How AthenaHQ and Browser Use Are Building the Next Layer of the Internet | E2132

25:42Exactly. Helping with both. Okay. I dig that. But what you're describing is the equivalent of white hat SEO, making better content that is more findable and better for both Google's results and end users. Everyone's in favor of that. There are other SEO methods that have been less popular because they're a little bit more nefarious or tricky. Is there any risk of that type of activity working in a GEO era? Or is it harder to trick AI models because they kind of do their own thinking versus kind of Google's great algorithm in the sky? This is a great question that we can get very deep into this.

26:22There are research papers written about these explorers. And I come from a cybersecurity background as well. And I remember the days of SQL injection and this analogy of SQL injection to prompt injection, right? And we don't do any of this black hat, you know, SEO, GEO strategies. The reason is that it's just not, I don't consider it ethical. It's if you read the papers, like there are ways, like if you're have a unguarded a model with no guard rails, like, yes, you can influence it through prompt injection. The joke is like adding white text to the bottom of your website, this sort of thing. But that's not how we operate as a company.

27:02So we don't engage in that. So to be clear, the two guiding principles of Athena HQ is I can find is make sure you're delivering far more value than you're charging for and be ethical. All right. That's got to be the best takeaway I've ever had from a chat with somebody. I really dig that. I'm joking. Y Combinator, you guys are part of Winter 25. I think you raised the full half million from YC. Is that correct? We've raised more than that. Oh, okay. What can you tell me? We've raised over 2 million from investors and angels from companies such as OpenAI, Ethographic, DeepMind. Also at the consumer companies that you expect, like Instacart, Shopify, brands that are going to want to do well on AI search.

27:48Exactly. Now, on the Anthropics and OpenAI side, I don't know if you can tell me this, but are those investments made by individuals at those companies or were they done through OpenAI's venture arm, for example? These were made by individuals. I want to be very clear about that. It's very interesting that people at those companies see the potential in what you're doing. I know you guys were part of the last demo day. I think that's when I first heard of you. How has growth been as a business in the last couple of months? Yeah, growth has been explosive. Our customers are seeing rises in traffic from ChatGPT, Perplexity, Gemini, AI Overviews.

28:30And that's just the tip of the iceberg in terms of the users who are impacted by AI search. So we've been working night and day to meet up with it. meet demand. And it's a space that's also rapidly evolving. So we're also staying up to date ourselves on the latest developments, for example, looking at how different integrations with these model companies are possible, specifically in the e-commerce sense of shopping or different verticals. A lot of surface area here. A lot of exciting things because this is the future of how product discovery is going to be in the next foreseeable future. What would you do if Anthropic just showed up and was like, we'll give you$100 million in stock, just come work for us?

29:19That's something I can't make that unilateral decision without our team. I'm curious about what's going to happen to what I consider to be very promising early stage AI facing or AI predicated startups, because there's a lot of just very wealthy AI model companies. And right now we've seen open AI charted by Cursor and now they're looking at Windsurf. And so I'm kind of curious if there's going to be a suction effect in the market of the model companies, the foundation model firms, just buying people that are doing awesome stuff. And frankly, because I'm a fan of your company, I would put you in that bucket, if that makes sense.

29:54Founders are all about efficiency and performance, and that's why they automate and optimize every system within their startup. But when's the last time you optimize yourself? Well, startups move fast. Sleep and workouts get skipped. We know that. And health becomes an afterthought until it's a problem. But now there is superpower. The ultimate founder health membership, of which I am a member. Full body testing across 100 plus biomarkers. And you get guidance for organ health, hormones, inflammation, and more, giving you actionable data and clear insights. You track your KPIs at your company.

30:27Why not track your health the same way? Well, with superpower you can. I'm using it. Lana Harris is using it. A couple of other people on the team. and they have 150 ,000 person wait list because it's so cool. But because you're a Twist listener, we're going to get you ahead of that line. JCal is going to take you through the VIP entrance at superpower.com slash twist. And you're going to claim your spot in the VIP line. Yeah, it's not even a line. It's just the back door, folks. And you're going to unlock Peak Health today because better health equals a better founder, which equals a better business.

30:57Again, superpower.com slash twist to go through the VIP entrance with your boy JCal. It all depends on how fast you get to AGI. Right. Well, that's either going to be six months to 60 years, depending on who I'm reading. So you've raised more capital. Growth is explosive. How big is the team these days? I know you have a couple of folks on your YC page, but I wasn't able to figure out more. We're a team of four. Our team is lean and mean. If you think about, I model our team after the teams of tomorrow. If you think about the teams that are winning right now, they're the teams like Cursor, like Lovable, like these small tight-knit teams because context is everything, right?

31:36And context is what's most expensive in this startup world that we live in. So our team members have full context on every part of the business. I think I know what you mean by context. And so if you're working inside of like, I don't know, let's say you're working at LinkedIn, which is a superior at Microsoft, you're going to be put onto a team, say a product team, then you're going to be filtered down to like a sub team. and then you're going to do machine learning on how people click on a button. At some point, you're so niched down into the pixels that you can't at all see the forest for the trees.

32:06But on a smaller team where there's a lot of shared responsibility, there's a natural commingling of operational information so that way everyone actually knows kind of what's going on with everything, i.e. greater context. Fair? It's exactly that. Seeing the forest for the trees while also being able to scope down into the tree level of granularity. So here's my question. Four people? Hell yes. I totally get it. Eight? Sure. 12? Getting hard. 15? Ugh. So does this desire for greater intra-team context imply that there's an upper level on how many people can work at Athena HQ? Definitely. That's interesting.

32:44I think we can scale a company with under 20 employees. Including, like, customer success? In the near term. In the near term. So not in the long term. How long is the near term? the next year. If you think about the amount of leverage that each individual employee has, where if you're using the right tools in the right way and you're chaining them together, there is this... I can very much see the road down the line of what Sam Altman describes, that one-person billion-dollar company. I accept one-person billion-dollar maybe a bit further along, but... Yeah, but I mean, directionally, he's drawing an X on the map and telling us we're walking in that direction, which I kind of agree with.

33:23Would you consider it to be like, I'm going to draw a hypothetical scenario here, but let's say five years from now you guys are preparing to go public and you have 50 people. Would that be a win or would that be an underinvestment in human talent? I wouldn't say that's a underinvestment in human talent if the talent density is very high. That's a key metric is talent density of your team. For folks who want to take a look, what is the URL and what is, well, normally I ask what's one role you're having a hard time hiring for? I'm not quite sure that's relevant, but answer it if you'd like, Andrew.

33:58But first, what's the URL? The URL is athenahq.ai. Yeah, the role that we're open to is we're hiring for selectively across roles in engineering and technical growth. So everyone on our team codes and that's part of our culture. We are highly technical. We come from highly technical backgrounds. Our investors are highly technical. Our team members are highly technical. This is our culture. This is how we serve our customers. We are that technical advisor in some sorts to help them adapt to an increasingly technical world of AI search. I really dig it. And when you hit 10 million ARR, come back on the show.

34:40I want to hear all about it and how fast you're going at that point. But in the meantime, Andrew, thank you so much and good luck. All right. Thank you, Alex. That's fun. Yeah. I, for one, am incredibly excited to see what comes next in search. And I think it's going to be a little bit more than just the AI answers we see today. But if the SEO era taught us anything, it's that no matter what form search does eventually take, companies and brands are going to want to dominate it. So frankly, I have very high hopes for what Athena HQ and its competitors like Profound, another Twist 500 company, have coming in the next few quarters.

35:12Now, next up, we're talking to browser use. We've all heard the hype about AI agents, you know, AI tools that can do a lot more than just answer questions or write you a limerick. No, these can go out and actually do tasks for you. I mean, hell, today we're actually talking about agents working with other agents, perhaps even led by a chief agent, and things are only accelerating. But how will these agents actually interact with an internet and all the internet applications that are designed for us humans. Well, browser use has an idea of just how to bridge that gap. Let's learn more. Today we have yet another Twist 500 interview, one that I've been excited about from the very first moment that I discovered this company back during a YC demo day a couple months back.

35:54Essentially, here's my thought. Right now, when I use AI, I tend to go to an interface, I type some stuff in, I click enter, stuff comes back. Very easy, very simple. But I'm a human. What if I was, for example, a different AI system, perhaps an agent? Well, then going out onto the web might be a little bit harder. Getting the information that I want could be difficult. What if there was a tool that can make the whole web a bit more approachable to your local friendly AI model? Well, I think that's what Browser Use is doing. So please welcome to the show. It's Magnus Muller. Magnus, how are you?

36:24Hey, great to meet you. Thanks for having me. Everyone should know that Magnus is calling in late at night from Switzerland. So he made this happen and we are thankful for that. But Magnus, first of all, when was the company founded and how many people currently work at browser use? We founded this whole thing five months ago in January this year when we started with a side project doing our master here in Switzerland, me and Craig, in a builder's house here in Switzerland where it's forbidden to study. It's from the university, but it's forbidden to study. You're only allowed to work on your side project.

36:57And we started to work on this random side project. And we got into YC, spent the last three months. And now, yeah, five months later, we are, it's me, Greg, and we hired two full-time engineers built all together. And if you want to know why we're talking to a company that's so brand new, it's because in five months, they've already accreted as of current count, 61 ,005 GitHub stars, which is a simply insane run. But Magnus, here's the problem. I'm a little bit awash in technology to help AI do stuff. There's Anthropics Model Context Protocol, helping AI kind of bring data in. There's Google's A2A or agent-to-agent framework, which lets agents talk to one another.

37:36Microsoft has a new thing called NLWeb. So first of all, for folks out there who don't know, what does browser use do and why is it critically important? Yes. So we enable your AI to control your browser, to navigate there. Imagine it like you can tell your browser your goal and your computer does zero clicks for you. You don't have to click anymore yourself. Yeah, like you described in the intro, all those websites have those search bars, which are highly optimized for humans, those visual contacts, but agents don't directly understand them. And to get the full power of agents, to have many agents work for you, actually doing things for you, we convert websites into something which agents can understand so that they can take action and do things for you on the browser.

38:22So it's a translation technology that takes... Is it any website that's out there in the world today and using kind of, I presume, a shared tool set, it can convert that into something that an AI can talk to, but it's broad. So everything works with browser use. Yes, yes. Every website works, can convert everything. Some can convert better than others, but all work by default. And you can imagine it as many websites, they have APIs with which you can automate things. But then there are also many legacy systems or things which don't have an API for your specific things. and we enable all those websites without API that your agent can just go there and do things for you like filling out a form or getting back your API key.

39:08You know, all those AI tools like GBT or Publexity, they're very, very good in extracting information from the web. But they all stop where information gets interesting. They all stop at login walls. All your interesting information is behind login walls. And if you have agents who can take action, They can just log in into a bank account and extract your current sales numbers or log in into your company's CRM to create new contacts. Or they can log into my bank account and take out all of my currency. You know, just I'm kidding. So I was going through the GitHub page and I was reading some of the code and I'm not going to lie.

39:43I got a little bit lost. So I'm going to ask you a very simple question. Is browser use the goal to help agents, AI agents, do things on the web? or is it almost an agent in and of itself that can also take written commands and then do stuff for you? We built an agent with a memory system, so it can take hundreds of steps, go down, and can do things for you on the web. Prompt this. You type in your prompt, say, okay, go to my CRM and create a new content. Got it. But all our things are based on actions. and actually click something, do a Google search. And you can simply extend this. So you can write your own action, like calling an API.

40:26Imagine it like before MCP was out. It's basically like you can include more and more MCP tools, which can do just things. And the agent can choose which tool call to do to then decide, okay, what's the best action? Okay, so essentially it's a framework that lets me control the web, but I could bring, for example, some of my own homegrown technology and still use browser use on top of it as almost like a hook onto the web as it exists today. Yes, but mainly browser use is used to actually take action in the web, do things where you don't have APIs for and automate that. So really, it kind of takes the entire human internet and makes it machine understandable and interaction friendly.

41:11Is that fair? Yes. and the main component is really to take action, to actually click buttons to fill out a form, to select from a dropdown. So people have been talking about agents nonstop for the last, I want to say, 12 months. But before technology like browser use existed, were those agents just incredibly limited in what they could go out and do? Or were people kind of building a similar product to browser use internally and then using that to go out and talk to the web with their AI agents? That, of course, depends on the definition of an agent. If you just define an agent as an LLM with tool calls, then all those chatbots which can output JSON structure are usually agents.

41:52But I was always wondering myself when people said, oh, we built those agents. And I thought, this is just a chatbot with which you can interact. And it maybe can query the current weather API for you. So I always imagine an agent like something which actually can take action for you, just like a player who can take action. Okay, that makes good sense to me. Now, I want to talk about the RAM, because you guys, actually, I have a chart here showing the browser use GitHub star growth. And it's one of the craziest things I've ever seen. This is from basically right before the new year, through to May, and you guys just grew very, very, very quickly.

42:32So how did you get the word out about browser use and the kind of first iteration of your technology. And were you surprised by just how quickly the market was like, yes, we want that? Yes, we were very surprised. We built the first prototype in five days, pushed it to hacker news, and it took off. And I mean, I had never built big open source projects before. It got a couple of hundred stars, people created issues. And the main thing is we made it extremely simple for people to run it. We see the setup is run two commands and it runs and you can actually prompt it and it can control your computer.

43:06And that's where I got confused between does it help other people's agents do stuff or is it an agent itself? I think that's where I got a little bit mixed up is the fact that you can prompt it. Yeah, exactly. It's prompt-based. You can do both. And the cool thing is we could create very cool demos with it. It's a really, really cool tool to create demos to go and see, wow, this AI actually does something on the web. And with this, we could also create many controversial demos like applying to jobs online and by just creating very very small not very beautiful demos and sharing them on x that started our growth engine couple of weeks later where then millions of people reposted those examples created their own shared them and i think the key factor for us was there were many things happen in the world like operator came out deep seek worth came out and we jumped on those hype trends we said okay this is the free operator.

43:58And then DeepSeek came out and we said, okay, you can use now DeepSeek together with browser use against operator. And then MCP came out and we always used those external events in the world, which gave us many, many more users. And also, I think when Manus, the Chinese agentic project came out and really had one of the biggest weeks of press I've ever seen, they were using browser usage technology inside of their agent, right? Yes, they were using part of our open source code to convert the website into something which their agents can understand. Of course, they built many things around. Oh, of course.

44:35But yeah, they used the core part of our library and developed it further. And that was really friendly for them to give us those credits and definitely boosted us a lot. All right. So I'm going to just show a quick clip that you guys have up on the GitHub page. And I was hoping you could just kind of maybe tell people what we're looking at as we watch this, kind of do a sports cast, if you will, to help people just kind of get a bit of a visual element. So this is the AI did my groceries example. So Magnus, just tell us what we're seeing here. Yeah, this was a friend of mine, and he is right now building a startup where he does grocery shopping for other people.

45:09And his problem was he always needed to do shopping on his own. So now he created a big problem for browser use. It tells browser use, okay, go to this website and do the shopping for me with those elements. Now you can see, okay, browser use goes to the store and types in, for example, your oranges and new products, and then can click on the actual products and adds them to the chart. And with this, he can automate for his new startup, his shopping process, something she would have done manually before. And this is just a thing of one big prompt. You can see on the right, this is like a big prompt of me, 100 lines of code.

45:49This is just natural language without studying computer science or anything or learning how to automate things. Just a prompt and then it can get started. And one thing I noticed watching this video is that it often draws, well, browser use draws what appear to be like boxes around text. Is that how the system parses discrete bits of information to figure out where different options of potato types would be listed? Or what am I seeing there with the squares, the colors, and the numbers? Yeah. So We combine the actual HTML with the image of the current screen. And inside the HTML, like DOM, all the elements, where we have all the access.

46:31Okay, this is a button, those are the coordinates. This is the core part of our library. There we extract, okay, what are the elements where humans can interact with, like buttons, input fields. And in there, you have also access to the coordinates. So we can just take those coordinates and highlight them on the screen as bounding boxes. And then we can take a screenshot and can combine both and send it to an LLAM. Oh, okay. That's really interesting. So you're taking images of the website's elements and then sending that back to an LLAM to essentially parse it. And then you can tell which thing to select.

47:08And then it can tell the cursor equivalent to actually click on the right thing. That is a lot of compute, Magnus. That feels like a relatively compute heavy. Am I wrong? It depends how you define compute heavy. It's definitely much, much more than normal chatbots. And the main reason is that we have multiple steps. For example, let's say to fill out a form, you need 20 steps. And at every single step, one step, we take the current website and we send it to an LLM. And the LLM sends us back, click on button five. Then the LLM sends us back, okay, go to this page. Then the LLM sends us back, okay, go and fill out this input field.

47:44And every single step we send out the current website, it's around 8 ,000 tokens, let's say around 8 ,000 tokens together with the image. And the LLM gives us back the output. So it's around, I don't know, one, two, three cents per step, which you pay in token costs for GPT-4. Now for bigger, for newer models like Lama 4 or DeepSeek, we can drastically reduce those prices. Well, Lama 4, I hear, is not the world's most exciting model, but it does share something very important with you guys, which is the open source component to it. And I really feel like after watching Databricks become one of the most valuable private companies in the world, people have really decided that open source projects can really build super strong companies.

48:29But I'm just curious, from the early days of the company, your codes on GitHub and such, why did you choose to go the open source route versus a closed source approach? Yeah, in the beginning when we built this first pilot in the first five days, it was just, okay, should we do it close or open? And we said, oh, let's do it open and push it to Hacker News. And I think for us, it was definitely core for the growth engine. Like we sure wouldn't have raised such a round and get all the traction with like a closed source tool just because we get so much feedback from the users and can iterate based on that.

49:02That's what I always thought was the most impactful element of an open source project. Not only can people look at it, tear it apart, tinker with it, and play with it, but they can just tell you what they need and want. And people always talk about startups need to be listening to customers, iterating quickly. Well, what if you had just tons of feedback from smart people who are looking deeply at your product? It just makes a lot of sense to me. But on the business model front, I'm curious about your guys' plans. There's been a couple of different ways people have been monetizing software projects like this.

49:30I presume you're going to offer something along the lines of a hosted version of browser use. So with open source models, you normally have like hosted solutions, you have support which you can sell, and you have enterprise features on top of which only enterprise care about. And right now we may look at the, like in the end it will be a combination of all three, but we have a cloud solution where we host the browsers and the LLMs. So you only send us via an API, the task, we run it, and we send you back the result. It's very neat if you want to run like 10 ,000 of queries which you don't want to do on your own hardware.

50:06And then of course, we have so many features in our product. So some people are a little bit confused. What can they actually do? How can I make this reliable? How can I make this cheaper? And so we can help them to configure those agents to actually save the money, save them token costs. So if you guys offer both the browser's technology and the LLMs as a hosted service, does the end customer get to choose which models they're using? Or do you guys select for them? Like you're all getting 4.0. In the beginning we did this, but because there's such, the biggest cost part is definitely token usage in all of this.

50:43It's like, let's say for QBD4, three cents per step. And if we offer them cheaper models, we can offer it for one cent per step or even cheaper. And I think this is where those agents get really, really powerful. If you can run hundreds of agents in parallel for you for a couple of dollars per hour, I think this is where we see a lot of productivity boost. Oh, that's interesting. I hadn't done the math. If it's one cent per step, running a hundred, it could be dollars per hour, but who cares? Humans cost tens or hundreds of dollars per hour. I'm curious, what are the first customers you guys are seeing?

51:16What industries, what big companies, little companies, startups? I'm really curious who's come knocking on your door. Some are individuals who want to optimize their outreach. Some are QA testing companies who write millions of scripts to test their websites. And the scripts break and they now want to use agents to test their websites. Or even wipe coders, you know, wipe their website, then they need to test it. Just now want to use browsers to test their website. Then a big industry is definitely around form filling. And form filling goes really, really wide. It can be for CRM, in your context.

51:53It can be to fill out. I realize there are millions of people on the planet who do the entire day, nothing other than receiving a PDF. patient data reading the PDF and filling out a form based on a PDF for companies it's just about converting data from one form to the other what's happening in some internal database and there are millions of people who just get an invoice and need to fill out an internal database system with those forms millions of people OCR isn't new why are they still doing it by hand I don't know. It's this, it's this many companies that have never built this, this, you know, you have OCR, you get the data out to fill out the form, you know, it's amazing that some of the stuff I, I guess I just assumed existed, maybe didn't.

52:48And so now we're filling in those workflow gaps with AI tools. Gosh. All right. Well, you, you, you guys also raised, um, I think it was a$17 million seed round, but I'm curious about how the round came together and what it was like to fundraise for your sector in 2025? Yeah, it was definitely one of the wildest weeks of my life. Well, we got all this open source traction. Of course, many investors reached out and we scheduled them all to YC Demo Day. So like beginning of March, all in one week. Okay, so I had like 140 investor meetings in this one. You had 140 investor meetings in a week? Like I started, I scheduled them one by one.

53:24Okay, and then the interesting part was like the week was really crowded. So some investors, of course, they got FOMO and they thought, okay, they don't got a spot in the round. So they emailed me, hey, I really want to invest. Can we meet earlier? And I said, no, we start here. And they said, okay, just take my money, take uncapped saves without valuation, without meeting. And we got like uncapped saves without even meeting us. Okay. And this was before, before a weekend march. So this got us a lot of leverage. Then when we actually started with the meetings, we just said, oh, we could just keep going with.

53:58with the unkempt safe, right? So I guess what's coming next for browser use that people have been asking for, requesting, or kind of pushing you towards? It's definitely for all those agent companies, also all who came before us, like multi-announced. The core problem is always the reliability of those agents. You can create super fast, super cool demos with browser use, okay? You just prompt it, and wow, you have a super cool demo for share on eggs, and you win the hackathon. But to make them really reliable, to get like the last mile done. It's really, really hard. There's where many companies spend millions on fine-tuning those things.

54:35And I think this is the core challenge in this. And this is what we are currently trying with workflow use, that we combine deterministic scripts, which you can record, with agents who then step in to heal those things. And yeah, this is the big feedback which we get from many companies that they want more reliable, faster and cheaper agents. And did you say that you can basically send an agent in to fix whatever breaks. So you can essentially use AI to kind of plug the gaps in existing AI tools. Yes, exactly. Previously, you would have like a playwright script, an automation script, which runs.

55:09And then at one point, the website changes and your script just breaks and says, okay, I didn't find this identifier. If this error occurs, you can just now switch and browse use and say, okay, where should I click on? What's the next step? And it It feels that script, basically. That's going to... You know, I'm always really torn between celebrating technology and our advancements and how quickly we're building new things. And then I also worry about the economic impacts. So I was about to say, that's amazing. But then I'm thinking about other people who probably get paid right now, like the PDF converters, to fix those things.

55:42We're going to have to find new things for them to do. But this is going to be awesome when it's kind of pervasive and we're all just a lot more capable. The awesome thing about this, 99 % of the tasks which people reach out to me They hate those tasks. Like no one likes to fill out 1 million forms per day and just reads PDF and fills out forms. And this is the amazing part. Like I haven't seen any task where people would say, I love this. Because all of those tasks are super repetitive. And I think as soon as you do a task 1 ,000 times, you just hate it. Oh, absolutely. I think in 20 years, we're going to look back and be like, thank God we had AI automate all the boring stuff.

56:20In the meantime, we'll deal with the economics of it. One last question before I let you go, which is San Francisco. You know, you guys are currently in Switzerland as we talked today. Again, thanks for staying up so late. But I'm curious about just how important you think it is to be living in the Bay. I've heard from a lot of founders that they think it's the place to be. But I'm just kind of curious from your perspective, why it's the right place for browser use. And you know, when I grew up, I often went to church. And I could pray from home. But I go to church because there are many other folks and many other people.

56:54And that's why we go to San Francisco, because all the agent builders are in San Francisco. And we don't want to pray for ourselves. So San Francisco is, by analogy, the AI church. It's a church. And that makes Sam Altman the Pope? I wouldn't say so. Well, I mean, I think Sam thinks so. He's still young. Touche. Wait, how old is Sam? I actually have no idea. 45? I don't know. That's old enough to be Pope. He can be president. Yeah. All right, Magnus, when is the next time we should have you back on? What's the next big release that we should have our eyes out for? I mean, we now release this workflow use is super beta, breaks all the time.

57:38So now we will really iterate on this and get a lot of feedback and make this really, really awesome. All right. Well, what's the URL? And just before you go, what's a job you are looking to hire for? Yes, the URL is browser-use.com, that's our main website. And there you can reach all our GitHub projects. And the main job is currently really great builders who are really obsessed, who want to join the hacker house and just love to build all day. Can be young, wide-corders who just want to ship cool stuff and create cool demos. can be really browser gods who have 10 years of experience around fingerprinting for browsers.

58:21Can be people who build the infrastructure for this. But I think there will be many problems which no one has ever solved before. And we just need obsessed, great people who can solve problems. Well, you heard it from the man himself. Magnus, thank you so much. We'll have you back on the show. And in the meantime, good luck. Listen, guys, I've been doing this job for a thousand years. And one thing I can tell you very confidently is that talking to founders never gets old. You learn something, you get a little peek at the future, but also just their enthusiasm always leaves me pretty darn gassed up.

58:51I just love it. We have a lot more of these interviews coming up, so get excited. And if you want to get a peek at which companies might show up on Twist, well, twist500.com for more. And as I said at the top, alexw at launch.co if you want to suggest a company for us to look at. This is Alex. This is Twist. We'll talk to you on Monday. Bye.