How to Build Unique Apps in the Age of AI

How to build unique apps in the age of AI. Now, Ray, you've been building applications for many more ages than I am. How do you want to start with this? I didn't mean to call you old, but I wanted to call you wise.

I appreciate you hit both. And yeah, there's such new stuff that you can build now. And I think that we might just sort of you know, start with like, what, what's, uh, you know, what are you seeing, especially around Despia, right, around the kinds of applications that people are building today that are not like what you've seen either in your previous agency work or just on your phone over the course of the last, you know, several years? Like, it's so much changing so fast. Like, so what are you seeing today first?

I think it's impressive how useful AI consumer apps are now. I remember back then, like in 2023, I was building some AI consumer apps and was like flipping them on Flippa and reselling them on like Flippa and acquire.com and all that kind of stuff. Back then we were just like doing AI grammar assistant, and those apps were like selling for like, like $50,000, $60,000 on Flippa, like with pre-revenue sometimes. Like it was wild. Back then it was the Wild West. If you had any app, like, it could be terribly executed as long as it was AI. There was always some of the investors on Flippa who want to buy pre-revenue because it was hot. Now the stakes are higher. Now people are actually building useful applications using AI for AI. So they're using AI, they're building with AI, and then they're building for AI, right, for consumer applications. And I found that, like, at least 3 out of 4 applications are like primarily like AI, like AI to reorganize my closet or AI to, uh, track my habits or AI to like the weirdest application. Like, I don't know if you see those. I always see those ads like Tai Chi walking, the, the, the new health trend. Like it's like always on YouTube.

I don't know if you see those.

It's like those AI ads with those AI voices. It was one of the people that were like buying those Tai Chi walking ads. He was building an app with Despia for Tai Chi walking. I was like, I'm curious what it is. I just looked into it. Probably not GDPR compliant, but I was like, interesting. So those apps exist as well. But I found that most of those apps are like some sort of AI app that are actually useful. Like back then we were like doing like AI text, like it was all about text. Now it's beyond text. I think this is what you kind of like can take away from it, right? Because back then all the AI tools have been text-based, enhancing your text, making your text better, generating images. Now it goes a little bit deeper into the real world.

I think that's a really key point. I think there are, you know, a few things that are all built into that. Right? So one is the idea of being multimodal and, you know, how you can, you know, take data that was previously not available to you through AI, you know, because, you know, couldn't really deal with images or video. And now it really can. I think people sleep on like how much AI can do in terms of comprehending video. Like Gemini's got a service for that. And the being able to use that context to figure out, okay, what's really going on here? Not to say that's completely free or cheap, but it is, uh, but, but it's available to us and it's only getting better. The, um, the, a second aspect is, you know, the value of collecting data, right? You probably, you know, like, like Limitless was around, right? There was, you know, uh, the Humane app, right? That was sort of this little, little, little thing you could wear, right? The, the, the AI wearables, just like using APNs to send you push notifications. Well, some of those were, but I think that the— but, but was that kind of—

no, no, that was, that was not the Humane pin, that was the a little thing called Friend. It was like this thing you put around your neck and it was like just like an AI thing. It just like picked up your speech, sent it to the device and sent you like a push notification. Or, but actually it can't send it to the device. It has to send it first to the server, otherwise it can't send a push notification.

Well, and, and, and well, you can actually have push notification what it entirely look like, but the—

Right, right. But, but you wouldn't be able like the, the, the, the pin thing that they have. Like if the phone is fully in the background, you can't like just like execute a foreground, you would then have to have the thing call a server to then call a silent push to the device to pick it up. But by that time you already mentioned the sentence. So they likely had to have the device stream the raw data to their server, process it on there, likely send a silent push with the notification. So you had like a little like spy always with you.

Yep. And like what I think that points to is like the value of collecting data. So like if, if, if, if one trend's about multimodal, the second trend is about like if you have more data, AI can put that to work for you, right? And that means there's a big market for apps that can collect that data. And I think the, the thing you originally saw, it would collect the data and keep the data. And what's even more valuable is if it collects the data and then shares it with the other things that are sort of about, about you, you know?

And so like, that's kind of like, that's kind of like where like MCPs and tools kind of like get relevant because you can take this piece of data and you can share it with this piece and then get another piece and you could kind of like go through steps, right? Like back then you would have to build interface for it. You would have to have the AI generate the script and then generate the photos and then generate the videos. And you would have to build interfaces to have people put this together, right? Like I'm thinking of, oh, what was his name? The one with Animatics. AI. He was using Wist and Webflow and Xano from StateChange. What was his name?

I'm sorry, which one you talking about?

The one who did Xano and Wist and Webflow to build Animatics AI, or what was it called?

Oh yeah, yeah, yeah. Well, Animatics is out there. It's a, it's a, it's a successful business now that's doing, you know, AI video generation. Uh, yes, yes. And you're exactly right. Like the, the, what they're doing is composing the stuff together.

Uh, you had to kind of like build an interface for it. Now you can have like all of your tools, and I think he probably with your help already has it. But now you can take all of his API endpoints in Xano and you can put them in an MCP and you can like have OpenAI directly work with it so it can skip this or it can go from this step. If this step fails, it can go back and all that kind of stuff, right?

Right. So you're seeing lots of composability and that. So the thing is about collecting more data. Third is going to be about having more composable applications. And this is— there's a bit of like, well, API has always had this, and that's true to some extent. But the, but the idea is like to— there's actually something that makes documentation much more valuable because you can— the best way to think about MCP is equal to API plus documentation. And when you don't have the documentation, AI doesn't really know how to use it. But when the AI does know how to use it, boy, there's a lot more that can do with it.

That's a little bit like GraphQL.

Well, yeah, I mean, GraphQL is like a— that's a standard for getting at a lot of data.

Obviously not like— it's just, it's just for getting— GraphQL is just for querying basic data, but like the, the, the, the way how it is self-documenting is kind of like GraphQL, right? Because like with GraphQL, you can explore what data is available before calling it, right? Like an AI, you can't explore. I mean, obviously you have the open API spec, but besides that, like you can't really call it to explore stuff, right?

Well, you can call it to explore. Actually, that's a really good use of, of AI. I saw a comment about how, like, if you have an API out there, you know, that one of the best things you can do to make your APIs consumable by AI is to have it come back with verbose error messages. Because, you know, when, if the AI can find out what it did wrong or get some hints in that regard, it can go make another try. Whereas you just say, you know, 400 bad payload, like, to do with that. But if you say 400 bad payload—

Google is actually good at that. Apple is good at API— App Store API, App Store Connect API is really good at telling you what you're doing wrong. Like when you call it, the only thing it's terrible at is they did some changes. Now app developers can essentially do the same thing as app managers. So now we don't know anymore if an API key upload is an app manager or a developer, but it makes a big difference in Xcode. So I wish they would be a little bit more precise on that. Otherwise, Apple is really good at telling you what you're doing wrong with their APIs as well.

There's another state changer who works in finance, whose clients come in with these big complicated documents. First, he had an app that like, okay, here's a big complicated document. Now all you need to do is retype 10% of it into my interface so that I can use my service. Now what they can do is because of the MCP-ish approach, just go dump in your document. I'll go check out your document and rip out the pieces that are necessary to go fill in what needs to be the inputs that are necessary to make these things work. It changes the user from being the person to just being the AI. That has been a really neat shift in model because now for him it's like, oh wait, how do I get people to use this? Because it takes so much work for them to get to it, but then it's really valuable. But now it's like, oh no, I just dump in my document, I get the magic at the other end. I had a friend who used to call that a million bucks for farting. Uh, and I thought that was— it's a, it's a, it's a, it's just like, if you can make it super, super low effort and super, super high value, there's so much to it. And the ability to like say, you don't have to use our software, the AI uses the software for you. And that is like a big change in modality I've seen, particularly on the, like, the business to business front. And then, uh, and then, you know, as a result, nobody wants to use your app. They don't want to actually go to your website or you, or open up your mobile app or what have you, except in very specific circumstances. So that means there's like this where the MCB apps standard that just was finalized a week ago, 2 weeks ago, becomes, you know, really, really useful because now it's instead of saying you have to go to my app afterwards, my app comes to you where you work. You dump in your doc, the, you know, Claude or ChatGPT or whatever is sort of working to process your stuff, you know, by working with that app. Like you say, that's sort of how these things connect. And then it's able to present to you, all right, here's what you can do next, not just in the form of a wall of text, but in the form of a widget, right? It says, here's, here's what you do now. And now instead of you have your big app, you're trying to get AI to be inside it. You have your app that's inside the AI. And from a user experience point of view, that can be really superior because now you're just giving them the best 5% of your app from a UX point of view, and they don't have to be using the other 95%. And they can always go back to your app if they want to have more context. That's what I find to be really cool about like how that, that user experience can go.

Also, what, what I think is very important, you know, back then it was very tough to, like, if you have a sophisticated software and you have lots of buttons and lots of settings, it's been an absolute pain to try to condense this into a mobile responsive version. Like, I, I've, I think I, it's already my third attempt and to try to create a halfway decent looking user-friendly interface that looks nice in a mobile app for Despia. Like, it's impossible to get all the settings in that small of a screen. Like, it's impossible. You can't do it, right? But now I'm building an MCP. So the idea is that the Despia mobile app, you'll just talk to an AI in there, and the AI will tell you all the settings that are available, and you can just say, hey, make the splash screen pink, and then it does a splash screen and turns it to be pink. That way you don't have to like, like have a color picker and have to build a color picker. It's just going to look like ChatGPT and it'll just do all the settings for you. You can say, hey, do a new deployment, and then you have like a little card. I can reuse that component for the deployment and then it just shows the progress. I can build a little interface into it and I'm not going to pay for any AI credits. I'm just going to have people use the AS Web Authentication session to go to Claude to get their What is it like? Their access token will encrypt this and then we'll consume it, their token. So we don't have to pay for AI and you have to pay us 10 times whatever it costs. You just use your credits or your cloud subscription directly yourself. That's the goal. And that way I don't have to build all the interface and that way I don't have to like wait till I'm at usage tier 5 to actually be confident in scaling this. 'Cause like OpenAI doesn't sadly have an option where you can just pay them to get access to high usage. You have to like use it and it's hard to use it 'cause it's like actually quite hard to use all this stuff. I don't know if you have that issue.

Well, if you get these things to a little more scale, the other option you have is to use what's called, well, you can use somebody else's enterprise tier., right? This is what like OpenRouter really does, right? OpenRouter, Clarify, what they're really in the business of doing is being the mediator.

They're on the enterprise plan and then like they just like use this to just chip in and they rent it out a little more fractionally.

I think that's where these other tools become a lot more interesting is because OpenAI is all concerned about like how to be handling, you know, the variable usage, right, from these customers. But because, you know, these other resellers are standing in front of it, right, they're basically wholesaling, they're buying it wholesale and then retailing out, uh, there's an advantage, you know, to that. And I think largely it's pretty similar pricing as well. So like, that would be the way I would probably look to handle it, you know, in the medium term. And then potentially being able to have a relationship with OpenAI, but OpenAI doesn't really want to have 1,000 enterprise-level relationships because that actually gets very expensive to manage. So they sort of have this divide between the retail and then they have big enterprise. And for companies that are kind of in the middle, That's where the resellers become interesting.

But I also think it's very interesting that like ChatGPT and Claude and most other models, they all have now OAuth. So they all have like, you can tap into native OAuth SDKs and you can open and you can tap into OAuth flows. So you don't have like login with Google, you have login with your AI model, bring your context, bring your tokens, bring your plan, right? Like why do we have to pay for tokens and then charge them 3 times or 4 times as much if they can just like log in with Claude, right? I think that's going to be, that's going to be very interesting because I prefer to cut one another liability.

I think logging with Claude is a super, super interesting idea in terms of being able to bring your APIs with you, um, or, or doing it Oh, really? Yes.

If you go to Xcode, you can— they have an OAuth flow for Anthropic, and they have a public OAuth API, and we already added support for that to Despia. So when we— when I build the Despia mobile app in Despia, we can have login with Claude, so you can bring your tokens into it.

That's fascinating. I didn't even know about login with Claude.

Now, yeah, it's a whole thing. It's— they have a whole— they have a whole thing for that. Yeah. Uh, RevenueCat is doing the same thing now. RevenueCat, I'm still emailing them because you have to manually email them and they take forever to reply to you, even if you are like a partner of them. But, uh, they also have login with RevenueCat now. So all of those companies are now providing OAuth flows. So like you can have a mobile app and they can, or a web app, and you can just like one-click OAuth into RevenueCat, get people's access token, create subscriptions, create products on their behalf. And they built this for all those lovable no-code tools because then they can turn their APIs into an MCP being consumed by that model without having to copy paste keys and all that kind of stuff.

That's super cool, dude. Appreciate that.

You're welcome.

The, the, the, yeah, the, there's definitely something to that. There's also something there, there's part of the MCP standard where the the tools, the MCP servers, can actually just go back to their host chat to ask questions and just use those tokens directly.

That's probably going to be very expensive if you do it in Lovable, but—

well, right, that's all about like how you're paying for the tokens. But I think you make a really good point, like who's paying for that you know, who, who's paying that power bill, uh, is going to, you know, matter in the, in the near future. Certainly something I try to do is like cut my costs, right? So that I'm not, you know, sitting on as much.

I, I do prefer, like, with Despia, if we're going to add some sort of generative AI into it, which I am thinking, like, if we will have a mobile responsive version, we're not going to have a mobile app— like, I'm not trying to build like Webflow in a, in a mobile Like, that's not the goal, right? Like, there are too many settings for it to be mobile responsive. It should be a chat app where you can just have an AI do major settings for you, or we narrow down what you can do. But if it's going to be into the AI direction, I'm just going to have people log in with ChatGPT, log in with Claude, and use their existing subscription. That way I don't have to run out of credits and then I can't upgrade with OpenAI because I am not on tier 2 or 3. 3 or whatever tiers they have because it's— they make it surprisingly hard. I wish I could just put $1,000 on the account, but I can only put, I think, like $300 at a time. And then you have to wait till you're already in the negative and then you can put another $300 and then you have to do this 3 times to go up to $400 and then make it very hard. I don't want my service to be down. The best way is people already have some sort of ChatGPT account, some sort of cloud account. They can just bring their own account and get unlimited usage. Why would I want to double charge for that? They already paid for Despia, right?

Yeah, no, I, I, I, I think that's, that, that's gonna be something that's coming too. It's like, how do you, how do you bring your own, bring your own content?

I just want to be like WeWeb. I just want to charge for the editor.

Uh, okay. Well, right. And then the otherwise, well, actually, no, but, but WeWeb charges for the hosting. That's a big thing that WeWeb does.

No, I know, but, but, but back then you always said like WeWeb is technically charging for editor.

Well, right, because the hosting was so cheap.

Yes, yes, right.

Yes. And the, the, the question is like, where do your, where do your expenses really come from? And in the, in the world of it, I think that, that gets like another thing, like how, how AI apps are different, right? Like what, what's, what's new is like the, you know, AI apps today have marginal costs in a way that the old ones, uh, just did not.

The old no-code apps.

Yeah. Uh, and the, uh, well, or, or really any app, no-code, pro-code, whatever, right? I think the, the, the, in general, you know, the story of apps written over the course of the last 15 years has been heavy on networking and maybe storage, right? But like, the, the cost of trivial, it was all, you know, fairly deterministic stuff. Um, and now that it's just much more mathy, well, and they're, they're, they're charging, and like, the, the mechanism for charging is tokens, and tokens are crazy because compute and memory.

I hope one time, one day, we'll have something like Bunny CDN, because Bunny CDN disrupted the whole hosting because they were like the cheapest one and nobody has ever been able to beat their price. I hope we'll have one time something like this for AI where they're just so cheap you can just like, you can host like everything and never delete it.

Yeah, I think it is worth asking how the— oh, what I'm trying to say here. It's probably just worth asking how some of these other models can help as well. You and I have been talking about Claude. We've been talking about ChatGPT. And some expectation that the reason why things are expensive is because you get this whole stream of tokens, and tokens cost a lot. Opus is not cheap. ChatGPT 5.0 or GPT-5. 5.2 or Codex 5.3.

One thing that I found out is Lovable made this big announcement. I don't know if you ever noticed this in technology. They said now Lovable is 71% better, but they didn't provide any case studies, any benchmarks, nothing. It's 71% better. They make another update and it's 21% better. I'm looking at this. I'm like, those numbers, how does it always end with a 1 when it's better. It's like, why not 2 or 1.75 or something? But that's not the point. But after the 71% improvement and after their, uh, 21%, so after like a bunch of improvements, um, like I figured out that now I'm spending like twice as much tokens with it. So they made it and the results are worse. I found that now the first prompt is really good. It's really good getting you off the ground. And make back then lovable, you needed to have 5 to 10 prompts to get something from a structural perspective that's halfway acceptable. Now you can go to halfway acceptable in your first prompt. But once you're at your first prompt, it becomes incredibly hard to detail it now. Back then, it felt like driving in a Toyota Camry, kind of like you got full control over everything. Now it's kind of like a Tesla where if you try to do certain stuff, this car is restricting you because of all the technology, right? So back then it kind of was like manual. Now you can't really do detailed work Like if you try to do detailed work, the AI is always trying to like get you off the guardrails and it's like trying to put you on what it thinks is best. And it kind of like has a self-awareness now that I don't like because if I have a particular issue, I say you need to move this one pixel to the side. It's not doing this now. Now it's making everything pink or something like that, or just like putting everything one pixel to the side or saying, let me just restructure the entire route. Which you didn't ask it to do. So now it— but why they're doing it is because they are obviously aiming for the more beginner market because it's a bigger market. Like, the amount of people who understand what they're doing is a smaller market than the amount of people who don't understand what they're doing. That's why Base44 is so successful because they're only focusing on the end of the market, like the bottom of the barrel, but like the barrel, the biggest market available, right? And that market doesn't care about being able to detail things in their app. That doesn't care about having specific implementation. They want a database immediately. They want AI built in. They don't care about scalability, and they just want to have 3 prompts and the app is done. That's what Lovable is really good for now. But what I feel where it's falling short is when you need to really have an opinion on how it needs to look or what it needs to do, or you have some novel documentation, you give it to the AI and say, do it. It's now not following it correctly anymore because now you see like all those reasoning steps. It always thinks, thinks, thinks, thinks, and it thinks the thought broken, if that makes sense. Okay. That's just something I figured out. And it happens with a couple of those models because, and I had people actually tell me that because I got some emails because people had issues integrating like certain SDKs from this because now Lovable didn't understand it. And what I found is that now Lovable also disregards the system prompt more and it just like more just figures, does what it thinks is right. Which is something that's interesting because that's removing like the granular control from it. And that is kind of like a movement in the market that I'm seeing is because all of those AI tools, they're not trying to target people who are truly non-technical, like who don't even know how to use Excel or something. That's not what they're trying to target to. And it's kind of like Shopify, right? Where Shopify back then, it was like a landing page builder. You can do your integration and they always went easier, easier, easier, easier, easier. And then you have all the people making videos, how to become a millionaire using it, because you only have people make videos on how to become a millionaire doing something when it can sound very simple and it can give people the feeling of making progress with it simple and fast and easy and no effort required. So it's just like an interesting thing that I'm seeing. And I think that most people who are right now doing advanced stuff in Lovable, at some point, because those tools are trying to always more and more target the bottom of the barrel, will have to go to either Cursor, VS Code, Cloud Code, in whatever IDE you want, or we will have a new generation of web-based no-code AI builders that's going to be tailored for professionals. So I think we'll have at some point the Nordkraft of AI or the WeWeb of AI or the Webflow of AI where you can't compare Webflow with Squarespace. But like, if you have, and if you care about the details, yet you don't want to like be too deep into it, there is like an upper middle class type of market for, like, it's hard to describe what it is, but I think we'll have something like this for AI running in the browser soon. And I think those will be the ones that likely be adopted by agencies. Okay.

So that's, uh, well, yeah, I like your—

I do like Claude though. I do like Claude Code and I do like Cursor. I like Cursor. I like paying the idiot tax for using Cursor and not Claude Code because it's just so convenient. I do like Cursor, but, um, but I feel that And like, there's one called— I think it's Ona. I tried it today, it's really good. It's unopinionated. It's unopinionated. It used to be Gitpod, um, but it's unopinionated. It's like GDPR compliant, security compliant, and it deploys— it runs a virtual machine with localhost and it just proxies that through their hosted infrastructure. So it's not doing like what Bolt is doing or Lovable is doing, like having like their own client-side Node.js container situation. It's actually running a real virtual machine that runs VS Code on it, and you can integrate XanoScript and Xano in it and use the Xano MCP with their agent, and you can control their agent from your phone, and it gives you the flexibility of being able to use the browser, yet you have the native file system because it runs it like, I think, on a Linux machine or something, and it has a real VS Code cloud environment and you have localhost and all the testing tools and it's fully autonomous. I was able to prompt— like how I do my benchmark to understand if a tool is really good is I tell it build an application with Framework 7, Framework 7 icons, Consta UI, and Despia. And I'm not putting any links into it for documentation. If it one-shots it, it's a really darn good tool and it one-shotted the whole thing and you had the liquid glass from Consta UI and all that kind of stuff, all built with HTML, CSS, and JavaScript. And looked one-to-one like Apple. It's really good. You can do it from your phone with the power of the native local system.

Being able to teleport into your phone, which actually gets to— I mean, we talked about a bunch of different types of apps, but I think there's another type of app out there which is the autonomous app. Right? The one that's able to be running things for you on a regular basis. And like, when we talk about like running it from your phone, you're talking about like using Slack or WhatsApp or Telegram or something. And that actually means having a two-way channel. And you were talking about push notifications before, which become really important because now it's like, hey, I want to express my intent, but also want people to get back to me every once in a while. This gets like— by OpenClaw, I think is really popular. It's because now I can have— I can say, hey, I want you to go take care of me, and like communicate through messaging, right? And then it's doing things for you. The one of the, you know, like many people are deploying OpenClaw like locally, like they get a Mac Mini or, you know, have a second machine or something that is going to be the thing that—

Sounds scary.

Hmm?

Sounds scary. I would probably just like rent some— I would probably just like rent a Mac Mini on Scaleway.

Well, the reason to have it local or reason to have it all And of course, Scaleway is not cheap for those things either, is it? But the reason you have is because you want to be always on. And I think that becomes a key thing about, you know, we were talking about like—

Probably want to sandbox it though. Like, imagine it can like just like access all your information.

Well, that becomes the question. There are a number of people who don't sandbox it because they want it to have access to all your information because that way it can do more things as your assistant, right? Much as you might hire an assistant, like give them keys to your passwords and whatever, and like you're going to hope they're not going to do anything terrible with it. Then similarly using AI for that kind of thing. Then there's also the more constrained version of like, I'm going to have some things running up to the cloud and that has access to very specific resources, which gets to your point about sandboxing or certainly guardrails. But one way or the other is the issue of autonomy. The idea like I have something that is going to be always on running for me. You were talking about Xano before. It's actually something I think would be a big opportunity for Xano. Xano as a development environment is falling way behind the times, but Xano as infrastructure is actually quite ahead of the curve because it's, you know, they have thought to be instance-first in a world where everybody else was going serverless. And by being instance-first, you can have something where you can just have resources running all the time. Basically, instead of being a server that's providing services when people ask for them and having all the problems of spikiness, it's a robot that can be working for you all the time. And that, that model for Xano could be quite effective. And when I look at like who's out there, it's fairly limited. Uh, like what the, uh, you know, what the, the company that's gonna be making the most money from OpenClaw is Apple because people are buying the Mac Minis.

Um, but also, you know, I mean, to be fair, I, I'm, I'm curious why, uh, CodeMagic isn't going into the OpenClaw stuff because they have like thousands of Mac Minis lying around.

Well, yeah, but they're— which they're also using for their service, but I certainly take your point about But the, um, but, but the other thing is like just the, the DigitalOcean is getting into this game too. I think this sort of shows like DigitalOcean, which, you know, when they first got into it, they were like the best, easiest, cheapest way to go have your own server running in the cloud. And then that kind of became less popular, but with the rise of serverless and you could have, you know, Netlify or Vercel or whatever became the more popular way of deploying these things, or Amazon Lambda, et cetera. You can sort of keep on naming lots of different technologies, but like serverless was like the popular thing to do in the previous decade. And really heading into this one, like a lot of what we saw the no-code companies doing, doing was heavily about making use of serverless to control their infrastructure costs so they could scale to zero. Um, but you know, as you have something you want to have be always on, you don't want to scale to zero. And frankly, and you also, on the other side of it, you don't have a huge scaling problem if it's about the person or about like, you know, solving this particular zone. Like, if you are an agent, you want to have more agents, you can have more agents, you build multiples. But like, you don't need to have one that scales to infinity. It doesn't, doesn't, it doesn't scale in the same direction.. And as a result of that, I think there's a lot of opportunity to make, um, for these companies to make a lot more money.

I do like Xano. I also do like Xano as a development environment. I know I'm old school when it comes to that. I, I do like— I, I built my— I don't build my APIs with AI. I built them both hand. I take the block, I put it there, I put the next block under it, because stuff I'm building is already stuff I have in my mind, and I know how it's going to have to work and how it has to translate from Xcode to being able to run a function in Xano, right? And being able to explain that to an AI is just too stressing. Okay, that's interesting because like I have to explain an AI and it has to understand it, something that I already know. It's just easier to just like, you only have to talk. Like I thought like me talking to AI, it's like half of the things are getting lost in translation, the other half are getting lost in hallucination, and then it just doesn't do it. And I just spend more time explaining and screaming at the AI what I wanted to do, and it doesn't get me because I already know very specifically what it needs to do. It's not something I can look up on the internet. So I just go in Xano and build it myself. And then I just use Xano's assistant sometimes saying, hey, just add comments to this in Xano script. 'cause I'm too lazy to write them.

Guess, assuming it doesn't blow away half your code in the process. The, um, it's just been a problem I have routinely with the, with the SandOps system. Um, but the, uh, but, but, but you, you make a good point. Like, when you know what you want, how can you more easily convey what you want? And like, where does AI create comparative advantage? Also, the kinds of things I can do with AI and, and, and doing well with AI are very different in, say, January than they were in December. And, uh, and I'm expecting that to keep on changing in February too, because right at the end of December, right at the end of January, we had two new models drop, right? We have like, you know, an Opus 5.6 keeps on getting better, and then the agentic harnesses are getting better as well, like, uh, swarms, right? Being able to have teams of agents that are running on your behalf and running in multiple times, that's— there's a lot of juice in that, uh, and the, you know, combine that with, with the model itself getting faster, like Opus previously was very powerful but kind of slow And now, now they have Opus Fast. Oh my goodness, it really goes quick.

Opus 4.6 is actually really good with UI. I like normally, like I built all my UI by hand and now I use Opus 4.6 to do the wireframes for me so I can then like I'm not taking it one to one. I am absolutely allergic towards using AI generated interfaces because I can spot them and I know other people who care about as much as I do about UI can spot them as well. But I can use them as good wireframes and they work really well as wireframes.

Yep. And you make a really good point. How do you articulate your taste becomes something that becomes more valuable now, particularly if you're like yourself, like someone who's more strong on the front end. I'm probably more strong on the backend. I care a lot about data. Oh wait, UI, I want it to be as easy, in fact, as inobtrusive intrusive as it can be. That's one of the reasons why I like having it sort of built into the, built into the chat, right? Hey, now, now it's not my UI at all.

That, that's why you're so fascinated about MCPs.

Well, yeah, because, you know, if the best UI is no UI, then how can you, you know, make that more of this stuff go away? Because that will reduce friction. That makes sense. It also changes, but that creates other kinds of problems, creates kind of the guardrail problems you're talking about in terms of the way it's gonna look. It creates discoverability problems, like how does AI know what to use? Which has been a constant. And that discoverability thing, that's all about, that's what like skills are all about. That's where, that's what, you know, the context overload problems are all about. And then, you know, you take that to a bigger scale, like, hey, what tools are available out in the world? I don't know. Right? And then that's how do you help the AIs find more agents and more tools that they can use to solve your problems.

Well, I got good news from Xano. I just had issues like with Xano. This is not a paid advertisement, but I guess it's some sort of advertisement because they just gave me some stuff for free. I had issues with the Xano database being super slow and just got an email from Louis from Xano. He says, I hope you had a great start into your week. If it's only your issue, it might not make sense from a cost perspective to upgrade, something like this. I've gone ahead and increased the database memory allocation for the Despia instance. I hope this leads to some improvements. Please keep me updated. So I just got like free memory and free resources from Xano.

Look at that. Yep, we'll see how that works.

The, um, yeah, like it works fine. The only issue I have with it is fuzzy search because I'm doing fuzzy search on very big JSON objects projects. And when I use Xano's built-in search, it takes like 5 minutes for me to find the app I'm searching for.

Oh, okay.

Yeah, but when I use an index— when I use an index and do fuzzy search only on the name, is instant, obviously. But sometimes I need to do fuzzy search on the bundle ID or like some random name set in some— like, it's, uh, It's— Xano is probably not the best data warehouse, but, uh, I do like the convenience it gives me.

Yeah, I hear what you're saying about convenience. I mean, fuzzy search is just slow, and it, it, it allows it to—

it doesn't help that it's— it doesn't help that it has to go through very big files and like, like through like lots of JSON, like, or like 10 records in the database for the Despia app. So very big JSON objects, and Xano is struggling with fuzzy search on that. Yes.

Uh, and yeah, that's, that's gonna be a little ugly. Um, and you know, we start to learn like what these tools are good for.

Like, but I do like it. I like if they tell me you have to pay $5,000 to get good performance data warehousing on Xano, I don't mind paying for it. Like, I like, I like the convenience it gives me. I don't have to like go now to Google Cloud and like do all the crazy stuff with it and then bridge it back into it. Like, I think provides is greater than the debt from it. And you know, like, I don't, I don't, I don't mind paying a premium for that. And I think Xenom knows that, and I think that that's their business model. And then you can't get away, or you don't want to go away. I could go away, but I don't want to go away.

Well, you know, that's a really interesting point. Like, what is the thing that makes you not want to go away? What is the— what for you is the most valuable thing about Xenom?

I can change the entire backend 3 AM in the morning. Without breaking big stuff. Right. That's kind of like the big issue I have. Like, like if I have to write with code and all that kind of stuff, you do syntax errors, you forget a comma, and obviously there is AI. But I do like that I can just go into it. I can think about the logic I'm building and I can understand it and I can directly without having to talk to something and losing my thought, just do it. It's kind of like drawing. You don't think about what you draw, you just draw, right?

Got it. So, really, it's the function stack that drives you.

It's the function stack.

Yeah.

And also the NPM.

Oh, being able to do— when you say the NPM, what you mean is being able to have JavaScript get deployed in their— what they call their lambdas.

Yes, that's very convenient because, uh, you know, you can import all that kind of stuff. You can— obviously you can't do heavy stuff like Puppeteer.

I'm still waiting on your heavy Lambda, but did that thing ever ship, or— oh, I dropped it because Google Cloud Run was doing such a good job with it.

We would still appreciate a nicer interface for it.

I hear what you're saying, but like that, then, and that becomes the key is like, what are people willing to pay for, you know? And the, uh, and like the, I, I think, you know, Zeno's, Zeno's function stack is a good, what was a good technical insight, right? That a lot of what you build can be relatively straightforward. ZenoScript, I think, was a misstep. Um, but, you know, I think if what you're doing is building the way that you're building, then I can see how it creates a whole lot of value. The AI just keeps on falling down with it because it's trying to interpret, you know, this weird DSL that Xano made that nobody else uses. Just sort of very unclear why they went down that road. But the, um, but, but if you are building as low-code the way that you are, you know, it's, it's non-piratical. It's definitely the best way, uh, to be building, you know, the back end. The question is, is that the way you want to build? And that, that becomes, I think, the nature of the, the market challenge that they deal with and that other low-code players deal with, right? If they have built the best low-code way of doing it, but wait, now people don't want to build using low-code They want the agent to do it for them. How do you then make that transition? And the, there was an opportunity, right, where there's lots of trust that you have today that you can then use to get them to, oh, you should now sit with the agent. But then if you don't do that job well, you lose that trust and they'll cross the street. And I think that becomes the challenge that, you know, is the reason why a lot of people have churned from Xano and a lot of people haven't churned from Xano because they have that love of working with the function stack that you just described. And then I look at like, okay, Okay, what's the other half of this, which is what are you trying to build? And Xano, because of its instance model, I think has an opportunity to be retaining and building a lot of business that others would have a harder time getting into. I was like comparing it with like e2b or Fly.io, Sprites, which are very cool pieces of technology, but they can't be offering the kind of always-on or the kind of backend that like Xano has built. And I think that'll be be, you know, a question of like whether they can be capitalizing opportunity.

What could Xano do with the always-on? Now we're talking about unique apps. Like, how could Xano be an asset to someone trying to build an application that is unique where if they would have all the always-on features enabled that you would imagine? Because I assume Xano obviously has the function stack as an abstraction on top, so they can only expose those, the things that are visual so that you can do them, like MCPs that have the capability to do it. But there was a time where they didn't support it. What would be the features that they would need to add to make maximum use of their always-on capabilities? And then what would be the apps that you could build that would make you stand out of the masses?

Well, I think this is where autonomy comes into play. Like the, you know, the, oh, are you going to build an app on Xano? No, Xano is Here's your Xano instance. Here's your Xano instance. Here's your Xano instance, right? Here's the robot that works for you. And being able to have that layer of communication, like what are we really doing when we talk about Open Claw, right? Okay, well, you got a messaging surface to it. All right, so you have a messaging layer, messaging app or something that goes on top that talks to your instance, right? Now this is your box that's working for you and it's able to do things for you. So being able to, you can start to imagine how you could, you know, express your intent. And then it goes and builds out function stacks that run as cron tasks, right? That are going to be, oh, this thing should run every minute, this thing run every 5 minutes. You can have a little bit of a layer thing.

It would be kind of like being able to chat with Xano.

Well, chat with your Xano, your app, as a developer, but like just in terms of I want, I want this to get done for me. And then code gets generated for you, code function stack gets generated for you, which is now all auditable because it's in the function stack. Right? The visibility of it becomes like an advantage on the after the fact. Not to mention, like, instead of having it just show up inside my Xano, how about having it show up in, you know, an MCP app, right? That's going to be sitting inside.

That's what I just wanted to say. I'd say like you can probably like get this into ChatGPT or Claude and then talk with it and then have that insights provided back and get updates on stuff that's running.

See, now it's not using it to build software that's going to be an app that you're going to go serve to a bunch of people. It's just just you talking to your assistant. Your assistant is Claude.

It's like FileMaker almost with AI.

Uh, well, I mean, kind of, kind of.

Yes, yes, like FileMaker is the old way of doing things, but it would be, um, yes, that makes sense. You, you use Xano's architecture and you use Xano's execution layer to generate code and have that code be used and then also be able to recycle that code and adjust that code to then automate tasks. And you could technically even have an MCP like discover things itself and then try to build function stacks. Like technically you could just provide like API documentation to it. It's like looping through everything, it is discovering the APIs and then it can build the endpoints and possibly even the interface to, like, have it fully autonomously tested. And through the testing, provide the endpoint, set up the databases and all that kind of stuff. And then you probably want to have, like, multiple containerized instances for each of those tests.

Well, okay, we— you can start to come up with a bunch more things. And then— but the other thing is, you're not— that's not even building an app, right? I think that the, the model is not, oh right, we need something to build software so I can serve up an app so I can help, you know, 10,000 customers do something. The software is just for you. It's you. The, the, the, the, the, the people are willing to pay, you know, for a Mac Mini, right, to, so it can go host their system, as opposed to, you know, and then whatever, the other software will come along. Well, they'll, they'll be more turns of this wheel. But I think there's an idea of, look, we talk about like what's unique, right? And we started with what's unique from an interface point of view. I think we can come back to what's unique in terms of how it creates value and the ability for it to be autonomous and always on. Like, hey, every 5 minutes go check my email, which is a very common thing that people are doing with Open ClaudNet. That would be a straightforward thing for it to be able to do. Now, does anybody want to go through all the work of setting all those things up manually? No. But if you can be showing it back in terms of a, you know, oh, here's the control over it. And it can be done for you. Now the value of the machine that you're, you'll gladly pay a couple hundred bucks a month for is that it's providing you with all of this, you know, value as it is doing things for you. Code, rather than being the end, which is I think the way Zale would have been pitched, say, last year, yeah, is, you know, code becomes the means to being able to do value that's just executing things for you. And that that's where I see it being a whole lot more valuable, being valuable either as an individual level or at the business level, but being able to have these autonomous robots because they've already thought this through from infrastructure point of view. They have an auditability built in because of the function stack. I think the one mistake is going to be the XAML script because XAML script I think was designed so that people could code in it, which I think has been an ambition of theirs between what they call their expressions language and then the more recently the XAML script. But, you know, if you can—

as a programming language, is not like— I think it's cool, but like, it's like, who besides an AI would use it? And AI is going to be— it's going to shoot in the dark.

And AI doesn't want to use it. That becomes the trick. It's a domain-specific language for which there's no market. And worse, it doesn't work that well, and everything that's wrong with it needs to be fixed from inside of Xano, which means they need to spend their resources maintaining it. Contrast that with what if they just had it be an SDK? What if they just had an API and then you could have your TypeScript functions, you could have your Python functions, whatever, that, you know, your Claude knows how to write TypeScript, which can use to go express what it wants to have happen inside your agent. And now, you know, the way people build is through the way they want to be building today, and the way they can audit is just by being able to look inside to get the transparency to it and be able to— and it has the autonomy because of the structure of its, um, of the infrastructure. There's, there is a lot of value to be had in there, but they're going to do whatever they want to do. I do think there's a big market for that kind of, you know, always-on, easily accessible, auditable. Uh, and if, if our friends in, uh, Zeno don't do it, or frankly even if they do do it, uh, somebody else is going to do it. And but I think when we talk about like what's unique and what's valuable, I think that idea of always-on, being able to do things for you and being able to communicate with it using some sort of two-way messaging, I think that will be highly, highly valuable.

We got one from Johnny. Did you both try OpenClaw? What do you think? I didn't try it, uh, but Ray, how about you?

I've only tried it a little. I've gotten to see how other people have done it with greater depth, and it is— it's super cool. It's like it gives you a glimpse of the future. Um, the, uh— I think it's, it's something that's worth doing and playing with and getting that. In fact, I would say this about all AI technologies right now. You play with it, you get value out of it today. It is not a thing to invest in, right? Because both it will get better without you and the— and other solutions will come around. But I think, you know, everyone should try open-clawing a bit because it is what the future looks like, which is this idea of autonomy and always-on and being able to get things done for you and having this two-way channel So like being able to turn on the messaging through, you know, use WhatsApp or Telegram or Slack or whatever you prefer, uh, being able to have it, uh, you know, create basically because under the covers it creates its own, um, uh, cron, um, you know, and have it, you know, be able to do things for you on the regular. That's where it gets super valuable. Uh, and then it has an idea of heartbeat management, which is also very important. Because it can run long-running tasks and checks to make sure long-running tasks didn't just hang on you. And that, that's kind of the third, I think, leg of the stool that is like super valuable about it. And then outside of that, it's like got a lot of good package management or easy to set up, extend, whatever. Um, but like those, those three things I think are the sources of value that one should pay attention to in terms of like what can be done and like what that glimpse of the future is.

Yeah, very nice, very nice. Yeah, I mean, I'm right now looking into running Fastlane Spaceship in a somewhat secure environment, and I found the only way to do it secure is by being on the desktop, right? And many of those things like OpenClaw and all that kind of stuff, I think we will see more and more pieces of technology being moved from the cloud to the individual device that's running it. I just hope that we can run like more advanced stuff soon on mobile devices.

Okay. All right, I think that brings us in for, uh, for the end of another, another good episode, man.

I, I can't wait to see all your clips. Our, our episodes are made to be clipped. They're not made to be watched watched in one, even though we do appreciate everybody watching it as one.

We definitely appreciate everybody who's, uh, who's watching. Um, I, I appreciate the— let Ray talk. Um, and, and hello to, to friends Nick and Dominic.

Opus Clip, Opus Clip is going to put this all great together. This is going to be like— we're going to get like a, I don't know, 100 short attention span videos out of this with lots of movement and all that.

That is what the Opus Clubs are for. All right, well, let's, uh, I'll, I'll, I'll put those up. Thank you, everyone.

Oh, we got one question. What was that? Is it about being, uh, totally unique or making something that exists but better? Oh, that's a good one, Ray. You want to end this one?

Creativity is remixing, right? It's about taking the thing that is good today, taking maybe something that was good yesterday but doesn't exist today, and like bringing it back or bringing or combining with something that's totally new. You know, I think that the— we should all give ourselves permission to say that like we're looking for what is that unique angle on it, um, without having to say it needs to be totally, totally unique. I do get blown away by stuff like, oh man, I never would have thought that, that's totally, totally crazy, and then you discover this is like a 50-year-old idea. So, um, you know, I'm not saying there's nothing new under the sun, but you know, I would say we should all just give ourselves permission to to Remix because that's going to be the way that I think we make a lot more progress.

What do you think? I completely agree with that. I also think that most of the things when we have innovation happening in the market, like, there will always be the opportunity to take something that already exists, a concept that already exists, an idea that already exists, and to make it like, you know, like adapting to the current times. Like right now everybody's using AI. Now the next challenge will become how do you, you know, add more visual control to building interface in AI for people who care about that? And then you can dig up the graves of all the no-code tools again.

Yeah, it will all come back around. And to use one of my slogans, you'll build this again. You'll—

yes, yes. I am already at the 3.5 times of building something again.

So it was great doing this. You see, we'll do this again next week.

Thank you so much for your time, and thank you so much for everybody watching, and see you next week. Take care.

Bye.