Anthropic’s Felix Rieseberg: why AI wins a Nobel before a Pulitzer

Felix Rieseberg leads engineering for Claude Cowork and Claude Code Desktop at Anthropic. Before that he led desktop and web infrastructure at Notion and built software at Stripe, Slack, and Microsoft. In this episode of Summation, Felix and Auren discuss: The next AI step function: going from solving a problem to owning a responsibilityWhy AI will win many Nobel prizes before it wins a single PulitzerBringing AI into the physical world, from a $30 coffee-machine display to a Wi-Fi gara...

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

• 0:02Auren Hoffman?

  Hello fellow data nerds. My guest today is Felix Rieseberg. Felix leads engineering for Claude Co. Work and Claude Co. Desktop at Anthropic. Before that, he led desktop and web infrastructure at Notion and built software at Stripe, Slack and Microsoft. He's also been a core maintainer at Electron, which is the framework behind a huge share of the desktop apps that we all use. And for nearly a decade, he's been doing that. And he's also behind some of these like super beloved hacks like Windows 95 running in your browser, which I think is super cool. And by the way, fun fact, he studied German literature and poetry, not computer science. Felix, welcome to Summation. Very excited for you to be here.

• 0:45Speaker 01

  Thank you so much for the invite. I'm delighted to be here. It seems fun.

• 0:49Auren Hoffman?

  Now Claude writes 80% of its own code, which is basically up from close to nothing a few years ago. And then Co-Work, I assume, was mostly built by Claude Code. What does it feel like to run a team that is mostly kind of telling the model what to do? And I assume part of the beautiful thing of that is you're eating your own dog food as you're doing it.

• 1:16Speaker 01

  I think that's probably like the one aspect that is maybe unique to my team is that this notion of like, this is our product and we enjoy using it quite a bit. But I think the thing that is most interesting about having Claude contribute so much to the product that we work on is maybe less the thing that I think we all experience, where like the models are getting better. The thing that is most interesting to me is sort of the journey of the entire experience, where we used to be and where we are now. So for me, for instance, I had my first interaction with AI sort of like in the year 2013. I was at Microsoft. We had a thing called

• 1:57Speaker 01

  We had a thing called Project Oxford. It was it was an Ngram model. It was well before transformers, well before the models that we have today. And it was pretty dumb. You basically gave us two to two or three tokens. They were basically were tokens. You would give me worldwide and like, maybe my NPM package would give you back worldwide web, right? And that that was it. That was the

• 2:16Auren Hoffman?

  This was like pre Word to back stuff?

• 2:20Speaker 01

  Yeah, this was the within Microsoft Cognitive Cognitive Services. It was pretty simple. Like, we at no point did anyone expect that the language model part of it would be interesting. And in fact, like, I worked sort of on the marketing side of some of it. And like, my focus was entirely, look, this thing can detect faces and photos.

• 2:39Speaker 01

  And obviously, where we are today, right, like 10 years later, and I think the thing that is most interesting to me is like, if you consider the rate of improvement, that it feels like very close to an exponential. The thing that's most interesting to me is not that Claude is currently writing most of the products that we're working on. It's sort of like the, what is this ticket's next? And like, how do you organize your work, not for the capabilities of today, but for the capabilities of tomorrow? If that makes any sense? And that's, I think, probably the piece that's more interesting to me.

• 3:12Auren Hoffman?

  For me personally, and I would say, like, of people that I know, it seemed like everything changed in December, December 2025. To me, that was the month where like, my entire workflow changed, how I'm doing work personally changed. It seemed before that a lot of people in my team were in code, or they're doing, you know, or back in the cursor, all these other kind of things that were there. Whereas like, starting in December 2025, everybody was either in Claude Code or CoWork, or some equivalent of that.

• 3:48Speaker 01

  Yeah, I think I think we had a similar experience and this is sort of like where Co-Work came from, but I think the the step function that I've really seen is that we hit this magical marker where models got good enough to go from, I used to call it like an answer. So basically, people used to ask AI simple questions. They used it for like, hey, please help me like autocomplete the sentence. The sentence might be like in an email or it might be in cursor line of code. And we hit this magical marker where all of a sudden it was good enough to like solve a problem for you. And in my domain, in the coding domain, that was, oh, you can now, you can now tell it to fix, like make a PR, right? Like you would still have to intervene quite a bit, but it was like good enough to

• 4:29Speaker 01

  get there. And when we say that Co-Work mostly wrote itself, like what we really mean is that that is the step that we were at. So humans still thought about the features. They still thought about like the taste behind it. Obviously, humans are very good at like evaluating whether or not something works for the human computer joint experience.

• 4:51Speaker 01

  That is what we use for Co-Work, but I think the step function that I'm seeing right now, and that's why I'm so excited about talking about the future, is I kind of feel like we might be seeing like something similar again. Like we used to, we moved from answer to like solve problem. And I think we're, we're about to see the same thing again. We're like, you can go from problem to responsibility, right? Like instead of saying, hey, please write this PR, you can say, your job is to make a good app. The latter point is not where we were with Co-Work when we built it, but I feel like, and this is so wild to say, right? Because like we built Co-Work like, it's now June, we built it in January.

• 5:30Auren Hoffman?

  Oh, was it January? Okay, it was only in January it came out? Okay, I thought it was December. Okay, yeah, Daniel, okay, got it. Yeah.

• 5:35Speaker 01

  The time confession is wild. We recently celebrated Claude Code's first birthday.

• 5:40Auren Hoffman?

  Wow, that's crazy.

• 5:42Speaker 01

  Like if you think about that for a second, it's wild.

• 5:44Auren Hoffman?

  That's, that's, to me, there's also like a UI thing that's really important is just most people don't feel comfortable in the terminal.

• 5:56Auren Hoffman?

  And if you can just get it into something, so even like Claude Code, I mean, you are Claude coworker, which is sometimes I have some questions about for you about it because sometimes I'm not even sure where I should be or you know what, they seem to be very similar. But let's say you're in Claude Code, it just the fact that you're there with a kind of a nice UI as opposed to being in the terminal, all of a sudden, just like 100 Xs, the number of users that are going to like actually benefit from the product.

• 6:28Speaker 01

  Yeah, I think, I think code, you can look at Claude Code or the initial version of Claude Code. Today Claude Code is like a whole bundle of very powerful features, which by the way, and I think I should say this, right? Like I'm on a podcast, I should say that we also have Claude Code as a UI that I think I'm quite proud of and it's quite pretty. So for people who prefer a graphical user interface over the terminal.

• 6:49Auren Hoffman?

  I'm in the Claude Code UI all the time. I'm not, I don't use it interacting with the terminal. I prefer it. It's actually just like, it's very, it's a very nice experience. I'm literal. I mean, multiple hours every day and I just, the terminal, for me, it just wouldn't be as, as exciting.

• 7:08Speaker 01

  Yeah, that makes sense. I think it's like a question of personal choice, almost like preference, right? It's like very close to like, what is your favorite ice cream? But if you look at the first version of Claude Code, for me, there were like multiple pieces in there that I found quite interesting. And one of the things I found very interesting was this agent loop and the underlying agent harness that we discovered to be quite powerful at not just coding work. So the, the initial goal was if you give Claude the ability to write files, to read files, to execute a little bit of code on your computer, suddenly it can be dramatically more powerful than if it only has the chat interface with you, right? And that I think to me makes very immediate sense because I often like to think about Claude as just another coworker. That makes it easier for me to build the environment in which Claude

• 7:53Speaker 01

  is powerful. And if you think about yourself being successful in a company without the ability to run code or like use a little bit of a computer, you're really just constrained to like questions and answers, right? And the example I often give people is like, imagine you join a company tomorrow and the only way you can interface with the rest of the company is like a fax machine.

• 8:12Auren Hoffman?

  Yeah, crazy. Yeah.

• 8:14Speaker 01

  You're not going to be very good. And I think in December, as you, as you like pointed out, like, one thing we saw is that a lot of people discovered that they can use this like very developer focused interface and this like made for developers agent harness and do a bunch of work that is really not about development. Because a lot of knowledge work happens at a computer. A lot of knowledge work involves opening files, reading files, and then doing things with them. So the initial thesis of cowork was, how do we make it easier for people to like do these like flows that they're already trying to do in the terminal, like, keep a tomato plant alive, organize, organize a move, like, try to

• 8:52Speaker 01

  prepare, like, memos, that kind of work. And we took the piece that was most meaningful, I think, which is the agent loop. And we moved that out of the, the like, environment that was very much made just for developers. And the entire thesis of cowork is how can we make this work for people who are not developers, but make it as powerful as Claude Code?

• 9:14Auren Hoffman?

  Yeah, yeah. Okay, that makes a lot of sense. My partner in May, so he really wanted to read a book about Dario, your CEO. And at the time, there was no good biography about him. So my partner used Claude to write a 200 and 200 plus page biography, I think it's a 253 page biography, and just went and it took, it took the entire weekend. So it took him literally the entire weekend to go do it. But usually you don't write a whole amazing book an entire weekend. And he published on Amazon. It became the number one biography book for a while. It's still, I think, the number three technology book on Amazon.

• 9:58Speaker 01

  That is wild.

• 10:00Auren Hoffman?

  Yeah, it's a really well written book. Like, it's actually, and it's like, it's like, they, they hook you every chapter, like hooks you in, and it gives you like the teaser kind of thing, and it gets you excited, and it talks about him and his sister, and growing up in San Francisco. And, I mean, it's like, it's a, I really enjoyed the book as just a pure biography. Now, is it as good as Walter Isaacson? No. But it's still pretty darn good. And again, took him a weekend to, to, to do it.

• 10:28Speaker 01

  Yeah, and it's not just it's not just speed either, right? Like, I'm always super interested in these use cases that involve writing. And like, I love talking to journalists about how they use the product. Because for a long time, I think people who are in the industry have correctly assumed that AI would win many Nobel prizes before it wins a Pulitzer, which I think is the right approach, right? Because you can go deep into the philosophy of it. You can ask yourself like, how much of art requires a human experience, right? Like, how much do you need the artist to like, be able to to express a human experience? But at the same time, especially in the production of professional content, right? Like, be that journalism or like a Walter Isaacson book, there, there. And you mentioned that I studied literature and poetry.

• 11:12Speaker 01

  There's a surprising amount of rules that I think is often surprising to people. There's like more structure in literature and poetry than than is maybe apparent. And like, even simple things like hero's journey, right? There are some things about, if you take the average person out there, you can probably, with a few simple rules, make them much better writers. And I think we're now entering this interesting point of, this interesting point of model intelligence, where for most, my experience, my personal experience is that for most of my writing, Claude makes it better.

• 11:45Auren Hoffman?

  Yes.

• 11:46Speaker 01

  Right? It's probably not hitting Walter Isaacson yet. That makes sense, but I wouldn't myself either.

• 11:51Auren Hoffman?

  That's right. Yeah.

• 11:52Speaker 01

  And the only question is, does it make meat more productive and better? And I think we've surprisingly hit that point.

• 11:58Auren Hoffman?

  Almost certainly. Yeah. It really can make, what I use it for, which is kind of, which I've been using it for for the last three or four years, is making it funnier. It's actually very good at that, because there's lots of rules about how to like, make things a little bit funnier, make things witty, and you can kind of guide it. It's the type of humor you want it. So if you, even if you're writing a pretty technical piece, like, lightening it up a little bit, and it just makes it so much more readable if it's, if you throw in some jokes here and there, you do some other types of things. So I have done that quite a bit in my pieces. And so almost all the jokes are either written by Claude or some competitor to Claude.

• 12:39Speaker 01

  Oh, man, I'm happy to hear that it works well for you, you know?

• 12:43Auren Hoffman?

  Okay, I've got some suggestions, or I don't even some suggestions for Claude and CoWorks. So if you don't mind, I'm gonna dive in with you. I'm sure a lot of our audience is using it every day. Okay, so this is super nitpicky, and I don't even know if you have these kids, but I use multiple machines. I've got a laptop, I've got two different desktops, one at home, one at work that I use. They're all Macs, but they're, you know, I mean, and it just like, it just doesn't sync well. It's like still got this like a local environment. Like, it doesn't seem like it would be that hard to kind of like add that. And I was like, and then if I was using some sort of shared drive, whatever you wanna call it, who, you know, whatever. Or if you wanna use it, it was, it was still not working very well.

• 13:25Auren Hoffman?

  And lots of corruption was happening and the skills weren't transferring and those types of things. It does seem like that could be like a very seamless experience.

• 13:34Speaker 01

  Yeah, I think I agree with you. I think there's a few pieces about Claude Code and Claude Co-Work that we would consider hard constraints. This is not one of them. I've had this conversation with a lot of people, especially in public, because I'm, I'm, I consider myself like one of the few people in Silicon Valley who like puts so much value in your local computer. But you're perfectly, you're perfectly right that it could be multiple computers. That, that makes absolute sense. You're not the first one to bring this to us. We're aware, and I'm sure that eventually we're gonna solve that, you know?

• 14:05Auren Hoffman?

  Okay, all right. Now, the second thing is almost the same thing. It's like, okay, I wanna collaborate with people. And it could be something very simple, like I have a skill and I want somebody else to share that skill or they have a skill and I don't want, and like, like right now, sometimes like where, well, like, I mean, we've been like putting the skills on Slack or sometimes people email skills or we have like a shared GitHub, which is kind of janky and you gotta grab it and then you gotta remember to check it in and stuff like that. Like, it seems like there should be like a much easier way to collaborate. And then you kind of have to choose who you, it, it does get complicated because who do you collaborate with? And you might wanna collaborate with some people on this and some other people over this thing. So kind of just like, just like GitHub, you have to have like some sort of rules around all those things.

• 14:51Speaker 01

  Yeah, collaboration is like such an interesting problem. I have two answers for you and like, one of them is maybe just. Slightly boring and maybe obvious like product placement, but we just launched the Claude tag in Slack.

• 15:02Auren Hoffman?

  Yes, we use it. By the way, it's great. I love the product. I think we've been using it at least for the last few days or something and it's very good.

• 15:11Speaker 01

  We've been using something like the product that we launched, including the product that we did launch in Satanthropic for a long time. And it's kind of hard to like really communicate to people why it's so good. Our recent Hayek party took to Twitter to be like, this is the new way to use AI. And I think people on Twitter were, didn't fully understand what he was trying to say because they were like, what do you mean? It is Claude and Slack. Like, how could this be a big step? But it's, but it's like really about this point that you mentioned with this collaboration with AIs. You have this like multi-user experience.

• 15:45Auren Hoffman?

  By the way, the other good thing about Slack is that you start to see what your colleagues are doing.

• 15:53Speaker 01

  Yeah.

• 15:53Auren Hoffman?

  With it. And so it just becomes like much more public. And then you can learn from your, you're like, oh, I didn't realize like, oh, they're doing this thing. Like, oh, that's cool. Like I'm in this channel with them. They're interacting like, oh, I'll start interacting that way. Well, I found like all these things are like, it's, it's like a virus that it's super, it just super infects what you're doing. It's very, very contagious. You see someone else doing it. You see them and somehow they're benefiting from it. You're like, I want to do that. And all of a sudden like, and then everyone starts upleveling each other in the organization.

• 16:27Speaker 01

  Yeah, it's super powerful. And I think collaboration is really interesting because I think it's actually less of a technical problem. Like all of the, whatever ideas someone might have out there about like how they would like to collaborate with the humans with Claude. It's probably pretty easy for me to build. The thing that is interesting about collaboration is that the way humans organize work and the way humans collaborate is like, so interesting. I spent so many of my years at Slack and.

• 16:54Speaker 01

  Slack too, there were some things about it that were technically difficult to build. But when I joined Slack, my friends were like, why are you joining a chat company? We have like 50 chat apps. There's something that was really, really hard about going into organizations and trying to convince people. But that there's a different way to work that involves more transparency. And like, what if you worked in a channel and yes, it's like more people than you put on this email thread, but there's a value here. But that took people a really hard time to adjust. It's like kind of hard to, it's hard to remember for people who now live in Slack. But back in the day, the thing that most companies had before were like these email lists and like email threads and like things just all email.

• 17:32Auren Hoffman?

  Yeah. Or you had it, I mean, you had like HipChat, you had all these other kind of things that are, that you would use internally and stuff like that.

• 17:40Speaker 01

  Yeah. And whenever, whenever I talked to my users, I was, I was responsible with Slack desktop apps. The biggest challenges that I had at that job were less about how do I build the app? They were much more about how do you understand human collaboration? And how do you build something that is a really good experience universally for people working together?

• 18:02Auren Hoffman?

  What did you learn about human collaboration that is non-obvious?

• 18:06Speaker 01

  I think the most interesting piece is always that, like, especially in Silicon Valley, we like to forget how little attention people have for tech, just like they don't have time. People out there are incredibly busy. People have like their lives to juggle and they're really not interested in like trying too many of your new toys out there. So you need to get very, this is not news. Like you asked me if this is like obvious or not, I do think it's like obvious to a lot of people. But I think Silicon Valley keeps making the mistake of like giving people 50 different buttons and being like, oh, this is all an A B test, you tell us what works for you. And I think like a huge part of the population is just too busy to like be your guinea pigs. You have to actually execute a lot of like really good taste and be very mindful.

• 18:49Speaker 01

  And Stuart, the Slack CEO, I think, I was always impressed by how focused he was. I built a lot of things that Stuart shut down. And it wasn't even on the grounds of like, this is not good or bad. It was more on the grounds of like, this is not necessary, shouldn't be there. This is a. This should not be in this.

• 19:05Auren Hoffman?

  I do think like there's just still so much complexity and like if you're building, I built this simple app for myself. It was like for me and some friends to collaborate on our pickleball games and stuff like that. And you know, it's like, okay, build it in Claude code. And then I had to like deploy it to Versal. And of course, I'm like putting the code in GitHub. And then I've got to like, I've got to like, well, I had to do off and I had to do some other stores. So then I use Superbase for that. And then I had to send emails. I got to use resend for the emails. And I wanted to text people the thing. And I had to send this other thing for texting. And then I had to use like these three other APIs. And every time I had like, it wasn't the problem. I had to go to each, each, all the, all these different places. That was fine.

• 19:47Auren Hoffman?

  But every time you have to like go login, it's like, you know, you have to go, you have to find, you have to go through the UI, you have to like click three buttons to set up the API key and store it there. And you have to, you know, is like, I wonder like, why can't I just do this from my Claude code? Like, why can't it's great? Like, you need Superbase, great. Just, I should just press a button and it automatically does this thing for me. And if like, you need my credit card, like, fine. Like, I'll put it in one time and it just gets it all, the flows through it. So I do think there's still a lot of time that we could take out of just building like even some of the simplest stuff.

• 20:22Speaker 01

  Yeah, for sure. I think we're honestly just at the beginning, right? I'm like quite convinced that the apps we're going to use and the way we interact with AI, like collaborate with AI, like either just one human, one agent, or like multiple humans, multiple agents. I really think that's going to look so different even next year. Yeah.

• 20:40Auren Hoffman?

  One of the things I like about, like, let's say, let's say Claude code and Codex, which would be like a competitor. But also, it's interesting on the business. So I, I, I move things sometimes from one to the other. Sometimes they're both collaborating on the same repo, right? So they're both going on and they both can, they, it, it, there's, there's like no switching cost. And I love the fact that both groups have decided that there is no switching costs. For me, as a user, that's great. And it's just kind of like, let the best product win. It seems, it seems like that's kind of the mentality, rather than like, let's lock in the user.

• 21:22Auren Hoffman?

  Let's make it so hard for them to switch. Or, you know, the classic SaaS stuff from 10 years ago. Is that, was that like an intentional thing when you guys were designing it?

• 21:32Speaker 01

  It's an interesting question. I think it's maybe just not in our interest as a company. Generally speaking, Anthropic is working very hard to build highly trusted, reliable, enterprise-ready AI systems, right, that keep you productive with your work. It just doesn't serve us right now, right? Like, the things that I think about most of the time is less about like, how do I, how do I keep you from leaving? And it's more, how do I make sure that I build a product that is a very delightful experience and really has you more productive? And I, as an individual, like, as a product builder, I just have always believed that if you build really good software that makes people happy and like makes them more productive,

• 22:15Speaker 01

  as long as it provides value, that's really all you have to do.

• 22:19Auren Hoffman?

  Yeah, absolutely. I mean, there's so many software applications that people hate using. And they're stuck using for a lot of various reasons that they, you know, very hard to switch and stuff like that. And then, of course, every once in a while, you have an application like Claude Code and Claude Co-Work that is, to me, purely delightful and getting better every day. And so it's like, it's just something like, great. It's something you want to use, something like you look forward to kind of, you know, going there. It's like having this like friend that's like always makes you happy. And then you have this other, you know, person in your life who's like, it's like a family obligation. You're like, oh, I gotta go meet with this person, but like, I don't know if I like kind of doing that. And I feel like if you're a software engineer, you want to work on like the cool products.

• 23:04Auren Hoffman?

  Like, that's just more fun.

• 23:06Speaker 01

  We also have such an abundance of latent demand, right? Like, if I wake up in the morning and I'm like, okay, I think I've solved any given domain. I feel like I've solved it as well as I can. And I think within the area of AI, we're pretty far away from like saying that at any point, but let's imagine I wake up in the morning and I feel like I need to acquire more users, more value, generate more revenue, all of those things. Right now, we're still, we're still at a time where there's so much latent demand, which presents to us in the way of people using our products in a way that was not intended, in a way that is actually quite painful. And you mentioned this at the beginning, right? Like, in December, we saw all these, I saw so many tutorials on the internet that were like, you have never used the terminal, but you've heard from your coworkers that Claude

• 23:51Speaker 01

  Code is great. Here's how, here's a terminal tutorial. And that's obviously, that's like, that's like a goldmine for me, right? I'm like, oh, there's all this friction, there's all this pain that I can take away from you and make it easier for you. When I do want to acquire more business, that's really where my mind goes. And currently, we have so much of that of like people crawling over glass to like get to the value and we can just make that path much easier.

• 24:14Auren Hoffman?

  Yeah. And I remember like, I remember installing OpenClaw, and this is, I don't know, maybe in December, I can't exactly remember when it was. Frankly, it was a pain in the butt to do it. And it was a lot of work and I just don't think like very many people would have like taken all those steps and every one of those steps like you could mess up on and something could go wrong. And, and today, whatever, that equivalent of if you want that same thing today, there's, there's already many different applications, including doing it in coworking code that just is so easy for the user and delightful. Still could be easier, of course, but so much easier in the last, in the, in six months than it was six months ago.

• 24:56Speaker 01

  Yeah. And hopefully also more safe and more secure, right?

• 24:58Auren Hoffman?

  That's right. Yeah, also very important. Yeah, yeah, yeah. Absolutely. One, okay. So usually, I would say today, I'm personally, and when I look at my colleagues, even my non-technical colleagues, we're probably doing like more stuff in code than in cowork. So even the people have never written a line of code before in their life, for whatever reason, sometimes I don't even know like when you should be in one versus the other. I will say that the scheduled tasks in cowork, just the UI of it and just how it works and how you can click into it and everything like that is like so much easier than the routines in code, even though they're kind of the same thing.

• 25:43Auren Hoffman?

  So it's just like that whole UI and like, you can't actually like click into the routine. You have to, it's like three clicks to get into the routine, but it's one click to get, you know what I'm talking about?

• 25:51Speaker 01

  We'll make it better. We'll make it better. I think I have a good answer. This is a question I get all the time, right? Like people ask me, when should I use cowork? When should I use code? And I have like multiple answers for you and we're going to see which one maybe, maybe gives you, gives you the best answer. You tell me. But the first one is that Claude code is really powerful for developers. That's the core thesis, right? And it's using your computer to help you with your development tasks. And CoWork assumes by default that you may be not a developer. And what it assumes there is that you probably don't have development tools installed. Your computer is not set up for advanced development.

• 26:27Auren Hoffman?

  Ah, yeah, okay, got it. You don't have like the NPM or whatever you're going to do on your.

• 26:32Speaker 01

  Exactly. And maybe also you just like either don't have the skill or the time or the interest to figure out. And when Claude is trying to ask you, hey, do you want to install this NPM package? You might not be the best person to ascertain whether or not that is safe or not. So CoWork takes all of that, CoWork takes all of that as Claude code and puts it in this little environment in which Claude can be the most productive developer we can envision, which currently involves like a full Ubuntu machine. It involves thousands of development tools. And it's really the difference between like,

• 27:06Speaker 01

  like if you take a very powerful human developer and you give them two computers and one of them is like locked down for, say, the legal department inside a financial institution. And the other one is a developer machine. And then you ask them to solve a problem. They're going to take two different, two very different approaches. And CoWork gives Claude a development environment, even on machines that otherwise might be locked down or like in a secure environment and usually don't allow development. That's the first part. That's like one way to, that's one way to think about it. The other way to think about it is, and this is like, I think, like an easy way, we really think about it as two different audiences. We think about CoWork as being very powerful with general knowledge work.

• 27:49Speaker 01

  And to that end, we have built a lot of features that make that easier. CoWork is much better at creating good documents. It has a harness that like, that's you easily edit and preview markdown files or docx files. It's like good at making presentations, doing long-running research tasks.

• 28:06Speaker 01

  So there's a lot there in that area.

• 28:08Auren Hoffman?

  And do you think it's better than code for those things?

• 28:12Speaker 01

  Yeah, it's just additional features in the harness that we have not brought to code.

• 28:16Auren Hoffman?

  Okay.

• 28:16Speaker 01

  Yeah.

• 28:19Speaker 01

  It's just like slightly, how do you evaluate the harness and how do you improve the harness? The other way around, if you are a developer and you actually do want Claude to like mess around with your local computer and you want it to like work within with your actual hardware and like use your Git credentials and like do things as you like. If you wanted to SSH to a machine that you control, Claude code is probably a better choice because CoWork is really meant to like have its own developer environment. So it's less like you as a developer and more like someone else's developer. Now, when it comes to the actual tasks though, and this is maybe the important part, when it comes to the actual tasks,

• 28:57Speaker 01

  it's really more about the machine it's running on and the human it's running on. So my wife, for instance, is a TPM at a big AI company too. And she's using CoWork regularly to build tools, right? Like if you give Claude a prompt, whether it's in CoWork or code, it will do it and it will try to solve it to its best abilities.

• 29:21Speaker 01

  But I might build some of the things that she's building on my computer in code rather than CoWork, because my computer is already set up as a developer computer and I do have all the development tools. And I have my developer environment set up just the way I like it.

• 29:33Auren Hoffman?

  But it's still like, but building tools might involve writing a Python script or something like that. Or, you know, and she's just doing that in CoWork.

• 29:41Speaker 01

  Exactly, because like her computer actually, I assume, doesn't allow the execution of raw Python by default. Like if you're in an enterprise, you probably, many enterprises have the computers correctly locked down so that unsafe software can't be run. It's like the airbag approach, right? Like you should not be hitting a wall, but if you do, you shouldn't die. That is how most enterprise computers are set up. And they sometimes have this carve out for developers. Developers by nature write software that has not been evaluated, gets run all the time. It's the same thing on your iPhone, right? Like your iPhone installs from this trusted app store, but for a developer, it runs anything. That's the idea.

• 30:20Speaker 01

  But Claude as a software developer needs Python, it needs Node, it needs the ability to like fetch code from the internet. And CoWork is very good at keeping that safe and also making sure that if you do accidentally download bad stuff from the internet and over the course of your development, it's constrained in this little sandbox and it doesn't like break out of it.

• 30:37Auren Hoffman?

  Yeah. Now, you are, one of the nice things about working at Ambropic is you can work, where you can play with some of the newer models, the newer capabilities. When you, when you have a big jump, let's say, Opus to Fable or something like that, like what unlock internally surprised you?

• 30:57Speaker 01

  I have such a hard time shutting up about this, but because I'm like so excited with this year, because I really think, I really think this year is the year where we're going to like move from problems to responsibilities. Some people now call this loop or whatever you want to call it. And whenever you make predictions of what was going to happen AI, you always sound a little unhinged, because it always sounds like, it always sounds like the kind of stuff that like should be 10 years away, but humans are very bad at like actually picturing an exponential.

• 31:26Auren Hoffman?

  It does seem like every prediction, pretty much that everybody I know has made, including myself, is wrong. Like every one year out prediction, and often wrong, and not always wrong in the, like, this won't happen, like, wrong in the other direction, like, oh, by, you know, we won't have any customer service jobs in three years, or, you know, or something like that. And, you know, it's still, there's still a lot of customer service jobs and, you know, all these other things, yeah.

• 31:52Speaker 01

  Yeah. So I think the, the big unlock that maybe a surprise, I think is a lie word, because this happened to me much sooner than I expected it. So we've been talking a lot about like last December and how we could give AI, AI models, like entire problems to solve. And within my domain, maybe one of those problems is a crash, right? Like the app crashes if you hit this button. And I used to just tell Claude, okay, go and fix this crash. Like I would be in the loop. I would like find the crash. I would look at the crash. I would say, oh, this is probably actually a mistake that we made. Claude, please help me fix this. And then in December, if you'd asked me what the future looks like, I probably would have told you that, oh, we're gonna, I'm gonna keep doing that. I'm gonna be faster at it, right? Like Claude is gonna make better fixes and the kinds of crashes I can give it can be more and more complicated

• 32:37Speaker 01

  and like, well, figure out more of this out. And where we've been now for like honestly months at this point is that I don't do any of this work. Claude looks at all the crashes. Like Claude has access to the crash database.

• 32:47Auren Hoffman?

  It automatically knows it crashed and then it can look at the logs and it can jump in before you tell it to jump in.

• 32:54Speaker 01

  Exactly. And can I independently verify, okay, this is code that runs on like millions of computers and it only crashed one person. It's probably a case of a broken computer rather than our app being broken.

• 33:04Auren Hoffman?

  Or, yeah, and sometimes, yeah, you have to figure out, it's like, oh, in this, why did it crash on these 1% of machines? It's like, well, they probably had some other thing that it got was installed on their machine that's somehow interacting with us in a bad way. And how do we figure out what that is? And all these, this is stuff very good for AIs to figure out.

• 33:26Speaker 01

  Yeah, and even there, right? Like, we now have a Claude running, whose job it is to keep the app crash free. And even there, like, if you come across this thing that Claude will just like, spin up a virtual machine, install the software, everything might be at fault, try to figure out what is the actual problem there. So the step away from Claude helping solve this problem towards Claude, your job is to do the following things, please just take care of this for me so I don't have to, is maybe the thing that surprised me the most because it came much sooner than I expected it.

• 33:58Auren Hoffman?

  Interesting. It generally only on the, when, I remember when, when the, so it like, let's say, January timeframe, I was using the Chrome extension for Claude a lot to drive my Chrome. And, and it wasn't working for me at all, like when it first came out. And it took me like a long time to debug and turned out it was like, because like one of my, like my password manager was messing it up somehow. And like, if I got rid of the password manager, it worked great. And then I tell, within like a few weeks, there was some sort of Claude fix that like, then it worked, it worked great. But yeah, these are the things that like, those are the super frustrating for a user.

• 34:43Auren Hoffman?

  Because like, you know, it's working for your colleague, it's not working for you. You're not sure why. You have to look at like these 50 things to figure it out. And of course, if like Claude could just figure it out itself, like that would be so much more delightful for the user.

• 34:58Speaker 01

  Yeah. And it's, it's, the other thing that's fascinating is like how it changes how I interact with Claude in the first place. Because it's not just, it's not just that Claude now runs for longer independently before I check in. It's also the kinds of things that I check in on have kind of changed. So last year, I would probably help Claude in like, Understanding what the crash looks like, how to actually fix it, or whatever opinions about how to do it. And these days, I spend most of my energy on the verification side. We sometimes go with like this brain hands metaphor, right? It's like we now have this very smart model. Have we given it all the tools it needs to get this job done? And like

• 35:40Speaker 01

  a good example here is like, does Claude have an ability to run a computer? Does it have the ability to like, run multiple operating systems? We spend most of our human engineering time these days, not on the actual problems we give Claude, but on if we hired a person to do this, what kind of things would they need to do a good job? And the things that I do as a human to like, figure out whether or not Claude did a good job, does Claude have the same ability to look at those things? And that is, that is such an interesting, such an interesting switch, because we, we start really thinking about the model experience, right? Like, if you are the model, you have good tools to get the job done, which is

• 36:22Speaker 01

  very job dependent. And like something that's pretty hard for us to bake into the product a priori. It's just so interesting, because like, the way I worked with AI has changed so dramatically over such a short amount of time.

• 36:36Auren Hoffman?

  Has it changed the way you work, just in the things you can't use AI for? Is it changing the way you interact with your colleagues? Is it changing the way you're interacting, doing work that you can't use AI for in some way? When I give AI instruction, I, I do kind of think about it a little bit. I kind of have a sense of like, if I'm kind of creating a skill, I have a sense of like, what the guardrails might be and stuff. And in the past, I realized when I was doing something similar with colleagues, I wasn't, I wasn't thinking it through. I was expecting them to think all those things through, but

• 37:17Auren Hoffman?

  I, in some cases, I may have had more context than them. And so, I'm actually like, whenever I'm interacting with a colleague, I'm trying, I'm almost trying to write out a skill. For them, in a way, you know, that is, I mean, it's obviously a different type of skill than than I would than I would write, you know, for Claude or something. But if I if I could do a little bit more work and think through a little bit more rather than just writing them like a one line email, like that could make them so much more productive.

• 37:49Speaker 01

  Oh, that's a good point. Yeah, I think you might be right. Like, I often tell people that like technical writing is such an art form and technical writing, not just in like, you know, Silicon Valley style of technical, but technical as in like instruction based writing. Like, anyone who's ever assembled furniture knows that like writing instruction manuals is actually quite hard. Like, like, all of us have assembled furniture that was like, super annoying to assemble, and maybe also that was like, weirdly delightful to assemble. And it often comes down to like, people are really thinking through all of the tools you have and all of the things that you know. And my interaction with AI and humans is probably very similar. I think you're probably, you're probably right that

• 38:32Speaker 01

  like, writing good instructions now being such a core part of my job probably made me a better collaborator with humans to some extent. Because I do think whenever I have a problem, I probably show up with like, a little bit more prep work on my own, whether that is I've done more research myself, or I make it easier for to like, hand off the work. I think that's true. But the interesting thing is like, maybe we're already on the other side of that curve, because increasingly now, if I do it in Slack, I'm like, you know, I saw a thing today that I didn't like, and I talked to my designer about it. And I was

• 39:14Speaker 01

  like, hey, Peter, what's up with this button? Can we make this button better? And Claude already like chimed in proactively. If anyone hasn't used Claude in Slack, like the new Claude in Slack, it's like really cool. But Claude jumped in proactively and was like, by the way, this is this button from this design system, we've already applied these things. I've prepared three different options for you if you just want to take a look. So I, I think on the bell curve of how prepared to show up, like, we might be coming down on the other side.

• 39:44Auren Hoffman?

  Yeah, that might be interesting. I actually recently have been starting to write skills for things that like AI just can't, that I'm personally doing for myself, that AI can't do yet, but I feel confident that AI will be able to do, let's say, in the next 18 months. So like, basically, like, right now, AI, at least today for me, it has failed almost every time I ask them to like, buy something for me. Even something simple, like buy these shoes or buy this charger or something. AI has often bought the wrong thing or something happened. And then if I go to like a flight or something, it's like, it's really messing up for me. And because a flight could be very complicated. Okay, like, what seat do

• 40:29Auren Hoffman?

  you want? Like, we all have different preferences and how do you want to, and you know, and what kind of leg do you want? And are you willing to pay a little bit more for a little bit more convenience? And, you know, it's like, we're all so weird. So I just kind of, I've just been like, putting it in and writing it down. Every time I go through it, feeling like, pretty confident that in the next 12 months, it will, but so I just want to put in the work now as I'm like, learning these little things to go do. I don't know if like, other people are doing this stuff.

• 40:54Speaker 01

  I think it's a great idea because I think we have like mountains of management books and productivity books that were written well before AI that highlight the value of like critically thinking about the process of your work. Right? And if I can take you on a detour real quickly, this comes up within a valuation of of Claude quite a bit. Because people often ask me, how do I make sure that Claude is doing a good job? Like, and sometimes coworker does a better job than other systems. And I think there's one step that I think often I can see a real aha lightbulb moment for people, which I love. So I just want to like, do it here real quick. If you ask Claude what the temperature is outside, and it gives you a number, I actually don't care too much about that number.

• 41:39Speaker 01

  That is not, that is not what determines whether or not it did a good job for me. If it guessed the right number, that is a bad job. That is an F. It should not guess. If I, however, look at Claude's process and I see that Claude was like, ah, this person lives here. They're probably asking me about the temperature in this location. The best sources for like getting the temperature right now, probably like these three services. I'm going to go look at those services, figure out what the temperature is. Now, the user asks about what is the temperature? Like, do they, do they mean like within this neighborhood or within the city? So the process is more meaningful to me than the actual outcome in many cases.

• 42:15Auren Hoffman?

  Oh, by the way, and like, when you're asking for the temperature, what are you actually asking? Like, are you asking like, you know, if it knows, are you asking like, well, if it's, if you're asking for the temperature, but it's raining? Or it's gonna rain in the next three hours or something like, maybe that's what you care about. Or if it's like super humid, you might care about that. Or there's, there might be other things you want to understand. Not just like the fact that it's like eighty-two degrees, like what, what does that even mean? Something, you know?

• 42:42Speaker 01

  Yeah, it's a super good point, right? Like, and if, if anyone who listens to this lives in California, like coastal California, it's like a famous phenomenon that like the beach can have its own microclimate. That's completely different from like ten minutes drive. And if my last five conversations were about like prepping a beach party, I, you know, you should take that into account. And doing the same introspection for your own work can be super insightful. Because I think most humans, certainly myself, have not actually critically thought through like for all the things that we do, what does success really look like? Right? It's like an old management question of like, what does success look like? What does that actually mean? Like, there's many different philosophies. I'm, I'm not gonna, I'm not gonna tell anyone what the right or wrong one is.

• 43:25Speaker 01

  But I do think doing that like introspection of like, what are your goals? Like, what do you want? Can be super powerful if you can articulate it because you can articulate it not just to an AI agent. We can articulate it to yourself and that probably makes you much better. Which is like weirdly philosophical, but I, I do think it's true. So I think the exercise that you're talking about of like trying to figure out, okay, what do I want? What is it like, what is a good flight? What does that mean to Orin? You know, it's like a useful exercise.

• 43:55Auren Hoffman?

  How do you use it in your personal life? You mentioned you're married and kids and like, what are you doing in your personal life that maybe the rest of us could learn from?

• 44:05Speaker 01

  I did move recently and I think

• 44:10Speaker 01

  none of this is like grandiose, but there's like a few things I'm really enjoying. The first one is the level of optionality I have and like how I set up this house and like how I did the interior design, like how I moved the furniture around. That was, that was really fun. Cowork built me. Let me back up. I use Cowork a lot in like evaluating whether I should move here, what this house should look like. And my real estate agent gave me this like marketing blueprint and I wasn't here and I was kind of wondering what furniture to buy. And Cowork had so many suggestions about how I should do it.

• 44:49Speaker 01

  Cowork first made me like this little 3D furniture planner where I had like the blueprint, I could move furniture around, which I made not because I was able, that is super fun. And also it's super fun how I did it because it like the, the marketing blueprint didn't have any units in it. So I was like unable to see how big is a wall, but it had the garage in it and they found the permit for the garage on the website. I was like, ah, I can figure out what this looks like. But the reason I'm telling you the story is because I myself constantly forget that I can just give Claude jobs. So I sat down and was like, okay, cool. Now like put in my bed and I have this bed and Claude was like, don't worry. No need. I have your emails. I've seen all the furniture you've bought.

• 45:32Speaker 01

  I will just put all your furniture in here. You don't need to tell me all the dimensions.

• 45:37Speaker 01

  That like constantly going one step up is an experience that I use in my personal life all the time that I can heavily recommend. Which again is like, you know, what is your actual goal? Like I just want this house to be free and like to be functional and serve my family well. And like sharing a lot of that with Claude was really helpful because I often make this mistake of actually going too small and like the task I give it. And the task I gave it that was too small was like at this bed, right? But I really wanted was I want to play with all my furniture in this house. And the the other tip I have is, um, this is more for like personal joy, but you can bring AI systems into the physical world. And I have two examples for you. The first one is that, um, so I moved, this house is old.

• 46:20Speaker 01

  And as a result, a lot of the the plumbing, the electrical work, like a lot of this stuff is pretty old and in need of a lot of work. And I constantly have Claude just like make me step by step diagrams.

• 46:35Speaker 01

  If if anyone has not used like the cowork visualized tool yet, just ask Claude to like explain something visually. It is awesome, right? So if I need to fix some plumbing, um, I just have Claude write me manual or print it out. I take that into the crawl space. Beautiful. I I I'm.

• 46:51Auren Hoffman?

  You you actually print it out and not look at your phone or something? Or, okay, I love it.

• 46:56Speaker 01

  Yeah, because like it's like pretty gross down there. You know, I don't need to bring my, I don't need to bring my pristine technology down there. And the other thing, and this is like a puzzle, this is like, um, product productivity is a stretch here, but hardware components have gotten so cheap, so cheap. And I constantly walk around my house and I'm like, this appliance should have better software. I don't know. I don't know. You feel that have feelings about the software in most appliances?

• 47:26Auren Hoffman?

  Oh, my God, so terrible. Yeah.

• 47:28Speaker 01

  Could be better and.

• 47:30Auren Hoffman?

  So you just hack it?

• 47:32Speaker 01

  Yeah, I'm like, um, I I posted this on Twitter the other day, but, um, my coffee machine doesn't have a display and doesn't need to, it's fine. But it's like one of those espresso machines where you have to move like a lever and then coffee comes out and then it stops. And they sell you for, I think, $600 or something. They sell you like a scale that has a timer in it.

• 47:55Speaker 01

  I'm unwilling to spend that much money. So I I bought like, uh, I bought like this little device. It's like a self-contained ESP32 chip. It has a display. It has a battery. It has Wi-Fi, it has Bluetooth.

• 48:09Auren Hoffman?

  What does that device cost?

• 48:12Speaker 01

  I think I spent like thirty dollars.

• 48:15Auren Hoffman?

  Okay, and it has a, it has a display on it.

• 48:17Speaker 01

  It has a display.

• 48:18Auren Hoffman?

  Oh, that's cool.

• 48:19Speaker 01

  It has all the sensors.

• 48:20Auren Hoffman?

  Okay, that's awesome. Okay, so you could basically do whatever you want with that. Okay, yeah, okay, that's so cool. Okay. And it's got a little Android system running on it or something, or Raspberry Pi or whatever.

• 48:30Speaker 01

  Great question. It's just, it's a raw hardware component, but Claude is very good at running firmware.

• 48:35Auren Hoffman?

  Yeah, okay, got it. Yeah, perfect. Yeah.

• 48:37Speaker 01

  And I just told Claude, Claude, your job is to like turn this into like a beautiful companion display for my coffee machine. And now when I move my lever, the thing displays like a pretty clock, like tells you what the temperature of the water is.

• 48:49Auren Hoffman?

  Oh, that's awesome.

• 48:51Speaker 01

  Yeah, and doing that for like, I, in my software there, just have like a bundle of like components. And I think we often forget how cheap components are. Like, if you want to build your own Kindle, you maybe spend fifty bucks. If you want the kind of hardware, I want to say like sort of the first Apple Watch, right, like a beautiful little display and like a tiny, it's like maybe thirty dollars to ship to your house. And you just connect it to your computer, tell Claude what you want, and it's gonna build it for you.

• 49:18Auren Hoffman?

  You know, it's interesting because I think everyone I know is doing that with software today, and it's so delightful. And even in the last few months, people, I've got like, I've got like four different Chrome extensions I wrote recently that helped me like with my own personal, just for me, like my own software that does this just for me. But like, I don't know that many people who have done it yet for the hardware, and I could see that coming really soon. Like, it just seems like that's, it's so obvious, the next step. It's like, oh, I want this little thing for my, my cooking thing. Or I want this little thing next to my bed to do the thing, or I want this like thing that just changes it a little bit. Like, you're right, like, why would we not be doing that? Or, or I want these two things to interact.

• 50:02Auren Hoffman?

  And so somehow I need to get the, somehow the, okay, maybe one of them already has a Bluetooth chip. And so I just, somehow they need to interact with one another in some way, and then Claude can go figure all these things out. But that's gonna be cool, because then we're really gonna be able to like personalize it for ourself, and it's gonna be really, it's just gonna be much more fun.

• 50:18Speaker 01

  It is so fun, and it's like especially fun for.

• 50:23Speaker 01

  I think about, I don't really know when it started, but there was like the point in time where suddenly a lot of people had 3D printers and like sort of like sold it with their own hardware and Raspberry Pis came out. And I think I may have bought like one Raspberry Pi, but I didn't really do anything with it. I was busy, had other things to do. I don't have a 3D printer. I wouldn't consider myself, I wouldn't consider myself a hardware hacker. I don't really, I just want my espresso machine to have a display and it doesn't, right? And that, I used to just live with like bad appliances, but now it's like two hours later and it's plop and it's wireless.

• 50:58Auren Hoffman?

  I'm already at the point where if the software isn't beautiful, I just don't want to use it if it's not like delightful because like it, my bar for using software has gone up so dramatically over the last six to 12 months. And I could imagine my bar for using software within hardware, display within hardware, which right now my bar is still so low because it's all so bad and so annoying. Like I imagine like in the next 12 months, like that bar is going to get so much higher as well. Like you're just, you're buying something and it's like you're fiddling or you're, you can't set the clock well or you're just all these things that are like so frustrating sometimes.

• 51:38Speaker 01

  Yeah, it's like so liberating. Like I think I really like the feeling I want to convey is really like that of liberation. Because for anyone with an old house, like I don't have a lot of options for like changing things here that would not involve a massive remodel. Like the garage opener, right? Like it's one of the ancient Liftmasters made in, made in the eighties. But yeah, Claude helped me like buy a few components for like a few dollars. And now the, now the Liftmaster is like wifi connected.

• 52:08Auren Hoffman?

  One of the, it feels like this own goal, and I know you don't want to like speak ill of competitors, but okay, with Gemini.

• 52:17Auren Hoffman?

  You know, like a lot of companies, we are, we are on Google Docs, Google, we're on Gmail, or on all these other types of things. And frankly, I have to use Claude with Gmail and with Calendar and with Docs to get anything done. It's kind of weird because like, Google already has access to all my data. I should be able to use Gemini. If I want to ask something on my calendar, I should be able, if I want to ask a complicated question on my calendar, like, I don't know, what Malaysian restaurants have I been to in the last 10 years? Or something like that. I should be able to ask that with Gemini, but I can't. But I can ask that with Claude, and it gives me an answer immediately, and it's great. It doesn't seem like it's that hard for Gemini to do that. It's already native.

• 53:02Auren Hoffman?

  I would just do it with Gemini, but for whatever reason, they haven't done it. Like, do you have a sense of why they haven't done some of those types of things yet?

• 53:11Speaker 01

  Oh, I think that's a better question for Google. Maybe the one thing I will say, the one thing I will say is that anyone who makes software is like, very aware of, like, a long list of questions that begin with, why haven't you just, right?

• 53:28Auren Hoffman?

  Yeah, right. Like, there's legal reasons, there might be like, creep in the whatever, you know, yes.

• 53:36Speaker 01

  Yeah, like, we recently released our Linux app, like, yesterday we launched a Linux app for Claude. And for the longest time, people would like, listen to me talk about how easy it is for Claude to write all this stuff. And then they would come to me afterward, like, I run a Linux computer, where's my Linux app? Like, clearly, clearly, you can just press, like, just tell your Claude to build me this app, please. And I had like, real reasons for why I was a little hesitant to like, release a Linux app. I've gotten over those reasons, we solved those problems. I'm now like, quite happy to release a Linux app. So I think I share your frustration for a lot of the tools and services I use in my life that I want to be better. But I think I also have empathy for those services and like, even my competitors when it comes to like, why haven't they gone all

• 54:20Speaker 01

  the way? And like, making, making things better.

• 54:24Auren Hoffman?

  Okay, that's a very diplomatic and good answer. All right, two personal questions before we leave. We mentioned, I think, at the top of the show that your background's in German literature and poetry. In some ways, like, when I first started thinking about this before the episode, I thought that was very unusual for a deep tech leader. But then I started thinking, like, actually, it seems like more of a feature than a bug. Like, it actually does seem like it actually would be very useful. Talk about why that studying was useful for you.

• 54:55Speaker 01

  I will say that I'm probably less of a special snowflake than it might seem. Like, especially at Anthropic, we have so many people who come from, like, all kinds of different fields and sort of ended up here. I sort of entered the tech field, like, through this, like, maybe very different lens, which is like a very humanistic lens and, like, thinking about what people want and how to be, like, you know, how do you make people more productive and, like, taking this very human-centric approach. And I think there's a part of tech that really benefits from, like, trying to understand the human more than maybe really deeply understanding the computer. Not to say that that doesn't help me, but I think that's like a component that often makes people very productive. And tech is like thinking more about the human than maybe the computer, right? That's like maybe my answer there.

• 55:39Speaker 01

  And like, certainly Anthropic has a lot of people who think a lot about human impact, society impact, human productivity, and all of this stuff that is like maybe a little less about, do we write this on Rust or C? Like, and I've been doing this long enough now that I also have opinions about memory safety. But I do think that sometimes starting from first principles around what do the user, what do the users really want and need is more helpful in getting you there. The other thing I will say that I think there's like very different schools on, schools on thought on how to grow very good product builders and good software engineers.

• 56:20Speaker 01

  And I've not gone to the University of Waterloo, but I keep seeing their praises because I've hired like a decent chunk of junior engineers in my life. And certainly in Germany, when I went to school, the reason, I was always like really big into computers, but the reason I didn't go and study computer science in Germany was because it was such, basically informatics, like theoretical mathematics that were very far away from...

• 56:43Speaker 01

  Any of the work that you do in in software engineering, the way often explained is like biology versus medicine. We just like in Germany at the time did not have medicine. We only have biology. And if I can't do medicine, I might as well do something entirely different. Right. And I think schools are increasingly getting, at least the schools that like people go to to become good software engineers, increasingly getting better at like really teaching people the skills that are necessary there. And I think the University of Waterloo is like such a good example.

• 57:11Auren Hoffman?

  It's interesting though, because like, I mean, Waterloo has been the same, like, it's been like this for 25 plus years, where everyone always sings Waterloo's praises, where, okay, you gotta, you've gotta, you, you take a semester, you go work, you then you go back to school, then you go work for a semester, you go back to school. So it's very integrated with work and school. Everyone has been talking about this for 25 years. I still don't think anyone's copied them. Like, I don't, I don't even know of another good school that does that today. I don't know why, like, it seems so obvious that somebody else would copy it at some point, you know?

• 57:48Speaker 01

  You think we should, but I think this might be like a bit of a Western phenomenon sadly, which I, I'm like quite sad about. But one thing that's like super fascinating to me is like the number of high schools that are like on the grounds of TSMC factories.

• 58:02Auren Hoffman?

  Oh, really? I didn't know that. That's cool.

• 58:05Speaker 01

  That is super cool, right? Like, if you, if you ever want to go down like a very deep rabbit hole, like go down the rabbit hole of like, what would it take for us to like replicate TSMC in the United States? You, you can't just start like in college, you have to start much.

• 58:20Auren Hoffman?

  So you're learning it from like 16 years old about like how to be like an incredibly talented engineer there or something?

• 58:29Speaker 01

  Yeah, yeah, exactly. And I think this like kind of integration. I think if you think of AI as a moment where education really needs to like maybe take another beat and like rethink how to restructure education, which I'm not an educator, I'm not the right person to like have those opinions. But I, I do think I've seen enough to like at least have the personal opinion that AI should result in some changes in education somewhere. I think this could be another opportunity to think about like integration with, like thinking critically about, okay, what is the theory and what is the application? And how do we bring a little bit of my application into education?

• 59:07Speaker 01

  Would be nice, from my personal opinion.

• 59:09Auren Hoffman?

  That's really interesting. That makes sense. All right, last question we ask all of our guests. What conventional wisdom or advice do you think is generally bad advice?

• 59:18Speaker 01

  As an engineering manager, I maybe go with my pet peeve. You know, I have a pet peeve about engineering management books, like management books in general.

• 59:27Auren Hoffman?

  Sure, there are a lot of, I don't know if I've ever read, I mean, that many engineering management books.

• 59:33Speaker 01

  If you want a book recommendation real quick, like An Elegant Puzzle by Will Larson is great, great literature.

• 59:39Auren Hoffman?

  An Elegant Puzzle? Okay, I will read that.

• 59:41Speaker 01

  Yeah. But like, what conventional wisdom? The amount of times I've heard either speakers or people, like, written books or articles or like essays, talk about how

• 59:53Speaker 01

  here's what you can learn for your team from the Navy SEALs.

• 59:59Auren Hoffman?

  Totally.

• 1:00:00Speaker 01

  And I think, I think the piece of conventional wisdom that I strongly disagree with is that there's much to learn. Like, it's such a different environment. Your startup is not the same. It's like a war zone.

• 1:00:11Auren Hoffman?

  So like, great football coach or something like that, they often write these books that often get passed around to management as well.

• 1:00:19Speaker 01

  Yeah, and I think, I think that's like, maybe goes into a similar vein, that there are

• 1:00:27Speaker 01

  there are teams that are extremely well functioning and have like, I want to say, the most extreme of commitment of every member on the team. And I think the conventional wisdom that I disagree with is that there's too, too much management literature and too much like business literature that sort of glosses over the commitment part. And one of the reasons it's such a pet peeve of mine is because I've worked for, I spent my entire career in tech and like, most people in tech show up, it's a W2 job, right? You show up, you do work. And Anthropic is a company where people do have like a deep commitment, we're much more aligned on the mission and the reason why we work at Anthropic than we are maybe on what the product looks like or what the right path to solution is.

• 1:01:09Speaker 01

  And now that I work again at a place where people are like so deeply committed, I completely changed how I manage. Compared to some of my other past like management leadership positions. And it reminded me of like how useless the advice is.

• 1:01:20Auren Hoffman?

  In the past, you had to like motivate the, your employees, whereas now that that's not is necessary part of your job? Or how have you changed?

• 1:01:32Speaker 01

  Yeah, for this like very simple example, right? Like motivation is just not a thing I worry about at all.

• 1:01:36Auren Hoffman?

  I mean, obviously, a typical thing a manager would worry about is is people on my team learning? Are they growing? Because if they're not growing, probably the best people are optimizing for growth. And if they're if they're not growing, they might be unhappy and they may end up looking for something else. So I'm always trying to figure. I assume that's something you'd still be worried about in Anthropic as well. Or is it just like the work? You know, the work is so interesting that by default, they're going to be growing.

• 1:02:00Speaker 01

  The way I structure my teams and the way I think about team leadership now, like has if you if you like picture any team and then you take every single team member and you imagine them as like a Pokémon, maybe. And you take the slider on like, how deeply committed are they? And how much do they burn for this? And if if they weren't working here, they would, you know, how likely is it that they would still do all the same things? Just with like fewer resources? And if you take that slider and you drag it all the way to 11, the way you run your team chain is quite a bit. You think about different things. And I think I spent quite a bit of like my youth on like, on like, like pre-Olympic trainings for main teams.

• 1:02:44Speaker 01

  Where like similarly, it's much more about like, how do you, as a leader? How do you like unblock people? How do you give them the resources they need? How do you handle the emotional impact of like the wins and the losses? Both of them, right? Like, I think it's a very different job from how do you motivate your people to like really care about this B two B SaaS application? Which is different. That's a different job.

• 1:03:06Auren Hoffman?

  At a, I imagine a professional basketball team, if you're managing a professional basketball team, you have these like incredibly talented people. They're extremely well paid. They know they're good at their craft. They could go work somewhere else, essentially, another team if they want to. I do, but there are probably some analogies to like something like Anthropic, where you have these like super incredibly talented people who are highly compensated. And they may, you know, maybe they don't have quite the same ego or it's quite the same Instagram. Following as a professional basketball player, but there's some, there's, there might be some analogies, right? Or no?

• 1:03:48Speaker 01

  No, there, there are, but for me, this is the first time I'm experiencing that. And it wasn't quite the same. Like, I'll give you an example, maybe at Slack, my teams, people had different reasons for working at Slack. For some of them, it was just a prestigious company to work for. It was a good salary. Maybe they had some passion for chat applications. Maybe they, some of them, like a very small fraction was maybe motivated by idealistic ideas around transparency in organizations. But we were much more aligned on, we enjoy working here and we were much less aligned on like, why are we here?

• 1:04:22Auren Hoffman?

  Yeah. And also like, I like my colleagues. I like by, I like the people I work with. It's, that's a huge motivation.

• 1:04:29Speaker 01

  Yeah. And as a result, I think as a manager, I spend a lot of time about motivation and like, I spend a lot of time thinking about the why are we here question, right? Like, do we have a shared understanding of why are we here? What are we even trying to do? Like, what does that look like? What does success look like? That kind of stuff. And I now, to your point, you're right that Anthropic feels a lot more like this elite basketball team. But for me, as someone working in tech, it's like really my first time that I work in a group of people who, for whom the why question is like so crystal clear. And it's not just upside, which is the other interesting thing. It's also like, if you're super aligned on like this one thing, you're probably less aligned on the other, right?

• 1:05:05Auren Hoffman?

  Yeah, yeah.

• 1:05:06Speaker 01

  What does a good cloud look like is like a thing that we debate all the time.

• 1:05:12Speaker 01

  And that's like, that's like an interesting shift. So to summarize this like very long answer, if, if someone tells you that you should do your, you should manage your team this way because that's how the Navy SEALs do it. I strongly believe that's bad advice.

• 1:05:26Auren Hoffman?

  All right. That's been anything. Thank you, Felix Rieseberg for joining us on Summation. I follow you on X, at Felix. Tell me again, I want the German pronunciation of your last name. I want the German pronunciation.

• 1:05:41Speaker 01

  Rieseberg. Okay, Rieseberg. It's really hard, but like, Google is quite good. You can, whatever spelling you put in there, it's going to find me.

• 1:05:48Auren Hoffman?

  Okay, Felix Rieseberg on X, you're great, great there as well. So thank you again for, for joining us. This has been really fun.

• 1:05:56Speaker 01

  Thank you so much. This is such a fun conversation. Appreciate you.

• 1:05:59Auren Hoffman?

  One more thing before we go. I read a blog called Summation. It's the same as this podcast. The blog is about non-obvious idea sharing on business, talent, data, longevity, and random.

• 1:06:11Speaker 02

  ...intakes. If you like the conversations on this show, you'll probably like the blog. It's free. New content comes out twice a month. You could subscribe at orin.substack.com. That's orin.substack.com.