#20: Develop apps by just prompting this AI

Hey, this is going to be a fun one. I wanted to get a little more nerdy because I'm trying to meet a lot of entrepreneurs in the AI space and understand what the heck is going on in AI. And I saw Marvelism on my friend Jason Calacanis's podcast and that caught my eye. And, and venture capitalist said, I, I gotta see what you're doing because you're doing some cool stuff. So let's talk about Marvelism today. Who are you?

Hey. Yeah, thank you for having me. Yeah, it's great, great podcast. So I'm Ulric. I'm one of the co founders of Marbleism. So essentially it's a platform for developers where you can generate any software from a single prompt. And before Marbleism, I was running Vauban IO, which was a platform to launch investment funds such as angel syndicate or like VC funds. We got acquired by Carta in, in summer 2022 and, and yeah, basically we, we stay, you know, we spent some quality time at Carta and then with my co founder, we decided to launch. Yeah, another startup. So my Marvelism.

Yeah, my friend Didi, who's a venture capitalist at Menlo Ventures, was tweeting about you and, and said, I just saw the coolest thing you, you talk to it and it writes code and it's like, and it writes projects and it's like, whoa. And he was real excited about it. So congrats on getting the, you know, that's good.

I mean, he didn't mention the name of the company in the tweet. And the tweet did really well. Like one million impression. Yeah. And it was funny to see people like, tagging. Mar was like, oh my God, we have some kind of brand awareness at least like a start of brand awareness.

Yeah, that, that's how I, you know, it's funny. You're right. He didn't even use your name and I had to DM him. I'm like, what? Who are you talking about? I want to see him. But that. Congrats on you because that's, that's the goal for an entrepreneur, is to get everybody excited when they see your product. Right. And, and so tell me a little bit about what, what it does because why is it so magical? Why, why have you gotten so many people like, of this note, so excited? And why you? Why not Devin? Why not, you know, Google? Why not, you know, what is it that, what's your secret sauce here that lets programmers sit down and just start talking to it and have, well, maybe we should get A demo so that we understand.

Yeah, we can do a demo. I should start by talking why we created Marbleism and then extend a bit on the Secret Sauce. But essentially we are building the dream product we would have wished to have when we started all of our software projects in the past. I learned how to code code when I was 12 and so I created maybe 20, 30 different like mobile application or website or everything. And basically when you launch a software, it's always the same thing, right? Like you have your front end, you have your back end, you have your authentication, you have a module to generate PDF, like you have plenty of different modules and it's always the same thing. And so with Serial, my co founder, when we wanted to launch another startup, we had this long list of ideas that we wanted to do, right? And first what we did was to build an engine to quickly launch all of these ideas, right? Because we had like ton of ideas and so we wanted to launch all of them. And then we, so we spent some time like you know, coding this engine and we were like, oh, actually that could be the idea, right? Like that we could make an engine. So it's very easy to launch an idea in just a day. And so essentially why it's working so well is that we don't use AI to generate everything. And what's counterintuitive is that for example, you don't want the AI to generate your authentication or your stripe payment integration or your permission level access because these are critical modules of an application. And if the AI makes just the tiniest mistake, then you have like a big security flow or your stripe payment doesn't work and you can't take payment anymore. So essentially it's like pre made module that the AI is going to customize and integrate into the final application so that we can guarantee that the application is secure, it's stable and it's customized of course as well for the use case. So that's kind of the secret sauce. It's counter intuitive, right? Like you want to use AI for everything, but sometimes you need to pick right, what the AI is good at to, to use it for this particular use case.

Yeah. Well that's, that's brilliant. What, let's go back up to 100,000ft. What are you seeing happen in AI right now? What are you expecting to see happen and how is this whole market? Because there's like I have, I have lists on x.com of 4400 AI companies and I think 600 of them are developer tools. So there's A lot of competition to get developers to try a new tool, new thing from 600 companies. It's crazy how much competition there is. So maybe go up to 100,000ft and then I'd like to get a demo of what this is because like I said, people I trust are real excited about you because when they get the demo they're like, wow.

I mean, so first, like, personally, what I'm the most excited about in AI is speed. Like if we can get like a 10x faster model, like what Grok is doing, right? Like on llama 3, where it's like 200, 300 token per second.

And you're talking about Grok with a queue, which made their own AI inference. Yeah, it runs the model from Meta very, very fast. Right?

Correct. Yeah. So their only focus, I guess, lies on speed. They can make a model like 10 times faster than the standard because they develop their own chips. And I think this could really change all the AI ecosystem. For example, for Marblism, right now, when you type your prompt, it's going to take a minute or two to make the change on the code base. And so it's a bit annoying. If it's a tiny change, you may do it yourself as well and go in the code and make your change, but if it takes one second or two seconds, then it becomes really magical and it's really important. So this speed for me is what I'm the most excited about for the near future. Like in the next six months. Then on your question, around the coding space. Yeah, there are a lot of different tools. It was really funny as well to pitch VCs, like when we were fundraising, because they were essentially, you know, every kind of startup, like dev tool startup in AI would pitch their approach and they would like, oh, you know, we, we do this and it's a better approach because, you know, we're going to do this, et cetera. And so the VCs are a bit lost, right, because they don't know what's the best approach by definition. So they just need to make a bet, you know, to place a bet on. On something and hope for, for the best. So it's very funny, right, because even in terms of user experience, like you can see a lot of different user experience being developed, right? Like maybe people don't like the chatbots so much and they want maybe a graphical interface. So it's really fascinating to see what people try to do, especially in code generation, like our space.

So, yeah, you just mentioned user interface. I just was at a user Interface, an AI user interface event on Sunday and I saw dozens of startups that are using AI not to have a chatbot, not to just do text with you, but to do menus, to do outlines, to do graphs, to do all sorts of different user interface ways to interface with the user, with the customer. Right. How are you seeing that evolve and how are you trying to innovate on the user interface of how you work with a pro, how your system works with a programmer?

Yeah, so it's a really important question for us because essentially the software that we generate, it's like web application with a graphical interface, but for the long term vision, we want to be able to generate any kind of software. So for example, if tomorrow or in five years, everything is in VR, right? Like all of your software are in VR, you are not interacting with a computer anymore. Like we should be generating VR software, right. So from the get go, we architecture and we design every. Oh, here we go.

Right? Yeah. You know, we all want to be Jarvis, right? And work in a VR headset with infinite screens around us and being able to talk with our voice and maybe touch things with our hands and move them around, right?

Yeah. And so for us it's really important to design everything in marbalism so that we can generate any kind of software. Right. And if tomorrow everyone wants VR app, we should be able to generate VR apps. If it's about, I don't know, like voice based assistants, we should be able to generate this voice based assistant software. So that's what we had in mind when we were designing the platform, that it should adapt to whatever technology there will be in the future. I don't think software is going to disappear. It's just going to be different in terms of graphical interface, voice interface, whatever. But there would still be a software software in general.

Could we get a demo of it? And so it. Yeah, that'd be fun.

Let's do it. Let me share my screen.

There we go.

Yeah, here we are. So it takes 10 minutes to generate a project. So I've just generated this one like an hour ago. So that's what I put as a description, Web application for managing project and task. Like similar to Asana or Confluence. So as a user, I can create a project, I can create tasks in a project, I can view a list of projects, I can view a list of tasks and I can update the status and details of a task and then what to create.

How did you start and how are you talking with Marbleism to create this?

So essentially I can, I Can what.

Kind of skills do I need to have to get to this place?

Right, so essentially Marbalism is for developers in a sense where we can get you to the 90% towards the finish line of your product, but you're going to need to spend a couple of hours tweaking things or working on the design. So you need to have a bit of technical and coding skills. But our goal is to turn you into like a 10x engineers right. Of in a couple hours, in a day you can have like a fully functional product and you can go to market with. Right now our target is not really like non technical people for the moment, but it could be, you know, maybe next year if the AI, you know, like works better, etc. And we can do like the, the full 100% like it could be possible underneath.

What? Tell me about your developer stack while it's deploying or compiling.

Yeah, so essentially what we created, like the kind of secret sauce, is to have all these different modules and then to have an orchestration layer that's going to have a lot of different agents. So you're going to have an agent that's going to work on the product requirements, analyzing your project and defining some user stories associated to to it. Then you're going to have an agent that's going to work only on the front end, so on the design on the front end, then an agent on the back end. So all your data models. For example, I have a project, so I need a name, I need a description of the project, then I have a task, et cetera, and then it creates a backend and then you're going to have an agent that's going to come and fix the code because sometimes the AI can make mistakes. So you need an agent able to fix the codes to make sure that, that the project is always working and compiling. So it's kind of this orchestration of all of these different agents that's going to work on the code base and output the code base. And if you put your GitHub username here, you can have access to the code so you are in full control of the code. We don't hide the code or we don't do anything with it. You are free to.

And what languages do you support? All of them or just certain ones like Java or C or something like that?

Yeah, so it's funny, right, because in this space a lot of people are trying to be super generalist, so they support all kind of languages. So you can do Python, you can do Java, etc. On our side, we decided to focus on one single language and one single framework. So it's TypeScript and React and Node JS for the backend. And so we want to go extremely deep in this particular framework versus doing everything the bit average like. Do you see what I mean?

Yes. If you needed a general tool.

A.

Lot of developers I talked to use like Gemini and they dropped their whole code project in a Gemini, Google's Gemini or Microsoft's Copilot or whatnot, their general tools. How does this compare to a general tool like a Gemini or ChatGPT or something like that?

So essentially on ChatGPT or Gemini, et cetera, you can code, right? Like you can ask to fix something or code a page, but it's going to be very difficult to make it generate like a fully functional application, because in an application, in a code base, you've got 500 different files. And so if you ask ChatGPT to generate all these files one by one, ChatGPT or Gemini is going to deviate way too much. Right? And so you're not going to end up with the same function used across different file and it's going to be a mess. So with Marbleism, we are able to really output a unified code base that share exactly the same context. That's the power of it.

Got it.

So, yeah, so here you can see.

It again, it looks like you got it loaded up.

So yeah, yeah, it took a bit of time, but we got it. So here is our app. So we generated like this kind of homepage where we can see our notification on task updates. Then we have the project page where we can see all of the current project. We generated some dummy data as well so that the app doesn't look empty when you arrive on it. So it's already pre filled in the database. So let's check here. So we have our project detail. It's about exploring advanced space technologies.

So when you compare, do you call it compiling it? When you build this, it builds a database for you, it builds the back end, it builds the front end, it builds the whole app for you.

Correct, it does the whole app. So here it's fully working. If I wanted to add a task, I will call it a new task. Hello to do. And here I can see it has created the task. So there is a back end, there is a database behind and there is of course the front end. And then you can see the code as well here. So you can work on the code while seeing your app in the preview, but you can as well have A copilot to update the page. For example, here in the homepage I can say change the title to hello name. I can upload a mock up as well so that, you know, I can take a screenshot of a particular website I like and it's going to reproduce it. So that's what we. Yeah, that's what the kind of V1 looks like where you have the preview, the copilot and you can as well, you know, go in the code as well to tweak some stuff. If you prefer to change it yourself, that's cool.

If we're building this app and we wanted to add on a new feature or change something or change the layout, how do we go back to the development environment to do all that?

You can change it right here. Essentially. For example, I asked to change the title. It's going to output. I can change the title in the homepage. So I'm like ok, continue here. It's going to code and change my code base. Wow. Or I can do it directly myself in the code. So for example, if I wanted to change the title, you can see it's taking some time, it takes two minutes. So to change the title, maybe I would do it coding myself. But if it was a bigger feature, it would be quicker to have the copilot doing it. Essentially why it's better than the other copilot is because it has all the context of your app, so it knows all your backend, it knows your data model, it knows already everything so you don't have to put everything in your prompt right, you can just put like change the title and it knows already everything about your project.

If I was working at, I don't know, Hertz rental car as a developer building systems, they already have databases and infrastructure. Can I hook up, hook this up to my existing databases so that I can start building applications on top of, you know, SQL Server or Oracle Server, whatever database they're using.

So it's a vision for now it's not possible. Like we only handle like new projects because we generate everything around. But in the future, yeah, we want to integrate with external. Yeah, external code bases or external database. But right now it's for new. Yeah, new software projects. So you are launching your startup and you want to do it in a day versus weeks for your mvp. Yeah, you can use Marbalism to do that. And so you can see here the changes have been implemented and he has running some checks on the code to test it. Everything is okay. I'm going to update the code so we can see here the code has been updated and then it asks me, okay, should I commit and push the changes to the code base or do you want to discard? And yeah, so we can, yeah, that's how it works.

If something doesn't work right, I assume it works right most of the time. But how do I debug this? What's the debugging tool? Can I get into the code and start writing my own code to maybe do something really special or add on a component that I built in C C or something like that? How extensive.

You can go in the code base and change everything about the code base. We're going to just read the code base again and synchronize the context. So you can absolutely go and code everything yourself. And then essentially when you code something it runs some checks. So if you have an error, the AI is able to auto fix itself or to auto fix the error you are seeing on your screen. But it's one, it's what's really difficult about building an AI startup versus a normal startup like I did before, is that you always going to disappoint people. You know, the AI is never going to work 100% of the time and that's the most frustrating things, right? Because you know you're going to have some issues, right? Like for example, you go on ChatGPT and you know I would prompt like okay, make me sound more, make this email sound more native, like Native American speaker. And it's not going to create a good output, right. And so I'm going to be pissed off about it. And so as an AI company, you always going to disappoint people. And so the goal is to disappoint. Yeah, the least people. But it's very, yeah, it's very tricky to, to, to, to manage. And it's funny as well, right? Because it's when you develop like a software, the goal is that it's extremely scalable, right? And so no matter of the number of people using it, you're always going to have a very deterministic outputs. And so if it works, it's going to work for a lot of people versus like an old school company with people and people make mistake and so you're never sure you're going to satisfy the clients. Right. But with AI, what's funny is that we are going back to these kind of things where the AI make mistakes like a human.

Yeah, yeah, they hallucinate, they make things up, Right? Yeah, exactly. As a programmer, this is a new world, right. Where in The C world or the C Sharp world or the Java world. You're very specific about what you want to get in and what you want to get out. You're deterministic with this. You're talking to an LLM sometimes with English and then it gives you an answer back and sometimes it doesn't give you an answer back or sometimes it gives you a bad answer. How do you handle this mushy or how do you look at this new world of AI programming? How is it different than the old world?

So it's definitely more difficult because essentially, let's say you develop an AI startup, an AI solution and it's working but like 60 or 70% of the time you have two options or you pray, you pray that the model is going to get better in the future. And that's a valid option, right? Like it's pretty valid. It's like Sam Altman, right? Like he gave an interview and he was like, you need to build an AI software, an AI startup, knowing that GPT5 is going to be 10x what GPT4 is.

Yeah.

So you have this option and then the second option is to create a lot of.

You pray that that OpenAI company fixes the problem and makes it better and it doesn't hallucinate as much or gives you a better answer. Now that's a viable work.

Option which is essentially to create a lot of, to add a lot of checks in your code like deterministic test or like to have a processes where you check like every time that this is the right thing that has been output by the LLM. But the problem with this is that you lose a bit of the magic of the LLM. Right. Because you need to build so much like deterministic code around it that you can waste a bit your time if in six months the model gets better. So essentially it's your two options and you need to decide which one you want to go for. So for us it's a bit of a trade off where we pray a lot every night that is going to get better, but we add a lot of checks as well around the LLM.

Yeah, that's smart. Give us some, some idea what does, what does this do really well and what does it not do well? You know, let's say I'm a corporate programmer. What am I going to be really happy using it for and what, what is it not going to work well for?

Yeah, so it works well on like if you want to launch a simple like SaaS software as a service solution, a Marketplace or a social network. For now, it doesn't do like mobile application. So we're going to, it's going to be available like in the coming months. It doesn't do as well like a desktop software or like a Chrome extension. So our like Marblism right now, it's good at web application in this specific stack, right. With TypeScript. But of course, you know, the vision is to add more, add more frameworks, add more, you know, software platform to be able to do Chrome extension. Why not VR apps if the Apple vision, you know, ends up working well. But actually I've seen the number of apps decreasing on app. So I don't know. Yeah, I don't know how it's.

Yeah, developers follow the consumer, right? If, if there's a lot of heat on the consumer, then all of a sudden developers are, get much more excited by. And right now it's sort of flatline sales, right. I'll be honest, this thing's too heavy for a consumer, right. Even if they had the $4,000, it's too expensive and too heavy. If Apple fixes those two problems, then all of a sudden it's going to start taking off. Because everything else about it is amazing, right? When you're working in it, it's like, whoa, this is a much better way to work than staring at a 2D screen.

But do you use it every day or no?

Most days, yeah. It's, it's heavy to work and it, it doesn't integrate well. Like this doesn't work real well with it because, you know, I have to talk with a camera. I shoot a lot of video with it because it's a great camera. So when I go to a, like I'd shot a robot company and when I was shooting that, I was wearing it, shooting it. So it's good for watching movies, for content, you know, watching. Right. Looking at web pages, that kind of thing. AI art in it looks spectacular, right? You know, stuff like that. Look, movies in it looks spectacular, but it's heavy. And so you, you have to decide are you going to look at your crappy TV or put this thing on to watch a movie? Watching the movie is better, but it's heavy. And we all know like lightweight glasses are coming. To me, this is a developer environment for what is coming, right?

It's.

It's to let Apple work with developers to build things for the glasses that are going to come in whenever. 1, 2, 3.

The ultimate dream would be contact lenses, right? Like that could be proper.

Yeah, that's, that's further out. I mean, you know, you go completely weird and you get a neural link, you know, jacked in on your brain. And that's, you know, that, that'll be later. We have to go through the glasses phase first. And there's, there's several lightweight glasses being developed by Apple and Meta and others. You know, the, the Loomis company comes to my house every year for four years now, showing me how the glasses are getting better. And they're, they're getting pretty, pretty good. The displays that are in, in front of your eyes are getting pretty good now. So I expect in the next 18 months we're going to see some really new, really interesting new products from Meta and from Apple. Yeah, ready, man. You'll have the developer environment and I'll be like, Jarvis, programming it up. Hey, build me an app that does this right? Boom, boom, boom, boom, boom, boom.

All right.

Now that would be super cool. Yeah. How much does. Let's add five people working with this tool. How do, how do you. How much does this cost? What's your business model? How do you make money?

Right. Right now it's almost free. Essentially, people buy like pack of marbles and essentially for $10, you get 100 marbles and you can generate two applications or make 10 changes. And when you sign up, you get 100 marbles for free. So it's really like, you can make an app for like, yeah, $5 and you have like a fully functional web application. So it's really cheap. And yeah, we are just one thing. We are called Marbleism. And we have like all this thing around marble because essentially the pitch or the tagline is that we generate a code base as beautiful and robust as marble. So that's why the name and that's why all this thing around marble, that's our business model for the moment. I think the space is evolving super quickly, so it might change. We are doing hosting at the moment where you can host your app on our infrastructure. So you need to spend some marbles on that. So, yeah, that's, that's the business model all around the marbles.

That's awesome. Tell me a little bit about the company. How many people are working with you on this and how did you get. What's your funding? You went through a round of funding because I know VCs were seeing you. That's why I'm talking to you, because.

So we did yc and it was very funny because during the interview, like, the group partner asked me, like, why do you need to do yc? Because you're an exited founder, like you could raise capital in the UK and just go from there. And I was like, yeah, I had a small European success, but now I want to win the World cup. And NYC is a ticket to have the right to play in the World Cup. And so they really like this answer. And so we got accepted.

When I was interviewing at Microsoft, somebody asked me almost the same question. Why are you interviewing at Microsoft? Did I give them this almost the same answer? That's good. Yeah, Play on the big stage, right?

Say it again.

You want to play on the big stage.

Exactly right. And the big stage is San Francisco, America. So yeah, that's why we went there. But yeah, the YC experience was great. I think what most people don't know about yc, I mean of course they know that it's good for investors, right? Like everyone wants to invest in you. You get like great exposure but you get as well a lot of exposure to potential users. Like when you launch on yc you can launch on Hacker News. So that creates a lot of traction. They have 1 million followers on Twitter so they're going to retweet your tweets, etc. And it's good traction for your product. So it's really fantastic, right, because you get accepted to IC, they give you 500k, they give you some investor connection, plus they give you your first like 10k users, right? And so they are good for funding, good for distribution as well. So it was pretty good, really good, really good experience. But yeah, for now the team is pretty small. Is the two of us with Serial plus two new team members that just joined us last month.

Very cool. Anything else that programmers need to know about you and using the tool? What should their first 10 minutes be trying something? What do they need to know about you that I like? Because I'm not, I'm not, I'm not a daily programmer, you know, I don't program. So I worked at a programming magazine. But I. So I understand a little bit about what they do. But you know, that's not my job every day to be in the code.

I mean the best way is to, you know, the best ways to experiment by, by yourself, right? Like go on marbleism.com and try to, to. To. To build a software with Marbalism. I want to manage expectation as well that you know, sometimes you're going to ask something and it may not exactly does, you know what you ask. So it's still day one. We are a lot of things to build, right? And a lot of work ahead, so. Yeah. So bear with us. You know, stay tuned. It's already good, but I think it's going to become great in the coming months.

Very cool. It'll be fun watching what you do. And congrats on getting the kind of praise you have already gotten for your tool service. Where did we learn more about it? And where did we learn you?

Yeah. So we are on. On X, right? Like Marblism, AI. And I'm on X as well. Uh, so am I.

That's how I found you. Thank you so much for spending a little bit of time with me. This is a really interesting new company that you've built. So congratulations. And I'm looking.

Thank you for having me.

I'm looking forward to seeing what you do in the future.

Yeah, let's do it.