Anthropic’s hunt for the next Claude Code

With Fin, we've built the number one AI agent for customer service. It solves up to 90% of queries for businesses, tops all the performance benchmarks and the G2 leaderboards, and it comes with a million dollar guarantee. Check it out at Fin.AI. How many third-party vendors does your company use? 20, 200, thanks to AI. Someone on your team probably added three more this week. And your security lead has no idea. Traditional third-party risk management can't keep up. Vanta gives you continuous coverage across every vendor, automatically, so you actually know what's in your stack, and what to do about it. AI on, risk off. Vanta.com/TPRM You're never going to be able to fully replicate the feeling of two or three of us against the world. We have X amount of dollars, and if we don't manage to make something work, we're going to have to go wind it down, return the money to investors. That's like you're Kevin back in the day. Exactly. What's Anthropics' next big product after Claude Code? This week on the show, we've got Mike Krieger, the co-founder of Instagram, who now leads the team at Anthropic working on its moonshot bets. What are the dark days of my Anthropic 10? You have to name me at three, five, eight, two, and I can tell you why I waited up that way. Mike and I recorded this conversation in person on the sidelines of Anthropics' recent Claude Code conference in San Francisco, where the company announced a big new compute deal with Elon Musk. Once you guys apparently go into space, you see what? Exactly, right? Yeah, finding new sources of unexpected sources of compute. We get into what Mike is working on now, and Anthropics' fierce competition with OpenAI, and the human work that Mike thinks will continue to matter, even as AI gets better and better. This is Access. Mike, it's good to see you here at Claude Code NSF. I was thinking back to our last conversation, and you were more recent in your labs 10-year, but it's been a few months now. Yeah, it's been almost four months now. For almost four months now. For people who don't understand labs that I actually like to start there, because it's a pretty unique setup. We were talking about it when I was at the office a couple of months ago. What is it? What is the mission of labs in Anthropic? Yeah, in a nutshell, the way I think about labs. I would call this labs V2, and we can go into what labs V1 did and what labs V2 is meant to do. But I think of it as doing two things. One is closing the gap between what Claude is theoretically capable of, and how that actually shows up for people in their everyday lives. What products or prototypes or projects do you build that illustrate how you could unlock more of that potential, like close more of that gap, unpack what that means. And the second piece is where the advanced scouts figuring out where the model is going to need to go in the future to meet different needs for people. So a successful labs project might look like a prototype that we build and then say, "You know what, the models actually aren't good enough to do this yet. Let's put this on the shelf and we'll revisit it next model or release. Or we'll use that as an evaluation for future model development and basically iterate from there." So versus a product lab at a pure product company where maybe the success criteria is, "Did you ship a product or not?" There's other ways in which this labs initiative can inform the direction of where Anthropic is going. And labs has produced some hits, right? It's Claude code is one. MCP, what are the other skills? Ancient skills is another big piece from labs as well. And then maybe I'll talk about one other one, which was one we didn't ship at the time, but actually was one of those research and forming ones, which was computer use. So I joined Anthropic May 2024 next week as my two-year anniversary, we call it. Is it anniversary? Yeah, everything's ants. Okay. I resisted at first, now I'm just surprised. We don't dog food, we eat at food. I'm like, "Two-year anniversary." And when we started, we were spinning up labs and one of the first products that was pitched was, "Hey, why don't we try having Claude use a computer?" I do computer use. This was in the Sonnet, what era? This was Sonnet 35. So the first model I was sort of part of launching was Sonnet 35. That was my third week. Like we joke that Anthropic doesn't do onboarding projects. They just stick like a very hard project with you. And mine was like, "Hey, great. We've got like, launch a model. That's a model." And actually 35 is a really interesting model because it's one of the first ones that really unlocked some coding use cases, not quite full-agent to coding, but you know, the beginnings of that. So we put Sonnet 35 and we built a computer use product around it. And it had a lot of problems. Like it was too slow at doing the computer use piece. It was not accurate enough. Like it couldn't see very well. And so it would see the screen and say, "Oh, I need to click on that button." And then doing click on something else. That was actually very helpful to have built that sort of not quite working harness. Because then when Sonnet 3, at the time we call it 35V2, we can get into that. That was one of the dark days of my Anthropic 10. They would name me at 35V2 and I could tell you why I waited up that way. And then we've wrecked it to 36. 36, still not quite good enough. Some sense of life. And then 37, I remember this day really vividly. So it was in New York on a trip to visit the New York team. And I got a ping like, "Hey, we think this actually, that thing that we had built in labs, at that point it was nine months old, is showing real signs of life with Sonnet 3.7. And we think it's time to actually open up computers as a capability that we talk about externally." That was nine months of trying out new models every couple of months and trying it in the harness. So even though at that point labs had kind of let that project go by the wayside, it still was very useful in being an evaluation set for how computer use is evolving in the models as well. When you joined, I mean, you were a chief product officer when you first joined. And I remember going, "Huh, Mike, who, you know, co-founded Instagram, who I think of as a very consumer, forward founder, is joining the Enterprise AI company." Yeah. And we may have even talked about it at the time, but I was just like, that's an interesting bet. In hindsight, a correct bet. And you obviously, the timing was incredible. I'm curious, you know, from joining as CPO, where you're overseeing all of products, and product in AI is kind of a nebulous term and it changes quite a bit, right? Yeah. How did you get to Forish five-ish months ago going to labs? Where my understanding is you're, and I see, do you actually manage any of these? Yeah, I have never actually entered our performance review cycle. And that's what you wanted to get out of performance reviews. Exactly. I went into the portal, you know, where you get your list of all the reviews you're going to have to write. And it was like, you've to write your own review. Yeah. Review your manager review. Then that's it. Now Cloud does the review. Cloud actually does help with reviews, which is helpful. You know, I don't write all of it for you, but it's useful in even being like, what did I do the last six months? I think the company's been through phases and it's in different parts of it. I think it's been more or less suited to like what truly drives me. So when I joined all of product engineering was maybe 30 people, maybe split evenly across product and engineering. Like product engineering. So we had an engineering team like working on research infrastructure and some scalability pieces, but there's like core people working on building at the time, Clot AI, and what we called the API, we didn't call it the Cloud Platform at the time. Maybe 30 people, maybe 35, but very, very small. Still felt very much like in the beginning. And there was still so much, I still think there is, but there was still much at the time, shaping of what the product even meant. You know, at the time, Clot AI had no projects, no artifacts. It was basically a list of conversations that you had with Clot with effectively no frills on top. So it felt like joining a startup that was trying to figure out, you know, it had tailwinds for sure, you know, at the time I joined the kind of opus on it in Haikut 3 series had come out. That was the first series where Anthropic had put out something that was sort of at least frontier adjacent or close to the frontier. So much left to do in terms of like what is this product going to be? And the reason that, you know, even though my background was more on consumer, the reason I got very excited was the time between Instagram and Anthropic, one of the things that Kevin, my Instagram covener, I did, has spent a lot of time doing investing. And we had a whole sort of set of investments that we were looking at in terms of what is the future of work and how is work going to get done. And Anthropic felt very much like it could help unlock a lot of that kind of that theme, which is what happens when you have a really intelligent assistant helping you. And I don't think I even forced how transformative it would be at the time. You were like, oh, this interesting little AI company. This may help with my investing. Right. This might help with some of the, you know, some of the themes that we've been thinking about. But of course, it transformed it much more. So that was phase one where a tiny team, you know, you could count the number of initiatives you had going on one hand. Then we can talk to some of the different chapters in depth, but you fast forward it to an end of last year. And the product team is hundreds of people. We have portfolio of projects. And a lot of the work is deploying, understanding, customer facing, management, layers, all of these things that just learned over time. There are people that love it and thrive it. And I'm super respectful of those. And it's one of mine where I have a great coach and she calls it either zone of competence. It's like the things you are good at. And like, you're fine. Like you're probably doing well, but it's not the thing that ultimately like feels you and drives you. And that's a good, that's almost the most dangerous place to be because you could keep doing it for a very, very long time, but it is not your place of maximum sort of like fire. And so I entered Q4 last year and started having cars as your Danielle. I like this, you know, we've grown up already. We've we've we've speed run a lot of the growth that usually would take five years, you know. So it only been about two years or-- >> Yeah, it's like Rospind decent. >> Yeah, Rospind, you know, and the team growth and the product portfolio, I was like, I think I want to go and do another new company. And she was like, well, you know, is it about leaving Anthropic? Is it about shifting what you want to do? And I was like, I love the company. Like the people here are amazing. I love the technology, the mission, et cetera. And so that was also when we were spinning LOWS back up. Because LOWS we want us so successful that all the projects graduated and left nobody behind. And so we kind of wound it down and had it kind of on ice. So let's spin LOWS back up. And I'm going to go back to being a builder. And it's, you know, everybody who sees me in and out of work is like, like, you look so happy. >> Some of your colleagues were telling me that earlier today. >> Yeah, it's really nice. >> My SMICS have the best time. >> Yeah, and you know, it's, I still, I am my own harshest critic. So I'm always every day, I'm like, how do I like, you know, what can we do? What can we build? What are we proving out here? So it's definitely not easy, but it is definitely much more aligned with like the things that truly drive me. >> We don't have to spend too much time in this, but I am fascinated by tech companies doing these kind of moonshot, zero to one labs things. I mean, Alphabet's probably the greatest example, but I mean, history is littered with ease. And some of them work, some of them don't. V1 of labs in Anthropics certainly worked. If you look at the products. >> Yeah. >> I imagine that puts a lot of pressure on you right now to go, okay, a cloud code is like the thing that I have to live up to. >> Right, well, it's funny because one of the, like, sort of internal emissions data is like, oh, we want to find the next cloud code, which was already a high bar when we start labs V2, but since then it's like growing so much higher. So I think there's a few things that you need to do. Like, and I think it's helpful that I've had the start of experience a couple of times because you're never going to be able to fully replicate the feeling of there's two or three of us against the world. We have X amount of dollars. And if we don't manage to make something work, we're going to have to go wind it down, return the money to investors. >> That's like you and Kevin back in there. >> Exactly. >> Yeah. >> And like, I remember that existential sort of question, I've lived in every single day, like this doesn't work. Like, you know, I'm not going to go, we have this independence, I'm not going to go to work on this thing that I really care about. That's hard to manufacture inside a company, unless you want to create some very complicated structure. And I've seen that as well. >> Yeah. >> Before we kick started labs V2, I went through and studied as many of these sort of company labs initiatives. And you had ones where they get equity in the thing that they create. >> Or, you know, and all of them that still does. >> Yeah, and they're all trying to patch around the sense of it. The thing that we have found works well, one because I think Anthropic is attracted just a lot of really high agency, ambitious, but also mission line people, is that the thing that, to watch out for is less, make sure that people are incentivized to like do their best work or, you know, put, you know, everything they can into the project. And it's more, how do you make it not just a comfortable place for okay ideas to get like pulled out for months at a time. And so the way we've gone about doing this is actually shortening the timeline. So labs V1, we are on effectively four to six week for view cadences where a project would be sort of funded. And then they would have about six weeks to prove some sort of things. Here we actually run two weeks, and every two weeks, every project goes in front of the whole lab's leadership. And it's not like a murder board type meeting. It's usually, you know, a reasonable meeting. The most of the project is really going off the rails. But we really hold every project. And we truly do it to this question of, what have you learned the last two weeks? And have we learned enough? Like sometimes you've learned it off in the project, you say that project was great. It did what it needed to do. And we don't actually need to spend another two weeks on it. And right now, things are moving so fast. And you can build so quickly with these models that the opportunity cost of letting a project language for an extra four weeks is actually quite high. And that's, I think the best thing we've done with labs V2 is better sort of turn the crank so that we're at least measuring our learning if not the external impact every two weeks. And most of the people in labs are former founders. A lot of them are former founders. Yeah. So that's a criteria you have. For, we have sort of two rolls in labs. There's the, we call them all bets, 'cause they're all, you know, pretty high variance, right? So we have bets, and then we have bet leads. So, you know, if you're the kind of spiritual, directly responsible individual, those tend to be founders almost exclusively. And then we have just people on labs. And those might just be builders that were maybe the first employee at a startup have seen that zero to one, or just love to build one person to join. She was early at a, her last startup, and then, you know, early to labs. So she's got that sort of, I can span the, you know, stack. I can do whatever needs to get done. And that's as important, I think, as the, you know, just pure founders, you also need that complimentary like founding team energy too. My coach, Elk, can be here, but he sent me some questions to ask you. And one of them was, is this founder supergroup, essentially, with labs that you formed? Is this a new model for company structure in general, especially with AI, or is this a thing that you think is unique to a company like Anthropic? - I, it's a really good question. I think there is going to be much more, many more teams like this, where a small group with the models, the models are not perfect. So you still need, you know, founders with taste and direction. And there's, you know, I spent, there's a yesterday example, I spent about two hours debating this lab's project on working on, like what the shape of, you know, in this case, like what multiplayer looks like within this product. That was time that was really valuable and very human. It was all like in a room, us hashing stuff out. And then I had cloud work on all the ideas, they synchronously for the next, like 12 hours, right? But to have that taste of what are the things we need to debate at front, what are the things that we should just go and implement, right? Because the models will make a lot of decisions for you if you don't specify them. So I think that's still a useful, very important role that like a founder type can play. But once you get that right structure, you can do a lot, like I was looking, I did a little retrospective of what we'd build and lab so far, and it's been four months. And there were projects that we sort of conceived of in January and like shipped to testing in February and wound down in March or conceived of in February and are now cloud design, right? Like there's all these different things that have already happened and the velocity is really, really high. And I think more companies are going to ship to this model of smaller teams, give them autonomy, have a person that feels accountable and sort of can drive that. And then don't overstaff it to begin with. We've really learned that with that was read too. Like you don't want to add five people to a project up front. You want to start with two or three, much more like a startup. - This is the fun part where I try to get to tease out for you as much as possible about these projects. So in March, when we met last, you told me that you were applying cloud to running for a really long time, this kind of long horizon problem. And you were leading that project, are you still leading that project? - No, so that's a great one where we actually put that project on ice, but the learnings graduated into outcomes, which we announced today, and managed agents. So that whole idea of rubric-based sort of autonomy for cloud to be able to perform a task towards an end or a goal rather than a sort of just an individual prompt that you gave it very spiritually based on the things that we had been exploring and the labs. And so that, I was mapping out exactly that project and where all the tentals became, you know, and it's a couple of the two themes that you saw on stage this morning that were kind of direct derivatives of that labs project. One was outcomes. And the other one in Boris's demo, you had the cloud code actually verify its work by taking screenshots and testing out what it was doing. That was a big part of what we were pushing on with that project because it's one thing for cloud to look at its code and say, "Looks good to me." It's another one for it to fully explore. And we, in the project I was leading, we were even looking at things like having cloud-tank videos of all the work that had done and then have it scrubbed through its own videos and say, "Oh, that animation is broken." And it wouldn't have been able to tell just from pure screenshot verification. So yeah, that was one thing. When we chatted in March was in development, we've since basically put on ice. It still runs, so internally we still have facilities for some demos, but it's mostly been useful as a sort of source of upstream energy. The other one was cloud choosing its own form was I think the way you put it and then introspecting and being like, for this next conversation instead of a command line interface, I'm going to be a website. Was that cloud design or is that something else? - So there's a lot of spiritual, it is something else. Although we're the kind of question we're pulling on now is like, cloud design is an instance of effectively like agent plus canvas, right? And you can imagine a lot more of those. And even within cloud design, you know, I've done everything from writing technical specs. My new favorite way of writing technical documentation is actually in cloud design, 'cause you can have it visualize like, how's the information flow work and how could it work differently and actually just watch it evolve? And that's actually might probably like number to use case behind slides, which I use cloud design all the time for. But you can imagine other formats that could live in that canvas as well. And so that's one labsie thing that we're pulling on is like, like cloud design, but for even more sort of types or applications and things like that. And that's I think an exciting thing again. Just more generalized productivity software, is that the idea? - I think that's like, yeah, that's definitely a theme to pull on, which is like, and productivity software that hopefully can be like pretty custom to you. I think it's an interesting trend to think about. What are the other, just kind of white spaces that you're thinking about right now, NAI? I mean, I think as the models improve their ability to be useful partners in things like life sciences, I think is really interesting. I've started to see more and more, there was a great thread on X of somebody that did their full genome sequencing at home. And actually, he's now a professional professionalized it, so he'll come over to your house and do it for you. I'm very interested in this because I'm like a, you know, self-knowledge geek, so I'm like, wanna figure out how to do this. And what I've told from people that like know the space is that even the gap between models three, four months ago and something like Opus 47 and being actually useful sort of parsers of, you know, genetic data and like finding inferences or ingesting your labs like, it's now actually useful versus like, oh yeah, cute, like yes, that is the thing that a doctor would have said if he looked at it. Or, you know, we already knew from heuristics. That's one white space, interesting area. Like I'm very interested in that whole area of like personalized medicine and where those things are gonna evolve to. I feel like we're at the cost there. You know, and I feel like there's still so much overhang, you know, what we call like the gap between like what's now possible with that. Well, there's all these startups popping up like, you know, super human or super power and row and like all these companies that are function health. I mean, that are doing stuff like this. Yeah, exactly. And we'd like in January when we did our, our like initial cloud for healthcare. One of the things we did is we partnered with functions so you can bring your lab results into cloud and do that. I started using a service this week called SupCo for supplements. So you've put all your supplements in. It, it knows how they interact and then it kind of gives you a recommendation. Like you don't need to take all eight of these. Oh, that's actually really helpful. You can get it down to four. And now the next step which they don't have, but like I could imagine somebody doing is and you know, based on your last run of your, you know, labs or based on your genetics data, we think you might be a super responder to this or that. So that whole area like there's so much interesting. And that's even just on like call those that like optimization side think about all of the sort of, like areas that are like currently underserved by medicine either because they don't have like local access to good healthcare and like we can help supplement that. So a lot of interesting white space there that I think is super interesting. The other area that is funny, I was on stage maybe a year ago at a consumer AI conference that four runner put together and I was talking to, you know, founders like what, what is the going to be the consumer AI break out case beyond chat endless assistance? And I saw haven't seen it, you know, maybe health feels like it's not quite in the theme of what I'm describing, you know, and like AI powered dating feels like there's maybe something there, but there's kind of a natural - Oh, we're gonna say a cloud dating service? - Yeah, I don't think we'll see a cloud dating service, but maybe a, you know, downstream cloud, you know, one of our API customers might create something like that. - Okay. - I don't think we'll build that one in-house. But yeah, like I think it's still interesting white space around can AI actually be useful in helping us either understand ourselves or our world or our communities in a way that is connecting rather than distancing. And as you said, I've been getting interested in civic engagement and how you get representative groups together for civic debate. It's not an area, I'm an expert in it, but I've gotten interested in it. And that's where, you know, AI could be helpful in sourcing a representative group rather than trying to make decisions for you, for example. Like keep the humans in there, but make sure that we're listening to the right voices for you. - Yeah, I was wondering if you still have the social media bug in you or not? - Less the social media bug, you know? And it's like probably the most interesting issue, idea in that has been something that they actually explored in Sora that that was interesting was describe the algorithm you want and you will like degenerate it for you. - That's coming everywhere now. I mean, that's in threads, X has it to a degree. - Yeah, I think that's gonna be standardized. - I agree. - Spotify. - But I think is, yeah, I use that in Spotify. - Yeah, Spotify, I create these really cursed playlists where my daughter and I can't agree until this end too. We just make a Spotify playlist that is the intersection of what you like. So we end up with like pavement plus frozen music. It's so good. - Oh wow. - Okay, I like that. - Spotify, the Spotify DJ has some thoughts. - That's great. - But yeah, I think that that, I think is one way in which it's personalizing us, you know, an idea I was kicking around more recently. I don't think this is quite lab shaped, but maybe somebody should build it is, can, you know, can AI be a useful sort of like, filter of all the noise out there. I've started using dispatch and co-work for this reason. Like I have, otherwise I am a like, incurable news junkie and I will like check every news site and like read a lot and there's value to that, but at some point I'm like, am I reading the same stories over again. So I have my like, now kind of channeling artifact my previous start, I've like my daily brief of the places I tend to check. And often I'll click through and read those stories. But it's a useful sort of, I don't feel like I need to go check these 12 websites, you know, first thing in the morning. It can kind of give me the trends and I can go deeper on what's, what's, you know. - As long as you're so reading the sources newsletter, I'm like, you can synthesize that. - Exactly, that goes straight to the embassy. - Yeah, yeah, yeah. I mean, I did actually read, I don't know if it was a leak on X or what, that you guys are thinking about something more proactive. I don't know, maybe it's orbit or something where you may do that, which you're talking about, more proactively, bring things into, I don't know if it's co-work or another part of the interface, but for the more consumer audience. Do you think that's an opportunity? - Yeah, I think that proactive sort of, we, anytime you see your power users create those kind of use cases, it's an interesting question to ask, like what would it mean to build more of that built in? And that for sure is, you know, the scheduled daily briefing or, you know, proactive, you know, monitoring of either, you know, we are very slack, heavy shopper and topic. So sort of proactive slack monitoring or email. I've started doing this as well, like where my morning cloud routine is the news. When I just talked about it, I'll say I have it scanned, superhuman, which is what you're saying. I'll launch to MCP. So now it'll go through and it's great, because it knows about the split. So it kind of, okay, you really should look at this one and like some of these are like things you could read later. And it's actually when I check first before I open my inbox. Oh, and then like there's like some shopping thing that I'm looking at where I'm like waiting for the thing to pop up and like rather than like obsessively check every morning, I'm like, the cloud can go. - Is this a sneaker drop? - I should use it first day. It's probably not even fast enough for sneaker drops. You need to use it fast enough. Or it's all the rabbit holes, you know? And there's always that balance. You don't want to remove the joy of, it's nice it's fun to get lost in internet rabbit hole every once in a while, but maybe do it a little bit more consciously rather than the thing that you're just like noodling on in the background. - Yeah, I am a co-work power user. I use it every day. I use it to work on this interview actually. And I use it to write, I use it for everything. And it's been pretty transformational. And it's something that I've like personally gotten my wife on and my friends on. And I do feel like it's still early. Like I think most people are probably not doing what I'm doing. And co-work is still fairly, I don't know, from a product perspective, it's a little bit divorced still from chat, classic cloud chat and cloud code. These are three things that I don't think really long term will be three things. - I agree with that. - And the thing some of your colleagues were telling me when I was in the office a couple months ago was like, maybe co-work is the front end interface for everything, eventually. And co-work was a labs thing before you got to labs, right? So I'm, but you were a CPO in that ship. So I'm curious, what was the insight with co-work when you guys ship that? Like was there a moment where you were like, we have to abstract away the coding for normies and that's an opportunity or wasn't something else? - It was, so I give Dario a lot of credit here. He, you know, I thought because such a diverse set of things that we're thinking about at a given time, like the research side, the products side, the commercial side, compute policy, societal impact that it's not a typical company where like the product and the go-to-market are the two things that matter. And that's like what the CPO and just spending a lot of time on. So Dario's attention kind of shifts between different areas depending on what's most urgent. But shortly before co-work ship, he was like, look, I've been seeing this interesting trend of people using cloud code for personal use cases. It's very hard to do that for most people. They don't have to get a terminal. That feels like, and it's like, what would it mean to build cloud code for everything else? And that was a very useful, and I think it's like how he operates. He wasn't like, here's the mockup or here's the like, probably, that was your problem. - Yeah, okay. - It was like, that's a product he felt like going, go and solve that. And it actually took a combination of like two, of course, that the many people were involved here, but I think of like two personalities that really like combined really well. One was Felix who was the creator, he was like one of the main electron maintainers, like knows desktop software, at least like desktop web software really, really well, and have been like, newly on this problem of like, how do you enable people to do work on their computers? And then Boris, who obviously like deep cloud code knowledge. And once we put the two of them together, plus their teams, you know, working together, it was only a couple of weeks before you were able to ship co-work, because not because all the work happened in those tweaks, but because a lot of thinking had been happening both on the cloud code side and the desktop side around, how do we enable this use case for more than just coders? Because the models are powerful, agentic engines, and how do we give that engine to more people? But I completely agree with you that right now, it feels like it's less that we're shipping our org chart, because, you know, it kind of feels that way a little bit. But we are shipping our harness strategy or something, like in a way that I don't think makes sense to most people as well. I think it's also a bad product experience because like my co-work thread can't sink to my cloud mobile app. Yes, so like I couldn't take the interview prep doc, I had been spending time and co-work with me on my phone. Right, and some of these projects or products naturally suggest their next evolution. Like I'm a huge dispatch power user, so dispatch lets you access your co-work remotely, but you gotta run your computer, right? So your next logical thing is like, well, this would be great if I didn't have to leave my computer on all the time, so how can we evolve it in that direction? Whereas, Cloud Code has been on that journey, but you know, a year or maybe six months ahead of it, where Cloud Code remote, you know, I use all the time where I'm, you know, out on the go and I can kick off the coding task. And by the time I'm back, it's often put up a pull request. So you can imagine similar things for code. But I agree, we're definitely in the, we like a lot of, you know, things proliferate, which is great for innovation, but now I think there's even for like context and continuity value, the fact that you had a conversation here versus here, shouldn't be something you should ever think of in my wife is all the time it's like, I don't remember it. Did I, was that a coding thing? Did I have to chat or was it going to work? Like that feels like a broken abstraction that we need to fix. - And it makes a lot of sense when I'd spend time with the OpenAI folks and their super app strategy and what they're going out with Codex because it is the end state of kind of what you guys got to first with Co-Work. It's unifying all of that. And I'm wondering if A, you think that super app approach is the right one for AI. And if that puts more pressure on you guys. 'Cause I think OpenAI has realized that this is a huge opportunity to bring coding, the power of coding to people who aren't coders. And they're going at it very fast. And Codex is good. The new model's good and Codex is doing well for them. So I'm curious, you know, Claude had that huge surge with coding and is still considered, you know, widely I think leading in code. But Codex is really nip and it's real seals, isn't it? - I mean, it's gotten, I always had the most interesting times that Instagram, when we had like a snap, for example, like I say, you know, interesting competitor in the same space, but with a different approach in some ways. And you know, in a fast developing market, you end up like having some ideas that you borrow from them. Some ideas they borrow from you. But I think-- - So answer the snap in this analogy. - What's that? I don't know which is, which is, wait, I mean, the market dynamics are different, but at least like we're-- - Market caps are also different. - Yeah, a little different. Yeah, and actually the personalities of the companies are very, very different. But at least in terms of, you know, the way I always thought about it was, I don't know that every competitive situation resolves at this, but at least that one felt like it did, which is we each had our own strengths, and we each saw the future in sort of, not exactly the same way, but directionally similar ways. And it was a really interesting question of like, how each company evolved to get to those different places. For Instagram, for example, we were known as like, the really precious once a week photo, and we wanted to get to a place where people were more free to share, and that's where, you know, being able to adopt stories really made a difference for us. And then I think for Snap, it was, they wanted more of the influencer celebrity piece as well, because that, you know, those grow interest in there, and so like, how managing both as well. Maybe an overestimated metaphor, but I think about for us, like, it is really interesting, I think, to have opening it in the space as well. And I think for us, it's the two things I think are most important. One is what we were just talking about, is like reconciling and making sure that like our product makes sense, you know, whether it's on the web or whether it's on the desktop. I don't think it needs to be a single app or super app, but at least that you're sort of building blocks, I'll talk to each other and all makes sense together. And then two is continue to close that gap, because even coworker, any of the products right now, the gap between, even inside Anthropic, I was just talking to one of our recruiters, he is like the most extreme cowork power user. He's actually our top cowork user, who's not a coder at the company. And you see his workflows and they're amazing, but that is not how most of the recruiting team operates, even at Anthropic, right? So that is, I think, whether the shape is super app or not, like that is the goal, is can you, again, affect labs? Like that's probably what we need to do, is can we create the shapes, the form factors that actually make it so either that one recruiters knowledge spreads really quickly and it's very easy for everybody else to adopt that workflow, or for everybody to get to that level of sort of confidence in using cloud. >> I was hoping to hear more about coworker today. I mean, I know it's a developer conference, but I'm curious, how is cowork growing relative to cloud codes are growing faster, anything on that? >> I think index two, I don't know relative to like overall coding and coding is growing really, really fast, but definitely on the trajectory that cloud code was on or faster, which is really exciting. I feel like every couple of days, one of our slacks that is more like a productivity-oriented will show some new milestone of where cowork is. And it's very exciting to see that grow. I think personally, very exciting, because I always wanted us to reach beyond kind of pure code in terms of like impact. So the growth has been really, really good. And we talk to enterprises and they'll say, great, I adopted cowork, and I'm seeing that phenomenon. I've some people really get it, some people still need a lot of handholding, different departments are using it differently. It's a good problem to have because you know the product is capable of the right things in the right sort of shape and with the right person, but there's definitely still a delta there. I think the remaining, it'll accelerate growth a lot further when we can actually feel like for most people, it doesn't take a lot of ramp up to get to that level of, like AI is actually really helping me. - Yeah. (gentle music) - AI is transforming customer service. It's real and it works. And with Finn, we've built the number one AI agent for customer service. We're seeing lots of cases where it's solving up to 90% of real queries for real businesses. This includes the real world complex stuff, like issuing a refund or counseling in order. And we also see it when Finn goes up against competitors. It's top of all the performance benchmarks, top of the G2 leaderboard, and if you're not happy, we'll refund you up to a million dollars, which I think says it all. Check it out for yourself at fin.ai. Support for this show comes from Odo. Running a business is hard enough. So why make it harder? With a dozen different apps that don't talk to each other. Introducing Odo. It's the only business software you'll ever need. It's an all in one fully integrated platform that makes your work easier. CRM, accounting, inventory, e-commerce, and more. And the best part? Odo replaces multiple expensive platforms for a fraction of the cost. That's why over thousands of businesses have made the switch. So why not you? Try Odo for free at odu.com. That's odo.com. (soft music) Speaking of competition, this was a conversation you and I had, God, I wanna say, year and a half ago at a conference was competing with customers. And this was before you guys had launched Cloud Code, the product. I think this was, and we were talking about, well, you know, you have some of these, like, cursor was a giant customer, like, how is that gonna work? And you had to make that call. And you don't have to do that as CPO anymore, I assume. But as not as not the CPO anymore. But how does that these days? And we've danced around it, but there was cloud design and the Figma controversy, and you were on the Figma board, and you had to step down and Dylan has said nice things about you, and it didn't seem like it was necessarily you. But that was an example where I think the market and everyone broadly looks at what's happening, and goes, oh man, Anthropics, just like coming for every key vertical. And that was, the Figma one felt a little dicey. It's definitely more complicated than it was when we had the cloud, the cloud AI, I came in and I was like, well, this is not really competing with anything. It's like a brand new form factor. And we're still, I think, careful and thoughtful about the products that we build. And I think it's only when there is something, the value that we are bringing by shipping something, 'cause we have amazing customers on the cloud platform, right? And a lot of people inside the company who are really invested in these customers, sort of success as well, is are we showing some direction that they think we think the industry could go? And hopefully it's like a rising tide where a lot of companies could have adopted that way if thinking Cloud Code being a great example where the attention wasn't in the terminal, it was more in the editor, right? And it's really shifted how almost the entire industry has gone. Even though, yes, there were people who were thinking about the terminal beforehand, right? Like that, I think shipping Cloud Code and the way that we did and giving the attention that we did did that as well. And I think that's my hope for any of these products not that they become the only product in their space that would be a tragedy. I like the diversity of products to exist. But instead that they are, you know, one, they make sense within our product portfolio. So, you know, the fact that you've got all your MCPs connected and in Cloud AI should hopefully make these other products better, for example. And two, that they showed the way forward where, you know, with Cloud Design, for example, it was being really agent first and giving the agent like a lot of sort of control over what it outputs. It leads to a really specific product experience. Like for me, it's really useful for things like, you know, slides where I can kind of think along with the agent but also be producing alongside. You don't see it as cannibalizing figment? I think there are different use cases. Like I, like there is production and there is like refinement and that there is collaboration and that like is like real figment that I love those folks, right? I think that like that product is really well tuned for that use case. And then there is what I tend to use Cloud Design for because I'm not a full-time designer, you know? So I'm doing more either production of, you know, a, you know, visual communication, kind of artifact, lowercase AI artifact, or some exploration of interactivity where the purpose is not the polish or production. It's the sort of early mock up feel of it. I was mocking something up in our iOS app in Cloud Design and the pixels looked quite different than the things that we would eventually ship but it pointed the way at what I thought we should be able to do in that space. And so I think they'll both, I mean, use Cloud Design and Figma as an example but in any of these, like they're all continuing to evolve and again, I think good ideas from us will hopefully make their way in other places and vice versa and I think it's more around we're all together figuring out what shape of product we need to build to make the most of the agents. I was having a good conversation with one of our researchers this week and they were saying, like the more you constrain the agents behavior, try to be over specific on it, the less sort of naturally emergent magic is gonna happen in there as well. And like that was, I mean, Cloud Code, I think, channel that, even co-work channel that a little bit versus Cloud AI where it was more like you are chatting, you are doing these really specific things. And I think we'll build products of that shape and that thing that that primarily does is have more people rock that and build really great experiences into their own products, like great, 'cause hopefully those are Cloud Power and how the cloud remains. best or one of the best at being a great sort of agentech underpinning for these sort of next generation AI products. But the headline is, it's definitely more complicated. But I think the same principle as we try to apply then, which is being thoughtful about where we enter. And then making sure that the building blocks are still available on the platform, like manage agents, like the things that you can now build on top of manage agents are as powerful as anything that we can build inside the house. And we're not, like, other than for safety reasons, like, not releasing a model because, you know, it gives us an edge. That's been a very principled approach that we've taken. And then we're only going to ship a fully featured product if we think it says something new. So you and Dylan are good. I love Dylan. I-- You know, we-- He's a former guest, so I got to just make sure-- I mean, I love Dylan, and I like-- I have some extra spectra what they're doing. This was another Ellis question on this topic. Everyone, especially in the early stage communities, debating the viability of, I think, consumer startups more right now, especially with just, like, what is Anthropocanidunex? What's the next cloud design? Or what is up an AI going to do next? And Enterprise feels a little safer. Dario was talking today about, you know, he thinks there'll be a billion dollar one person company this year. I bet that would be an Enterprise company. But what do you think about, yeah, the state of consumer startups? And if you were yourself, when you were starting Instagram with Kevin, and you were looking at what's happening, what would you have done differently now? Because the world's changed quite a bit. And it seems like a harder time to be doing this. I think a radically oversimplified version of consumer breakouts is there needs to be some new capability or form factor. For us, it was the camera phone, right? And nobody calls it the camera phone anymore, because that's just what phones do. They have cameras. Or maybe it was more rich media like the TikToks and Reels of the world, because there was now the ability to stream, right? And then some distribution mechanism that lets you break out. And for us, it was the fact that it was a little bit wild west all like you could share Instagram photos to Facebook and Twitter and Tumblr and Posturist, just to date how-- Posturist, Posturist, Posturist, Posturist, Posturist, Posturist. Posturist, Posturist, Posturist, Posturist. And all of those allowed you to freely link back to your product, so you could create that cycle such a different time these days, right? And so when I talk to founders, like Matt, who founded Lockit, they have found ways of like, all right, you're in the conversation on TikTok. And therefore, that leads you to some interesting growth moments. I think it's a much harder-- like it's a far less predictable sort of distribution ecosystem than we were working with. And so I think more than any AI player, I think that's actually probably the biggest challenge right now is how do you break out when you have-- before maybe you were competing with Facebook and maybe their time spent, I'm kind of making this up, was like 15 minutes a day. Now you're competing with TikTok, which I bet their time spent might be over an hour a day, which is frankly kind of crazy. But it is the way it is. You're just trying to just place some of that. So if you have those two components, on the former side, maybe it's around some novel interaction with some AI-powered being, or the opposite, where it's AI helping you get outside and touching grass and having more of those real-world interactions, whatever that is. But you still don't have to solve that second piece. And if you think about where a lot of attention is shifting these days, it is to these chat agents. And I don't think anybody is really figured out what it means to-- what does it mean to go viral on chat? It's kind of a nonsensical question today. Maybe there's something there where we launch MCP apps as an open standard. Maybe the Sony App Store kind of distribution potentially. And you can bring that in. And maybe that will be the first-- is the first MCP app to go viral in something consumery? That would be interesting. Right. That's a novel distribution mechanism. I was talking to somebody-- Oh, it was somebody who works at a nonprofit explorer. And we featured them in a-- we didn't article about how MCP is reneged for a new thing. And he's like, oh, it was really interesting seeing constitutes viral growth for a nonprofit explorer thing. We're just like, future numbers. But he was like, I checked the connector list. And we were in the top five next Gmail and Slack for a couple of days. And he's like, I got a ton of attention. So maybe I'm thinking out loud here. And this is very much exploring thoughts with you, which is fun. But maybe we're not on land on the perfect answer. But maybe that is a thing that we'll start seeing is can you provide enough utility within the frame of a cloud or a chat GPT or a Gemini? And then figure out how that spreads in some interesting way. I love to see that. I think that would be interesting. You did a pod with a Dan Shipper from every recently. And I was listening to it. And you were talking about using cloud to rebuild bourbon, which for people who don't know, was the app that predated Instagram that you guys turned into Instagram. And it did it in like a few minutes. And I was wondering, A, how did you feel emotionally at that experience saying, an AI, I just like build that thing that you and Kevin toiled on for, I don't know how long, in a few minutes. And also, if 2010 Mike had access to cloud today, would Instagram still happen? Yeah, that's a really good question. Because I would imagine the AI would have maybe not gotten there by itself. Yeah, I think that was my reflection having built it. It was one I knew exactly what I told it to build. And actually, and I told Dan this was really funny. Because it overbuilt, it actually built filters into bourbon, which bourbon did not have. We launched that with Instagram because bourbon was totally web based. And the time there wasn't anything like WebGL to do, filter in, right? So it's pretty interesting to see what it came up with even in that I think so much of product going back to our conversation on the founder teams inside Anthropic. So much of the work is still asking the hard questions, putting in front of people. I love both love and hate the feeling of getting smacked in the face the first time you put out a product in somebody and they're like, what? I have no idea what this is. Or I'm so confused, right? My first job out of school is a UX researcher. We can wake up, bring people into the lab. I still love that. You bring people in. And the fact that it was built using cloud or handwritten doesn't matter to the end user. The digital care is the product useful and usable. And it's a great, some kind of moment of delight. That's so hard. Like, cloud is not going to solve for that. So I mean, if we had cloud back in 2010, for sure, there were ways in which we were slower at delivering on things than I would have wanted us to be. Especially when the thing to build became really clear, right? We launched, we knew we wanted to build app mentions. And that was probably a week of wiring it up from the user interface, fighting with the text layout engine at the time, like persisting on the server. Could have done that faster and delivered value to users quicker. But I don't think that the journey from Instagram from Bourbon to Instagram is something that cloud would have changed much of other than in some of the coding sports that we did. But zero to one coding, at least the time I didn't feel like often like that was the limiting factors, the thinking and the exploration process as well. And so I think the hard things are still hard, I guess, is maybe how I would put it. And a fear or a concern would be, if the LLM provides too much out of the box in terms of decisions that are being made, does it actually prevent you from finding the orthogonal shape? Now, you can use it to your advantage and say, great, I wanted to generate three alternatives and then see the edge of these fields good. And I've done that before too. But yeah, I think it does definitely does not absolve you of making hard product decisions. And in fact, it's important that you do so. Because it's more important. Yeah. I feel that in my work, which is not even just building software, but just building media, is I could give myself over to Cloudmore. And I could just let the skill file hopefully like remind it of what I like and all that stuff. But then I do that and then I really read it. And I'm like, and then I'm talking to it like I would talk to a reporter when I was in Ettergard in an isroom. Yeah. And I feel like that's more important now as it gets more capable because I don't want to lose my intuition in the process. So I feel like a lot of people maybe are with how coding has taken off or starting to maybe give it into little too much to it. And I hope that we kind of shift back towards that. I think the human instinct is-- I like the reporter editor piece. Unless the reporter has really gone off the rails. I imagine-- What happened most of the-- And I did in the day. Most of the time, the editor is like, OK, this is like 90% close to what I want. And maybe with some edits, I'll get to 95. Or maybe it's like 80 to 85 versus if you're seeking it from the beginning, you might try to get closer to 100%. Not to these things. They're perfectly measurable. I have that experience building even with the labs where I've learned now that it's much better if I sort of hash things out with Claude and much more of a conversation before it writes a line of code and then say, all right, this feels good. We've collaborated effectively on a spec and now go and implement the spec versus if I just give it the high-level feature. It will build the feature. It will work. All the verification stuff for building will mean that it functionally works. And then I'll look at it and be like, I would have done that a little bit differently. But now it's done. And I don't like that feeling of it being-- So a hollow is not quite. So moving to a place where somebody put it really well internally, which is our whole job when coding with Claude is expressing the North Star with clarity and then helping make sure Claude converges on that North Star efficiently and effectively. And that is the role of a guide/in some ways manager or the managers, at least in software tend to be more on the human development side. So it's really architect. There's not a great word I've found for it. But it's that sort of Sherpa of project and Claude towards that end versus expecting [BLANK_AUDIO] one, it's just gonna do it out of the box, or two, that you don't have a hand early and actually shaping that North Star. - Yeah, I mean, there's so many interesting parallels with Instagram and I think I remember you guys had suffered with a lot early and Facebook help with was compute and like serving the app, right? And your servers, you didn't have the servers you needed and there was a moment where like you had to do something. And that deal came together very fast. And now it's funny like over a decade later, like compute is still the problem. And I imagine, you know, I know you're not working on compute directly, but I imagine compute is a thing you have to think about a lot internally. In a way that you probably didn't envision when you decided to join a little AI lab two years ago? - It was a maybe two sort of vignettes that really illustrated for me. One was and showing like, you know, I've learned a lot in two years, but when three five launched, you know, day of even we're so much smaller than we are now, but day of like people were adopting it, we were starting to get like near like the, you know, like fully maxing out our, like whatever chips we had allocated to the time, and I remember asking the infrared team like, great, so what do we do when we're out? Do we just like add capacity? Because in my world like from the Instagram days, like unless you were doing some very specialized hardware on AWS, there was more hardware out there to do. And they're like, no, no, no, when we run out of this, that's like, that's it. Like that, those are all the GPUs fully allocated. And like, yes, we're working on getting more. And I was like, oh, this is a very different environment. Because, you know, they, you know, you can't just, those naive, right? You can't just like hit new instance and do that. And so that's, it's you guys apparently going to space or seeing what exactly, right? But that was a very quick lesson learned. I learned that it turned out as that immediately. But yeah, we think a lot about, you know, even within a given product, you don't want to sacrifice intelligence. So how do you deliver as much of that as possible without being sort of wasteful on computer, being thoughtful about what you do asynchronously or being thoughtful about when you need the biggest model versus small. So it is a, it is a consideration. And I think it's a healthy one for a bunch of reasons. But I think it also brings us closer to the, I mean, you were talking about customers before, you know, they're operating in the world where they're buying tokens effectively and then reselling it in some, you know, sort of product. And I think we're very aligned with them in a lot of ways in wanting to make sure that that's a healthy ecosystem for people. So you've seen a bigger shift towards token-based, sort of a usage-based pricing that has been a shift over the last year. And I remember looking at, you know, the ecosystem a year ago and saying, look, our customers want to charge more, or they want to, they want to deliver more value and they're kind of constrained by their current pricing models. And now so many more of them are able to actually deliver on intelligence or at least make it customizable, right, where you can say, great, I'm trying to optimize for cost of many use, like, let's run to your model, but know that I'm going to get this value. Or I really want to get this done. I'm going to use fast mode. I'm going to burn tokens, but it's because I have like that knob that I can tune. But yeah, compute definitely way more of a consideration here than IG. But in a similar way, like, Instagram felt like, and probably was, if you do the math, like, exponential growth, at least. Yeah, Gargoyle said you guys had like a ADX Q1 in terms of cloud code. I don't know if Instagram had ADX Q4. No, I think we definitely-- I mean, other than maybe week one, right? Yeah, from 100 to 100,000 V1. But yeah, we had steady, very good percentage growth, but nothing like this. We can land it here. I mean, you've been very early at Anthropic. I mean, I say early, it's been two years, but that is early in AI world. And I think the culture of Anthropic is very fascinating. And it's something that I've been learning more as I spend more time with the people in the company. And I think people look at the headlines. They look at mythos. They look at Darjald Dario talks about, maybe 50% of jobs will go away. And a lot of people are like, are there just fear of mongering or there's tactics here that are about regulatory capture, all these things. The culture is very unique, I find. And I'm curious to hear from you. If you could identify a disconnect between how a lot of people externally perceive what you guys do and how people internally believe about what you're doing, is there something there that you could put your finger on? It's a hard thing to convince people of, because I can say all day, we are a profoundly transparent company. When we say things, we are saying what we think. And Darjald probably more so than any person I've ever met, speaks his mind, not in a calculated, oh, this will help me fundraise. And I've just gotten to know him and seen him operate both internally externally. And it is not the motivating force behind the communications or the rollout strategy. Of course, it's one thing to say that. It's not a thing to live that. But I think if it's a helpful illustration, even internally, this year has been great from commercial growth, great from usage growth, et cetera. But I think the reason the company has stayed grounded is that the point of anthropic is not to build a large commercial enterprise. The point of anthropic is to do our best to nudge the world towards a good future where they are. And once you see everything through that lens, it explains all the downstream decisions. Now, I can't ask people not to be cynical. It's outside. So ultimately, we'll have to continue to hopefully demonstrate to our actions that we're not trying to fear a monger, or we're not trying to hold things back, because we think it'll be perceived as more frontier. I guess maybe I'll illustrate one thing, which is the internal belief is we should be able to release a mythos class model safely. And the fact that we haven't yet is bad. It's not-- we're not proud of it, right? It's not-- if we do it right, then all of the positive use cases can be had as well. We talked about life sciences earlier. There's definitely some interesting things that could happen in mythos there. I built a lot of software using mythos internally. It's really good at it. There's already no cybersecurity risk for your internal software. Exactly. Well, actually, we figured out how to do it as safely as possible. But that is how we see the world. We want to give people these capabilities. We don't want to hold them back. And we're working really hard so that that's not the case. So it'll be a-- prove it. I think that's fair enough. But when I see people, they're just trying to inflate their next valuation, or they're trying to make themselves sound cool, or they don't really believe this. That is the extent that you believe me. That is not the motivating force internally. But again, I think it's a-- our journey is to show that over time. OK, lastly, when people come to you, especially because you work in Anthropic, and they say, Mike, I'm worried about my job. I'm worried about my family member, my kid, my grandparent, their economic security. What do you tell them? I don't tell them not to be worried, because I think that they're going to be big shifts. We don't know exactly how fast. You can disagree on the timeline. But it is already shifting and happening. And I get actually a whole class of evil that I've gotten recently is parents of college graduates. People I know, either socially or professionally, who have kids who are graduating. And what do we do? What do we do? And I think the thing that I tell people, which is what I believe is that it is not one company or one branch of government is going to solve this. It's going to take the kind of societal level conversation. If there's anything that we've been trying to do, it has been to try to spark that conversation. It's going to get rid of fear of memory sometimes. Whether that's everything ranging from figuring out like different tax structures, rescaling-- it's actually going to take all of these things. And one of the things that you've seen from us is being more specific about some of those policy proposals might look like. This is a long answer that I would tell somebody. But it's a complicated question. But yeah, I think what I tell folks is you're not alone. This is a shared kind of complicated thing. I think the things that will still remain human and ineffable and important are still relationships and curiosity and creativity and the ability to organize people towards an end. That's still going to be really, really important. I don't see AI replacing that anytime soon. To the extent that that's a skill that you can pick up or at least develop or at least continue to nurture, I think that's really, really important. And then also not get as things will continue to shift. So even if in the current moment with all the uncertainty, maybe our friend's kid doesn't land in exactly the job that they wanted, things will continue to shift. Nothing is set in stone. And I think if people remain curious and actively exploring what the frontier looks like, they might then either be part of creating a whole new category of jobs or progressing to a different place in their own companies. I think the landscape is still going to continue to shift. So not treating the current moment as uncertain as-- and in some cases, challenging as it is is a static thing that will lock people in. All right. Michael, let's you get back to the lab. Thanks for doing this. Great to see you. All right, thanks. Thanks. Don't go anywhere after the break. Ellis is joining me to recap my conversation with Mike and talk about how old we are. Support for the show comes from Core Weave. Everywhere you look, AI is expanding what we thought was possible, and at the center of it all is Core Weave. Medical research and diagnosis, education, complex visual effects for movies, science and technology breakthroughs. Core Weave powers AI pioneers around the world with purpose-built tech, building what's never been built before. Core Weave is the essential cloud for AI, ready for anything, ready for AI. To learn more about how Core Weave powers the world's best Go to coreweave.com/ready for anything. Support for the show comes from ReTool. Too many companies run critical operations on duct tape spreadsheets, slack workflows, and whatever else they could cobble together. Not because they want to, but because building internal tools means weeks of waiting on someone else's backlog. That's where ReTool comes in. Build custom internal tools just by describing what you need. Build something like "Build me a revenue dashboard on our Salesforce data" and "Retool actually builds it" on your company's data in your cloud with Enterprise Security built in. Go to retool.com/vox. We all need to retool how we build software. And we're back, surprise. It's me, the hamburger helper here with Alex Heath, to discuss an event I was not at for a variety of. Extemporaneous, extenuating, spontaneous circumstances that may or may not involve founder clients in two-year-olds. But I'm very jealous that you got to go. How was your trip? It was good. I feel bad for you that you can't do the Burbank Tessfo Day Trip every week like I do. It's something that I wish on everyone just kidding. But no, this was great. It was great to go. This was the Anthropic Second Developer Conference that they've ever done. And the scale of it maps to how Anthropic has grown. So I didn't go last year, but it was only a couple hundred people. And Anthropic has now obviously become like the fastest growing company of all time. You know, value around a trillion dollars. And this was like 2,000 people with three satellite events in like different cities. And I think they'll probably go even bigger next year. So it's actually kind of cool to just see this culture of this company. The events are this really cool physical manifestation of how these companies are changing and evolving, especially in tech. I mean, gosh, I just finished with therapy 30 minutes ago. So now I could borrow some language and say, can we just pause and sit with that for a second? How far Anthropic has come in just a year? Oh, this is good. So you should always do therapy before you record. Yeah. Because it's an absolute 180 is and not, I mean, gosh, I feel like OpenAI was just absolutely on a saleable in terms of momentum, product. And now all these companies are kind of seemingly trying to focus, bring back maybe some of the apps that built over the last year that were distracting them from gains as the market enters turmoil so much has changed. I feel like with divide, there's been a vibe shift, one they say. There has. And Anthropic is causing the turmoil. So I think they feel pretty good. But it was also kind of interesting. I saw a lot of chatter on social media about people like the day before seeing the cloud with code conference signage going up in SF going like, wait, this is happening. And people also didn't know that this was happening. And it was also reflected in the reporters who were there. There were the typical, you know, the tech reporters that I see all the time at these events. But I was amazed at all of the publications that did not have people there. And it was clearly just like not on their radar that, you know, maybe the hottest company in the world. Certainly, the most influential tech company right now was doing this big event with their co-founders and, you know, Dario and Daniela were there and did a chat and all of that. And the headlines poked through because of the Elon thing where they announced that they were working with Elon on compute. But other than that, I don't know if this would have really poked through into the mainstream, like I'm sure Anthropic hoped it would have. Most importantly, where do all the reporters sit now? I've always been a goody two shoes sitting at the front of the class. So back when I was reporting, I used to sit at the front right next to our old friend, Josh Constin from TechCrunch. Where do you sit these days? They put you in a press room, you know, to kind of toil away away from the normal attendees, right? Because part of that is they also don't want you interacting with the non-press attendees as much as possible. So you were more harder for the students. The creators don't have any hard questions. Don't poison their minds. But no, it was very nice. The Anthropic folks did a great job. The food was actually quite good, which it usually is not at these events. So I give them props there and they did have reserve seats for us in the keynote, which I always appreciate. It's nice to still get something as a member of the media these days. One of the other reasons I missed it, I was thinking back to my interview, which we can bring up on screen for those watching on video with Mike Krieger back in 2013, I think. Oh, wow, you beat him. Kind of a crazy amount of time ago, 13 years ago. 2013? Yes, they just got bought. Yeah, just acquired by fart book. And he has a quote about the last year being crazy because Instagram grew from 30 million users to 100 million users. Just incredible. And it is incredible. Yeah. As Instagram grows, he goes, we got to think about the world here. We got to optimize for a lot of phones, not everybody has an iPhone 5. It's just like, wow, so much has changed. Man, of course you beat me because he and I started before we were recording reminiscing. And the last time I did something on camera with him for like a podcast type thing was in 2018 and it was for the New York office opening of Instagram. And it was the year before Kevin Sister and Mike, the two co-founders, very dramatically. I'm sure you remember that dramatically left Facebook, right? Over like the power clash was Zuck. If you had told both of us where we'd be in 2018, I mean, Mike, you're a little bit more accomplished, but you know, it was definitely, definitely wild to think back to that. So this is funny. I'm actually looking at the URL for the article right now. It was almost exactly 13 years ago. It was May 13th. And when this episode airs, it's going to be May 14th. Yeah. It would have been, that would have been a funny Kuhnke link. It makes you feel old, doesn't it? Yeah. And that's the segue into my next question. Well, it wasn't really a question, but now I have a question. So I know that Mike, since he's been back at Anthropic and just recently is kind of, you know, let go of the reins for the chief product officer. Now it's kind of back to an IC role. I feel like that's always kind of the dream for a lot of engineers, but I was also reading back in this, in this interview I did with him that he was talking about waking up around 2pm and coding until 6am. Is he still in that mode? Or is this a different type of IC mode now being the world's most famous co-founder, Emeritus? Right. I mean, Mike is looking very healthy and well-rested. So I would be shocked if he was doing that still. But who knows, man, a lot can happen these days. Like with modern advances in health technology, you can do a lot. It's a very different job, though, isn't it? I mean, I just appreciate like he's always been a very kind humble guy. And I also interacted with him a bit when he was on the board at the browser company when I was there for a couple of years. He's just kind of a real one, I think. And it's fun to see people in that position who, frankly, have all the money in the world just kind of want to get back to work and help work on some skunk work stuff and actually be able to bring the urgency that in some ways only founders can if they can retain that energy. And I was talking with the Anthropic Cons team too about this. He's also like the Dream Spokesperson for a company, especially a company like Anthropic, because he is legit. He's personable. He's charismatic. He'll go and talk about things that maybe considered passe for maybe you're more cerebral AI co-founders to talk about, but at the same time, he can say a lot without saying too much. And that's a rare skill. Well, did he say too much when he made the Instagram first Snapchat comparison? No, I think that's great. I think that's like that's podcast gold, right? I think. Of course it is. Yeah, yeah. Did the PR people bite their nails? I don't know. I don't know. I don't know. I mean, you saw how the way he reacted to Figma, which was very diplomatic, I had to ask him about that situation for, and I don't know. Mike's good. He's good at what he does. And he's a great interviewer. And I really appreciate it the time with him. I think it was nice to get into his head and hear kind of the areas that Anthropic is exploring next. That was the goal for me because we talked about this. If the team he's on, the last big thing they did was literally clogged code, literally the thing that has made Anthropic the most valuable company in the world, that's a tremendous amount of pressure. And even for the co-founder of Instagram, that's a lot of pressure. And so I wanted to hear the areas he was exploring. I do feel like he gave us a lot of good detail on that, which I appreciate it. Yeah. Now that it's been a little bit of time between the interview, what do you feel like you've been thinking about the most and connecting the dots in your sources, personal knowledge management system? What are the big themes? Right. I need Mike to build me that. I think what I've been thinking about is it's clear that health is a thing they're going to go into next. So if you were an investor type, you'd be like, "Oh, I'm going to be a big deal." would short all the health stocks, right, given how things have gone with cloud design and Figma or what have you. It's very clear they're focused on health. There were two things he said that really stuck out to me about anthropics culture, which I think is very unique. And I think people discount how unique it is and how much it drives the decisions anthropic makes. But he said at one point that the point of the company is not to build a large commercial enterprise, which is ironic given that it's the fastest growing commercial entity of all time, but to quote, "nudge the world towards a good future with AI." I think there's reasons to shine a light on that and a magnifying glass and go, "Well, like, okay, then how does that make sense partnering with Elon or the Department of War stuff or whatever?" But I do think a lot of people, the vast majority of people inside anthropic do feel that it. Talking about this, actually with another senior leader there, just off the record that day, that the culture of anthropic really reminds me of Apple in a lot of ways. The good and the bad. It has the good of really focusing on the craft, the commitment to detail, the world building, the kind of reality distortion field of what's going on, the focus on product. But also, you get a little bit of that chip on your shoulder, a little bit of that reality distortion field, which can also potentially be negative. It has a bit of an insular culture like Apple, where it's very much like you're in, I don't want to say cold, but it is kind of like a cold, like you're in it or you're not, and you get it or you don't. It's not really like a normal big tech company can just employ, you can just show up to anthropic and necessarily succeed. I thought that was interesting. What do you think that's due to? I mean, I'm a believer in kind of two schools that thought one is heavily documenting very specific beliefs and philosophy. And I think they've done that better than anybody else today. I probably, I feel like I have like three to four browser tabs open that I just can't seem to make time for that are all the different essays that they've written. I think the most recent one was kind of like teaching Claude why it's been a lot of great stuff. You know, they have the Constitution, obviously Darios essays. But then on the other hand, it all just still comes back to people, hire people like them, and it just comes back to the top and the founders, you know, hiring people all the way down. I think that's what it is. I think it's very much a safety aligned culture and the culture interview part of the hiring process is very serious. It's way more intense than those companies. Specific questions to. Yeah. It's about like what are the hardest like moral trade-offs you've had to make? Like when have you, you know, put your morals ahead of like financial gain? Would you be willing to do that at Anthropic? That kind of thing. Did you ever go to detention in middle school? And so why? Right. Right. Right. But yeah, I mean, the culture, the culture is super unique. And it's probably a rare thing for someone like Mike to come in and succeed in a culture like that because he was, you know, in the Facebook machine for so long. And he joined, you know, two, almost two years ago, two years ago, which is an eternity in AI, especially for Anthropic. Anthropic was like not on the map really at all two years ago. So he obviously saw something and made a bet on something that he didn't need to make a bet on because he's, you know, he's the freaking co-founder of Instagram. Like he doesn't need to do anything. But he clearly saw that there was something happening here and yeah, not that he needed another break, but he certainly got one. I mean, it's a testament to him too, just that he stayed in touch with everything happening in the world and in culture, which isn't always or maybe even usually the case for people who've kind of been around and had such big success. I mean, I appreciated, you know, his recognition of kind of like the distribution challenges of building product today like back in the old days, Facebook had 15 minutes of time per day that you had to compete with, you know, and now it's an hour from TikTok and it's the most addictive content format ever created. And he's clearly still plugged in, which I really genuinely appreciate, you know, you meet founders or people in this industry all the time who may even celebrate not being in touch anymore. And just kind of moving beyond it, you know, whether it's with politics or otherwise. And it seems like he's he's retained that even going to the point of rebuilding bourbon from scratch, which was really fun. And then coming away from it with an actually cool insight, which is that even though it helped us build it so quickly, I think he said something along the lines of like if you just use LLM for everything, it might actually prevent you from finding the surprising or non-obvious or culturally edgy things that might make a product work. Certainly speaking of Snapchat, I mean, what AI in the world would have ever suggested that ephemerality was going to be the solution to the pressures of social media? Like that was just a crazy one and every single person said Evan was wrong about that until it started working. No, I mean his point about how human intuition and relationships and creativity become more valuable, I think, was really good to hear and I do agree with that. I think the people who are really plugged in on AI and where it's going are all saying that and it was good to hear that from him. I don't know if you also caught he's very focused on distribution still. Having built something that obviously has tremendous distribution billions of users with Instagram. I thought his thing about I don't know if you caught this like MCP is a distribution mechanism was interesting like an MCP leaderboard and hearing from partners on MCP that you know, they get tremendous inflow of traffic via MCP if anthropic features them. So it's almost like this new app store model being created where instead of downloading apps in an MCP store or something, I could see them productizing that as well. It sounded like he was really focused on that and that's his consumer mindset right being brought to bear on AI. Yeah, I'm trying to picture what that might look like and I mean if MCP turns everything into a pipe and there's a lot of competition, then the fundamentals underlying those pipes, whether it has to do with the fees that you get charged by platform or even descend to transaction are just going to get lower and lower. I think which is definitely a good thing for consumers. Obviously, the motes are very thin and we'll see what happens to venture capital and consequently to myself as a result of that. But yeah, it seems like. I think storytelling is going to be all right, Ellis. That's my takeaway of the last episode. Why? I just feel like everyone feels that way. I think marketers and CMOs and brand people are going to become more valuable because that's how you're going to stand out. When anyone can mix anything, that's what matters. I guess that's why our friends at TPPN are over at OpenA. I know. Have you chatted with them? How are they feeling being back in the slack of a big company? I need to check in with John and Jordy about their X-Views cratering since the deal and see how they're feeling about that. That's a good reminder. But yeah, no, look, man, this was a great episode with Mike. Bummed you couldn't do with me, Ellis. When we were actually charting out early guests that we wanted to have on the show, Mike was someone you put forward that you wanted. I tried to do you, I tried to do you right and ask a few of your questions. You sent me. I don't know if you caught one of them. I did. He took a moment and said, "That's a really good question," when I said it was also a question. Which one? It was the one about his lab being a broader structure for companies in general. These founder, lean-based companies that don't have a ton of rank and file. The answer was basically yes, of course. That's the future, man. All I see is. I'm picturing in 50 years, we have 500,000 clones of Brian Chesky that are all spitting up one billion dollar companies simultaneously with a fleet of AI agents. I mean, Brian's cool, like, come on the pod, Brian, but Brian? He's got that, you know, New York hustler, mindset, tone, body. Who's the perfect apex founder these days? I feel like a lot of people say that. You heard it straight from Ellis folks. How could any of us compete? Well, this episode will be dropping the day that we're having our live show and launch party at Notions headquarters with the COF substack, Chris Bass. LSD, you have your toast for the room prepared. Have you written the speech? Oh, Jesus. No. But I did buy this suitcase. You told me to get in a way suitcase instead of my radio travel pros. So I did upgrade. Good. Good. And I did get some nothing headphone A earcans so I could look cool on the plane as well. Oh, yeah. Shout out, Carl. So yeah, I'm trying to pump myself up. I'm excited to have several former guests there and it's going to be a great group and hope we can do more live stuff down the road. But yeah, hopefully they show up instead of just talking to their agents. Yeah. I mean, that's the one thing I think that may last in AI's is events because what is the joy of sending your agent to an event? You can't taste the drinks that we've creatively named. You can't have the apps, the past apps that will be delicious. You can't hear the DJ will have. You can't potentially meet your future partner, you know? What kind of partner are you talking about? Business or love partner, either one? That would be sweet. I'd be quite an accomplishment if we had people like making out on the dance floor at our party. That'd be awesome. We'll see. All right, well, should we wrap it there? And that's it for this week's show. Thanks to Mike for coming on. If you like the show, don't forget to like and subscribe everywhere you get podcasts. We are access.showOnline, so you can find all of our links. And you can find us in video, lovely high definition at access pod on YouTube. If you really like this episode, we'd love a five star review or even share it with a friend or agent. You can find my newsletter at sources.news. And you could find me @Hamburger on Twitter and at meaning.com for your startup storytelling needs as long as they remain relevant. I think they will. Access is part of the Vox Media Podcast network and the shows produced by Hoek2Creators. Bye. [MUSIC PLAYING] Support for this show comes from Odo. Introducing Odo. It's an all-in-one fully integrated platform that makes your work easier. And the best part? So why not you? That's odo.com.