Welcome to the Vergecast, the flagship podcast of pointing an LLM at just a bunch of text files to see what happens. I'm your friend David Pierce and I am sitting here getting ready for the next season of Version History. Version History, if you don't know, is our tech rewatch show about the most interesting good and bad products in history. It's a very fun show. And for this season, I have had to do research that has taken me down rabbit holes about Apple history, like deep into Apple's history and into the history of the monopoly that AT&T had for decades over the phone business in the United States. And that in particular is a story I just frankly knew nothing about. And I found myself reading a bunch of tech history books, which is delightful. First of all, I should read more books. We should probably all read more books. At this moment in time, my information system is just insane. I'm on social media. I'm scrolling through apps. I'm on Reddit. I probably read more words than I ever have. But it's this like discombobulated galaxy of just stuff all the time. And to sit down and just open up a book and stare at it for 3 hours has been like genuinely cathartic in some really interesting ways. So all of this is to say go books is the official stance of the Vergecast in 2026. But that's not what we're here to talk about on this episode. We're going to do two things on this episode. We're going to talk actually a bunch about AI. The first thing we're going to do is talk to Boris Cherny who created Claude Code at Anthropic. Claude Code came out a year ago today, Tuesday, February 24th, as you're hearing this and I think has kind of become the single most important AI product out there. So, we're going to talk to Boris about where it came from, what happened at the end of last year that really made it take off, and where all of this goes from here. I also have a bunch of like product support questions that I'm going to make him answer because I can, because he's coming on the podcast. After that, the Verge's Hayden Field is going to come on and talk to us about how to think about your own interactions with AI, particularly as it pertains to data privacy and security. We talked a bunch about this stuff with OpenClaw and Moltbook a couple of weeks ago, but I really want to get into this idea of like if I'm going to turn one of these things loose on my computer to build software and interact with my apps, how do I think about that as a person in the world with data and privacy and secrets? Reckoning with that feels important. We're going to talk about it. We also have a really fun hotline question about gadget buying in the year 2026 and why it's about to be so complicated. All of that is coming up in just a second, but I have a chapter of this Macintosh book to finish. Insanely Great by Steven Levy. Highly recommend. And I have to go get Claude Code to finish something before Boris gets here. This is the Vergecast. We'll be right back.
All right, we're back. So, for all intents and purposes, we're about a year into the vibe coding experience, and I think vibe coding to me is the most interesting piece of the AI equation right now. I'm continually skeptical of the idea that chatbots are the future of anything. I think there's a lot of interesting technology in a lot of these LLMs. I think agents are a cool idea whose time has not yet come and maybe never will. But the idea that you can use AI to write good code is just true. That thing has found product market fit. And all of the external questions about, you know, the way that these models are trained and the energy that they consume, all of that is real. But the idea that you can just write code by prompting is here and it is real and it is powerful. Let me just give you one example in my own life. So I am constantly switching productivity apps which means I have a bunch of notes in like 10 different apps. This is a terrible system because I can never find anything but I take notes on meetings. I have like interview transcripts. I have all kinds of stuff just sort of scattered around. And over the last couple of days, I've been using Claude Code to pull all of that data out of all of these different apps, put it all into one place in this app, Obsidian, and then actually structure it in a way that makes sense. So, I have without any manual labor or moving stuff around or messy copying and pasting on my own, I have just been able to tell Claude Code, in this case, Cowork, which is a version of Claude Code, where my stuff is and just have it go do the busy work for me. That's powerful and meaningful and a big deal and is a thing that would have taken me a much much longer amount of time to actually do and that's just the tip of the iceberg of what tools like Claude Code promise. So Claude Code launched like I said earlier a year ago today Tuesday as you're hearing this and this felt like a good moment for a variety of reasons to check in on where we are with Claude Code in particular but also with this idea of giving everyone the tools to write software in general. So Boris Cherny who created Claude Code at Anthropic, by accident is probably too strong but certainly not imagining that it would become what it has. He and I talked about what vibe coding means, where it's going to go from here, whether or not there is a future of something like Claude Code that is actually useful and usable for most people, and how we're supposed to feel about the end of people writing code at all. It was a really interesting conversation. I really enjoyed it. Learned a lot about how to think about Claude Code and other things like it in my own life. I think you'll enjoy it, too. Let's get into it. Boris Cherny, welcome to the Vergecast.
Yeah, thanks for having me.
You've talked a lot about kind of the history of Claude Code and where it came from and how you made it and now that it's a year old. I think the thing I'm particularly curious to talk about is your relationship with coding now. One of the things I saw in all of those interviews I've been watching is everybody does the YouTube thing where they like grab the splashy quote at the beginning and do it as sort of the cold open and then they get into the interview and over and over it's you saying I don't write any code anymore. Claude Code does 100% of my coding. And this is like this is a big revelatory statement to have made. And I want to get into what that actually looks like. But over the course of the last year or so as you've been building it, have you undergone basically like a complete reidentification of what it means to be a coder and developer at this point?
It's surprising how little of a change it's actually felt like as someone that writes code. I think part of it might be that in some ways engineers are used to change because our tech stack changes all the time. There's always a new technology. There's always a new framework, a new language. It's just kind of part of the job is always relearning and kind of re- I don't know, it's like every few years there's a new stack and a new language that's popular and so we're just used to kind of, you know, figuring it out and learning the latest thing. In some ways it's felt like a big jump because, you know, the big change over the last year is I don't work with source code anymore, like I don't look at the code of the program as much as I used to. I don't write any of it anymore and that's been kind of a big change. Back when we released Claude Code originally in February, that was like Sonnet 3.5 New or 3. I forgot what terrible name we gave that model. I think it was 3.5 New. We should have called it like 3.6 or something.
Yeah. AI model names not famously great in the industry right now.
It's not our strong suit. Not our strong suit. But so we released it and back then, you know, Claude Code was writing maybe 10% of my code. When we released Sonnet 4 and Opus 4 in May, I think that jumped to maybe like 30% or something. It kind of creeped up over time, but back in November when we launched Opus 4.5, that's when it just suddenly jumped for me from like 50% to 100%. And that was actually very sudden, but it also just felt very natural.
What does that change look like? Like do you just wake up one day and realize, oh, this thing has stopped making mistakes. I don't need to do it anymore.
Yeah. As an engineer, the way that you would code maybe like I don't know, like middle of the year last year is you kind of start the work in an agent and an agent does the first pass but then the code isn't perfect, there's a bunch of stuff that doesn't work so then I have to go in, I have to test the code, I have to open it in a text editor to make some final changes to it. And what I realized around Opus 4.5 is one, Opus is now testing my code. So this was kind of cool, like you know, it's running the tests but also it's able to open the browser and it's able to kind of verify that, you know, the website works correctly. It can click around. If something's off by a few pixels, it'll kind of move it over and fix it. And then the second thing is the code is just really good. So I don't have to open a text editor anymore. I don't have to fiddle with it by hand. And that was actually kind of nice because that means I can move on to the next thing and just write a little bit more code a little faster.
It really does feel like that Claude moment sort of happened overnight. It was like everybody went home for the holidays, got bored, used Claude Code, went, 'Oh my god.' and we were sort of off and running. But it seems like you as the person who pays incredibly close attention to it all the time also had that big a kind of overnight shift in how you think about it. Was it just a big new model all of a sudden had this new capability that no one was expecting it to do this well? Like what accounts for that big a change that quickly?
For the longest time, coding has been a thing that we just want the model to be really good at because, you know, essentially the road to safe AGI, like this model is going to be very, it's going to be intelligent at some point, it's going to be super intelligent. Our job at Anthropic is to make sure that goes well and that it's done in a safe way, so the model doesn't do bad stuff and so, you know, this is kind of aligned with the interests of what the users want and of humanity broadly. And the model is software and the way that it interacts with the world is through tools and through other software that it writes. And so for us, for the longest time, we've had this belief that the way to safe AGI is through coding and then kind of tool use and then computer use. So this kind of increasing capabilities to interact with the world, but it's always mediated through code. So it always goes through code. When you do model training, you try a lot of stuff. There's a lot of experiments. There's a lot of new ideas that people are trying all the time. A lot of times it just doesn't work. But sometimes it does. And, you know, for Opus 4.5, the direction was kind of set early on because we knew where we want to be headed. But it just turned out that a bunch of good ideas worked and there was just a big step change. It was just as surprising for me as it was for everyone else.
One of the things I have been trying to figure out and one of the things we've talked a lot about on this show and at the Vergecast is ultimately who the end user of something like Claude Code is. And I think right now it's fairly clear, right? Especially for a product in the terminal. It is a developer product for developers. Is that fair to say right now?
We designed it as a developer product for developers but even from the earliest days all sorts of non-developers started using it and this was just the craziest surprise but also, you know, the best possible thing that you can see in product is people want to use it so much they jump through hoops to use it.
Yeah, that is definitely a thing we've seen with a lot of these tools. I mean I was playing around with some like OpenClaw and some of the stuff like that and the amount of work you have to do as a just normal lay person to get some of these things up and running is pretty remarkable and yet people are willing to do it. I suspect there are a lot of people who had never heard of their terminal until Claude Code started to happen in their lives.
Yeah. That's right. And, you know, like now all the biggest companies in the world use Claude Code. It's like Spotify, Shopify, like Vamp, Netflix, Nova, Nordis, like Nvidia, Snowflake, Salesforce, everyone uses Claude Code. The small startups use it. But also a thing that we're starting to hear is even at these bigger companies, a lot of people that are not engineers are using Claude Code. And so I think like Ramp just tweeted about this pretty recently that they have a bunch of product managers, data scientists, a lot of people using it. So even at these biggest companies, this is kind of what we're seeing. And this was also by the way like the reason we launched Cowork is we see people using Claude Code for things that are not coding and we're like all right I think we can do better than a terminal for you and so we build a thing that we think they would actually want to use and this is a thing we're still learning about and we're seeing how people actually use it.
Yeah. So talk to me about that early signal a little bit when you start to see people who are not developers who are not traditionally people who would be in an IDE and thinking about code and thinking about the terminal start to use this product.
I remember walking into the office and Brandon who's our data scientist was using Claude Code in a terminal to do data analysis and he had like little charts in the terminal and stuff and I was like this is just crazy like there's no way this is the best way to do it and he was like no it's great and the next day he had like three Claude Codes running at the same time doing like data analysis in parallel and then all the data scientists started using it but I actually still didn't really get it because I thought there's something weird about, you know, maybe people that work at Anthropic, maybe they're very early adopters more willing to try these new tools because, you know, it's like engineers are always the early adopters and, you know, they try a thing and then eventually everyone else tries the thing. But I think by the time that I think now like half of our sales team uses Claude Code every week. I think when that started happening that's when I really started to get it that this is a product that's not just for engineers and we got to make that easier.
So yeah I would think that realization would lead you in one of two directions. One is to say, okay, actually we're giving people access to a developer tool and maybe we should do it in developery ways, right? That maybe have people understanding what the terminal is on their computer is not the worst thing in the world. And if people are willing to go through these hoops to do this thing, maybe we're on to something. Maybe we don't need to sort of radically rethink the UI because people are figuring it out. Or you look at that and say we need to radically rethink the UI because these people are having to jump through these crazy hoops just to do the work that they want to do. Do you have a stance on which one of those is the right reaction?
Yeah. So I mean look like we started in a terminal but pretty quickly we started experimenting with other form factors too. So, you know, we have like IDE extensions for like VS Code, Cursor, JetBrains IDEs. There's we have like iOS and Android apps. I actually do like probably a third of my code on the iOS app nowadays.
Really? I never would have predicted that but you know that's where we are.
We have like a web surface, there's a desktop app so like, you know, the same desktop app that has Cowork also has Claude Code in it so, you know, you can use the exact same Claude Code so we're just like always experimenting with this but yeah it's like the surface is just a little bit different for different kinds of users. So Cowork under the hood is just Claude Code. It's like the same agent SDK. It's, you know, it's an awesome agent and it's the same exact one that's running everywhere. But for people that aren't engineers, we want it to be a little less foot-gunny. Like we don't want people to mess up their system and things like this. So we actually ship like a whole virtual machine. We have deletion protection built in. There's a whole bunch of things that we built for less technical users that engineers would actually find kind of annoying and they wouldn't want in the way versus for engineers, there's something a little bit different about the tool because engineers love to customize everything. If you talk to like two engineers, they're going to use their tools totally differently. There's no two engineers that have the same setup. And so the way that we build Claude Code across every surface across terminal, IDE, desktop, everything is we want it to be the single most customizable dev tool that anyone has used. So it's very very configurable. You can hold it however you want. You can customize it however you want. There's hundreds of ways to configure it. And what's also kind of cool is because Claude Code is Claude Code, you can just ask Claude Code to configure it for you. So you can just be like, you know, change the theme or, you know, like change the setting. It can just do that for you.
See this is one of the things I have really enjoyed about my own experience with Claude Code is there's so much of it that is sort of relearning what's possible in certain ways. Like the idea of asking Claude Code to reskin itself because I don't like the color scheme. It had just never occurred to me. I don't like the color scheme and I would like a different one, but it literally had just never occurred to me to ask this thing that is writing code for me to write that bit of code for me. And I feel like this is kind of why I'm curious about your own relationship with writing code is it just, yes, there are certain things you have to do, but I feel like I don't know, I would think of learning a new coding language as like learning how to play a new kind of instrument, right? Where it's so a lot of the behavior is the same just pointed in new directions with new details and new systems to figure out. But this is like, you know, you used to play the violin and now you're on the soccer team. It's just like it's a completely different way of thinking about how to use your body. Do you know what I mean?
Yeah. Yeah. Yeah. The way I would think about it is like you used to play the violin and now you're conducting the orchestra.
Okay. That's a good way to think about it.
But it's also, yeah. I mean the hardest thing for me is just changing expectations every time a model comes out. It's just so quick, you know, like this thing that just never would have worked for Sonnet 4, Sonnet 3.7, now with Sonnet 4.6, it just works and I just have to constantly relearn this. All the stuff that I would have thought, you know, didn't work, I just assume it'll work at some point.
Do you have like a list somewhere of all the things that are broken that you try every time a new model comes out and just check some things off the list?
Essentially anything that I do by hand.
Yeah. Yeah. So for example like Sonnet 4, Opus 4, they were and even like 4.5 it was okay at this but 4 was like not great at it. If we have like a feedback group, we have this like Slack channel where all the Anthropic employees get feedback about Claude Code. We also have a lot of external feedback channels for customers and GitHub and things like this and before the model is not very good at looking at the feedback channel and deciding what to do and what to fix but now actually a lot of the code that we ship for Claude Code, you know, like Claude Code is 100% written by Claude Code at this point but also I would say maybe 20-30% of that is Claude Code just looking at the feedback group, figuring out the kinds of things people are reporting and then automatically fixing it and this kind of proactivity just would not have been possible with older models but with like with Opus 4.5, Opus 4.6 it actually just started working.
Yeah. What have you learned with Cowork in particular because I would assume like you said there's a different set of person coming to Cowork than to Claude Code with a different set of expectations and a different set of knowledge. Are they using it kind of radically differently and making you rethink this whole system all over again?
You know the most surprising thing, so at this point Claude Code, there was some study that it writes like 4% of all the commits in the world of, you know, like all the code in the world. I think the number is actually quite a bit higher than that because it's not including private code and also our growth has inflected since that study, it's actually going up even faster than before. So I think it's actually quite a bit higher. In the early days though Claude Code did not grow very fast. It was not a hit originally. It took like a few months to catch on because it was just such a new idea. Cowork on the other hand has been a hit immediately. So like as soon as we launched it, it's just been, you know, exponential since. And this is what we like to see because we also we think in exponentials. So I think the biggest thing that's been surprising is just how quickly it's been growing, how quickly people have figured out how to use this.
Why do you think that is? What do you feel like you got right about Cowork?
There was just a pent-up demand. I think that was like the biggest thing like just for something a little more understandable.
Yeah. More understandable.
Like you saw all these people on Twitter that are using like Claude Code to, they were like growing tomato plants, like recovering corrupted photos off of a hard drive. Like someone used it to recover wedding photos. Pietro I think who actually used to work at Anthropic used it to, I think it was like genome analysis, like he got his like genome sequence and then he like Claude Code to look at like, you know, like specific sequences and stuff. Claude Code not intended for medical advice but he did use it for the, there's someone that using it for like for MRIs. So I think just this like this pent-up demand is the single greatest thing that you can see in product because it just means like people are knocking down the door and they're jumping through hoops for, you know, this terminal thing that wasn't really designed for this. So it was pretty obvious I think that it would have been a hit.
One of the most interesting things about Cowork in particular to me has been that the product itself is really focused on sort of busy work I guess is the way I would put it. Like it's you open it up and one of the first things it offers is to organize your screenshots, right? Where it's not it's not build a dashboard of your entire life. Like one of the jokes we always make on the show is that everybody looks at AI tools and the first thing they say is I want to build a daily planner because all of my information is ever, this is like the first idea everybody has about what to build with AI. It's just a thing to tell me what matters in my life. But I think the real truth of like software forever is that this stuff all starts by just sort of solving relatively straightforward relatively simple problems for people. Like I need to do math and so spreadsheets exist, right? Like this is what it is. And I think to me one of the most eye-opening things about Cowork was it just has a bunch of ideas of little things it can do for me that would take me a long time to do on my own that aren't hard. They're just this is a tool that will automate away a bunch of my busy work on my computer. And my sense is that is the kind of thing that has just every single person in the world resonates with the idea of that in a way that strikes me as very powerful. And it's not as open-ended as you can build any kind of software you can imagine. Or you can talk to this chatbot about anything. It's organize your screenshots. And I think that is like a surprisingly powerful bit of product to put in there like that.
Yeah, absolutely. And if you want to build something, you just, you know, hit the code tab in the desktop app and you can go build whatever. But if you want to organize your desktop, like I actually I use Claude Code for a lot of stuff. Like I used it to pay a parking ticket the other day. I was up in Seattle. We went clamming and I used it to purchase a clamming license. So that was pretty awesome. Like I just did something else and it navigated this like actually kind of annoying government website to do it. Someone on the team is using it to pay their taxes right now. So also not financial advice, but it's actually like quite useful for all this different kinds of stuff. This is one of the things that's like also kind of hard to explain to people is like people ask what do I use it for? And my answer is well kind of everything. It's like all the toil, like all the stuff you didn't want to do by hand it can just do so you can do the stuff you actually want to do.
Yeah. So I think okay, let's talk about doing taxes, which I know is just an example off the top of your head, but I think is a useful sort of middle ground of the kinds of stuff that I think about a lot with AI where organize my screenshots is relatively low risk, right? Like the idea of it might delete a thing that I didn't want it to delete, but in general, it's just going to put things in places and delete stuff off of my computer that I don't want. And I think you can get people comfortable with doing things like that on their computers fairly quickly. Have Cowork do my taxes just has naturally more consequences, right? And I think part of, I know a question you get asked a lot and also a thing that I think is tricky with a lot of these tools is it's one thing to have it write code that I can then go check even if I don't, right? The responsibility is back on me to check it and make sure that I understand what it is and code is legible to me as a developer. But if I'm just a person and I'm like, 'Cowork, go do my taxes for me.' How much faith is it reasonable or fair or rational to have in Cowork or Claude Code or any tool to go just execute that entirely on my behalf?
At this point, the tools are not perfect and it's still early, but they are surprisingly good at things that people often expect they would not be good at. And again, it just improves with every model. For something like taxes, I would definitely, I'll definitely double check it. So like have it do the taxes and actually the thing that I would do is say do the taxes, but then triple check your results and just have Cowork do that work for you and then by the time you check it, there's a very high chance it's just going to be pretty good.
Yeah. Actually, to your point about you can have it test itself, that actually I think that there's something very powerful about that too. But part of the reason I bring up taxes is because the last innovation in tax software was that it will scan your W2s for you. I don't, this is, I remember this being a very big deal in my life where I didn't have to type out that I could just upload the PDF of my W2 and it would just pull in all the information. And I remember for a minute it was like okay you have to check that because the scanners, the scanning system isn't perfect. The software won't get it exactly right. But now like I don't remember the last time I double checked the numbers. I just, you just upload the W2, it shows up in the field and you move on with your life. And I wonder, it feels like we are just barreling towards that with all these tools too that it's like there's going to be a beat of, I mean I guess it's like your experience with Claude Code. There's going to be a beat of I need to check its work and then a beat of well I'll spot check it and then we get to I'm just not worried about it anymore. And that's like that's the right end state. It just it doesn't feel like we're quite there yet.
Right. Right. I mean it's like two things that happen at the same time. It's like the model gets better and the product gets better and then also as users we get more comfortable with this thing and like both things kind of happen at the same time. Before we released Cowork I was using it to do all of our project management for the team when I was like first testing it out and I still actually use it for this like every week. So we have a spreadsheet of kind of all the things the team is working on and we ask the team to just like fill out their status every week. So just say like is it on track, is it off track? And so I just have Cowork like ping people on Slack if they haven't filled it out. And so all I do is I'm like, 'Hey Cowork, open the spreadsheet.' And then for anyone that hasn't filled it out, message them on Slack. It'll just do it perfectly. There's actually one person's name that it for some reason can't figure out on Slack. So I have to do that. But otherwise, it just does it. And I was actually like kind of taken aback because I didn't even realize that it would be able to do this. So I would just like experiment with this. Double check until you're comfortable, but I think we'll be there pretty soon.
In a case like that, does it message people on Slack as you or as like a bot?
I asked it to sign its messages as Cowork.
Yeah. And Cowork actually, no, it supports this thing. In Claude Code, we have this idea called CLAUDE.md. It's just like a special file, but essentially it's like all the instructions you want Claude to take into account every time. So Cowork also supports this now. So you can just say like whenever you message people on Slack, sign yourself as, you know, sign the message as like Cowork bot or something. It'll just do that.
Yeah, that's smart. Yeah, I think that kind of there's a little bit of transparency there that I think it's interesting. Like I remember you said in one interview I was watching that Cowork occasionally in the course of doing stuff for you go and tweet on your behalf and that always felt kind of strange.
Yeah. Yeah. Yeah. It's funny actually. Claude Code does this too pretty consistently now. When I'm like debugging something sometimes Claude will be like hey this code is kind of weird let me like look at the history so it'll look at the history of the code in git. Once in a while it sees a really weird change by someone and it'll message that engineer on Slack just to get context. It'll wait on the response and then I've also seen it push back so like the engineer is like yeah I did this change for this reason and then Claude Code is like well I don't think that's a very good reason and I think you actually introduced a bug so let me go ahead and fix that.
How are you thinking about the rest of the UI around this stuff? I think like so much of Claude Code and Cowork are very chat-based still. Does that feel like the right UI to you going forward or is there more work to do there?
We are constantly experimenting with new ideas. I think the UI of the future has not been discovered yet. So we have a lot of experiments in flight. I would expect it to change there. There's going to be a lot of things that we test. The single most important thing is just seeing what people want. And so, like, you know, I'm on Twitter and Threads all day and so is a lot of the team. We just love talking to people. We love getting the feedback because, you know, we have a lot of ideas, but the only way to figure out what the right ideas are are to see what people say and to see what people enjoy.
I agree with you that the UI of the future has not been discovered yet. Do you have a hypothesis at this moment in early 2026 about what it might be?
I don't yet. I don't think we found it to be honest. I think there's a lot of ideas around like proactivity and Claude kind of jumping in when it knows that you're going to need help, but it's kind of hard to get this boundary right. Because you don't want to end up with something like Clippy.
And it speaks to I think the progression you're talking about a little bit from, you know, playing the violin to conducting the orchestra. It's just a different set of tools that are available to you when that's what you're thinking about. But also you mentioned going from basically code to tool use to computer use. Can you just walk me through what that progression looks like as we go through because I think we've heard a lot about agents to the point where I think the word agent essentially means nothing. Agent is just like magic that happens on your computer and it's like sure whatever. But I think you're thinking about this in a much more sort of practical how do we give this thing more powers kind of way. Why does it go code, tool use, computer use?
Yeah. Oh my god. Don't get me started. Like the word, okay, I will get started. The word agent I feel like I feel like everyone just misuses it. Like it has a really specific meaning when you talk about like AI research, when you talk about engineering. So an agent is an LLM that you talk to, but the LLM can use tools. This is the thing that makes it an agent. It's like it can use tools. And so if you think about without tool use, the agent can write code. So let's say you give it a prompt and it can kind of write some HTML or something. And then as a user, you take this and you kind of copy and paste it into like an IDE or something like this. So this is just like the coding capability. And as the model gets smarter, it gets better and better at working with big code bases. But there's still kind of this problem that you hit where at some point you just can't give it all the context it needs. But you know the model actually does know the context that it needs because it's able to search around and it's able to look throughout the entire codebase. It's able to look at Slack. It's able to look at like the history of the code. It's able to do all of this but it's just too much information. Like you wouldn't be able to give it all the information up front. And so the answer is tools. You give the model tools and it can use a tool to look at the code. It can pull in more files. It can look at history. It can do all this stuff. And so this is why tool use is important. It's, you know, the same as a person if you don't have tools like you actually can't do a lot like just with your hands, right? You need like keyboards, you need shovels, you need like if you're cooking in the kitchen, you need a whisk. Like these things are just very, there's not a lot you can do without it. So it's kind of the same thing for a model. And then when you think about computer use, there's just like a lot of things that are kind of hard to interact with just with tools. So if you think about like what can you actually do with a tool on a computer, it's something like MCP or it's an API or it's a command line interface but not everything has that. So you know like if you have like a, I don't know, like this clamming thing I was getting this like clamming license and you know there's no API for that but there's a website and to use the website you want the model to be able to use a browser, you want it to be able to use a computer. And so this is kind of this natural evolution so you start with coding, then you move on to tools and this is the way to interact with the world and so you don't have to spoon feed the model context. It can just use the tools to pull in context and then computers are kind of the last thing because then the model can just use everything.
Okay. Do you think as AI continues to grow and if it sort of takes over all of software and computing the way that a lot of people think it's going to that the computer use part eventually becomes sort of obviated? Like if there were enough tools and enough MCP access and enough of the stuff that you're talking about, does computer use just sort of an elegant hack that gets around the stuff that maybe will exist later and we won't need it?
Early on, in the early days of using the model for coding, people were talking about designing special programming languages to make it so the model can code better, right? And I always thought this was kind of silly because the model can just figure it out, you know? It's not like us where, you know, there's like a programmer that likes Python, there's another one that likes JavaScript and like won't touch Python. The model's not like that. It can just write whatever language. It doesn't care. So I think it's kind of the same thing here. I think over time, the model doesn't care. Whatever tools you give it, it will be able to figure it out and it can use those tools to do, you know, things for you.
Talk to me about how people should think about kind of their own risk profile in giving access to their data and their computer and their files and their photos and whatever to a system like Claude Code or Cowork. I think you have, you know, lots of incentive to tell me that it's totally fine. You can have all this stuff on my computers. We're putting the safeguards in. But how should people think about what it means to give Claude Code access to a folder on my computer even something like that?
Yeah, totally. So I would think about it on a few levels. So the most basic level is like why does Anthropic exist? We exist to make safe AGI. We, you know, initially we have a bunch of founders that, you know, like left a different AI lab and came and started Anthropic.
Ones that people have heard of. Yeah, I'm familiar.
But this is the reason we exist. And there's a lot of corollaries to safety. You know, like the security is actually very important if you want to get safety right. Privacy is very important if you want to get safety right. And all of this stuff we sort of have to do. We're very lucky that we care about safety and so does our most important target customer which is enterprise and companies. You know, there's a lot of like consumers that use Anthropic products. This is awesome and this is something we love to see. We will build for you. But actually like the main market we care about is enterprise as a company and we're very lucky and we picked this market on purpose because we know enterprises care a ton about safety and security and privacy and so we built for them and so like if you look at the product it's actually kind of annoying for me because like if someone has like a Claude Code bug report or something I literally cannot see your data so like I need you to like give me reproduction stuff so I can like reproduce it but I literally can't access the data to see this issue. So there's a lot of like controls like that in place also because we care about safety a lot. There's a lot of work that goes into just making the model inherently more aligned and interpretable. And this is also just very important and also very related to this. And yeah, I mean the final thing is like there's just a lot of stuff that we build into the product like Cowork can only see the folders that you give it access to. It cannot see anything else on your computer. We put an entire virtual machine in Cowork to make sure it's a really hard security boundary to, you know, so it can't access stuff that you don't give it access to. The biggest thing to worry about is attacks like prompt injection anything like this that would kind of exfiltrate your data. We have a lot of protections in place for this and Opus 4.6 is just the most aligned model that we've ever built for prompt injection in particular and there's also a lot of like runtime classifiers and kind of safeguard that we put in place for this but this is the biggest thing that I will think about is as you have Cowork as you have Claude Code interact with the internet just be thoughtful about what websites it is using and it will ask you for permission but it's a thing to keep an eye on because this is not a solved problem yet it's quite good but it's not yet solved.
That's a good one. All right, give me one like normal human Cowork activity that lots of people should do that you've either done or building or have heard from people that not everybody might expect that they should go do and then I'm going to let you go.
Oh, a normal human. Okay. One is just like responding to email. Just like open my Gmail, look at the top three things I should respond to, draft responses. So you can do that quite well. A second one that I do is just like canceling subscriptions. So I actually use it to cancel like I canceled like a TV thing that I wasn't watching.
That's the most unbelievably annoying thing to do. I'm going to make Claude Code unsubscribe to all of my email newsletters that I don't want anymore. This is going to work for me.
Yeah. Yeah. I love this like dual track. Like you can use it for your write the emails and also unsubscribe for email.
Yeah, exactly. I just never want to look at my email ever again. If Claude can make that happen, we will have accomplished something.
All right, Boris. Thank you so much. I really appreciate you doing this.
Yeah. Yeah. Thanks, David.
We'll be right back.
Support for the show comes from Bilt Rewards. Unfortunately, no one will pat you on the back for paying your rent on time every month. But you can earn rewards on your rent with Bilt, which is like a pat on the back, but for your wallet. Bilt is the loyalty program for renters and rewards you monthly with points and exclusive benefits in your neighborhood. With Bilt, every time you pay your rent, you earn points that you can put towards flights, hotels, Lyft rides, Amazon.com purchases, and so much more. And now, it's not just for renters. Bilt members can earn points on mortgage payments, too. Being a Bilt member also unlocks exclusive benefits with more than 45,000 restaurants, fitness studios, pharmacies, and other neighborhood partners. Join the loyalty program for renters at joinbilt.com/verge. That's j-o-i-n-b-i-l-t.com/verge. Make sure to use our URL so they know we sent you.
Support for the show comes from Shopify. Starting a new business, it could be a lonely endeavor, especially in the beginning. And if you're just starting out, it's more important than ever to make sure you have the right tools at hand. If your business includes e-commerce, a great next step is to try Shopify. Shopify is the commerce platform that millions of businesses around the world rely on to sell their products online. You can get started with your own design studio with hundreds of ready-to-use templates. Shopify helps you build a beautiful online store that matches your brand's style. If you're asking yourself, 'What if people haven't heard about my brand?' Shopify helps you find your customers with easy-to-run email and social media campaigns. And if you get stuck, Shopify is always around to share advice with their award-winning 24/7 customer support. It's time to turn those what-ifs into. With Shopify today, you can sign up for your $1 per month trial and start selling today at shopify.com/vergecast. Go to shopify.com/vergecast. That's shopify.com/vergecast.
Support for the show comes from Upwork. When you run your own business, the to-do list can feel endless. Well, that might be the sign that it's time to grow your team. And for that, there's Upwork. Thousands of growing businesses already trust Upwork to hire flexible, high-quality freelance talent for everything from one-off projects to ongoing support. Upwork also cuts down operational hassle by handling things like contracts and payments in one place, so you can spend more time running your business. Upwork is free to use, but if you decide to upgrade to Upwork Business Plus, you'll get access to the top 1% of talent on Upwork. And with AI-powered shortlisting, you'll get
Matched to the right freelancer in under 6 hours. You can visit upwork.com right now and post your job for free. That's upwork.com to connect with top talent ready to help your business grow. That's upwork.com.
All right, we're back. Hayden Field, Verge senior AI reporter is here. Hi, Hayden.
So, you were here recently and we were talking about Moltbook and OpenClaw and all of the insane things that you can do on your computer with AI tools and AI agents. And we talked a little bit about privacy and kind of how to think about whether or not you should engage with these tools and install them and what kind of data you should give to them. And I've realized I've been having kind of varying levels of existential crises about AI tools. Starting with like I had a real experience with OpenClaw where I downloaded the installer for OpenClaw onto my computer. I have a Mac Mini and I have a MacBook Air and I was on the MacBook Air and I downloaded OpenClaw and I was like, I'm going to use this, get into it, see what it's like, try the whole thing out. And I got literally halfway through the install process and was like, this is so stupid. Like this computer is full of all the information I care about in the world and all of the stuff that I know about everyone that I know, including like important confidential information as a journalist. Giving this unknowable AI agent access to this is insane. So that's one level. But then even like I mostly use Claude for AI stuff and one thing Claude really wants you to do is connect your Gmail and connect your Google calendar. And I've had moments of being like, is this an irresponsible thing to do? Like am I being stupid giving Claude access to my email? So, what I want to do as best we can is just try to think through sort of a how to think about your data and AI framework. Does that seem reasonable?
Perfect. I've been asking the same questions.
Okay. So, let's just start kind of big picture. I want to get into some sort of nitty-gritty. I literally want you to tell me if I should give Claude access to my Gmail, but we'll get to that. You've been reporting on this a lot and talking to experts and trying to think through this for yourself. Do you have kind of big picture guidance on just how people should be thinking about this stuff?
