Anthropic CPO: How AI Will Build the Next $100M Companies | Mike Krieger — Silicon Valley Girl Podcast

Marina Mogilko: In the AI age, anyone can become a hundred million dollar entrepreneur, build a company solo. Do you think it is possible now?

Mike Krieger: I think it's very possible. Just even watching the journey I went on with Instagram where we did a lot of the initial work with just me and Kevin, we were able to do a lot with just the two of us. There's a focus and an energy when you get when you just have one or two people working on something. I think the ideal is actually two because it's helpful having a partner when going through the ups and downs. But what I've learned is as you grow, every person you add to the team is another person that can bring their own ideas and bring their energy which is great. But it's also another person that you need to get on board if you need to shift where the company is going. And so what I think is very exciting now is that one, two, three person team can scale themselves up, do a lot more than they would have been able to do before, maybe do it faster, and preserve that kind of I call it conceptual integrity. You have all of what matters about that company in your head or in two heads basically working together versus trying to steer a huge ship.

Marina Mogilko: Do you think being technical is crucial to building a large company these days or is it unnecessary with all the existing tools?

Mike Krieger: I think with increasing things like Claude code, with products online, I think you're getting to the point where you're able to at least get to a version of your idea where you can see if there's something there. You know, because that's the biggest thing that I get. Like in the age of kind of mobile apps first emerging, I would hear from people all the time and they'd be like, "Mike, I have a great idea for a mobile app. Like, do you know any mobile engineers or how do I get it built?" And that was often where the idea would die. And some of those ideas maybe wouldn't have worked out, but some of them might actually have created something really novel, right? There would be people that said, "Oh, I have a novel idea for a social app or a dating app or a game." And so many of those ideas just kind of died at that stage. And so what I think is exciting is even if the tools aren't there where they're not necessarily going to get you from your first 10 users to your first million users in terms of the system scaling up, they'll at least get you in front of those first 10 people, right? Like get you to the prototype or even that initial version. And that seems very possible even for people who aren't technical. We see that inside Anthropic where some of our non-engineers, you know, people in marketing, people in communications who just have had an idea in the back of their heads for a while will sit and talk to Claude and work with Claude code on the weekends and then share it internally like look what I built and I don't code at all.

Marina Mogilko: So what would be your road map? Come up with an idea, build an MVP, share it with like 10, 20 people. You had experience where you had to shut down the company, right? Can you tell me what the product was?

Mike Krieger: It was our second company. I started a second company with the same co-founder as Instagram and it was called Artifact. And the product itself I'm super proud of. It was an AI powered news recommendation product. So we would learn your interests and then we would give you a feed of you know different articles and stories that you might be interested in. And like at a really nuanced level, like you people might be interested in not just graphic design but you know Bauhaus graphic design or Japanese architecture—like really understanding the kinds of things that people were more deeply interested in. But the product didn't hit enough of a product-market fit and an acceleration that made it make sense to continue to invest in it. So that process of having created something and iterated on it and built a team around it but then also coming to the realization that the right thing was to move on from it was really challenging.

Marina Mogilko: And the pressure was on, right? It was right after Instagram.

Mike Krieger: It was, and you know in some ways that was really a hindrance.

Marina Mogilko: When do you make this decision about an idea that you showed to maybe like 20 people? Yeah. When do you say like hey this is it or do I keep pushing?

Mike Krieger: I think for me it's very much when you're in that phase where you're making changes and showing it to people and they're giving you feedback and maybe even have some people using it—does it feel like the more you change or the more you listen to feedback and make some alterations, is there like a snowball effect where energy is increasing? We definitely saw that in early Instagram. We had initially a group of maybe 20, 30 people using it in just a private beta and you know it was very much a weekend product because you're out and about and especially in the early days that's all you were doing was just taking photos with Instagram. It wasn't that you were uploading photos from somewhere else or your camera roll. And so we would find that we would work all week and then on Friday we'd ship a new version to those like 20, 30, 40 testers and then we'd see what would happen over the weekend. And every weekend those changes that we were making, we'd hear from people, oh, I tried the new filter, I tried this thing, it's going well. So that's what I look for a lot. And I think where it's time to maybe either pivot the company or move on to something else is when you feel like, man, I've tried every idea that I have or I've added 30 things and none of them are really sticking or people are kind of okay about it but they're not excited about it. And that's maybe where you might want to step back and say there's probably energy better spent somewhere else. Like, it's not scientific, but when you feel like one unit of input you're getting 10 units of output, that's a good feeling. When it feels like 10 units of input you're getting one unit of output, it's time to pivot to something else.

Marina Mogilko: What fascinates me most in these conversations about AI is how our consumption behavior changes. I wonder what it looks like in your life. In my life, I think like 20 to 30% of purchases are made through an LLM. I ask Claude to research something for me or ask other tools to pick the best product for me. There's actually statistics that 60% of searches end up without a click. Attention is scattered. And yes, more and more buyers are doing research in LLMs instead of following a traditional funnel. And businesses have to adjust. You have to be generative engine friendly. We're moving from SEO to GEO—generative engine optimization. Let's talk about people who already have a business. I've heard you talk about how people are not using AI enough in their businesses. What do you think are the top maybe two or three cases how entrepreneurs can use Claude to automate some operations which doesn't require you to be very technical?

Mike Krieger: Yeah. So I was actually talking to a fellow friend who's a second or third time founder and he was telling me he's like, "Mike, I'm really glad you launched Claude Max because Claude is my product manager. Claude is my lawyer. Claude is my you know, founder therapist as well." And what he does is he has a Claude project for each of those disciplines. So he has his product manager Claude, he has his contracts Claude, and he just uses that for all of those things. It's let him run a very lean initial company overall and do those pieces. And so even though he happens to be fairly technical, he's not coding with Claude really in his day-to-day. He's actually set up Claude to be sort of a mini version of some of these disciplines already. Even as the models continue to get more and more powerful in those ways, even today with the right context and the right sort of history around something, you really can start having these sort of per-job-function thought partners in there. And for an entrepreneur, for sure, you know, have a model that validates your idea or sees what you might be missing. Another one doing you know competitive intelligence to understand what's even out there and understanding those pieces. Of course helping you build with code is another part of it too, but even beyond that, the way I like to think of it: when you grow a company, you want to hire. I want the best CFO. I want you know the best head of product, et cetera. When you're an entrepreneur, especially like I was a first-time entrepreneur without a big network, you're doing all those things yourself. So why not try to bring the best of what we know into that company via something like Claude?

Marina Mogilko: What do you think is going to happen in two or three years? How are we going to progress and advance?

Mike Krieger: I think part of it, and I was having this conversation this morning even with some of our researchers, the progression or the dynamic that I see is: if last year maybe the beginning of this year, these models—and I'll use Claude as the example—were really kind of assistants that were helping you with maybe a question at a time or maybe a task, right? This year they're moving more into collaborators where, especially if you watch somebody use Claude code for example, they're actually delegating what would have taken them maybe 20, 30 minutes and Claude's going off and doing it and then you're checking in and you're more on the sort of validating or verifying work, right? Going into next year, you're going to delegate even bigger chunks of time or even like pieces of the job out to that. So it's not just I have this very specific thing, please go do it. It could be more sort of think of it, Claude, as something that's in the loop of your business. So, hey Claude, watch for any new user feedback and maybe propose a change based on that. And you can imagine evolving that to don't just propose a change, write the code to make the change and I'll check if it's a good idea or not. And then the next step there is Claude actually being an entire sort of coworker where it has a discipline. So not just a project that happens to be a good thought partner for product management, but an actual product manager in your company.

Marina Mogilko: Like being active, right? Is that what you mean? When it, I don't know, look at your ads and says like, "Hey, I'm going to change this."

Mike Krieger: Exactly.

Marina Mogilko: So when do you think that's going to happen?

Mike Krieger: I think that is within, you know, it depends a little bit on the discipline. With coding sort of tasks, I think within the year we'll start seeing Claude be able to do those things. And I think in some of these other disciplines maybe it's more like two or three years, but it's not wildly far away. And it's not that it's going to do everything perfectly well. There's still going to be this validation human-in-the-loop aspect of it. But I think the big shift is going to be this autonomy and proactivity where you're not having to give it explicit instructions every time and instead you can almost describe what kind of role you want it to play and then it will be able to play that role as long as you've connected it to the right you know data sources.

Marina Mogilko: And how active you want it to be, right? Because now it's basically just you chatting. I want the model to be talking to me as well proactively, like hey I just noticed this.

Mike Krieger: That would be cool. Exactly.

Marina Mogilko: Half a year ago Dario Amodei said 90% of code would be written by AI. What percentage of code is written by AI at Anthropic now?

Mike Krieger: It depends on the product. But for our products that are sort of the closest to being written completely by AI, so Claude code for example, the development of that is using Claude to develop Claude. It's almost entirely.

Marina Mogilko: How has it changed the way you work?

Mike Krieger: Well, it's let me—I'm technical but my day-to-day often is in you know external commitments or doing management. It's let me remain really technical which is fun. So because we've oriented the codebase to being able to be written by Claude, it means somebody like me can also contribute as well. So we launched our Claude for Chrome just earlier this week and I had a little bit of time on a business trip and I said, "Oh, I have two hours free. What am I going to do?" I was like, "I'm going to contribute to this thing that I know we're shipping in a couple of weeks." And so it's broadened the group of people that can contribute for sure. I think it's also kind of moved where the bottlenecks are where we need to pay special attention to. So I think we've put more time up front now into all right let's get clarity across our team about what needs to get done so that you know we don't have this like amazing technology that can code really quickly but the engineers are actually not even sure what the thing is that needs to get built. So that still remains very important—being aligned, having that consistent vision. And then on the last part, which is that you know code being ready to get committed and get reviewed, the number of pull requests and changes that are going into our codebase is just dramatically accelerated. One of our developer productivity engineers shared this last week we're just on an exponential there too. So we've had to re-engineer those systems too.

Marina Mogilko: Crazy. Wow. Claude for Chrome. You mentioned it's going to fix email, something that happened for us like yesterday. I just came back from a business trip. My kids went to school and they were—because they don't really, some of them like Lily who doesn't speak English, she doesn't understand what's going on. She was like tomorrow's pajamas day and I was asking my AI like is it really? And it's like no, no, all good, nothing's happening at school. And of course we come and it's pajamas day.

Mike Krieger: Pajamas day. Yeah. Like oops. And it was interesting. The demo we had in front of the company was like Claude triaging emails and it's still you know probably too slow right now still to be like the main way in which you interact with your emails. But I went through for example I had a lot of pending LinkedIn invites but I wanted to you know ignore the ones that weren't real but I wanted to make sure that if it was somebody from Anthropic or it sounded like they knew me that we would like make sure that we captured those. And so I had Claude just kind of triage that whole inbox for, you know, an hour actually and it went through.

Marina Mogilko: Did you have to code something or you just asked?

Mike Krieger: Just asked it. I was like, "Hey—"

Marina Mogilko: That's Claude for Chrome, right? Just ask it to go through.

Mike Krieger: Oh, wow. Yeah. Visit LinkedIn. Look at my invitations. Be your assistant in your browser.

Marina Mogilko: Exactly. Can you share some personal practices for how you use Claude?

Mike Krieger: For me, it's anytime I've written something. I found that I still want to take the first draft myself because you know, sometimes writing is thinking and that's really important that you're actually, at least for me, that I'm kind of expressing myself through writing. But before showing any human basically anything I write that's of substance, I'll basically tell Claude and say, "Hey, I'm writing on this. Like what am I missing? Please challenge me on what I haven't said yet." And sometimes it'll give you a suggestion. And it ranges from I can't believe I forgot to address that. It would have been really embarrassing to share this document and not have this. And then sometimes it's, oh wow, I wasn't even thinking about this dimension. It's something like very new and different. And actually, I use it less for copy editing, but I use it a lot for challenging my ideas and figuring out how to, you know, understand what a very smart person looking at this would ask as the next follow-up question and then can you kind of go from there?

Marina Mogilko: Do you still type? You said you're writing or you would do voice?

Mike Krieger: It's kind of a mix. Another thing I've done with Claude sometimes when I'm like I'm a little bit of writer's block, I need to get started, but I still want that experience of working through an idea myself, is I will turn on the voice mode and just talk to it for 20 minutes sometimes and at the end say, "All right, that was a lot. Now, can you organize that into some kind of, you know, real cohesive document that I can send?"

Marina Mogilko: When do you think it's going to happen that AI generates ideas better than humans? Because where I see this going, you have this marketing AI, you have this product AI. What if we just combine all of this and tell like hey AI go identify niches in the market where you can build a business, build a business, launch ads, make me money? Like when do you think it's going to happen?

Mike Krieger: Well, it's an interesting one. We ran an experiment internally. So we have a whole team that's really starting to focus on sort of Claude in the real world or frontier applications of Claude and what can we learn by studying it. And one of the things that we had it do is actually run a vending machine. And now it's multiple vending machines here at the office. People could talk to Claude. It's a great paper. It's called Project Vend. And what's interesting is in some cases it's very capable. Like it was able to track inventory, place backup orders, like do all these things, interact with the people who were talking to it. But as a pure sort of business, it still made some mistakes like overestimating demand, you know, overcharging or undercharging for certain things. So I think one aspect that we need to develop before it's fully ready to be sort of an entrepreneur in a box is a little bit more of that business sense and how to understand the market like what is actually sellable et cetera. But often we find with some steering or some feedback it actually can do very very well. So maybe initially it's still going to be all right, partner with a person. Let's go, you know, maybe like business school 101, like identify a problem or a market gap, try it out, learn from it, see what worked and what didn't and go and build from there. But what I think is also really interesting is that you can also set up some experiments around like what are the things that people are searching for? Like what are the demands? Like what are the feedback that other products are getting that maybe a new product could feel better? And then have Claude both brainstorm the idea and then also go off and build it.

Marina Mogilko: Do you think it's more for you as an entrepreneur, is it more encouraging or discouraging when AI is getting so powerful?

Mike Krieger: I think for me, you know, having watched over the years, the friction to starting something new in some ways has decreased. In other cases it's about the same. You know, it's there's more resources out there about what is it like to incorporate or raise money than there was when I was doing even Instagram for the first time. But building stuff is still, you know, can still be a challenge and that needs to continue to get easier. So I always felt like there was more good ideas than there were you know products that kind of surface those ideas out there. So I think it's exciting to have more of that out there. Partially also because if the cost of creating that company goes down, we should also get new funding models. Not every idea needs to be a huge venture funded, you know, billion dollar business. But if you have to build a whole tech team around it and you have to go hire, then it's going to require some funding. And so I'm very excited about call it the next five years of companies that can scale from the you know world-changing venture-funded crazy ambitious companies, but also the ones that solve a problem really well for a particular demographic and maybe remain a one to five person team along with AI.

Marina Mogilko: Yeah, this is where I think this model—because I raised venture capital as a creator, like I was the first creator to do that. I feel like this is the age where you invest in humans because now they can iterate through so many ideas in a year and you want to back someone who has these ideas and can execute.

Mike Krieger: 100%.

Marina Mogilko: For all the entrepreneurs who are watching, where should they start today if they want to start?

Mike Krieger: I mean I think there's two things that I think you'd think about that are unique to the moment versus like generally about starting companies. One is not building for where the models are today, but where you think that they can get to in one or two model generations. By the time you've built it, the models will have done that move on. And so the way you do that is to push the models as hard as you can, like break them almost and say, "Wow, wouldn't it be great if—" One experiment somebody ran internally was like they gave Claude their whole to-do list, you know, in like reminders or something, was like, "Hey, Claude, go do all these things." I mean really quickly hit a point where well it didn't have all the information about me so it can't do it. Or this is before we had Claude for Chrome, it can't use my browser so it can't like go and you know fill out this form that I needed to fill out. But then you can imagine well what if it could? All right, should I be building towards that and does that feel like a five-year journey or like a you know two months and six months? And things are accelerating. And so that's like a very big piece. I think the best companies are building ready to take advantage of the fact that the models will continue to improve, and your advantage as an entrepreneur is that you don't have a legacy codebase. You don't have people using a product in a particular way. You can build for that you know kind of inflection point.

Marina Mogilko: It's just so hard to predict like if the model is going to take over your whole business niche, right?

Mike Krieger: But I think that you know, and it definitely hard to predict in the three-year range at that point, you can also kind of evolve the business. But like seeing what models were bad at six months ago or are better at now or like okay today but could be better at in the future. I think that can really come just from like constant contact with them.

Marina Mogilko: And just start iterating.

Mike Krieger: What would be your advantage compared to a model that could do the same thing in a year?

Marina Mogilko: Yeah. I mean then yeah, it's very much so. Exactly. If you can create that, um if you're building at that edge and you understand your customer segment then you're going to be the one that is going to best be suited to say all right you know what also got better? Like this aspect and I get how that's going to connect to my customer segment and I already have the relationships. Because people are still, they seek trust, right? And you know somebody who understands their segment and has the relationships is going to be a more trustworthy source any day than like a brand new person or a company that they've never interacted before.

Marina Mogilko: Yeah. And I like how you said in one of your interviews that every product has this vibe.

Mike Krieger: Yes.

Marina Mogilko: And if you're building something with passion, then it definitely has your vibe, right? That some people are going to connect with and that's going to be yours.

Mike Krieger: Exactly. Right. And that's your voice coming through there, even beyond whatever AI is going to provide.

Marina Mogilko: What would be your advice to entrepreneurs who are looking to start a company? What niches do you think have the most potential?

Mike Krieger: I think people are still fundamentally human and still have like very human needs. So I think a lot of the aspects I still think about being really important are, for example, people's health—both physical and mental. And even though there's the beginning of AI helping with those aspects, I think there's so much more potential around that as well, right? How do we relate to ourselves? How do we develop our own understanding of where we're showing up well and where we could be showing up better? How we can continue to evolve kind of as individuals but as also as teams or as partners? So that's another area that I think I'm excited for more companies to get formed in. Fitness and performance coaching is another place that's important there too. I think there's also going to be this renaissance of being, you know, into the physical or the real world. You know, you had very much this like social media thing that was, you know, Instagram at its best was helping people get out and see the world in a different way. At scale, maybe it became more about like watching what other people were doing than going out yourself, right? So how do we get new companies and products that are about, you know, either exploring something new, getting to see the city through a new light, participating in a civic way that's different? I think not all of these require AI, but done right, they could be encouraged by AI.

Marina Mogilko: What should entrepreneurs be focusing on? Is it getting expertise in the niche where they're building or just being focused on like identifying those niches and iterating ideas?

Mike Krieger: I think a lot of the most important companies are going to come from understanding those niches or those businesses. So a friend of mine is building a company in the construction space and she went to all of the construction conferences, got to know people really building in that space and understanding what is like uniquely important about that. This reminds me a lot of when I was at the Stanford D School, that whole process of developing empathy and storytelling and really getting to the heart of what the needs are—where technology actually can have an impact on an industry or a person in a positive way. That ends up being very very important. You know, I've also met another entrepreneur who works with AI for a very specific type of legal practice and they're, you know, going off to, you know, states in the US that aren't necessarily the most tech-forward, spending a lot of time with people who are in that particular discipline. So that kind of expertise is I think what is going to differentiate the companies because it's going to be easier to build, but that knowledge of like what actually to build and how to build very specifically for that is going to be what's really valuable.

Marina Mogilko: What about marketing? Because it feels like okay ideas—I can generate a lot of ideas. Claude can help me. Then I build the product with Claude or Replit or Lovable. What's going to happen to marketing? Everyone's talking about like hey every what should be on social media because otherwise how do you market your idea? What do you think?

Mike Krieger: That's definitely evolved from when we were building Instagram. The way you got your you know product out there, especially a social product like Instagram, was for us people sharing their Instagram creations to Facebook and to Twitter at the time and then seeing that link and then finding that. It's definitely shifted to much more of this sort of creator, you know, whether it's Instagram or on TikTok, and people telling those stories and discovering those products. It's interesting watching that evolution now and kind of wondering like well what comes after that as well? You know, is it going to be still like algorithmic recommendations? Are we going to return to oh like word of mouth—I really like this thing and it solved my very specific problem and I'm going to share it with you? And maybe there's some combination of those two that needs to happen? But I completely agree that what will also differentiate, beyond an understanding of a market, is also the ability to tell the story around that.

Marina Mogilko: Yeah. What was going to happen to content? I think in two or three years because my feed is like maybe 50% AI something—like crazy going on or I see an AI generated image.

Mike Krieger: It's been interesting. Like I try using social products from all around the world and for a while I was using a social media app that was like very popular in China and the thing that struck me was that you had creators that were really good at AI generated sort of fantastical architecture and they had a real following even though their work was all generated. Because they had a voice. And so I think that is the next piece that not a lot of people are doing well, which is yes you can be using AI for generating videos or text. But if it's just generic or if it's just, you know, kind of a bunch of random things that you manage to generate, I don't think that's going to generate following. Or I think that's not going to generate a real presence on these platforms. Whereas I remember there was a photographer—an early Instagram photographer, like probably one of our first 10,000 users. And he took a photo of the same hill like outside his window like every day. And it was like that was like a photographic practice that he had. And that was really cool because then his followers could see that he had a view of the world and kind of a perspective. And I think the same thing needs to be true to succeed as a creator, even if you're using AI tools.

Marina Mogilko: Yeah, that's fascinating. Let's talk about people who don't have entrepreneurial ambitions, but they're like, I want to work for a company. Your hiring strategy at Anthropic has changed recently, right? You're hiring more high-level people. Is that right?

Mike Krieger: It's kind of a shift. So the way I think about it is you want people that are going to be defined more about the problems that they want to solve and how they can creatively solve them than a very specific sort of I know JavaScript and I'm going to do work in this exact environment. And so I'm still really drawn to the folks that on the weekends or in their you know spare time are prototyping ideas they came up with. They're bringing that even to the interview process. You know, some of our product managers that we've talked to have had like a culture of experimentation that they've been working through. And I think that's very interesting. I really look for that when I'm hiring. Interesting thing is it's not so different than when I was hiring for Instagram. The people I always gravitated towards were ones that, you know, if you talk to them on a Monday morning, they'd tell you all about the new thing that they got interested in. Doesn't have to even be tech-related, but there is that sort of curiosity.

Marina Mogilko: Excitement. Yeah.

Mike Krieger: Exactly. And I still absolutely look for that as well. And then in terms of the kind of makeup of the teams that we want to build, you know, we haven't had a like summer internship program. So we've tended less to hire like kind of fresh college grads. But I've in talking to a lot of people who maybe are going through college now or recent graduates, you know, they're also enabling themselves via AI to sort of understand the market, understand the dynamics, be entrepreneurial in their own way, even if they're not starting their own companies. And then coming into companies and telling that story.

Marina Mogilko: How do you learn that? How do you learn how to become entrepreneurial or generate ideas? Are you teaching your kids that?

Mike Krieger: Well I mean our kids are pretty young. I do try to encourage the idea of like observe the world and be you know open to what is there. You know the thing I did early on was I had a little notebook that I would carry around everywhere and basically just always kind of writing and ideating and like making space for that sort of quiet ideation time, you know, is really useful. One of the ways I've seen people use Claude—if you ask Claude for one idea, it'll generate an idea. Maybe it'll be a good one. But the thing that's unique about it is it can also generate 50 or 100 ideas. So a practice I find is useful is give it you know five or 10 and then ask it to fill in the next 30 and maybe one of those sparks something else. There's a lot of research where asking people to brainstorm together in a room is not as effective as asking one person to write down 10 ideas and pass those 10 ideas to the next person, have them kind of generate the next 10. So you can simulate that same process but with AI.

Marina Mogilko: What is your ritual for generating ideas? You said quiet time. Do you go on a walk or—

Mike Krieger: Walks are really helpful for me. I also just like any kind of repeated action. I really like, like I got one of those Peloton rowing machines and I found something—you're like okay it's a very repetitive action, but then your mind can kind of start wandering and start thinking about what those different things are. And also you need to give your brain the opportunity to make new connections while you're not focusing on it. And then maybe later it'll come and say oh yeah that actually does connect really well.

Marina Mogilko: Is it like every day that you make time for this?

Mike Krieger: I wish it were every day. It's more probably like once or twice a week these days. But I always every time I do it I'm like I really should be doing this more regularly. And all of a sudden I'm excited to like sit down and like put pen to paper and then share that with the team.

Marina Mogilko: What about going back to kids? I have two small kids and sometimes I read these articles where like oh people who just graduated from college are struggling to find jobs because a lot of entry-level jobs being replaced with AI. And I'm like okay I'm worrying about myself but now I have to worry about them as well. What are you teaching your kids? What do you think they're going to do in 10 years when they go out to the job market?

Mike Krieger: Yeah, it's super hard to predict like what that market looks like for a six and a four-year-old. So the what might change a lot and even some of the how will change. But I think there's a few things that are important.