't allow you to do kind of big UI in using our API. That's totally intentional, right? Because we care about stability and performance. Whereas you could not do a co-completion UI using our extension API, right? That's just not how it is. And this has, of course, a little bit of tension, right? You wanna kind of change deep in the core, but of course, we don't allow that because you want stability and evolution. And, but by now, I think things have converged and evolved, so I find it a missed opportunity that there was no collaboration

• 38:28Scott Tolinski?

  Yeah.

• 38:29Speaker 01

  on that, right? Because we're a small team, the other forks are small teams, and I think most talented people, so we could have maybe something even better for the whole thing, because the whole value is in the backend in the end.

• 38:42Scott Tolinski?

  Yeah, in that regard, was it a surprise when these forks started popping up, or was it something you all had anticipated?

• 38:50Speaker 01

  That was a surprise, right? And it's, of course, no fork for an open source steward like Kindai, or it's not a pleasant experience, right? And when you see about, and of course, you wanna understand why, and of course, one is the API thing. Yeah, but I must want to say, even though the API is not really made for deep UI innovations, what we see is the community innovate a lot on top of our existing APIs. Right? This is a nice AI editors, extensions also. So this always blows me away how much creativity our community has to do innovation on top of our UI. But it's totally true. Deep innovation below it is challenges. Yeah, that's one reason.

• 39:32Speaker 01

  The other thing, of course, is what always made me jealous about them, this, they have, don't have AI bolted on, they have it built in, right? And this is what a fork allowed them to do, which what we didn't have, because we had to restructure our code that all the AI stuff is in this closed source extension, right? And to me and for the developer team, it's a big relief now when we no longer have to do this kind of indirection. We can also go to a really AI-first mindset.

• 40:03Wes Bos?

  And and that means that like other extensions can now like use the AI features as part of their their feature set, you know, if you're making a a totally unrelated to Copilot extension for for something like a a database tool. Is is there APIs or will there be APIs that like, let's say that database tool add-on could then use AI APIs?

• 40:28Speaker 02

  Yeah, there's actually a couple of things here. The the first one is we actually made that available before, right? So there is, for example, in the VS Code API namespace, right? There is an LLM namespace. So you can create, for example, chat participants and then those actually can participate in in a normal chat conversation, right? The tools, for example, that models can can call, right? Those tools are already public APIs, so extensions can contribute tools, MCP servers can contribute tools, right? There is a way you can just, as an extension, just read all of the tools that are available. You can craft your prompt, send it through the model, through our APIs.

• 41:12Speaker 02

  That is actually stuff that that already exists today.

• 41:15Wes Bos?

  Okay.

• 41:16Speaker 02

  Right? And there are also, there's a good ecosystem around this, right? When you look at client, for example, right? They they use some of that API, right? And their dropdown, they can say, yeah, use the models that you get through Copilot, right? But they also do some some other custom pieces, right? But will there be more APIs going forward? Probably. Do we know right now what they will look like? No, not not quite yet.

• 41:44Speaker 01

  But the mindset, right? It's still like in VS Code. In VS Code, you can contribute by writing an extension, which is an extension API. You can contribute by making a pull request if you wanna find an issue or find a problem, wanna improve in AI, I think, a prompter. So this is kind of the same thing. What we now give through this open sourcing, we give kind of the community two things, right? Source code access. If they have a problem with their extension, they can go to the source and look deeper. That's that's one thing. And we give them the tool to make pull requests, right? Which is really make code extension. That's that's the thing, how to think

• 42:24Speaker 01

  about it, right? It's not that we already have extension APIs, as Kai said, and a very important one, right? How we can contribute new tools to our Dr. Chantik flow. And that's already there. And this will continue to be there. But what we see from from VS Code, developers love to have source access when they do extension authoring. Well, at least me, I do that.

• 42:46Wes Bos?

  Totally. I agree. That's I'm excited about it because I don't know, a couple of years ago, there was a bug in Copilot where it was adding like an extra curly bracket at the end of all the tabs. And I was losing my mind over it. And and I was like, if we just had access to this thing, and I couldn't even fathom like, like, how it was working, or what was going on, or how edits get applied. That's all like, I've kind of black box to me. So it's I'm very curious now to actually dive into the code and see how that stuff works.

• 43:20Speaker 01

  I think we should say, right? We announced that we will make it open source, but there is work ahead of us. Because it's different, right? When you open source VS Code, you have a totally closed source. And then we, when you're ready, we through the switch. Right now, what happens is we already have an open source core. And now we have to unit refactor it to make it open to

• 43:45Speaker 01

  To learn these AI features. This means the community will see we make changes in the core and they will ask questions, right? What's going on? You will be surprised, no? They watch every commit, right? And of course, we can do fake commit messages, no? Improve the naming when you really do kind of an improved app, right? So we have work ahead of us to do. So today, we have an announcement, we will do that and we're excited about it. But we have work ahead of us. And I don't know, kind, whether you wanna, since you drive the whole thing, whether you wanna say timeline things? But for me, you have big problems, right? That we have to solve and I'm not a fan of timelines, but if you want, you can.

• 44:28Wes Bos?

  Software engineers are famously great at giving timelines, so let's have it.

• 44:33Speaker 02

  No, the, uh, let me put it this way, the, um, Eric is right, there is a rock ahead of us. There, there are a couple of things to consider when you open source something, right? So there's a legal aspect to this, right? And that you have the right copyrights everywhere, that, you know, uh, that you make sure that this is all properly done, that you don't have, uh, internal references to internal issues and these kinds of things in your source code, right? All, all these kinds of things that are kind of obvious, right? Then there are other considerations about code quality, right? Uh, developers, there's no secret here.

• 45:13Speaker 02

  If no one's watching, maybe your code is not exactly the same quality of if the world is watching, right? So there, people might be concerned about this, right? But we'll make a point here saying code quality is not, is not an, an aspect that we, we care at this moment, right? When we make this open source. But there are other, other bigger things, like what do we do with the internal issues we already have? What do we do with the, like, for example, we invented this whole mechanism of what we call S test. Think about it, this unit test that can deal with the stochastic behavior of models, right? So you don't run the test just once, you run it several times, right?

• 45:53Speaker 02

  Uh, you, you cache responses from L L M and, and so there's very interesting what we did there.

• 45:59Wes Bos?

  That's cool.

• 46:00Speaker 02

  But now there is a lot of-

• 46:02Speaker 01

  S test, S test, Emma. So as a, as a person who worked on unit testing, right? So the big problem is when you have in a test, anything with L L M. Then the test becomes stochastic. And testing, the worst thing in unit testing is flaky tests, right?

• 46:16Speaker 02

  Yes.

• 46:16Speaker 01

  You hate flaky tests, right? But as soon as you have an L L M in there, your test is by definition flaky. Right? And that's what we try to address with what Kai explained there. This is simulation test, right? Which is pretty cool technology, but also that will require work to to package and make available and think about that.

• 46:37Speaker 02

  Make available then the actually test cases. Alright, so how do we get good test cases, right? That is usually from internal kind of flights, right? Where we can actually real look at the at the concrete prompt that was being sent, right? With the concrete source snippets and so on, right? I'm saying, oh, in these particular cases, it failed, right? Then we added it, for example, to our S test suite. Right? Now we've got to clean this all out, right? We cannot make these things available publicly. So there's a lot of work to be done, right? So we'll start pretty much today, right? With the announcement that that work is starting, right? And and then we'll make sure step by step by step, right?

• 47:21Speaker 02

  That we go, I think our main point here is we make the announcement, we don't throw it over the fence. Right? That's not how we work, but we will provide transparency in every single step that we that we take here. Right? So that you know exactly what to expect, right? And you know all of our iterations plans are public. Right? They are, they are even pinned on the issues, right? For for every single month. And GitHub, GitHub repo, there will be entries in there. You can follow along, you can ask questions, right? That's that's how we operate. So we will make sure that there's a fully transparent process. Will you see everything right now?

• 48:02Speaker 02

  No, I just mentioned why why there's things that we need to do first.

• 48:08Wes Bos?

  That's cool. That S test stuff is, that's really neat. I've, I know that there are some other like just generic AI tests that help you answer the question of, is this getting worse? You know, because you always hear that from people being like, oh, the AI is not as good anymore. They, they turned some knobs down and you'll often wonder like, is that just my perception of it? Am I, am I expecting more or is it, is it actually doing it? And this like non deterministic output of AI is, is, is kind of hard to do. So that's really interesting to see that you, you have some sort of testing against the, the AI output.

• 48:48Speaker 01

  Yeah, the life is no longer red, green, right? The life is percentages and you have also that a baseline, as you said, right? It's very important you compare against the baseline because you don't have a red, green anymore. And that's what the S test infrastructure ties to do. And it uses of course, cloud infrastructure for doing that, which is another to do item, how we make that.

• 49:10Wes Bos?

  I do have to ask, just like switching things, because you're in this AI space, obviously. Do you have any opinions on like what future of software engineering looks like with AI? Are we out of a job or like, what are your personal opinions and thoughts there?

• 49:28Speaker 02

  Oh, if he only would knew, huh? No, um.

• 49:33Speaker 02

  I had this conversation just the other day with someone and um, I, I made a statement there where independent of what it will look concretely. I, I think you will not have just people who are just at one particular level of abstraction, right? Really deep, just do right, uh, I don't know, optimization for, for one particular aspect. Maybe a few of them will survive, right? But you could make a good living as a, as a developer by being quite specialized, right? And I don't know if this will remain true, right? What I believe is that. As a developer, you need to be, or what could, what a developer will be in the future,

• 50:17Speaker 02

  or creator, maybe we should just call them creators, right? Will be going across many different levels of abstraction in, in pretty much speed of light, right? So you think about a product idea, you've, you scuff, scuff out something out, you, you run it, fails a little bit, so you then all the way down, right? Just to fix a tiny little issue, then you go up again and think about your customer, right? And think about changes and think about design, you go down again, right? And this, I, I think this flexibility, right? The thinking about your users, I, I think developers will become a little bit more entrepreneurial in their whole way of how they think about this holistic, how they have to approach a problem, right?

• 51:00Speaker 02

  And I think that's, that's where that, I would guess where that really goes, right? Where you live on all of those levels of abstractions very, very quickly. And today you could make a comfortable life to, to kind of just sit here and never go down, or just be a PM and sit up here and never really, you know, go down, go down. So it's like, I think that, that will go away, right? I, I think that the most powerful people will be those that can think in all of these levels of abstractions, pretty much fluently going up and down.

• 51:31Wes Bos?

  Yeah, you're not the first person that said that to me.

• 51:33Speaker 01

  So what's definitely true for me is there will be more creators, right? Given all that, I think there will be more creators, but I also think, if I look, I, I like to extrapolate by looking back and looking forward, right? It's incremental, right? First we got ghost texts, then we got chat, then we got snippets, now we got then multiple file edited, then we got agents, now it, it's, it is an evolution, right? It's not that all of a sudden this is changing, I think. That's why I'm not too nervous, because we can follow along.

• 52:03Scott Tolinski?

  Yeah, it is interesting, even like, I mean, we've noticed, uh, Wes and I are both kind of tinkerers and creators and it, it's enabled me to get into. Man CAD modeling and all kinds of stuff. Wes was hacking his Roomba with it. Yeah.

• 52:18Wes Bos?

  C++ microcontroller for the first time in my life. And then like on the flip side as well, like you get a little higher into the 3D space and trying to figure out what customers want. Like you said, so you're not the first person that I've talked to that said that like you both you both go both deeper and and higher. All right, we got we got five minutes left. Is there anything we we didn't touch upon that you would you would like to get across?

• 52:47Wes Bos?

  What's your favorite model to use? I'm curious.

• 52:50Speaker 02

  Oh, that's a loaded question. It's not loaded. It's relatively straightforward, right? Because and that is clearly a point of time, right? Statement. It's always a point of time statement with all of those models, right?

• 53:04Wes Bos?

  Yeah.

• 53:05Speaker 02

  Right now, Sonnet is is just nicer to use, right? Because it's it's more predictable and the output that it generates. It's more tracking, particularly in agent flows, more tracking, you know, on a particular task. And it also creates code that is nicer. Right? And what I mean by this one is I look at the code that Sonnet proposes and and I like it better than the code that some other models propose, right? I personally didn't have that much exposure yet to Gemini 2.5 Pro, right? I hear really, really good things about it. So the few experiments I personally ran looked kind of good, right?

• 53:47Speaker 02

  But yeah, then there are the more aspects to it, right? So like, for example, speed, right? So do I like a model that is not as great, but it's much, much faster? Yeah, it depends on the data, but I like fast, right? So so it really is. It's quite a fluid field, right? And what I really like is the competition, right? What I like to see is that new Sonnet models is here and everyone like new Sonnet model, then OpenAI comes back and say, oh, here's our new model. Try this one. This is awesome. So then it kind of goes back and forth, right? So.

• 54:24Wes Bos?

  Yeah, everything gets better, faster and cheaper, which is is good for us.

• 54:29Speaker 01

  And you get more choices, right? Bring your own key. You can select a model and switch it. Like things become multi multi model, right? That's what I really think is such a great accomplishment recently.

• 54:39Wes Bos?

  Yeah.

• 54:40Speaker 02

  I mean, because we're talking open source, right? What what is really nice and now when I talked about the models, right? The I really just took the the close weight ones, right? But what is really beautiful and I think will be way more important going forward, right? Is pretty much open source or open weight models, right? And you've seen what DeepSeek had gone and what kind of stir that actually created in the community and so on, right? And now think about it, right? You have VS Code as an open I editor and you have open weight models. And that's really a combination that that's quite powerful, right?

• 55:20Scott Tolinski?

  Yeah, well, um, man, this is just so exciting overall and thank you both for coming on and sharing all this with us. You know, I I think people are going to be very excited about this news in general.

• 55:32Wes Bos?

  Mm hmm.

• 55:33Scott Tolinski?

  Just as you mentioned earlier.

• 55:34Wes Bos?

  VS Code, man, like, I honestly think like, yeah, VS Code is probably. If I were to like, look at my entire career, like VS Code and TypeScript are probably the two most important pieces of software in in my career, which is is absolutely massive. So like, it's like, thank you for for putting it out in the world. It's it's, I think it's made software in the world a much better place in general. I know that people in tech tend to say that, but man, it's it's been huge.

• 56:05Speaker 02

  Yeah, and I mean, we we started with this, right? Thank you for for helping us making VS Code great, right? I I remember the the one feature that we just built for you.

• 56:15Wes Bos?

  What was it?

• 56:18Speaker 02

  Oh, that the I think because you I forgot where it came from, maybe from a tweet or something. Eric might remember it was like, oh, you know, when I present in my courses and so and talk, then people cannot see my cursor. So I need a little bit of a wider cursor.

• 56:33Wes Bos?

  Oh, yeah.

• 56:33Speaker 02

  We made an option to make the cursor wider and that was just for you.

• 56:38Scott Tolinski?

  I love it.

• 56:38Wes Bos?

  Yeah, I was. I appreciate that. I was like, I think I was hacking it with CSS at the time because like my students needed to see where I was with the cursor. So I just made a huge, big, blocky one and threw it in there. So I appreciate that.

• 56:54Speaker 02

  And then we gave you a setting for it.

• 56:55Speaker 01

  And we were tracking you on Twitter, right? We were following you. And as you said, once you had something you don't like, we jumped on. We want to make you a happy customer, right?

• 57:05Wes Bos?

  All right. Well, let's like, while I have you.

• 57:07Scott Tolinski?

  Yeah, what else you got?

• 57:08Wes Bos?

  What else do I I want on this type of thing?

• 57:14Speaker 01

  Now you're convert. Now you're convert, right?

• 57:16Scott Tolinski?

  Yeah.

• 57:17Wes Bos?

  Yes. Yeah, now you don't have to worry about me.

• 57:21Wes Bos?

  That's great. Great. Well, thank you so much for coming on. We appreciate all your time. Really excited about this and and hope to talk to you again sometime.

• 57:29Speaker 02

  Thank you for having us.

• 57:30Speaker 01

  Thank you. Bye bye.