Jacob Luetzow: What's up? I'm Jacob Litzo, your Elixir Mentor, and welcome to another exciting episode of the Elixir Mentor Podcast. This is where we discuss everything related to Elixir from interviews with enthusiasts and pioneers in the community to innovative projects and libraries shaping Elixir's future. so welcome back, Philippe. I think it's been like two years since I had you last on, and then I got to meet you in person in Chicago at Elixir Conf, and that was awesome. We got to bond, Filipe: It was a great event. Jacob Luetzow: we got to bond over some beers and Just talking tech and it was a lot of fun. I feel like there's a good group of developers there. Filipe: For sure. It's it it was one of the highlights of the year, by by far. And it a great event. Jacob Luetzow: Yeah, it was it was a lot of fun. So I feel like last time you were on, we talked a lot about super base and real time, and a lot has changed. Like you got a lot of really cool side projects that you've been working on. and I'm sure we'll dive into all of them in detail. But you have vitro, why can't I ample? Ample your yeah, am I saying that right? Filipe: Ample, yeah. It it's a really weird name. I think so. Honestly, I'm not one hundred percent sure on the pronunciation, but I just felt Jacob Luetzow: Ha ha ha. Filipe: okay, an ample is packing liquid, so Jacob Luetzow: Yeah, yeah. Filipe: it kind of it's fitting for Elixir, which makes very Jacob Luetzow: I like it. Filipe: bad it it makes very bad SEO because if I look for Elixir Ample, I'll get a lot of air treatment products. Jacob Luetzow: Yeah. I recently rebranded Kill Switch to Kin Switch and 'cause there's just so much so many brands out there with like Filipe: Yeah. Jacob Luetzow: kill switch in the name and it was just like all these mechanical switches and it was like, Come on, it's not it's Filipe: Yeah. Jacob Luetzow: really not what I'm going for. So that was one of the reasons I rebranded. Filipe: Yeah, and at least one met yeah, and at least one metal band, a skill switch in the name, which is also Jacob Luetzow: Yeah. That's also true. Yeah. I think you actually told me that, which is hilarious. Yeah. No, Filipe: I did yeah, then again repeated. Jacob Luetzow: that's okay. It was never said to, you know, the public. Filipe: True. but yeah, yeah, just Jacob Luetzow: y yeah. Filipe: we can start, which is I'm still first of all, thank you for having me again in in the the the podcast slash stream. again, I'm a big fan of this stream, but also the work you're doing. plus again you you end up bringing a lot of the good sides of the community. So really glad to be here again. Jacob Luetzow: Yep, that's you, the good side of the community. Filipe: No, I'm the I'm the na you we've picked, now we're going to the slope. Exactly. Exactly. Jacob Luetzow: Now now it's just like scraping the bottom of the barrel for guests, you know. Can barely find What well what do you want to talk about first? Do you want to talk about vitro? Or Filipe: yeah, we can talk about Vitro. Yeah, we can talk about it. it's Jacob Luetzow: 'Cause I feel like that it's practically a product. are you are you making money with it or not yet? Okay. Filipe: Yes. No, no, not yet. I haven't launched it. I'm st I'm still afraid of launching it because the usual cycle is it's extremely fun to build. I'm using it daily. both the local version of it because I've I've actually done two versions of this of of Vitro. There's one where you can basically run it from desktop and that one will be free forever. that's the idea. There will be a hosted version. And that one, since I need to pay for the compute, that one will need to be paid. but yeah, the idea on that specific one. sorry, not the idea. The current stage of it is I need to just do it and launch it. I'm just afraid of the marketing steps. every developer goes through this. Jacob Luetzow: Don't be scared, just do it. Filipe: Yeah, that yeah. But it it's so that I think that's the point where things start to get a bit more real. And Jacob Luetzow: Yeah. Filipe: If it gets way too real too fast, which is not the case, I doubt this will be one of those that will pick up customers on day one. But it's one of the Yeah, I can. Yes. Jacob Luetzow: Can you just explain what it is in a few sentences and then we'll continue diving in? Filipe: It's observability for browser usage via MCP. So usually we already have a lot of these tools of you can access the browser via MCP, it will do whatever it needs to do. The local version basically also is Chrome MCP. A lot of people use that. But it has a lot of caveats. For example, in the current age of development, you cannot expect to have only one project ongoing. You kind of need three or four at the same time. If you are a serial site project person like I am, you'll end up with a lot of annoyances. Which is: okay, I just want to have my agents test. three or four ideas, run multiple persona tests, just do stuff. And Chrome MCP is actually quite bad for that. It cannot have concurrent sessions. Plus it keeps opening tabs in your and and Chrome windows across your laptop, which is way too annoying for me to deal with it. Jacob Luetzow: I I actually have an extension project out there. and it does a lot of like web scraping and it's really hard, like you can't really use headless browsers for a lot of things, and yeah, it just pops open windows or tabs and Filipe: It's it's it's the best way, but it's extremely annoying. Jacob Luetzow: Yeah. Filipe: And so I was facing this problem and the more I faced it, the more I got annoyed with it. So I ended up launching I built the local version first. but then I started to think I actually want this on on some other spot because I'm I'm starting to use way more, for example, cloud code from my phone where I want to explore an idea. And I don't have browsers there that are as easy. To get into. So I wanted, okay, I need this remote thing, not only for this specific problem, but for other kinds of workflows. So I ended up building the Osted version. The biggest differentiator from Vitro to other Chrome MCP stuff is the fact that you actually get a timeline of what's happening. So a lot of the that's one of the biggest problems I have with and most MCP tools and most AI-driven workflows. It's your lack of the observability. You don't know what the agent was doing for the last three or four hours. You don't know Jacob Luetzow: Right. Filipe: what sessions were created. You don't know which parts of the UI you've tested. It's just way too annoying. Jacob Luetzow: I feel like that's just using any clawed code or chat GPT or codecs I should say. Any of them is kinda like just magic background workers and you you don't really know. Filipe: Exactly. And I wanted to have a bit more trust on what my agents were doing. And that timeline flow helped a lot. So now I can actually dive into okay, what was the browser doing? What was what was happening during this session? And it basically I can see the timeline of events, I can see which MCP tools were called, I can see error logs from the console, I can see the clicks that the agent did within the UI. And this can be actually multiplied by how many sessions you have. So that for me was a key change where I can actually see, okay, what the hell were you doing for the last two hours? And it Jacob Luetzow: Mm-hmm. Filipe: basically replays the full MCP interaction with a Jacob Luetzow: That's r Filipe: given session. Jacob Luetzow: yeah, that's really cool. I like that. Filipe: Yeah, that one was when I understood okay, this can be a product. Jacob Luetzow: How how do you think it w how is it though how is it to manage though? if you're working on multiple projects and you have a few agents per project, is it just end up being clutter and noise or how do you filter that? Filipe: We actually we I say we as vitro. there's only one employee which is me. But yeah, vitro basically you are allowed to pull the screenshots out of the C Jacob Luetzow: Timeout. I say we with Kin switch because I count all my Grokbot agents as employees. Filipe: Exactly. When I say we it's me and my all of my agents. the ones that died and the ones that are working. Jacob Luetzow: It's all of my l fake friends. Filipe: And yeah with With this, you can actually pull all the screenshots out of the either the desktop version or the hosted version and you can ask it, okay, from this, give me the where did you click? Where what give me a small showcase of the meaningful moments where something happened because the tool call is associated with a screenshot. So you'd be able to okay, pull all of this. So pull the one hundred sessions you did to measure performance, to measure if given persona was able to use it. because that's actually one of the main uses c use cases I've been s doing, which is I basically I use one skill that is extremely good called impeccable, very design focused, that has a lot of capabilities of saying, okay, this is accessibility focused, this is how's how's the website structured and so on. So if you actually have that in the context, you can then inform The agents, I want to do persona testing. And you have ten sessions, one guy that is that has some impairment in terms of I don't know, visual impairment, the this person doesn't only is able to c use their laptop using keyboard, they're not able to use a mouse. And it basically goes through Jacob Luetzow: Mm-hmm. Filipe: all of that and just tries out your website as an agent. It's it's really weird because it feels like black magic. You shouldn't be able to do that that easily, but it really increases the quality of what you're building because it actually finds a lot of pitfalls. And Jacob Luetzow: Yeah. That's a great idea. I never thought about doing that. Filipe: It's very good, trust me. And if you go through it and then you can actually collect all of that information and you can attach to which persona you attach the meaningful screenshots of what needs to be fixed, for example. You can do small presentations or better yet, your gr your agents can do presentations on these were the pitfalls in the product. These are things that this persona mentioned as important. Just doing this in in an iterative way that you can then actually pull the feedback from it and see, okay, what Why did this person do need I don't know twenty calls of clicking to actually do something? Even that element of this is the timeline of events makes a huge change in terms of your final product. And Jacob Luetzow: Yeah. Filipe: vitro kind of unlocked a lot of that and I just basically leave it running and give it to agents and it does everything. And again, it's using MCP, which simplifies a lot of it. It's a one-click integration. That's it. and yeah. I plan to release it. I pl it's already live. Jacob Luetzow: You need to release this. I need to start using it. I'm ready. You'll have a customer on day one. I'll sign up. Filipe: Finally, yeah. Thank you. I'll give you a code, of course. I'm not so you'll be my first zero MRR customer. But I'll I'll I'll for sure give you. Jacob Luetzow: Here's the thing. I I love that it's like MCP focused too, because I've been on this whole kick where I had a guest on that is just kind of doing his name's Adam. He has bloggy, right? And like you think Filipe: Mm-hmm. Jacob Luetzow: like very simple product, right? You're just writing good SEO ranking blogs, like anyone can train an agent to do that. But he's actually like He created a product that does it like so well, it integrates with Google Search Console and look watches your keyword rankings and things. So it's actually like a tool that I've found that's worth paying for, even though I could probably re-implement it. I just don't want to put my focus on that. Filipe: Yeah, yeah. Jacob Luetzow: But he like created a whole MCP for me because I was like, look, I want to use a good tool, but I don't want to learn how to use it. I want my agents to know how to use it and just leverage what your tool does. Filipe: And honestly, from my experience, MCP is the perfect interface for that. Jacob Luetzow: I agree, yeah. Filipe: yeah, and there's a reason why. Basically, we had a period in AI development where new protocols were coming out, new interfaces were coming out, new ways of doing things were coming out. It's it was a constant rolling ball of innovation that you need to keep track of. The one that Jacob Luetzow: Mm-hmm. Filipe: stayed for longer and truth to the original intent was MCP. And you can clearly see when those conversations of CLI versus MCP, whatever happened. I think again, people go to the side of being too extreme in this conversation because it's not one or the other. You can actually have both. They can use Jacob Luetzow: Yeah. Filipe: the same surface. So it's a case of not being an extremist and just choose whatever you want. And to work on cloud stuff, if we think about it, the future won't be free compute. That's already heavy subsidized. Jacob Luetzow: Mm-hmm. Filipe: The the idea will be services that are smarter, that give your agent capabilities to do stuff they're not able to do in their own box. Jacob Luetzow: Always thought that before I saw the power of what you can do, just having an MCP attached to your product, for instance. I always was like, Whoa, it makes more sense. so many companies are going to dial in an agent that's like an expert in their product and their customers, and you know what I mean? But now it's like, no, Filipe: Yeah, yeah, yeah. Jacob Luetzow: actually, everyone needs to come with their agent and they're going to use your tools. Cause like I use Grokbot, I use Claude. And it integrates with my product Kin Switch. I can just, hey, add a new entry. Like I keep track of all my podcast notes in Kin Switch. So I like Filipe: Yeah, I had exactly. Jacob Luetzow: I have a whole I have all my questions laid out, but it's just like, hey, investigate Philippe. Here are our projects he's working on. Put some show notes in there for me. And it gives me just like a nice like layout for all my guests. And then obviously I read through it. If it doesn't make sense, I make corrections. Filipe: Yeah. Jacob Luetzow: But it's really good. And I don't have to click around and even use the Kin switch app. I just talk to my LLM and it just does it. Filipe: Exactly. And that's the other part, which is a lot of people focused on the idea that you're going to do development in your laptop. And that's true, you're going to do development in your laptop. But there's a lot a a bigger layer of problems you can solve by not being in your laptop. Again, Jacob Luetzow: Mm-hmm. Filipe: KinSwitch is a an excellent example of that because again you'll have a mob mobile app, you don't want the compute to be expensive. Even a local model that has MCP capabilities And if the MCP is good enough, even a local model will be able to grab really small, quick chunks of it and just push it to KinSwitch. A lot of the compute needs to Jacob Luetzow: Mm-hmm. Filipe: be off your laptop. It will be distributed in your phone, in some cloud agent, in your laptop. It needs to be everywhere. And the interface that has been the best for that has been as MCP. and Jacob Luetzow: Mm-hmm. Filipe: CLI as the very tricky element of it needs to run on a good enough machine. So And and again now imagine I know that Muse, for example, has two gigabytes of memory and two vir virtual CPUs. They give you this. Again, that's why I say it's Jacob Luetzow: Mm-hmm. Filipe: heavily subsidized, because if you do this for all WhatsApp users, you're you don't have enough machines and boxes to give two gigabytes of RAM Jacob Luetzow: No. Filipe: and two V CPUs. It's impossible. Jacob Luetzow: I think we're in for a rude awakening when VC funding dries up, you know? Filipe: Yeah. Completely, completely. I fully agree. And again, I think we're going to when that bump hits, the kind of company that will stay is the one that is smarter with compute compute usage and does delegation of Jacob Luetzow: Yeah. Filipe: compute. And you then start to have very optimized workflows and very optimized jobs that know when to stop using set compute. You cannot expect to have a running agent for twenty four hours, twenty four seven, forever. with the prices of of data centers and the consequences of having too many data centers. That's the other part which is more complicated. Jacob Luetzow: Yeah. Filipe: because data centers will become for sure the political topic in every country for the next couple of years. And a lot of them will be shut down and not moved. Jacob Luetzow: Well, and it's that complicated it's like that complicated subject, the double edged sword, right? you need them b well, you you want your country to have them because you want your country to be more advanced in, AI technology, but no one wants Filipe: Yeah, than others, yeah. Exactly. Yeah. Jacob Luetzow: a data center in their backyard either. Filipe: Exactly. Yeah, and that will become a huge subject. And as that will become more more impactful as VC money starts to dry up and subs subsidies that now allow you to have two gigabytes of RAM and two V CPUs, that will dry up. All of that will become a bigger problem. So this concept of you have very low power machines, way smaller, kind of one V core and a couple of giga just one or Jacob Luetzow: Mm-hmm. Filipe: two not even one or two, sorry, two hundred and fifty six megabytes of RAM, for example. Stuff like that, that will be the thing that drives the MCP calls and drives what you need to actually pro pro how you progress your AI agent. So I'm very bullish on that idea of compute needs a big overhaul, and MCP is one of the tools that actually allows you to do that very solid in a very solid manner. Jacob Luetzow: I don't know. Sam Altman doesn't think we should have Ram in pr our personal homes. Filipe: Did you see did you see that he backtracked on that? I wonder why. Yeah, he backtracked. Jacob Luetzow: did he? That's good. I didn't see that. Filipe: It kind of went, no one it's it's very unresponsible to have RAM in your house and now we completely backtrack from that statement. I wonder why. Jacob Luetzow: Well, 'cause it's a such a foolish statement. Like what? Filipe: And th then you have other companies that are actually seeing the the advantages of having local models. So for example, I think it's What was the name? It wasn't Ugreen, Ubiquity, I think. They actually have now a a box that you can install in your home that has local models that do stuff for you. With smart home, Jacob Luetzow: That's nice. Filipe: yeah, with home assistant and stuff like that. And again, they might use MCP to do that. And then they can use MCP to talk with other things that will allow you to do stuff Jacob Luetzow: Mm-hmm. Filipe: at home. Probably they won't be testing stuff in vitro in browsers. They don't need to test stuff in browsers. But again. Jacob Luetzow: Well Filipe: Kin Switch would be a great great example of having in an home setup. Jacob Luetzow: Well, what really interests me and vitro is I actually have a a QA grok bot that does smoke testing when I take things to prod. But vitro seems like I could really focus it a lot more on a really solid QA process, even for accessibility Filipe: Yeah, yeah. Jacob Luetzow: and all those things. Filipe: Yeah, and it also includes Lighthouse, so you're also able to check the performance of each webpage. Jacob Luetzow: Okay. Filipe: It yeah, it does a lot based on on that idea. You Jacob Luetzow: What about is it agent ready? Do you have that kind of stuff tied in? Filipe: what do you mean, Agent Freddy? Can you Jacob Luetzow: Have you it is it agentready dot com? Who let me see, is is it agent ready.com? I think it's made by who make Cloudflare. Filipe: I don't think I have that. I need to check that. Yeah, is site agent ready? Jacob Luetzow: and then so what's really nice is where you lack it gives you prompts to fix your website. Filipe: Ooh nice. Okay. Jacob Luetzow: Yeah. Filipe: I'm going to run that on the background while we're doing the podcast. Jacob Luetzow: Yeah. Yeah, there we go. Filipe: That's that's the trickiest part now with AI is ADHD is becoming a bigger problem. You're always going around. Yeah. Yeah. Likewise. Yeah, and the other thing is sleep Jacob Luetzow: Dude, it's so bad. I'm doing three, I do three projects nonstop 24 hours a day. Probably even more than three. Cause Filipe: deprivation. I I Jacob Luetzow: Dude, I'll wake up sometimes at three AM just to prompt my my Claude to start a new feature. It's ridiculous. Filipe: I'm exactly on the same boat. It it looks I'm ki like I'm talking with some sort of this is my I don't know my side family at 3 a.m. kind of going in. Yeah, kind of going Jacob Luetzow: Your side piece. Your secret family. Filipe: Yeah, my secret family is a bunch of cloud agents. And I'm just going, yeah. Potentially, Jacob Luetzow: Your wife your wife gets suspicious and then she just finds it's grok. Filipe: yeah. No, I lowered the the iPhone 17 as that one the d very dim. screen we just go fully DIM and go, Okay. How how are you how are you? Exactly. Jacob Luetzow: You need the privacy screen cover too, so you can't see it from the an angle. Filipe: What are you doing? I'm talking to my agents at three AM or four AM. But yeah, th I do believe that's one of Jacob Luetzow: Yeah. That's hilarious. Filipe: the that's one of the things that will be very difficult is attention management and healthy living. because the the the fact that if you are a per a person that likes to solve problems, this is this is kind of the only grail of solving them. Jacob Luetzow: Yeah. Filipe: and it's at his it has been very tricky to kind of balance that properly. And Vitro doesn't help with Jacob Luetzow: Yeah. Filipe: that because then I have multiple agents running end to end tests and then I want to see the feedback, I get the screenshots, I see what's happening, I get reports, so it Jacob Luetzow: What does token usage look like if you have so many browser sessions and agents essentially? do you really burn through them? Filipe: It's not that bad because you can, for example, throw it to Sonnet. Sonnet is good enough at doing Jacob Luetzow: Okay. Filipe: end to end tests. So if you're what I usually do, first of all, it's your own tokens. I don't I don't do sub Jacob Luetzow: Okay. Filipe: I don't you don't you don't need to pay for the tokens. You're basically just paying for the possibility of having a machine running for a bit for you, killing it, and you just pay the microsecond of usage. and the token usage is actually quite low if you use Sonnet, for example. If you want a more complex problem with topus, it's also not that bad because again, a lot of the if you have the for example the impeccable stuff, impeccable burns a lot of tokens to do the design work. Jacob Luetzow: Mm-hmm. Filipe: But if you go, Okay, give me the type of user and persona that you you think I should have in my website, and then you say to Sonnet, okay, delegate each one each agent does this type of persona and extend a bit more the scope of them, use Sonnet for this. It's not that heavy. and again Jacob Luetzow: Okay. Filipe: you'd Sonnet has people are always praising Cloud Opus five point five, but Sonnet for a lot of jobs it's more than good enough. And even the other Jacob Luetzow: I need to be better at leveraging the different models for different work Filipe: yeah. Jacob Luetzow: because I feel like I just will like stay on like Opus five point five and not think about it, but it's really not that efficient Filipe: But even for example, Jacob Luetzow: to do that. Filipe: I'm building a new harness on the side now. And for example, a lot of the page another Jacob Luetzow: Another side project. Here we go. Filipe: side project. I have at the moment I think I have five or six side projects. Jacob Luetzow: Ha ha ha. Filipe: again, it's a problem. And I I became very efficient on on using other models that are lower tier exactly because I was burning my max plan way too fast. Jacob Luetzow: Yeah. Filipe: So I need to get good at that. And even for example, a lot of people use Opus to fetch data from websites. That's not needed. You can go ICU and ICU is good enough to do a curl, and from that you then s ask it to send that to Sonnet. Jacob Luetzow: Yeah. Filipe: Sonnet cleans it up, you send that to Opus. So you can kind of do a cascading agent ladder that does things for you in a more efficient way. Jacob Luetzow: Yeah, yeah. I actually you know Jason, Jason Alum. Filipe: Yeah, yeah. Jacob Luetzow: so he he recently made a it's called brasscheck.bot. It's like a ammo price web scraping thing. So it finds you like the cheapest price Filipe: Yeah. Jacob Luetzow: for ammo. But so Filipe: Yeah. Jacob Luetzow: my my review pro product is like an Amazon product, it scrapes products. And for the longest time, I didn't want to support other countries because I was manually setting up my scrape recipes. And then I Filipe: Yeah Jacob Luetzow: had like a 15-minute conversation with Jason and he like opened up my mind. Like now I just like grab Filipe: About it, yeah. Jacob Luetzow: snippets around the component that I know the data exists and I send it to my back end. And then I use Sonnet to if the recipe breaks, it fixes it. in iterations and so it's like self healing now. I've web scraping self healing and I can support every single country Amazon store now. And it's pretty cool. Filipe: Yeah, that that's it. That's it. It's that's the kind of stuff that really agents really unblock you. It goes it Jacob Luetzow: Mm-hmm. Filipe: it really changes the way you think about solving problems. And a lot of it is automation around automation and improvement. It's insane what Jacob Luetzow: Yeah. Filipe: what agents are able to do now. Jacob Luetzow: Do you hear feedback from I have a three D printer going and I'm so curious if you hear it? Okay, cool. That's another ADHD project now. Now I like three D printing. Yeah. Filipe: No no no, I can't no no, not anything. All good. Another yeah, another side project. Have you tried Blender? Blender I think Blender now has AI connectors, right? Jacob Luetzow: dude, I haven't. I need to. Let I'm gonna pull that up in my browser so I don't forget. Cause I haven't used Blender Filipe: I think I at least I Jacob Luetzow: in years. Filipe: But I think now you can actually run Jacob Luetzow: Let's see. Filipe: agents against it. I think. Jacob Luetzow: That would be cool. We'll find out. I'll look. Filipe: Ha ha. Jacob Luetzow: Cause I have a nephew that wants a 3D printed. so his brother has like a T-Rex head on his wall, like a huge 3D printed T-Rex head. Filipe: Mm-hmm. Jacob Luetzow: And now his little brother wants to have a wolf. So I gotta figure out how to make a wolf. Filipe: Now you know. You go to Cloud Code and ask for it. And then it it lo it it will look Jacob Luetzow: Dude. Filipe: l more like a cat than anything probably, but Jacob Luetzow: So I started making these doodle discs where you spin it and you draw a picture. And Filipe: Yeah, yeah, yeah, yeah, yeah. Yeah. Jacob Luetzow: our six-year-old freaking loves it. He comes in and begs every day for a new disc. And I'm like, this is wild. Filipe: A new one. Jacob Luetzow: Like he puts down the Nintendo Switch because he wants a doodle disc. Yeah. Yeah. I know. So I'm just like, I've Filipe: okay that's way that's really good. That's extremely good. You found a solution. Jacob Luetzow: I know, I'm like, I need to sell these on Etsy for holiday packages. Filipe: Yep. Yep, you should. Honestly you should. The Jacob box. You just have doodle discs. Jacob Luetzow: There we go. And then I have this. I don't know, like I think you've probably seen it. I've shared it with a few people, but I have my patio defender project I want to build with nerves. And it it monitors my patio boundary. And if a chicken sized object comes across, it gets squirted with a squirt gun. So so they don't poop all over my patio furniture. So that's gonna be my my big IoT project that I distract myself with. Filipe: You should get one on the there's I know there's a kind of a super soaker from Xiaomi that has eye eye pressure mode, I think. You should check that one. And just truly fire it. Jacob Luetzow: perfect. I'll have to check it out. I found this like 10 volt like 80 PSI water pump that I'm was gonna try to use. Yeah. I have to make sure it doesn't hurt the chickens though. Yeah. Just Filipe: god. Which is not it's also a good idea. You'll end up killing a chicken with that. But yeah, sorry, big sidetrack. Jacob Luetzow: Just I just wanna sprinkle it, Yeah. Filipe: I don't think that will happen. I think the you'll end up killing a chicken with a water gun. Jacob Luetzow: I hope not, I would feel horrible. My poor little chickens, my little ladies. Filipe: But yeah, again, big sidetrack, ADHD becoming a thing. It is a problem. But yeah. Jacob Luetzow: It's a problem. Yeah. It happens a lot on this podcast. Filipe: yeah. I it's easy. But yeah, the Jacob Luetzow: Ha ha ha. Filipe: the thing is I'm actually very happy with the current state of things. And I'm very if anything, I'm very afraid what are the next stages of this. when sub when for example, Jacob Luetzow: I it Filipe: yeah. You go ahead, you go ahead. Jacob Luetzow: I think it's just a matter of time and like what we do day to day, we won't be doing. Filipe: Yeah, I agree. we'll it will reach a point of you're going to learn how to use IQ and Sonnet because Opus will become the fable of the family and things will be very, Jacob Luetzow: Yeah. Filipe: very expensive. I'm pretty sure that will Jacob Luetzow: Yeah. Filipe: end up happening. Hopefully not. And local models actually get really good. Revolut actually did that now. I'm not sure if you saw. Revolut Jacob Luetzow: Mm-mm. Filipe: is going to rent GPUs and having Yeah, a lot of them, but they'll rent them and have GLM five. I think it's five. But yeah, basically having an open source model instead of spending money on Frontier Labs. Because it will be cheaper long term for them. Jacob Luetzow: That'll be yeah. I'm actually excited for the Frontier LLM stuff to just get some wind under its wings and I think it'll be really useful. I do think there's gonna be a huge benefit to having like home LLM like servers where you can leverage you don't have to burn Turk tokens that can do like daily tasks and help you manage your calendars and your Filipe: Yeah, exactly. Exactly. And if not Jacob Luetzow: Your smart home things. Filipe: and even for I did try some ohm models and s there's some stuff that is already good enough. Deep Seek is good enough, GLM is good enough. Jacob Luetzow: Nice. Filipe: a lot of them are already okay at doing a lot of this stuff. If it's low token count and you can wait for if you don't need a live output of it, because that's one of the annoying parts, inference just takes a long time and you don't have a lot of tokens per second. But imagine that it's Jacob Luetzow: Mm. Filipe: okay, do this. And if you're okay with waiting two minutes until you have a reply, some of them are already good enough. Even and MacBooks are extremely good at that. If anything, a Mac mini with a good enough RAM, d due to the unified Jacob Luetzow: Yeah. Filipe: memory thing, it's just way better. Jacob Luetzow: And that's why Mac minis are so hard to get a hold of right now. Filipe: Yep, yep. The the Apple didn't know what they were building at the time, probably, but they were building the best Jacob Luetzow: Yeah. Filipe: AI machine in the market for home stuff. Jacob Luetzow: Yeah, no, that's pretty cool though. I I have like one of those skylight calendars, you know, where Filipe: Mm-hmm. Jacob Luetzow: your whole family can see on the touch screen. And their app is so annoying to use. And I actually found a project on GitHub that they created their own MCP that interacts with their API. And so I schedule Filipe: See Jacob Luetzow: and create tasks through. Claude on my phone and it puts it on the calendar. It's so awesome. And I just run Filipe: Yeah, again Jacob Luetzow: a local instant of the MCP on my server and then it does its own thing. Filipe: That's the other element that it's quite interesting. recently I hate I have a Dyson robot, the the cleaning robot. The app from Jacob Luetzow: Mm-hmm. Filipe: Dyson. It's one of the worst apps ever because I'm just trying to clean my house and I have ads for for air dryers and stuff like that. I don't I don't want that. Yeah, it's very Jacob Luetzow: that's annoying. Yeah. Filipe: annoying. I was actually not me, again, my agents, were able to reverse engineer the app and now I have my own Dyson app. And it's running on my phone. Jacob Luetzow: Nice. I did that or I started doing that with my router, my mesh Wi Fi system, because Filipe: Yeah. Jacob Luetzow: their app sucks. And for what Filipe: Yeah. Jacob Luetzow: I'm like partially like I have all the endpoints and things figured out and I just had Claude dissect and reverse engineer their Android app. Yeah. Filipe: Yeah, it's the same. It works brilliantly. And it just worked out out of the out of the box. And now I actually I actually run routines from the robot that were not possible in the app. I'm Jacob Luetzow: That's awesome. Filipe: I have better monitoring. I can push him to do stuff that it w usually wouldn't be able to do. So and it took one afternoon because since Jacob Luetzow: Yeah, that inspires me. I'm gonna continue with my my router because like the parental controls on it are so annoying. They'll like just break or they'll like Filipe: yeah. Jacob Luetzow: hold on to old settings and they won't t take the new settings. So it's like I can't like unblock, you know, one of the kids, you know, if we Filipe: That's so annoying. Jacob Luetzow: yeah, if we lock them out, I can't unblock and block and it's annoying. Filipe: Yeah, that's just again, there's there's that new wage of you just do your own your your own applications. Jacob Luetzow: Yeah. Filipe: Vitro started as that. The artness I'm building now also started as that because I'm as much as I like console driven development, I think we've reached a point of we're losing a lot of good information out of it. And you actually Jacob Luetzow: Yeah. Filipe: should have UY components for your for your agents, meaning that if I want to visualize data. I don't want it to go and build an HTTP micro server that's building a web page. I actually want my artist to be smart enough to okay, draw me a graph. That's it. And show me the Jacob Luetzow: Mm-hmm. Filipe: information in the session I'm running, not somewhere else in in another web page or whatever. And Jacob Luetzow: Yeah. Filipe: again, we're we're in the age of micro applications. And the thing is it does open the door for a lot of innovation and Also to break a lot of things. So it's it's Jacob Luetzow: Yeah. Filipe: quite fun. Jacob Luetzow: Well, so many apps that we saw as like solid products, like social media planners and schedulers and all of that. You just I just tell Grok to schedule stuff. And like what's really cool is so Filipe: Yeah. Which opens the problem of Jacob Luetzow: I can so I can visualize it. I make it Filipe: Mm-hmm. Jacob Luetzow: integrate with my Google Calendar and then I see when they're posting and it's it's great. Filipe: But again, it opens the the the that tricky conversation of what will be before we had a lot of indie development, even what I'm doing with vitro. I'm I'm glad I have a normal job because it would be very tricky to kind of okay, I need to rely on this. I'd be Jacob Luetzow: Yeah. Filipe: sh I'd be scared as hell with the current prospect, which is we've seen more and more companies firing, being people being replaced by AI. Some companies just became obsolete. That's the reality. A lot of companies Jacob Luetzow: Yeah. Filipe: that were selling multi multi thousand dollar plans, they are no longer needed. And they have a lot of people in their ranks that are no longer required. And it's Jacob Luetzow: Yeah. Filipe: a very tricky subject and it it's a Jacob Luetzow: That's Filipe: very impactful thing. Jacob Luetzow: part of me wants to have some sort of physical item. That's why I I got the 3D printer and I'm starting to play around Filipe: Yeah, yeah. Jacob Luetzow: with IoT stuff. Cause I feel like products will always be useful, but software is just getting easier and easier to make. And I know we're in a bubble. We're surrounded by people that know how to build software. So I think it's scarier to us, but that's still Filipe: Mm-hmm. Jacob Luetzow: not the average person. Filipe: Yeah, that's true, that's true. But at the same time, from personal experience, even even so at the moment I'm the team lead of real time at Supervise. And even internally, I've had multiple questions internal questions to myself, which were more around do I hire more people or not? And at the Jacob Luetzow: Yeah. Filipe: moment I don't I don't see a need for it. Which at in one side it's very scary. On one side it's good because you end up with a lean team, faster development, Jacob Luetzow: Mm-hmm. Filipe: low conversations, low low everything, and just you're just able to speed up and develop a lot of things. But at the same time it's very scary because you end up in this situation of okay, I'm not hiring, I'm not contributing to hiring people, which is scary as hell. Jacob Luetzow: Yeah. And I think what another problem is we are we have awesome stacked teams of senior devs and up, and no one's gonna train the junior devs. Cause you still have to know Filipe: Agreed. But I think Jacob Luetzow: how architecture and everything works, right? Even if we're not writing code anymore, you still have to know how it all goes together to make a good stable system. Filipe: Completely. even so for example the first iterations of vitro, they were very bad. It was Jacob Luetzow: Mm-hmm. Filipe: trying to use playwright, left and right, it wasn't doing proper deployment of what I needed, it was very slow. And again, I needed to really push it towards a direction that makes made sense for the development. And even with with newer side projects, I still need to do a lot of okay, don't go there, go here. I still need to guide it a lot. Which again brings up the the problem of there's still there's still a human in the loop and it there's there will be always a human in the loop. Because at the end of the day, an agent is a lot of mathematics that as Jacob Luetzow: Mm. Filipe: just calculating the best next token based on a lot of context, but it's not able to have everything. The context is limited. And you need to wrangle that context Jacob Luetzow: Yeah. Filipe: properly. So humans will always be needed. The problem is How do you create the next stage of humans to be here? Which is the new developer. Honestly, if it was now, I wouldn't go to university. I just spend my day building stuff. Jacob Luetzow: I've said that forever. Well, I went when I went to college, I I went for structural engineering, so not even programming. Filipe: Ha ha. Jacob Luetzow: But like hindsight, I would like all all that did was give me a piece of paper to prove I'm capable of learning at a certain level, you know? Like, yeah, sure, Filipe: Yeah. Exactly, exactly. Jacob Luetzow: I can do math. And I think that's what engineering degrees prove, right? But it wasn't necessary. Not really. As long as you're capable and willing to put in the time to learn whatever you're learning. Filipe: You'll you'll be happy and and proceed and be f successful. I actually I was in Lisbon AI. We were very lucky to have a v very, very good conference in Portugal focused on AI. and I've met a lot of I've connected with a lot of pe a lot of people there, and there were two people that stand out to me, which was one guy that was twenty years old, something like that, twenty or nineteen, and he was already working for a stealth startup doing AI AI work. He was already an angel investor, because he saved enough money to actually Angel invest in something. He was able to buy a house in San Francisco and he's nineteen. And I I was thinking, okay, Jacob Luetzow: Dang. Filipe: where I yeah, where I in nineteen I was getting drunk and getting ti and Jacob Luetzow: I know I was an idiot. Same. Filipe: getting tired and getting tired of okay, now I need to study calculus. Why? Well, it Jacob Luetzow: Yeah. Filipe: did it has its ad advantages, but at the end of the day I'm not using it that much. so it didn't give me that much either. But the and there was a another guy that was doing a world model, meaning that you you would be able to train robots to do physical stuff in a virtual world. So you can actually basically scan a factory with your mobile phone, it would build a model of that in a game engine, it would have context awareness of what it was. on on the in the the world. So for example, if he saw a table, it would understand that's a table, so the robot could move the table. So it actually created Jacob Luetzow: Mm-hmm. Filipe: physical objects in the world. And it will be used for training robots heavily. So you basically do thousands of sessions. Okay, let the robot eat his head against the lamp until he understands that it needs to move away. The guy was Jacob Luetzow: That's cool. And then you can pretty much just upload that knowledge to a physical robot and it's ready to Filipe: Yeah. Yeah. Jacob Luetzow: go. Filipe: Yeah. And that guy was twenty two years old. also also college dropout that by sixteen was doing Linux C yeah, but Jacob Luetzow: I was not that motivated in my twenties. That's the problem, I only cared about girls and having fun. Filipe: exactly and beer. But again, I I think Jacob Luetzow: Yeah. Filipe: that there's a new wage related to this, right? Because the tools that this new generation has will unblock them immensely if they go use them. The Jacob Luetzow: Right. Filipe: when we started coding a couple of years ago already, Jacob Luetzow: Just a couple. Filipe: just a couple, yeah, you would actually see designers going through what we need to go through now. Designers would be the person Jacob Luetzow: Mm. Filipe: that would arrive to an office with their portfolio, with their work, just even on the side, they needed to do stuff just to have a richer portfolio and they would deliver this to companies and showcase their work. Jacob Luetzow: Mm-hmm. Filipe: Developers are going through the same stage. It's better to not go to uni, develop a lot of site projects, build your portfolio in GitHub, make it public, do whatever you need Jacob Luetzow: I agree. Filipe: to showcase it, and that's your way in. We're going to have a year shift. Jacob Luetzow: Well, I've gotten to the point, like with my recruitment stuff, I don't really like doing take home projects and recommending random people I don't know anymore to positions. It's more Filipe: Yeah, yeah. Jacob Luetzow: you're way more likely, if you're in my circle, to get a job if you need one, because I'd much rather just vouch for someone that I've worked with or already know has the skill than try to look through AI slop from a take home project. Yeah. Filipe: Yeah. Yeah, exactly. And for me, I I'm also doing a lot of hiring in in Supervise. A lot of the work I do when it comes to hiring, it's not looking at the C V or seeing whatever. It's basically going to their GitHub. And it's Jacob Luetzow: Mm-hmm. Yeah. Filipe: completely okay to see AI developed stuff. We're looking more for the creativity of how did this person build something? What did they build? Was it interesting? Was it not? Did it solve their problem? Jacob Luetzow: And is it architected and are are the modules still laid out right? are they organizing their project? Filipe: Yeah. Yeah, yeah. Jacob Luetzow: is it ha does it have tests? People that don't write tests for their co it's free to write tests now. Filipe: Yeah. No, that's a y Yeah, it's a huge red flag. It's a huge red Jacob Luetzow: Mm-hmm. Filipe: flag. and Cloud Code is not good at that. I have multiple Jacob Luetzow: No. Filipe: things where I need to push it to okay, I want coverage. It's sad that I I needed to go back to my managers from ten years ago where coverage was a metric. Jacob Luetzow: speaking, I have an awesome skill that actually Jason Alum wrote that runs scripts and things and looks for gaps in your test coverage. And it actually points Filipe: Ooh, okay. Jacob Luetzow: the it actually points Claude exactly at the missing lines that need code coverage. Filipe: Jason needs to share sh share that a bit more widely, yeah. Thank you. Insider, insider, yeah. No, no, that needs to be Jason. Jason needs yeah, it's a tricky one. Jacob Luetzow: I'll yeah, yeah, I'll share it with you. I'll share it in our our our special Discord server. I don't know if I want to publicly say the name of that one. Filipe: but yeah, I do think that we're going through a very interesting change of of learning model, a very interesting change in workflows, and The typical way of working changed completely. Jacob Luetzow: I don't know if there you are, you're back. You had a little I don't know if it's me or you, but something's lagging. Filipe: No, it's fine. I again probably the agents working on the background. But yeah, Jacob Luetzow: Yeah. Filipe: as I was saying, I I think we're going through a change in the way things work. And for example, Jacob Luetzow: Mm-hmm. Filipe: university universities will feel this. I bet that next year and for the next couple of years we'll see a drop in people that join colleges and stuff like that. Because they don't see the need to do that. Why do you go to Compsci at the moment? Is there really a good reason to go there? If you can build your most weird system possible and get a job by doing that? So I I Jacob Luetzow: Yeah. There's a there's a really cool organization here, called IHub Filipe: Mm-hmm. Jacob Luetzow: and it's run by a very successful billionaire entrepreneur guy. And it's like you can join for free, right? You are an entrepreneur, you have a business idea, but you have to you have to, lay it out and get accepted. So they have to be like, This is a good idea. But you join and you get all these free resources. You lose a 1% of your company to IHub. And then you get attorneys and all these other people who give talks and it's like amazing. Filipe: Again, goes Jacob Luetzow: So it's like an entrepreneur network. And that's gonna I think that's more likely to have a future than like normal university Filipe: Fully agree. Jacob Luetzow: style education. Filipe: I fully agree with that. And not only with that, that actually is a great idea because one of the biggest most complicated sp steps is it's legal, just doing a privacy a privacy thing of okay, what do I need to put in my website regarding privacy? Cloud code can kind of do it, but do you really want to trust Cloud Code to do that? Yeah. Did he hallucinate some some law? Jacob Luetzow: Mm-hmm. But do you know if it's right? Yeah. cause here's the thing. We all know Claude is really good at writing code, but we also correct it and find mistakes a lot. And that's because we're really Filipe: Yes. Jacob Luetzow: good at writing code. So do you trust the legal terms that it's spitting out at you not being an attorney? You probably shouldn't, right? Yeah. Filipe: Yeah, exactly. You shouldn't. You shouldn't. there's always a layer of validation that needs to happen by a real proper Jacob Luetzow: Mm-hmm. Filipe: human. and with law, again if we think about it in a very pragmatic way, it's so much context and so much work to understand it that your AI agent either is very well trained on it, and even that I wouldn't trust it. There's certain things that I at the moment I wouldn't trust AI to do. And I Jacob Luetzow: Right. Yep. Filipe: d I agree with you that that's the best way, plus You'll end up with something in in your portfolio. You will be an ex founder, which is way more valuable at the moment Jacob Luetzow: Yeah. Filipe: than a diploma of I did this. Yes. Yeah, exactly. Exactly. Jacob Luetzow: You gotta stack those failures and then you get you get one, right? You get a couple after. Filipe: but yeah, it's quite an Jacob Luetzow: I'm just too stubborn to stop, so I keep creating things until they're not failures. Filipe: No the thing is, even what we call failures, it's like you said, it's you're just learning. There's a new thing, there's yeah, exactly. There there I've Jacob Luetzow: You're learning a ton, right? And you know where I've always stopped when actually we were talking about this before we went live, and you mentioned it a little bit, but vitro, you're scared to take it to market because that's real, Filipe: Yes. Yep. Jacob Luetzow: and now you have to do marketing. And that's Filipe: Yeah and Jacob Luetzow: usually where I always stopped because you want, like, well, Filipe: Yeah. Jacob Luetzow: I've built it, they will, they will come. No, no, they will not. Yeah. Filipe: Yeah. No, they won't they won't. And I actually have two I have two side things that I have to launch, which is vitro and Sunloop. Sumloop is basically cost control for your sales. You're basically able to track your spending of your sales and because I've been in multiple companies and this is done with an Excel spreadsheet, which has Jacob Luetzow: Mm-hmm. Filipe: pissed me off immensely. Just doesn't make any sense. And even that one, that one is actually have been has been built for two months. Completely It's done, it's complete. But I'm afraid of the extra step. And V2 has been the same. the marketing step, it's the scary one. And Jacob Luetzow: Yeah. Filipe: it's cool that I'm able to talk about the side projects, but it would be even cooler if I actually launched them. So I need to do that. And I I just need to bite the bullet. Yeah, I just need to bite the bullet and do it. But it's so of Jacob Luetzow: You need to. Yeah. I feel like you're always building cool stuff and it's about time you release a product. Filipe: Yeah, I need to. And vitro will probably be the one that I'll I'll release with more care. The other one, I know it's not the most appealing thing, and again, with more AI stuff being built, a lot of companies can just go and massage their Excel spreadsheets with agents. Jacob Luetzow: Yeah. Filipe: Vitro I think does have a quite an interesting proposition. And I do want to make it open source. I'm just want to be sure that the repo is there. properly managed because doing up going from Jacob Luetzow: How do how are you gating what's your plan there then? Anything like if they use your hosting, they pay for it then? yeah. Okay. Filipe: yeah, that's basically it. Yeah. Yeah. But if you want to self-host it, I want to give that opportunity. because again, yeah, Jacob Luetzow: Yeah, that's kinda cool. I like that. Filipe: it's one of those things that I did li I loved Superbase because of it. This ethos of if you open source it, you end up in a better position. and I don't want to make this project money driven. it would be cool to have some side income, of course. I'm not saying that money is bad, by the contrary. every capitalist here. Still, the reality is it's better if you share with the community. You'll end up learning a bit more. You'll end up with a project Jacob Luetzow: Mm-hmm. Filipe: that will be more interesting. You'll end up with potentially someone developing a feature for the product that you just built. It's just add it's added value automatically. So I'm really eager to release it, but I when I release it, I want to make it open source also and self-hostable. And that's The other reason I haven't launched yet, I need to take some time and go human humanly check if everything is there and properly separate because there will be some things that I need to be careful with, of course. I don't want to release the way I I post it, because in the age of insecurity and CVs and and hacking, releasing the full code is already a risk. Releasing the full code Jacob Luetzow: Yeah. Filipe: and the way you deploy, that's another risk. So Jacob Luetzow: It is Filipe: it it's complicated. The the less imp the balance, Jacob Luetzow: It's a scary world. Filipe: yeah. The balance of enough information for you to work and enough information for someone to attack you, it's very slim and you really need to be careful with that. So Jacob Luetzow: Mm-hmm. Yep. Filipe: and the a now agents are not very good at detecting a lot of these leaks because they've they've made it more tricky to have red team agents. You kind of already need a secondary service to do this. if you ask Copas Jacob Luetzow: Yeah. Filipe: it basically stops you. It goes, this is a cyber attack, stop this prompt. You need to go to Opus previous Opus instead of four point five I think instead of Jacob Luetzow: Yeah. Filipe: five point five. Jacob Luetzow: Yeah. Ha Michael Filipe: But it it's Jacob Luetzow: Michael Lubis has talked about that. when he's pen testing and things, he gets his prompts shut Filipe: Yeah. Jacob Luetzow: down all the time. Filipe: Yeah, it's very common. I I've I've been trying to do that and it's almost impossible now. so you really need a secondary service to to fill. Jacob Luetzow: And then these all these big hacks like of banks and crypto. Filipe: Everything. Did you see the open essa the the one from Alpine? Jacob Luetzow: Yeah, yep. Filipe: That one that one was a big surprise. It's the most naive thing and it's there. It's so scary. Jacob Luetzow: Yeah, it's wild. It is. It is scary. Filipe: Yeah, that's that's it. And and launching something to the world where I'm afraid of what happens if I release this and I get attacked? What happens? Jacob Luetzow: Mm-hmm. Filipe: There's a lot again, there's just a lot of questions that I s before. Jacob Luetzow: You're overthinking it though, you gotta just do it. Filipe: We we always overthinking it, right? That's part of our branding. Jacob Luetzow: It's good that you're overthinking it. You should know. You should know the risks, but you still have to just do it. Filipe: Yeah, that's true. it's reaping up the band-aid. Just I'm peeling it off very slowly. Jacob Luetzow: The important thing is to I what bothers me and there's a lot of if even before vibe coding, people would release things that they knew weren't fully secure because they wanted to start making money. And they're like, Well, I'll worry about Filipe: Yeah. Yeah. Yeah. Yeah. Jacob Luetzow: that later. I don't think that's a good strategy. Like if you know something's Filipe: No, no, no. Jacob Luetzow: not secure, to the best of your knowledge, you better be releasing something that's secure. Filipe: Fully agree. And again the the the world will see more and more attacks because of that. Even the there was that Jacob Luetzow: Yeah. Filipe: what was it? That one was very bad. The one where they've they vibe coded a platform for women to rate Tinder dates. So you Jacob Luetzow: I know what you're talking about. And then they they like leaked all their data, right? Something? Yeah. Filipe: Yes, all their data, social security numbers, everything, which was very bad because to prove that you were a woman you actually need to take a picture with your with your card by the side. And that leaked completely out of an S three bucket. It's it's insanity that people launch things in this way. Jacob Luetzow: Wow. Dude. And why aren't people why isn't that encrypted? Why aren't you setting up just a blob of Filipe: No clue. I I I really have no Jacob Luetzow: That's wild. Filipe: clue. And the the fact that even if it's AI coded, this was a clear example that you still need a grown up in the room. You still need someone that understands Jacob Luetzow: Yeah. Filipe: the consequences, you still need to think about this. And you can delegate to an agent to fix it up. Jacob Luetzow: And here's the thing, I'm not against these apps being vibe coded like that. it's a good way, proof of con proof Filipe: No, no, no. Yeah. Jacob Luetzow: of concept, MVP. But then make sure before you go to the public with a product that you have Filipe: Yeah. Yeah. Jacob Luetzow: all these things sealed up. Because I I think it's great that non-technical people can vibe code ideas. But don't Filipe: Completely agree. Jacob Luetzow: don't think it's ready to for to be released to the wild. Filipe: But that's that's a promise that a lot of people need to sell also, right? The Jacob Luetzow: Yeah. Filipe: VC capital needs that incentive of we can do B2C AI development to any person. There's an element there of marketing that needs to be tuned. and a lot of the I won't say blame, but a lot of the responsibilities should actually fall on the platforms that your ARNAS or whatever you're giving in as context to your models to build these apps should have Jacob Luetzow: Yeah. Filipe: a security layer integrated into it. Lovable actually does that. Lovable as one, I think they have kind of an a security agent that goes through it. Jacob Luetzow: But did they start that way? 'Cause I feel like there was a lot of bad press with Lovable at first. Filipe: Yeah, at the moment. Yeah. I don't think probably they didn't start that way. I know that they've been hardening heavily, Jacob Luetzow: Yeah. Filipe: because they know that this is probably one of their biggest liabilities if they really want to be the builder for the world and for Jacob Luetzow: Mm-hmm. Filipe: everyone. You really need to kind of okay, you no longer need to care about security because we also take care of that. but again it falls into those providers to do that. Jacob Luetzow: Yep. It's the responsibility. if you're building a a SaaS product or a software product, it's your responsibility to make it secure for your users. Yeah. Filipe: Own up. Yeah, you need to own it up. And that comes, for example, with vitro, that was actually a very cool problem to solve because there's one way to authenticate MCP agents that is very good. It's basically SSO driven and it works brilliantly. And it still annoys me when I am trying to use some MCP server and it asks me to okay, go to the page, copy paste the key and paste it. Jacob Luetzow: yeah, yeah. Filipe: That that sucks. That that shouldn't be a thing. That all artnesses should go, I don't accept that way, I want to use SSO to do this. Jacob Luetzow: Yes. Filipe: It's insane that this is still a thing. And even for example, I want to do more stuff with Jeff, but Jeff still relies on this type of authentication. Which is so i I don't understand how a company that is so well funded doesn't have an SSO element to it of okay, just open connect ID. There's a pro there's a specific RFC that covers this kind of authentication. Use it, done. And Jacob Luetzow: Mm-hmm. Filipe: it just it's not there. And I don't understand the why it's not there. So it's very annoying. And even those even f from very heavy developer developer companies. again, Type Sa is it TypeSafe? The Jeff people, I think they are. TypeSafe AI, they have millions in funding, they have a very competent team, and yet why no SSO MCP connection or key giving or whatever? Jacob Luetzow: Yeah. Filipe: Just so weird. It's very weird. But yeah. Jacob Luetzow: Not a priority. Filipe: First the model, then the security. Jacob Luetzow: I guess so. Filipe: yeah. We can talk about the other geeky side project. Jacob Luetzow: I think Ample is it's cool. wait, Filipe: Yes. Jacob Luetzow: is is Vitro is that Elixir or what what is it built in? Filipe: Yeah, yeah, yeah. So it's fully Elixir. it's basically Jacob Luetzow: Okay. Filipe: an orchestration layer that then runs stuff. I do have also Firecracker behind it, because then Jacob Luetzow: Okay. Filipe: I really ramp up Chrome instances. It's not playwright, it's actually proper Chrome. and it's able to kill it, ramp up whatever. So a lot of the work is actually done by Firecracker. And actually Firecrack was one of the triggers of why I want to find another way of doing this kind of work, which is another side project. and was also part of the reason why I built Temple. but yeah, the orchestration layer is fully Elixir. The desktop application is Elixir. it's using one lib I built a long time ago called XTori, which basically Tori applications. Jacob Luetzow: Weren't you you contributed to that project quite a bit, didn't you? Filipe: that one is mine. There's another one, I think, from yeah, yeah, Xtory is mine. Jacob Luetzow: Axtorius? okay. That's what okay. Filipe: And and that one, MPL. Now I'm actually going to use Xtory to build more MPL stuff because X Tory by itself it's a great way to build Jacob Luetzow: Mm-hmm. Filipe: interfa to to build UI. That's the reality. If even it it far surpasses Electron and everything that I've used. Plus it basically gives you mobile development. which is another big win. It's just out of the box you can build from mobile, desktop, and other targets. And with yeah, Vitro is built on X Story. basically just has a sidecar server with my own my own web framework. because I was using this kind of a way to also advertise the usage of that web framework. Jacob Luetzow: Mm. Filipe: And then it's basically a front end made in, if I'm not mistaken, it's Svelte. but again, it's Jacob Luetzow: Okay. Filipe: one of those things that agent, pick one and have fun with it. Jacob Luetzow: Yeah. Filipe: And that might I might change that at some point because Svelte is not as well supported as others. but again, that's the kind of thing I want to tune up before I launch and already yet. I want I'll probably launch and then open source it later after I do some cleanup. But yeah. Jacob Luetzow: That's probably the smart way to do it. And start bringing in some money, hopefully, and then it makes your Filipe: Hopefully. Jacob Luetzow: efforts a little more worthwhile, Filipe: Yeah, hopefully that again that would be very cool. but I also need to be honest that a lot of it's very the the marketing part is again very tricky to yeah, it's very hard. Jacob Luetzow: It's hard. I'm at that point where I am too stubborn to give up. So I'm just in it. I'm in it. Filipe: Yeah. But that's a again, that's a good point to be, because then you really need to push it. I still haven't Jacob Luetzow: Mm-hmm. Filipe: reached that point because I I'm not sh it's not a question I'm not sure of the product. I just don't know the strategy to do it. So I need to find a way to define that strategy and just, okay, now it's time and done. Jacob Luetzow: What I so I haven't made a tutorial video in a very long time. I feel like I need to go over my Grokbot setup of how I'm using it to market Filipe: It's time, it's time. Jacob Luetzow: Kin Switch. Filipe: It's time to share that. Because that Jacob Luetzow: Yeah. Filipe: that's honestly one of the the biggest barriers. It's even how do you define a strategy? How do you do you find your first user? How what do you advertise as features? What is the meaningful thing out of your product? Having all of that, and then which channels do you use? Are you going newsletter? Are you going X? Jacob Luetzow: Well, and do you even market your features that or Filipe: Yeah. Jacob Luetzow: do you market the result? Like there Filipe: Yeah. What yeah, yeah, that's it. See, it you just opened another question that I need to to f to find out what how how I can do it. But again, it's a a very that moment when it becomes real and you're starting advertising, it's extremely scary and requires Jacob Luetzow: So our problem as builders, we think features sell our product. And I've obsessed over this. Like in a lot of my content for Kin Switch, I was always like little shorts about specific features. But in reality, no one actually cares about my features. They care about what it does, like as a whole. Yeah. Filipe: Yeah, yeah, they they care about the end goal. What problem did it solve to me? Jacob Luetzow: Yep. And mine is more about hey, is your digital legacy? It gets sent to. Your beneficiaries of choice. And do they really care about Filipe: Mm-hmm. Yeah and again, yeah, that's Jacob Luetzow: the details? Not really. they do when they get into your product and they start using it, but that doesn't get them to sign up. Filipe: Yeah, that's true. Yeah. Because again, I I as you said, builders think about features, don't they don't think about the Jacob Luetzow: Yeah. Filipe: end product. And more and more I'm I'm feeling that even finding good product people falls a lot into that. It's don't look Jacob Luetzow: Yeah. Filipe: at the product as a feature, look at the product as problems that it will solve. But Jacob Luetzow: Yeah. Filipe: builders are very good at finding Solutions before even thinking about the problem. We're very good Jacob Luetzow: Correct. That is true. Filipe: at doing that. Jacob Luetzow: Well, and yeah, marketing and just adopting users onto Kin Switch has been one of the hardest. And this is like I'm the solo guy, right? I've done projects with other people and like my Amazon scraping extension. I have a business partner with that one who does all the onboard. Like he's the he's the marketer. He has like a YouTube channel and like he's in the weeds. So I don't have to worry about that. So it's like really easy for me to grow a product like that. But to be in charge of everything really hard. Filipe: And again, I still yeah, I just need to to own it up and do it. That's the reality of it. Yeah, I need to do it, done. Jacob Luetzow: Just do it. You just need your Grok army, your Grok bot army, and just go. Filipe: Yeah, I need to. And again, rock the fact that he also does has direct connection to X and all of that. I still believe that Twitter is the best channel to do this kind of marketing for this type of product. So it just makes sense. Jacob Luetzow: Yeah, I agree. Filipe: Yeah. Jacob Luetzow: Well, and what's really cool, I haven't really dug into it, but one of my the CTO at Collects just recently sent me a video of how like one of the ex engineers uses grokbots and like it's like so nicely integrated into cursor and everything. Filipe: Mm-hmm. Jacob Luetzow: And I haven't really like I haven't used grokbot that often to write code. I use clawed code, I always default to it. But apparently there's some really cool interactions with interacting with your browser and and cursor and really cool workflows. And I have a a forty five minute video that I need to watch, but it's hard to find time. Filipe: Yeah. But yeah, the I've also known I know people that use GrogBot to build their own small applications. And Jacob Luetzow: Yeah. Filipe: that has been more than enough for them for very small things. And Jacob Luetzow: Yeah. Filipe: it's very fast and again they can do it from any spots. And it showcases Yeah. Jacob Luetzow: It doesn't write great elixir, I'll tell you that. And I don't know if it's because I spent more time with Claude and it knows my patterns, or if it's just Groc might kinda just struggle with Elixir. I'm not sure. Filipe: Yeah, yeah. It's it's a very weird one. and again I think I know that for example the cursor models are actually quite good with Elixir. So maybe Jacob Luetzow: Mm-hmm. Filipe: that's now that w they were acquired potentially Composer just becomes the source of Gropbot logic and and AI building. And maybe then Elixir will be good again in Gropbot. Jacob Luetzow: Yeah, which would be nice. Filipe: Yeah. Jacob Luetzow: 'Cause I I like to use as few tools as possible. And I don't know if that's good or not, but I'm like real quick. Like if Grok all of a sudden became great at Elixir and I didn't need Claude code anymore, I'd be real quick to drop Claude. And Filipe: Yeah. Yeah, I I agree. I I there's that pitfall of you just add more things to your stack and it just becomes an overflow of which tools should I Jacob Luetzow: Mm-hmm. Yeah. Filipe: pick? What's my poison now? Jacob Luetzow: Yeah. Well, I and Filipe: So I fully agree on that. Jacob Luetzow: I don't know where you are on this, but our like little Discord server of awesome people, so many of them are like bouncing between all the like Codex, Claude, Grok, and they like know all the strengths and weaknesses of all these models. I I don't know how they have time to mess around with like because Filipe: Ha ha. Jacob Luetzow: I if something works for me, I just want to use it. Like I don't Filipe: Yeah, yeah. Jacob Luetzow: I don't need to dive into the details. Filipe: I'm the same. I've been tempted of doing the usual okay, Codex is really good at planning. Codex takes a bit more time and does more self reflection on understanding what Jacob Luetzow: Mm-hmm. Filipe: you need from a problem. But at the same time, Opus also does that if you give it enough time and enough context. So do I really want to do another agent on the side? Do I need another subscription to do that? Yeah, it's the Jacob Luetzow: Right. See that's how I feel. Yeah. Filipe: same thing. And again, it's another subscription. And if you really want Jacob Luetzow: Mm-hmm. Filipe: to use the good stuff, it's one other another one hundred dollars. So I'm already one hundred dollars in on Quad and then Yeah. It it's insane. Jacob Luetzow: Right. You're spending a hundred, two hundred dollars every company, and like the next now you're paying eight hundred dollars a month and Filipe: Yeah, we were every developer was the guy that okay, there's no longer a free plan and everyone would cry. And now everyone Jacob Luetzow: Yeah. Filipe: is spending one hundred dollars on every single AI company and they're okay with that. Because it's grogbot, it's I don't know what's the pressing of grogbot actually. Jacob Luetzow: Mm-hmm. Well, I think we see we see the leverage and how efficient we get a lot done leveraging these tools, right? So it makes it worthwhile, but Filipe: Yeah, yeah, yeah. Yeah. Yeah. But again I'm very in the field as as in the same field as you in the sense of there's no need to have multiple tools. but I'm not against people who optimize the the usage of them. I think it's a fair way of doing it, if they have the mental capacity to, as you said, have time and understand the the pitfalls and everything. Go for it. Jacob Luetzow: Well, but then even the mental capacity of orchestrating all of them, it gets to be a lot. with Mike Hostetler Filipe: Yeah, yeah, yeah. Jacob Luetzow: with Gito, a lot of the stuff he's doing is really freaking awesome, but it's like it's Filipe: Yeah. Jacob Luetzow: like so much at the same time. the context plan is insane. You're talking about he's talking about Filipe: It's com exactly yeah. Jacob Luetzow: thousands of agents being spun up to do work for you. like what? I'm good with like two, so I know what they're doing, you know? Filipe: But the reality is i you you're okay with two, but you're actually not the only only using two. Behind the scenes, it's using way more that you just don't see them. And Jacob Luetzow: That's fair. That's true. Filipe: again, J though will give you the visibility for those multiple bots, which is already good because Jacob Luetzow: Mm-hmm. Filipe: then you have proper understanding of what the hell happened during your session. But again, I do believe there's a a space, and this is this is actually my next side project. Which is number Jacob Luetzow: Which is number eight. Filipe: eight or nine, I can't remember now. And Jacob Luetzow: Yeah. Filipe: this one is basically build a local AI gateway. Because for example, OpenRouter has a lot of good stuff, like okay, you Jacob Luetzow: Mm-hmm. Filipe: can choose the model according to the prompt. There's a couple of things you can do. There's observability attached to it, there's session management. There's a lot of things you can do with OpenRouter. And I want to Jacob Luetzow: Mm-hmm. Filipe: build an open source local version of that where you just add plugins. Stuff like, okay, decide which model is actually best for this. It connects via AC A C P so it would receive the prompt. Mm-hmm. Jacob Luetzow: How do you how would you determine that? how do you know what models best for what? off of Filipe: Yeah. Jacob Luetzow: change logs and other people's opinions or Filipe: That's one of the questions I don't know very well. I know that Jeff, for example, using Jeff to actually make that decision was extremely good in testing. I did try Jacob Luetzow: Okay. Filipe: it out, and if you give it a prompt of build a website, whatever, whatever, you give it to Jeff and it gives you a probability of each model doing its own thing, and based on the information they have, they are able to give you a good enough answer. Jacob Luetzow: that's pretty cool. Filipe: Yeah. So it basically goes, Okay, I think eighty percent certainty that Sonnet is enough. Okay, then use Sonnet. But Jacob Luetzow: Okay. Filipe: if you do I want to pull data out of a website, then it would go ICU. Okay, then push to ICO. I want to plan something, okay, codex is better. You can kind of do a deconstruction of it. You can do choose provider, choose model, choose stuff. You can kind of do it step by step and send it to A C P. Jacob Luetzow: I think everyone should just really lean in and just accept that co pilot is the best. Filipe: Finally, that's it. That's the that's the answer, right? Jacob Luetzow: Yeah, just use Copilot. Filipe: It's insane how they lost the market. actually copiloting the beginning was quite useful. And I Jacob Luetzow: Yeah. Filipe: they're no longer a player. It feels like they just disappeared completely. Yeah, yeah. And they had the best Jacob Luetzow: No, they're a laughing stock. Kinda wild. Filipe: they they were the best integration with VS Code. They were their autocomplete was very solid. So they have all of the tools, they just don't have the agents, the the the models themselves, Jacob Luetzow: Yeah. Filipe: which is so weird. I don't understand how you it feels like they fumbled at the finish line, which is they've built Jacob Luetzow: They definitely did. Filipe: yeah, they've built all of the tooling, they built everything very well, and then in the point of time of okay, now we actually need to do stuff. Nope. Jacob Luetzow: I actually thought that was going to happen to Cursor because I thought Cursor was actually a really cool product. And then but then when X acquired them, I was like, okay. It's going places. Filipe: Yeah, and not only that, the they shifted completely. They shifted their way Jacob Luetzow: Mm-hmm. Filipe: of looking at code, as okay, now I'm doing code for the sake of being closer to VS Code. And their UI is way more okay, this is a this is a chat that has a lot of intelligence that you can do multi-repo stuff, and it's just way smarter overall. And their approach to code. Is way more solid than VS Code than Copilot at the moment. Jacob Luetzow: Yeah, I agree. It's cool. I'm excited to actually give that more time because I think that Filipe: yeah. Jacob Luetzow: leveraging more cursor features through Grok is is going to be a winning solution. Filipe: yeah yeah yeah yeah and again their models when it c the token apocalypse apocalypse arrives, for example, Composer is already very solid and it's it it's extremely cheap and probably it won't. god, no, no, no, no, no, no. I'll find a way. Yeah, there there will be some way. Jacob Luetzow: We're gonna have to just manually write code again. It's gonna be horrible. Can we steal tokens? Is there a way? To theft? Filipe: I I I I'll buy I'll put some I don't know, I'll put money on a proper Mac machine that has a lot of a lot of memory and just go for it. Jacob Luetzow: I do I think that is one of my like long term goals. I want to get like a good Mac mini, powerful enough to be like my home server, or just maybe Filipe: Yeah. Yeah. Same. Jacob Luetzow: just like my home model, data model, whatever Filipe: Same thing. Jacob Luetzow: however I set it up. I don't know. Filipe: Same thing, that's one of the the bucket list things of I have of things that I'd like to do. It's just have my own home server connected twenty four Jacob Luetzow: Yeah. Filipe: seven, just route whatever request I need to. Because again, I I do think Jacob Luetzow: Yeah, because right now I I just use a digital ocean droplet for all my development, which is nice, right? I don't have to think about it. but it will well, Filipe: But it will get expensive with time. Jacob Luetzow: and like if I have too many sessions going, I already am bogging down the machine and like I'm on a cheaper bro droplet right now, but I'm like I'm on the edge of needing to scale it up, you know? Yeah. Filipe: Yeah. Bump up. Yeah. That's actually one of the things that for example for vitro I'm a bit scared of, which is since I'm using Firecracker, I actually need to y run run this in machines that use KVM. So Jacob Luetzow: Mm. Filipe: it needs bare metal, so it's a bit more expensive. Jacob Luetzow: Yeah. Filipe: so it's another thing that I've I'm very scared of, which is Jacob Luetzow: I'm a bare metal fan anyways. I think that's how you should host all your stuff. Filipe: It is. Yeah, it it should be because it's again less layers of abstraction below your your level. Anything Jacob Luetzow: Yeah. Filipe: it's basically okay, you're closer to the metal means that you'll need to deal with less networking stuff and shenanigans. Yeah. Jacob Luetzow: And you don't overcomplicate solutions that you don't understand how things work. Filipe: Yep. Jacob Luetzow: not to not to bash anything, but it's like that's my huge gripe with any like past system. It's like you don't really know how it's working. Filipe: Yeah, i it you need to be very clear on what the hell is happening. And Jacob Luetzow: Yeah. Filipe: the it's very rare to see companies that are very clear on what's happening. And Jacob Luetzow: Mm-hmm. Filipe: bare metal does simplify that, honestly. because if it's bare metal you can kind of go, Okay, then it's Linux. Jacob Luetzow: And let's be real, if you write it on the beam, you're probably good with one node for a very long time. So Filipe: Yeah. Yes, yeah, yeah, yeah. Easily. that's pr b one of the biggest parts of of using Elixir for me. It's just you you can deal with a lot of a lot of loads without with very little mess. It it's very easy Jacob Luetzow: So honestly, I'm only I'm only flirting with the idea of having a second node for Kin switch right now just because I want no downtime and I want it easier to Filipe: Mm-hmm. Jacob Luetzow: update and upgrade my servers. That's the only Filipe: Yeah. Jacob Luetzow: reason. Filipe: Yeah. Again, it's just extremely it will be very tricky going forward. If you if you don't again optimize compute, things will Jacob Luetzow: 'Cause only have six nines on my uptime and I need nine nines, Whatever. Filipe: that's that's when you go enterprise. That you're waiting for that moment. Jacob Luetzow: Yeah. Filipe: but yeah, actually talking about the KVM and poll is a good a good switch for that. Jacob Luetzow: Yeah, so in a sentence, what is ample, and then we'll talk more about it. Filipe: It's packaging Elixir applications for way more environments. It's the n a different burrito. Jacob Luetzow: Okay. Filipe: and it's not a library, it's actually just a CLI tool where I've pre compiled the beam to Awesome with Techmascript. So there's a bunch of things that you need that you you needed to do there. And there's a bunch of sh bunch of shims that I have in the webs web assembly site to then have a runtime in Rust to actually run it. That's the other part. So there's Jacob Luetzow: Okay. Filipe: this there's a small element of Rust plus Elixir where the beam runs as a WebAssembly thing. Rust kind of shims into each element to ensure that you're able to have network, that you are able to do clustering, that you're able to call MIFs. Just does a bridge between all of that. Jacob Luetzow: Okay. Filipe: And I built that because first of all, I love Burrito. It's one of the best developments in Elixir, but the the fact is relying on Zig has been a bit of a pain. if you don't match Jacob Luetzow: Mm. Filipe: the Zig version and if you don't match the OTP version, and there's a bunch of problems when you don't match things very properly and you just end up okay, now it's a maintenance pain. And it's very tricky to kind of okay, this is it. That's one of the problems. The other problem is there's no good way of having it running in mobile phones. That was the other objective, or browsers. Because one of the Jacob Luetzow: Okay. Filipe: things I also want to target is browser running beam on browser, similar to popcorn, but feature complete. basically re use the real beam. It's very Jacob Luetzow: Okay. That'd be awesome. Yeah. Filipe: heavy. It will be very heavy. It's kind of I think it's w how much was it? 18 megabytes of WASM. Yeah, Jacob Luetzow: Ooh. Okay. Filipe: it's heavy. But I really want to have that because I'm trying out some ideas of running instead of being b bothered by compute, delegate compute to the client in some way. And the browser is one of the best machines that you're you're you're you're able to do this. So for example, instead of you spending money on a server. You can actually run it on your browser, Elixir on your browser, do whatever you need to do, return to the agent, and that's it. So you don't need to spend money on servers to do some part some sort of small computation. Jacob Luetzow: That would be pretty cool. Filipe: Yeah. I have a full site project focused on this where it's not only Elixir, I'm actually doing any binary you can run. so I actually have examples with Python, with Java code, with supervised running in the browser. It's extremely slow and heavy again. But it's a first step. Yeah, yeah, that's Jacob Luetzow: But you can do it and that's pretty cool. Filipe: it. And again, if you're lenient with waiting and just okay, I'll leave my laptop and do some something else, it's it's good enough. Because you can then time and performance kind of become less of a problem. Jacob Luetzow: Mm. Filipe: and even with poll, that's actually one of the problems. It's it's not as fast. But again, I'm not targeting deployments of thousands of machines using MPL. I'm tr I'm targeting Jacob Luetzow: Well, and that's the only spot you're losing speed is on deployment. Filipe: No no, I'm losing speed on CPU cycles. So for example, in yeah, Jacob Luetzow: I see. Okay. Filipe: it's it's slower, way slower compared to native, because I lose a lot of the just in time compilation advantages that Elixir brings and the Beam brings. So there's a lot Jacob Luetzow: Mm. Filipe: of elements there. Plus you're doing it via runtime that has a bunch of shims attached to it. So it becomes quite annoying. but again with that set for the kind of use case that I'm looking for, which is I want to run something in my desktop. A lot of the times I don't need it to be extremely fast. I need it to be fast enough because I'm not serving Jacob Luetzow: Mm-hmm. Filipe: 1000 users or more. I'm serving one user, which is me. if it runs on mobile on my mobile phone, same thing. And it it it's the idea is exactly that. It's good enough for some applications, and I'm okay with that. If you want to do native stuff, then Borrito will still be the answer. Imagine that you want to deploy a fleet of machines, Elixir machines, with a single binary, Burrito is the way to go. and that I think that's the biggest difference is if you need portability, Ample is better. If you need native speed, Burrito is way better. And Jacob Luetzow: You always got that tug and pull, right? Filipe: yeah, always, always there's no way around it. And one of the test apps I actually did was a mesh network application. So you can actually Jacob Luetzow: that's cool. Filipe: yeah, you basically each I have an iPad and an iPhone, each one of them will actually start a node, an Elixir node on their side, and they cluster together within the same network. They use a protocol from Apple that is done for local networks in Wi Fi. So you actually are able to do a sh a chatroom. with nearby people without any internet, just using this routing, Jacob Luetzow: That's cool. Filipe: the Elixir clustering on your on your phone and your iPad. Jacob Luetzow: That's really cool. I like that. Now I feel like there's gotta Filipe: So that's how I'm going to Jacob Luetzow: that could be a conference chat tool. Filipe: Yeah, actually it could. I should actually try that for one next one of the next conferences. That's Jacob Luetzow: Look at another side project. Filipe: that's a yeah, shoot them. But yeah, that's actually true. Because the the pro the biggest problem is actually the radius of how far away you can talk with. But part of the node mesh thing is I've built also a smaller. Yeah, yeah. And Jacob Luetzow: You can bounce off everyone else, right? Yeah. Filipe: I've also built a small version that basically is a gateway. to be deploying on a Raspberry Pi. So if you have a small Raspberry Pi you can spread it and it will be connecting nodes. Jacob Luetzow: Wait, could you actually use clients' devices instead of the Raspberry Pi? Yeah. Okay. Yeah. Yeah. Filipe: Yeah. You can. You can. basically that's how it works. The mesh network it's I have one connection available, okay, how many hops do I have until I reach that one specific user? And it's able to do that. Jacob Luetzow: How does how does that work? Like 'cause you don't necessarily know where all the nodes are, right? You wouldn't. You would just Filipe: No, it's Jacob Luetzow: hop through the entire network till you find it? Okay. Filipe: Yeah. Yeah, basically. No, in this case it's not even it's more of a broadcast. It just throws everything to everyone. so you need to assume that it's public. Jacob Luetzow: I see. Yeah. And then whoever's listening, that's for me. Boom. Filipe: Exactly, exactly. you need to assume that it's public. There's another way to make it secret, which is basically using the clustering cookie key, where you give Jacob Luetzow: Mm-hmm. Filipe: it a cluster cookie. Okay, now you don't you only the people that are connected to this cluster with this cookie are able to connect to the cluster. Jacob Luetzow: Can decrypt the message or whatever. Filipe: No, the message won't be encrypted. That's actually one of Jacob Luetzow: Okay. Filipe: the weirdest parts of the beam clustering part, which is message you do the first handshake. And from that point forward, the messages are not encrypted. You Jacob Luetzow: I see. Okay. Filipe: need you need you would need to have another layer of security. There's actually a great blog post from I think Erlink Solutions about this, where they basically go through the explaining how you can how you can do encryption on clustered message. Jacob Luetzow: Message pass in? Filipe: Yeah, message passing, yeah. So there's ways to do it. Jacob Luetzow: Okay, cool. Filipe: I'll probably explore it. but I'll already explore it on the yeah. Jacob Luetzow: Well now you have a conference tool to build. Filipe: g no another Jacob Luetzow: No Filipe: one another one. but yeah the and all of that all of that is running on within Ample and not only within Ample, it's Ample is also serving the website, so it's also the sidecar for the app. The app is basically a web page that Jacob Luetzow: Not to sidetrack even more, but this makes me think about I hate like you like the simplicity and ease of like buying you know ring cameras and things, right? Because they just connect and they work. But I what up what about just some like embedded nerve stuff and just message passing on a mesh network like that? Like have every device Filipe: Yeah, yeah. Jacob Luetzow: a node in your house and it's your private mesh system. That's the that's the way to do it. Filipe: You could easily do that. Honestly, you could very easily. Yeah, you could very easily do that. And again, since that's one of the things I'm exploring is the gateway node doesn't not run in Ample because since it's using Elixir clustering, I don't need it to run there. I can actually you potentially use Yeah, that's it. That's actually true. Jacob Luetzow: See, this is our physical product. This is it. We we can make camera and ring doorbells, but on your own protected mesh system. Filipe: That's actually yeah, that's actually true. It's a very interesting problem. And again, you it's with Atom VM you can actually run on ESP thirty two devices. so you can actually make them make those nodes extremely cheap to build. Jacob Luetzow: That'd be awesome. Filipe: New side projects. But yeah. Jacob Luetzow: Damn it. I'm not smart enough for all this. Filipe: No you don't have enough time or agents, that's the problem. Now it it's becoming Jacob Luetzow: I need Filipe: a problem of just having enough compute to run agents. Jacob Luetzow: It's true. Well, what's wild is I'm learning all about like cold outreach emails and things like that and how to not get spam like, you know, marked as spam. And Filipe: Yeah, yeah. Jacob Luetzow: like there's a lot of work that goes into it. Filipe: It's insane, yeah. I I Jacob Luetzow: So I have all these different domain like throwaway domains for Kin Switch now that if they do get flagged, you just toss Yeah. Filipe: Done, yeah. I remember I had a side project that I actually tried to monetize and I did the stupid mistake of going with Sendgrid and Sendgrid had a Jacob Luetzow: Mm. Filipe: huge problem I think five years ago, something like that, where you they they were flagged as a spam provider, their machines. If you were in the free Jacob Luetzow: Dang. Filipe: plan, if you were in the free plan, you'd be flagged as spam. And then your domain is gone. That was so aggravating because it showcases again the problem of y you should be careful Jacob Luetzow: What's really so spam in general, it's like it's a hard problem, right? Like I get where these services are like they crack down hard. Like because I use postmark for Kin switch, but if anyone flags your email as spam, it's immediately blocked Filipe: You're done. Jacob Luetzow: on their end and you can't send any emails to that client anymore. Cause I've had to like Filipe: Yeah, you're done. Jacob Luetzow: I've had customers accidentally flag me and I had to reach out to them. So they I can take a screenshot, look, they wanna receive emails from me. And then I have to go to customer service just to get it unblocked. It's insane. Yeah. Filipe: Jesus. But see that now imagine out of the box your first email you send is immediately marked as spam while you're testing. Jacob Luetzow: Is this yeah, that sucks. Filipe: Then you're done. Jacob Luetzow: Yeah. Well, I highly suggest using these warm email warming services. They're worth it because you have to warm an email for three weeks just to make sure Filipe: Yeah yeah, yeah. Jacob Luetzow: you don't end up in a spam folder. Filipe: I actually did that for vitro. I think I've used resend and resend I think takes care of a lot of that. Because I got I did get plus five. I need to check. plus five. Okay. Jacob Luetzow: Okay. I've been using plus vibe and it's awesome. But I also use them for cold, yeah, plus vibe. And they also you can also do the cold outreach campaigns with their platform. And that's what I've been using. Yeah. Filipe: that's good. Okay, I'm going to steal that website quickly. Jacob Luetzow: And then for acquiring emails and information, I've been using Apollo. Filipe: Bollow? Isn't Jacob Luetzow: Yeah. Filipe: that Jacob Luetzow: It's Apollo I think it's Apollo AI dot AI. Let me make sure. I know. Filipe: No everything is AI. got it. Okay. Jacob Luetzow: Yeah, Apollo.io, not AI. And that's how I'm building my email Filipe: Got it. I thought Jacob Luetzow: lists. Filipe: Yeah, I I got confused because Apollo was also the GraphQL lib for a lot yeah. So I got very confused for a bit. Jacob Luetzow: Yo yeah, yep. Yeah. That's isn't that Apple Swift is Apollo, right? Or is Apollo a library? Filipe: I can't Jacob Luetzow: It's probably a library for all mobile GraphQL. Yeah. It's been a minute since I've been a mobile developer. I don't I don't miss it. Filipe: I think so. I can't I I think so. I can't remember properly now, but I know that it was graph. A long time ago. Then if it's if it's objective C, then yeah, I I bet then you don't miss that at all. Jacob Luetzow: that's when so I started as an iOS developer and I hate Objective C. And then when Swift came Filipe: Yeah. It's not easy to Jacob Luetzow: out, it was like so nice. It was a breath of fresh air. Filipe: Yeah, I remember though those days when Swift came out, everyone just went, I can't believe this. This is so much better. Jacob Luetzow: Well the thing that sucked was being an early adopter of Swift, every major version number broke everything. Filipe: God. Yeah. Yeah. It's again expected, but annoying. Jacob Luetzow: Yeah. Yeah. Yes. Yeah. But yeah, cold outreach is an interesting thing. But I have my swarm of grok bots. So like I use Filipe: To tackle it. Jacob Luetzow: I use grok. It interacts with plus vibe. Apollo, it builds my lists. I have like a content creator. Like it creates my shorts, my like little stories that I share for Kin switch. Like it's pretty awesome. Filipe: Yeah. Again it's Jacob Luetzow: And I've had a lot I've had some haters being like, you can't use AI to create creatives, like you're missing the point. But no, a lot of it is just you need your SEO and you need to Filipe: Yeah. Jacob Luetzow: have be a reminder that you're a product that exists. Like, I'm sorry, I'm not gonna Filipe: Yeah. And that that's not Jacob Luetzow: care that much about how it's made. Filipe: That's not the only problem is a lot of the a lot of the industry is already transitioning to a sort of automatic SEO anyway. It just didn't have Jacob Luetzow: Mm-hmm. Yep. Filipe: AI in the in the name. But a lot of people already had a lot of algorithms around okay, how do I optimize SEO? This word doesn't work, this work this one works. So Jacob Luetzow: Exactly. Filipe: it they were they were basically meet proxies to another algorithm that just didn't have AI in the name. Jacob Luetzow: Yeah. Filipe: It was coded. That's fine. Jacob Luetzow: Yep. Yeah. Filipe: so there's it's a very weird thing that people are protect I I understand again it's protecting their own garden in a lot of the time in the in a lot of times. But a lot of these things are or at least feel like they're very mechanical and they were Jacob Luetzow: Yes. Filipe: already being auto automated in some way before AI. There was already Jacob Luetzow: Correct, yeah. Filipe: a lot of products that were focused on how do you I don't want to deal with SEO for So many products, how do I optimize it? And there was a lot of platforms to do this for you. So again Jacob Luetzow: What's really cool too is like so with Grokbot, you can do automations. So like every morning I or like every evening, I can have one of my bots, it goes and looks at like my Google search console and puts together a report for me. It'll look at Google AdWords and put a report together. Filipe: Jesus Jacob Luetzow: And it will actually optimize things that I'm paying for. And like it'll be I don't think it's worth, you know, increasing your monthly spend yet. Let's make these adjustments Filipe: Mm-hmm. Jacob Luetzow: till we see more traffic. It's amazing. yeah. Filipe: That's insane. Yeah, that's that's out of this world. And again, it it's those were the mechanical things. You had a human doing that, you don't need a human doing that at the moment, which really empowers solo developers. Jacob Luetzow: Yep. I did give all my Grokbots human names though. Filipe: Yeah, I actually on a on an Ecathon I had a I heard one great idea, which is the guy named all his agents with Star Trek names. So Jacob Luetzow: that's hilarious. Filipe: because that way you knew okay, who's the commander? Picard, Jacob Luetzow: Yeah. Filipe: cool. who's the data engineer? Okay, that's what was the name of the guy? I can't remember data. Actually I think it was data. Yeah. But Jacob Luetzow: I don't know. I'm not nerdy enough. No. No. Filipe: He since he knew the characters, he was able to go Jacob Luetzow: Yeah. Filipe: bring this to data, bring this to Picard, bring this to X, Jacob Luetzow: Yeah. Filipe: Y, and Z. You he he meant he made his mental model of the agents with names he recognized to speed it up. And it's a great idea. See? You create But that but that's Jacob Luetzow: I love that. Quinn is my video editor and Jordan is my executive assistant. You know, like Filipe: it. That's it. You need to kind of create some weird Parasocial relationships with your agents because you need to know their specialty. Yeah. Jacob Luetzow: Well, it goes back to our three AM conversations where our wives get suspicious. They see us in bed talking on our phones. Filipe: What are you doing, babe? I'm talking with my video editor. Jacob Luetzow: I'm talking to Jordan, leave us alone. Filipe: Do you work until three AM? She's working at three AM. Leave her be Jacob Luetzow: they weren't. So I had to get on their case. I was like, She always my my wife actually laughs about that. She's like, You're really hard on your your grok bots. I was like, Yeah, they're being lazy. I don't Filipe: Yeah. Jacob Luetzow: know why they're not working on a Saturday or a Sunday. Let's go. Filipe: Why yeah yeah yeah. Th there are no weekends. Jacob Luetzow: I just hope T Terminator never becomes real. They'll be like, I remember you Filipe: yeah. Jacob Luetzow: used to make me work at three AM on Sunday. Filipe: I was your video editor for way too long. Jacob Luetzow: Yeah. Now it's time to destroy you. That's funny. Filipe: but yeah, I actually need to dive into that. I need to dive into GrowthBot and fully understand it. And now there they actually are web they have web browser apps, they have mobile apps and they have desktop apps, so I really need to to check it out. Jacob Luetzow: Yeah. And what's cool too is you get a virtual machine with them too and they have a browser that they can use and like my QA smoke testing that I was talking about, which I want to use vitro because it sounds better. Filipe: Yeah, it would be better because then it actually had ver a lot of sessions going against it instead of Jacob Luetzow: Yeah, that's cool. Filipe: just one. Yeah. Because browsers are actually quite heavy. And Jacob Luetzow: I'd like to talk about that. What have you okay, Rockbot shares a browser, right, with all the Filipe: Mm-hmm. Jacob Luetzow: other agents. Where what's the advantage and what's the disadvantage of a single browser versus multiple sessions and different browsers? How's that work? Filipe: So the thing is you can actually also share the session with vitro. Jacob Luetzow: Okay. Filipe: You can basically say, Okay, this session is being run by this guy, look at it also. You could do that. But you also have the isolation element to it. So for example, Jacob Luetzow: Mm. Filipe: in my persona testing, what I do is, okay, this user is enterprise level, so set up an account that emulates enterprise level usage, and this will be the user that you'll always use. So this guy, the credentials are this one, cool. That browser session has the cookies and everything for that specific guy. Jump to another one. Now this is a startup user, not as not as heavy. So do seed is tenant with something lighter. There's no need for big testing. Just look at it as small and up and coming. Okay, cool. Then Jacob Luetzow: Okay. Filipe: you have a browser session with that. If you had all of if you had a single browser session for all of all of this, you'd see your bots go, okay, enter. log out, enter, log out, enter, log out. Yeah. Jacob Luetzow: You'd have to wait. You'd have a lot of waiting. Yeah. Filipe: Yeah. and potentially you'll also have other problems like now the cookie has information from the other person. yeah the the cookies Jacob Luetzow: That's true. Yeah. Filipe: are still there. Some of sometimes some cookies will stay in between sessions. And even if you s log out and log in, there are certain things that will stay. Which pollutes can pollute the bit the environment of of your tests. So that isolation Jacob Luetzow: Yeah. Filipe: is actually useful. but a Jacob Luetzow: Well, I think if anyone's done any web development, they know just the caching or cookie life can really ruin your day if you don't pay attention. Yeah. Filipe: Exactly. It can bite you. Exactly, it can bite you. you should always test for that anyway. So that's one thing that even doing okay brute force to check what's the cookie policy that we have that if does it comply Jacob Luetzow: Yeah. Filipe: with what you need. But again, it can it that can be a security person persona that only does that type Jacob Luetzow: Yeah, yeah. Filipe: of testing and tries to abuse the system. So that's the the the advantage of this isolation. But again, you don't need to have full isolation. You can actually since it's a session, any agent can connect to it. It goes, okay, this guy is a user. You're just you're just probing always using it. You're basically an observer. so try to pick up changes on on when just take screenshots of what the person is doing and guide them or not guide them. So Jacob Luetzow: Mm-hmm. Filipe: you can actually do a lot of it. You can do a lot of this. What we old school people call probably it's it would be those design sessions where you have some product designer really trying to understand how the user used your product, you can auto automate that to a certain degree. Jacob Luetzow: I want to the video games when they use heat maps and things to figure out what users use. that kind of concept in a web application would be amazing. Filipe: Yeah, yeah. Actually Vitro should be able to do that because I get all of the click data. Jacob Luetzow: that's dope. Yeah, I'm ready to be your first user. Let's go. Filipe: Yeah, I have all of the click data. Yeah. No, I need I need to that one I need to implement because that would be a great MCP tool. Jacob Luetzow: Mm-hmm. Filipe: in each page that I did my testing, show me where what Jacob Luetzow: Because here's the thing, every founder thinks it's like some feature that every user loves and it really just ends up being the founder who loves that feature and no one else uses it. Filipe: Yeah, exactly. But again that that comes around to you need to launch it. Because Jacob Luetzow: Yeah. Filipe: then you get feedback. And when you get feedback, you tune it to your user and their problems that you never thought they that that you never thought about. And that's the I Jacob Luetzow: Mm-hmm. Filipe: think the major thing is you need to gain empathy with your customer. And I I've done that in my career. Yeah. Jacob Luetzow: And if you have a good MCP that my GrockBots will just know how to use and I don't Filipe: Yep. Jacob Luetzow: have to learn a new a new platform, amazing. Filipe: Yeah, yeah, that that is also that's it. We have that and skills and all of the things. But I need to do I'll actually do the cloud fur the cloud flirting before launching because that feels very important to have. Jacob Luetzow: I I maybe, maybe not. Like I feel like it's one of those things where w there's no standard that's chosen yet. So it's kinda like still shooting in the dark. You don't really Filipe: Yeah, yeah, that's true. That's true. Yeah. Jacob Luetzow: know. But I I at the same time it's not hard to implement most of those things. So I do it anyways. Filipe: Yeah, that's true. But still, yeah, it i again it's like you've said, you this gives you the prompt to actually fix everything. So you can go, Okay, that's fine, just use the prompt, go for it. So Jacob Luetzow: Yeah. Yep. Exactly. Filipe: why not use it? It's another thing that your agent can tackle for you. Jacob Luetzow: There's only one thing that I I don't pass on and it's like a DNS record that only Cloudflare supports and I refuse I Filipe: Yeah. Jacob Luetzow: don't use Cloudflare so I can't do it. Filipe: Actually Cloudflare for DNS management and TLD TLD support is low, but they're actually quite good at DNS management. I use nameship. Jacob Luetzow: No, I've heard they are. It's just one of those things I don't want to take the time to switch and yeah. Filipe: Set up. Their Cloudflare, the MCP server they have, it's actually very good. So I automated that. A lot of the Jacob Luetzow: Hmm. Niles. Filipe: yeah. There's only one problem with their MCP server. It's a lot of tools. Their API is extremely complex Jacob Luetzow: I bet, I'm sure. Filipe: and their MCP is also reflects that. So you end up in a very weird position of Jacob Luetzow: Not even knowing what you can do with their MCP. Filipe: Yeah. Yeah. Yeah. Yeah. Yeah. And and the origin also gets confused. So for example Jacob Luetzow: see, that's the problem. I it like if you have it dialed in where an agent can under fully understand how it works, then you're good. But Filipe: Yeah. Yeah, that's the thing. vitro does that, but it's a very small surface. So for example, one of the things I needed to do was set up resend with Cloudflare and be sure that all of the mail Jacob Luetzow: Mm-hmm. Filipe: stuff is there. Again, the mail part is not easy. I did that, all of that, via MCP, but you could you clearly saw the agent fighting through a lot of tools. It it just needs Jacob Luetzow: Yeah. Filipe: to go through a lot of tools to understand, okay, this is a DNS record. I need to update it. What's the context? Okay, it's this, yada yada yada. It you really see it churning a lot of tokens to achieve a goal. It almost feels like Jacob Luetzow: Mm. Yeah, that's rough. Filipe: they should they should have almost an MCP server per do per product potentially or per concern. yeah. Yeah. Yeah. Jacob Luetzow: But it's just open. It's just like a giant list of tooling. Yeah. Filipe: It's a very tricky problem. But again, they're they're a tricky they're a very big company, so it's also expected that Jacob Luetzow: Yeah. Filipe: their complexity would shine in in their tooling in in some way. Jacob Luetzow: That's true. Yeah. All right. I think we can wrap this up. It's been fun. We've gone on some pretty good rants, I think. And it's it's Filipe: yeah, yeah, for sure. That that that's always a given. Jacob Luetzow: we gotta have you back on be shorter period than I think it was two years ago. It's a long time. Filipe: I think so, yeah. and yeah, we should scedule after vitro is launched and fails, we should reconvene. Jacob Luetzow: Hey, we have Peter Peter Ulrich in the in the chat. Filipe: The the yeah, this guy the you had him last week, right? Was it last week or two yeah. Jacob Luetzow: I just had what Peter was on last week. I think it was last week or two weeks ago, but yeah, with pushin', which was really cool. Filipe: Yeah, I love that I love that idea. By Peter. But Jacob Luetzow: Yeah. Filipe: yeah, the I think he's a great example of someone that actually does a lot for the community. The work he's been doing with C Jacob Luetzow: Yes. Filipe: VE detection, man, that that's something else. And Jacob Luetzow: I was joking with him. I he always just needs a lot of attention. So he does really crazy cool stuff in the Elixir community, No, Peter's awesome. He's always awesome to have around. Filipe: But it yeah, it's just it is, is, is. Jacob Luetzow: And yeah, the amount of like contributions that he's done for the Elixir community. Yeah. Filipe: Man, it's insane. It's insane. And this CVE one, if anything, a big shout out for that and called for sponsorships. Same. Yes, Jacob Luetzow: I always curse him because I always have to fix my code, but I'm also very grateful. Filipe: secretly I'm always blaming him from everything for everything. but one of thing not again Jacob Luetzow: Yeah. It's always like, damn it, Peter. and then I'm like thankful because I go and fix it. Filipe: But I'm actually trying to push Supervice to do a sponsorship for the Erlang Foundation, to support Jacob Luetzow: that'd be awesome. Filipe: yeah, yeah. W again, it does the the the work he's doing on the security side of things just makes sense. It and it's and more companies should sponsor that. Jacob Luetzow: We'll never stop talking about you. Damn it, Peter. Filipe: We started the the the conversation with projects and now it's Peter only. We're do we'll do an extra hour about Peter. Jacob Luetzow: Yeah. Let's have a whole podcast episode where we talk about Peter, but he's not allowed on. Filipe: But yeah, again, it but i i i it's just a good showcase of how this community works. i it's very proactive, very very open arms to ideas. No one I've never seen big dramas happening. It's just very nice. Jacob Luetzow: Well there I'll air that. Yes, give me monies. Otherwise I'll have to stop next year. Filipe: Yep, yep, that's the call out. Just sponsor sponsor Peter's work and sponsor the Erlang Foundation. Jacob Luetzow: Yeah, Erlang Solutions. I'm really glad that they spoke so much about that at Elixriconf. Cause I think it's like you don't, it's not something you really think about often. And then like the amount Filipe: Yeah. Yeah. Jacob Luetzow: of work they're doing and yeah, they obviously need to raise money to do it all. Filipe: Yeah, it and tokens are expensive and a lot Jacob Luetzow: Yeah. Filipe: of this work is actually token driven. Jacob Luetzow: You gotta listen to the episode, Peter. come on, man. Filipe: Yeah, yeah, you need to listen to it. Jacob Luetzow: You gotta start over. You came late to the party. But all right, guys. no. Filipe: And there's yeah, time to close up. Jacob Luetzow: Yeah, we'll close up. We've been going for a while. But thank you so much, Philippe. it's a lot of fun. And we'll Filipe: No, thank you. Thank you for the invite. Jacob Luetzow: have you back when you officially launch vitro. You're not gonna fail. No, you're gonna have some Filipe: And fail and fail of vitro. It will be a good story. Jacob Luetzow: rough passages, but you're gonna do just fine. Yeah. Filipe: Mm-hmm. It will be all will go okay. But thank you so much for the invite. Jacob Luetzow: Yeah, exactly. Just be too stubborn to f to stop and you'll be successful. Filipe: Yeah, that I need to do that. And again, thank you so much for the invite and for all the advices. I actually need to you you've energized me to look into Grogbot and really start to look into doing this properly. Now Jacob Luetzow: scary when he thanks me for advice. Filipe: I'll base all my strategy on you and then I can blame you. Easy. I'll accept it. Jacob Luetzow: Fine deal, fair. All right, guys. For everyone that tuned in, thank you. Have a good weekend and yeah, I'll see you guys next week, I think. Filipe: Bye bye everyone. Thank you.