What we talked about
Join us for an episode featuring Zvonimir Sabljić, an engineer and tech entrepreneur. He founded AWW (A Web Whiteboard), serving 1.5 million users and attracting Apple as its largest customer. In 2021, AWW was acquired by Miro. Now, Zvonimir’s focus is on GPT Pilot, where innovation continues. Tune in for insights into his tech journey and the future of innovation.
https://github.com/Pythagora-io/gpt-pilot
Show notes
Building an application is not the same as writing code, that distinction sits at the center of this conversation with Zvonimir Sabljić. After selling his collaborative whiteboard company AWW to Miro, he spent a year studying machine learning, then turned a feedback loop from 200+ developer interviews into GPT Pilot, a tool that crossed 10,000 GitHub stars in its first seven weeks.
What we covered
- GPT Pilot reframes the developer’s role: instead of writing code all day, the developer becomes the team lead who reviews, corrects, and redirects an AI that handles the actual coding, running, and debugging loop.
- Software is unusually well-suited for LLM automation because code output can be run programmatically and fed back to the model, unlike legal or medical documents that require human review to validate, code can validate itself.
- The pivot from Pythagora (automated test generation) to GPT Pilot came directly from user research: developers trusted AI-generated code more when it was derived from tests they had written themselves, not the other way around.
- Zvonimir’s acquisition advice runs counter to standard brokerage wisdom, he fired the broker, let the exclusivity contract expire, approached Miro directly as a strategic competitor, and sold the company for more than twice the amount the broker said was the ceiling.
- On the emotional reality of founding: the lows are as extreme as the highs, and the only thing that separates founders who make it is the willingness to keep pushing through that cycle rather than any particular skill or strategy.
- The future-of-developers question remains genuinely open, Zvonimir’s honest answer is that in five years, software development will be “unrecognizably different,” but nobody yet knows what that means for how juniors acquire deep technical knowledge.
About Zvonimir Sabljić
Zvonimir is a Croatian engineer and serial entrepreneur who built AWW (A Web Whiteboard) to 1.5 million monthly active users before its acquisition by Miro in 2021. He is now co-founder of Pythagora, where GPT Pilot, an open-source tool that orchestrates the full app development process using LLMs, is his current focus.
- LinkedIn: https://www.linkedin.com/in/zvonimirsabljic
- Website: https://www.pythagora.ai
Episode 3 of the PreVetted Podcast.
Full transcript
[00:00:00:30 - 00:00:52:06] Federico Ramallo Welcome to the Irregular Expression podcast, also known as Irregics. Our primary goal is to delve deeper into the experiences and stories of our guests. As we uncover their insights, we aim to understand, learn and subsequently share these invaluable lessons with our larger community. Each episode is an opportunity for both us and our listeners to gain knowledge and grow together in our collective wisdom. Through Irregular Expression, we’re building a repository of wisdom, reaching the community with shared knowledge from leaders and innovators from diverse walks of life. This podcast is sponsored by Density Labs, the number one top IT staff of a mentation services company. If you can dream it, we can code it. So today we have Svoni Mir, I’m sorry to pronounce incorrectly your name. How are you doing today?
[00:00:52:06 - 00:00:55:14] Zvonimir Sabljić I’m doing good, I’m doing good.
[00:00:55:14 - 00:00:59:32] Federico Ramallo I apologize, it’s difficult to pronounce your name.
[00:01:00:34 - 00:01:03:04] Zvonimir Sabljić All good, all good. It’s Svoni Mir Sablech.
[00:01:04:08 - 00:01:09:01] Federico Ramallo Svoni Mir Sablech. Yes, almost. Almost.
[00:01:10:55 - 00:01:22:28] Federico Ramallo So I have the same issue with when people in the US pronounce my last name. I say, yeah, they pronounce it 80% okay, so I kind of understand.
[00:01:24:22 - 00:01:26:02] Federico Ramallo It’s a Rama-show.
[00:01:27:10 - 00:01:28:42] Zvonimir Sabljić Rama-show, okay.
[00:01:30:07 - 00:01:31:07] Zvonimir Sabljić I would never say it.
[00:01:33:16 - 00:01:37:48] Federico Ramallo So it’s interesting how all those different sounds we have on different cultures, right?
[00:01:37:48 - 00:01:38:45] Zvonimir Sabljić Yeah.
[00:01:40:26 - 00:01:42:57] Federico Ramallo Great, so tell us a little bit about yourself.
[00:01:44:53 - 00:02:06:21] Zvonimir Sabljić Sure, I’m an engineer. I have like 10 years of software development experience. I went to college for software engineering and after college I started my first startup. It was a collaborative whiteboard called OEP. We ran it for five years. We had like over 1.5 million monthly active users. And in 2021 it was acquired by Miro.
[00:02:07:34 - 00:02:19:34] Zvonimir Sabljić So yeah, after that I took one year off to get into machine learning. And last year I started re-rating on my last startup, on my current startup, which today is called GPT Pilot.
[00:02:20:51 - 00:02:40:22] Federico Ramallo Great, yes, we’re going to talk about GPT Pilot. That’s the startup that you were showcasing in TechCrunch, correct? Yes. Yes, yes, great. We’re going to get there. But before we get there, tell us a little bit about your experience as Director of Engineering at FaceClan.
[00:02:41:50 - 00:02:57:14] Zvonimir Sabljić Oh, FaceClan, sure, sure. I mean, at FaceClan I obviously led the entire team of software engineers. We were building a call from the internal tools to like working on a website.
[00:02:58:32 - 00:03:06:30] Zvonimir Sabljić So, you know, my idea with FaceClan was I wanted… At that point they were a pretty good brand. Are you familiar with FaceClan?
[00:03:07:57 - 00:03:09:43] Federico Ramallo I am not familiar with FaceClan,
[00:03:10:58 - 00:03:10:59] Federico Ramallo so…
[00:03:10:59 - 00:03:11:14] Zvonimir Sabljić Okay.
[00:03:11:14 - 00:03:13:48] Federico Ramallo I wanted to tell me a little bit more about what they do.
[00:03:13:48 - 00:03:17:40] Zvonimir Sabljić Why did you ask for that experience specifically?
[00:03:18:50 - 00:03:40:51] Federico Ramallo Oh, because I’m a software engineer as well. I’m a founder as well as you. So I find the journey from individual contributor to engineer leader to entrepreneur or founder very interesting. So that’s why I asked for that.
[00:03:42:06 - 00:03:53:17] Zvonimir Sabljić Got it. I mean, you know, my journey at FaceClan, you know, it’s an entertainment company, right? They do like games, like YouTubers and stuff like that.
[00:03:54:45 - 00:04:25:28] Zvonimir Sabljić And I actually wanted to do like big things with them. At that point they were a pretty good brand. Today they’re not doing that well. But I wanted to do so many things and I actually was pitching the president and the CEO, “Hey guys, can we do this and this and this and mod this and that way?” So that was my try to get into like LA entertainment scene. But they weren’t so open to doing many technical things.
[00:04:26:43 - 00:04:27:09] Federico Ramallo Right.
[00:04:28:41 - 00:04:38:12] Federico Ramallo And what drive you to go from director of engineering to starting your own startup? Was those restrictions or was there was something else?
[00:04:39:28 - 00:05:03:15] Zvonimir Sabljić Yeah, I mean, you know, I started my first startup actually before FaceClan. But you know how it goes in the beginning of a startup. You don’t know, you don’t have too much money. You cannot fund yourself. So that’s in the meantime, I actually went and worked at FaceClan. But then I decided to focus on my startup full time. So I focused on it until the acquisition.
[00:05:03:15 - 00:05:11:40] Federico Ramallo Right. Right. How was that day when you decided, “Okay, I’m going to quit my job to pursue my startup?”
[00:05:12:41 - 00:05:24:34] Zvonimir Sabljić It was really relieving. It was really relieving because, you know, at FaceClan, I struggled to move. How is it? Like what’s the expression? I guess, move the needle.
[00:05:25:44 - 00:05:41:01] Zvonimir Sabljić I always wanted to do things and really, you know, let’s try this. Let’s try that. And they weren’t so much open to innovation. And when you have your own startup and right there, then when I transitioned to full time being a founder,
[00:05:42:13 - 00:05:53:11] Zvonimir Sabljić at that point we were profitable and we were able to pay ourselves and the entire team. So that was definitely relieving to go there.
[00:05:53:11 - 00:06:05:16] Federico Ramallo Right. Right. And what has been the hardest part to become a CEO and founder of your own startup?
[00:06:07:37 - 00:06:38:07] Zvonimir Sabljić What’s the hardest thing about becoming one? Right. I mean, it’s easy. It’s super easy to become a CEO of a startup. You just need to get started, right? There’s nothing to it now, you know, persisting in it. That’s something else, right? So, you know, startup journey is like extremely hard, like, and not only hard, like, in like going from zero to actual success, but rather, you know, throughout the day,
[00:06:39:18 - 00:07:13:24] Zvonimir Sabljić every single day I get like so many ups and downs, like the lows and highs, the highs are extremely high and the lows are extremely low. So, you know, there’s also that emotional, I guess, you know, strength that you need to go through to push through all these, you know, ups and downs. So I would say, yeah, I would say maybe that. But starting becoming a CEO is super easy. You go on to Stripe, Stripe Atlas, you open up a company and there you go. And in one day you’re a CEO, nothing to it.
[00:07:13:24 - 00:07:17:12] Federico Ramallo You’re the CEO of your own company with zero activity, right?
[00:07:19:12 - 00:07:19:31] Federico Ramallo Yes.
[00:07:19:31 - 00:07:21:45] Zvonimir Sabljić Yeah, I mean, you’re still a CEO, so…
[00:07:23:22 - 00:07:23:58] Zvonimir Sabljić But, yeah.
[00:07:25:02 - 00:07:30:31] Federico Ramallo Right, right. Yeah, but becoming a, you know, building a successful company, that’s a whole new game, right?
[00:07:31:32 - 00:08:02:03] Zvonimir Sabljić Yeah, yeah. I mean, that’s what I meant by, you know, persisting, you know, you need to just push through obstacles and they’re like obstacles after obstacle. And I have a feeling that it just never stops, you know, even if you become like a unicorn, which I still haven’t been able to do, but even when you get there, you still have goals that you want to become like a $10 billion company or $100 billion or whatever you have.
[00:08:03:28 - 00:08:05:42] Zvonimir Sabljić I think you always have goals and you…
[00:08:07:13 - 00:08:20:20] Zvonimir Sabljić Goals never come easy, right? Especially when you’re a startup founder, you have like big goals, you want to really change something in the world and then never comes easy. So, you know, the road to there is hard and you just need to push through it.
[00:08:20:20 - 00:08:28:34] Federico Ramallo Right, right. Yeah, when people ask me about this, I tell them that it’s like an emotional roller coaster.
[00:08:28:34 - 00:08:32:04] Zvonimir Sabljić Oh, yeah. Oh, for sure. Oh, for sure. 100%.
[00:08:33:22 - 00:08:36:07] Federico Ramallo Yeah, with the ups and downs and all that, but yeah.
[00:08:37:49 - 00:08:46:27] Federico Ramallo And you know, a corporate job is, you know, it’s more like driving on a sedan, you know, on a smooth road, right?
[00:08:47:28 - 00:08:51:22] Federico Ramallo But you are much more limited, right, as you said it, right?
[00:08:53:23 - 00:09:10:25] Zvonimir Sabljić Yeah, yeah. I mean, you can try things, but if they don’t go, you cannot do anything about it. While in your startup, you can… And I think there’s the thing that I like the most about being an entrepreneur, that you literally have a blank slate, like you have a blank canvas
[00:09:11:26 - 00:09:19:48] Zvonimir Sabljić and you can do whatever you think is best and you can try and you can fail, but when you fail, you can get back up and try again.
[00:09:20:49 - 00:09:42:03] Zvonimir Sabljić And this is the thing that I like the most, you know, it just… For me personally, it’s like a huge game. I think that like running the startup for me is like the best game that I can ever play. Like you cannot give me a game on like a PlayStation or a computer that I can play that is interesting for me as much as it is running the company, right?
[00:09:42:03 - 00:09:42:50] Federico Ramallo Right.
[00:09:44:18 - 00:10:01:05] Federico Ramallo When I was a software engineer, I used to play, you know, these resource management games, you know, like these Tycoon rails and, you know, mini games like that, right? But basically you manage resources, right?
[00:10:02:56 - 00:10:12:32] Federico Ramallo And then when I become CEO and, you know, remember those games, I play once and I was like, “Well, I’m, you know, I’m doing my job. So yeah, this is not fun anymore.” Right?
[00:10:15:09 - 00:10:16:01] Zvonimir Sabljić It’s not fun?
[00:10:16:01 - 00:10:22:19] Federico Ramallo Well, I mean, it’s not fun to play that game. I would rather play, you know, an action game or any other game, right?
[00:10:22:19 - 00:10:22:58] Zvonimir Sabljić Oh, okay.
[00:10:22:58 - 00:10:38:37] Federico Ramallo When I’m playing and because it’s like, it’s the same as when you’re being a CEO and founder, you’re managing many resources and trying to achieve certain things, right? So, you know, I don’t want to do a video game of that, right? I’d rather, you know, do something else, right?
[00:10:38:37 - 00:10:39:30] Zvonimir Sabljić Yeah.
[00:10:39:30 - 00:10:42:47] Federico Ramallo Because I’m really working, doing that all day on the work, right?
[00:10:46:36 - 00:10:51:31] Federico Ramallo So how was the exiting of a web whiteboard?
[00:10:53:41 - 00:10:57:19] Federico Ramallo What do you think are the, you know, positive and negative aspects of it?
[00:10:58:20 - 00:11:40:19] Zvonimir Sabljić I mean, you know, the way that we got to the exit, so we decided to exit. We didn’t have someone approach us and then we said, “Okay, we might sell.” But rather, you know, we were in the educational space and it became for us, you know, we weren’t so motivated to really push through in the educational space where we had, you know, move to US and go to schools and sell to every one of them and, you know, build this kind of a relationship with the entire educational system, which neither me or my co-founder were into. And so we decided to move on to the next things.
[00:11:41:27 - 00:11:48:39] Zvonimir Sabljić And yeah, so we decided to sell a company. We actually found a broker.
[00:11:50:17 - 00:12:01:45] Zvonimir Sabljić My founder, like after the exit, my founder told me not to say their name, but they did like a horrible job. They basically, you know,
[00:12:03:51 - 00:12:59:13] Zvonimir Sabljić basically the best thing throughout the acquisition process that I did is not listen to them. So actually we went with a brokerage company, the relatively famous one. And this is, I was kind of sad that I couldn’t like review them so other people know what they’re getting into. But basically the whole, the gist was that they wanted to, they were telling us that we cannot sell a company above a certain amount. And we actually ended up prolonging the whole contract. So we are out of exclusivity because you sign an exclusivity context. So whatever you, even if you sell your own company, they need to get a percentage. But after the contract ended, we actually ended up finding a buyer ourselves, Miro. And we actually ended up selling the company twice the amount, even more than twice the amount they were telling us that we cannot go over.
[00:13:00:33 - 00:13:02:23] Federico Ramallo So, yeah. Interesting.
[00:13:03:42 - 00:13:44:22] Zvonimir Sabljić Yeah. But anyway, you know, my kind of, I guess, takeaway from the whole acquisition process is, you know, I always ask that I always encourage founders that, you know, you should go to strategic partners because, you know, Miro was our competitor, right? Just go and reach out to every one of them because if you have a good company, if you have a good product, a good project, you know, people will want to acquire you because they get a value. So yeah, that’s one of the things that I always say, you know, try to talk to companies, see if they are interested in some of the assets. You know, there are many things that can be acquired, users,
[00:13:45:24 - 00:13:53:06] Zvonimir Sabljić revenue, SEO, technical, like tech, many things. So you never know.
[00:13:54:27 - 00:14:32:07] Federico Ramallo Right. Right. And the mindset of a zero sum game, you know, that it’s either we win or the computer wins. You know, that’s a very short, you know, you have a very short vision mindset, you know, if you think like that. Right. So, you know, talking to competitors, you know, potential partners, you know, yeah, it opens up your experience. Right. And yes, as you said, you know, there’s so many things that you could do to work together even with competitors. Right. Sometimes there are opportunities for that. Right.
[00:14:33:08 - 00:14:46:37] Federico Ramallo So so, yeah, plus, you know, when I talk with when I talk with competitors, I know we usually talk, talked a lot about, you know, we share some tips and tricks and whatnot. Right. There’s something you can learn. Oh, for sure.
[00:14:46:37 - 00:15:02:55] Zvonimir Sabljić Oh, 100 percent. 100 percent. I mean, you know, people think that like competition is like, you know, you’re at war with someone. But in reality, you’re on the same market and the market, you know, you have the same troubles.
[00:15:04:03 - 00:15:29:05] Zvonimir Sabljić You have, you know, a lot of a lot of the same things. So I think from my perspective, unless you’re like really huge company who is battling for a market share, right, like Facebook or whatever, then then you have like secrets or something like that. When you’re a small startup, you know, the world is so huge that, you know, there’s place for everyone, at least at that stage.
[00:15:29:05 - 00:15:48:21] Federico Ramallo So, yeah, right. Right. The market is so big that, you know, yeah, it’s not like you’re going to compete for the same client. And even if you do, you know, there’s there’s so many different options and alternative and different ways of doing things. So it doesn’t make sense to fight for it. Right. Yeah.
[00:15:48:21 - 00:15:50:05] Zvonimir Sabljić Oh, yeah. Yeah. Yeah. Yeah.
[00:15:50:05 - 00:15:51:07] Federico Ramallo Yeah.
[00:15:52:43 - 00:16:21:10] Federico Ramallo So how was the, you know, the first day after you, you know, you sold your company and, you know, well, at least that the way that I see it is like, you know, when you when you hear the silence, you know, like, you know, when you’re a CEO of a company, a lot of people reach out to you, you know, in the company, they need something. And then, you know, there’s that time after you leave the company that it’s OK, nobody’s looking for me. Right. That silence. Right.
[00:16:22:34 - 00:16:38:07] Zvonimir Sabljić Yeah. I mean, you know, that went gradually. That wasn’t like the like when we signed the last paper, all of a sudden it was done. But, you know, we had the transition period of a couple of months and all these things. So it was a was a bit of a process.
[00:16:39:19 - 00:17:18:54] Zvonimir Sabljić But, yeah, I mean, you know, to be honest, one thing that I didn’t do good after the acquisition, I didn’t took really time off. Right. I now or next time what I’m definitely going to do at least to take like six months, six months and go and get rest and all these things after I sold my last company, I took I went on a spa vacation for, I guess, like three days. And that was basically it. And after that, I was back in a working mode. I started working on machine learning projects and this and that and this and that. So I. Yeah.
[00:17:20:04 - 00:17:29:39] Federico Ramallo Yes. Yes, I agree. Like, you know, these these activities are so intense that you actually need a time to cool off and just relax. Oh, yeah.
[00:17:30:41 - 00:17:30:59] Federico Ramallo Oh, yeah.
[00:17:32:12 - 00:17:38:46] Federico Ramallo I was reading an article that was talking about that the ideal vacation is at least one week. Right.
[00:17:39:57 - 00:17:53:12] Federico Ramallo And less than a month. Right. Because people eventually get bored of doing nothing. Right. Which I kind of understand because, you know, after two days in a hotel doing nothing, it’s just to get crazy about it. Right.
[00:17:54:55 - 00:18:21:57] Zvonimir Sabljić Yeah, I mean, I mean, I guess it depends, you know, for me, to be honest, I don’t remember when I was bored. You know, if I was, let’s say the vacation that I’m talking about, this not a vacation like I want to lay on a beach for six months, but rather I don’t want to have any constraints. If I want to, if I wake up today, I want to, I don’t know, cook something. I’m going to cook.
[00:18:23:01 - 00:18:33:21] Zvonimir Sabljić And if I want to research something on the Internet, I’m going to research. So, you know, whatever I decide I can do. So this is, I guess, the vacation that I’m talking about.
[00:18:34:23 - 00:18:42:13] Zvonimir Sabljić And, you know, do that for at least six months and just relax. Yeah. Before you start doing some real work.
[00:18:42:13 - 00:18:47:10] Federico Ramallo Right. Right. Not having work commitments, you know, through the day. Commitments.
[00:18:47:10 - 00:18:48:05] Zvonimir Sabljić Exactly.
[00:18:48:05 - 00:18:51:53] Federico Ramallo Yeah. Yeah. Yeah. I understand. Yeah.
[00:18:52:59 - 00:19:00:50] Federico Ramallo Right. So tell me a little bit about GPT Pilot. How did you decide to create this new company?
[00:19:01:51 - 00:19:20:03] Zvonimir Sabljić Yeah. So I mean, GPT Pilot is actually iteration from our previous project. So actually, and I guess this is, you know, I always looked at from the outside on some of the projects that take off. Right. And it’s always the,
[00:19:21:03 - 00:20:04:16] Zvonimir Sabljić you know, that saying that like overnight success takes 10 years to create. Right. Basically, the idea is whenever you see someone like shooting up overnight, that there’s actually like a big effort behind it. Maybe multiple year of effort. And, you know, for example, this is the case in GPT Pilot. So we released it just seven weeks ago. And just today, actually just on this call, my co-founder sent me a Slack message saying like, hey, we finally crossed 10,000 GitHub stars. Right. And we, yeah. And we released just seven weeks ago, which is pretty solid traction. Right. Like 10K stars in seven weeks.
[00:20:05:22 - 00:20:29:22] Zvonimir Sabljić But we’re working on our projects for like a year and a half. So we started iterating. We did this and that. And a couple of different iterations. And the latest iteration was Pythagora. Pythagora was fully autonomous testing. So basically once you create a code, Pythagora just scans it and uses LLMs to create automated tests.
[00:20:30:28 - 00:20:48:43] Zvonimir Sabljić And that had solid traction in like five months. We were like over a thousand stars. And we started talking to people and we talked to over 200 people. And one of the feedback that kept popping up, this will be really interesting for the technical audience.
[00:20:50:02 - 00:21:12:12] Zvonimir Sabljić And by the way, I’m curious, what will you think about this? So basically the feedback is that developers are more prone to trusting a code, application code that is generated from a test by using LLM rather than a test that is generated from an application code by using LLM.
[00:21:13:24 - 00:21:17:16] Zvonimir Sabljić Would you agree with that? Do you understand what I’m saying?
[00:21:17:16 - 00:21:18:00] Federico Ramallo Yes.
[00:21:21:16 - 00:21:44:16] Federico Ramallo So actually I agree with that because I had a lot of rejection to do TDD until I got it and it was like, wow, this is great. Because now you can write your tests and then you basically write what you expect to do in test form.
[00:21:45:41 - 00:21:58:34] Federico Ramallo So then when you write the test, the code, then you can validate with whatever test you wrote. So you treat it as a black box. So I found that concept interesting as a mindful process.
[00:22:01:40 - 00:22:19:50] Federico Ramallo And then if I can write the code and I teach my engineers about the code that you write should be aligned to the acceptance criteria, should be aligned to what is the expected behavior of the application.
[00:22:21:40 - 00:22:42:23] Federico Ramallo So if you can describe on the test what the application should do, then I think it would be more accurate for AI to build the code based on that than the other way around. Because it’s difficult to infer about the code what you’re actually doing. If you make a mistake,
[00:22:43:40 - 00:22:51:22] Federico Ramallo the entropy goes up when you’re doing from code to test. Exactly.
[00:22:53:38 - 00:23:10:35] Zvonimir Sabljić The whole premise is that there is a business logic inside a code that is not embedded in the code, but rather that sits in developers’ mind. And whenever you’re writing the test, you’re building tests based on that business logic as well and not only the actual lines of code.
[00:23:11:36 - 00:24:02:35] Zvonimir Sabljić So anyway, that was the feedback. And from that, we decided, OK, let’s create an experiment to test this out. And as we go down into this kind of a concept, we realize that the automated test is not too different. It is basically a description of what needs to happen. It’s not like a huge code base that is intertwined or whatever. It’s pretty much independent chunks of code that basically describe a functionality. And we thought that since we’re using LLMs and they’re processing natural language, we might just make so the developer can write a description of what they want to build and Pythagora builds it. But later on, we spin it off completely from Pythagora.
[00:24:04:02 - 00:24:26:12] Zvonimir Sabljić But as we were creating this concept, we were thinking how we are using LLMs or chat GPT in the past year because both Pythagora and GPT pallet are heavily developed by chat GPT. It just works so amazingly well when instructed correctly. So we started to think, you know,
[00:24:27:22 - 00:25:41:26] Zvonimir Sabljić LLMs, it’s a raw technology. It can create code. But when you’re building an application, building an application is not the same as writing code. So writing code is only one small part of ad development process. When you’re developing an ad, you need to write code. You need to run it. You need to look at the output. You need to debug it. If the debugging is hard, you need to put logs all over the code. And you need to rerun it again to see variables propagating. And it’s like this whole complex process, which LLMs just don’t know how to do. And so GPT pilot, what it is, it’s basically like a tool that orchestrates the entire ad development process by using LLMs to code, to determine if something was correct or not. And the basic premise behind GPT pilot is that it tries to offload as many coding tasks as it can from developer to the LLM so that developers actually become team leads who are reviewing the code rather than the coders who are writing the code the entire day and sending the PRs to their team leads. But rather than AI is the one who’s coding and developers is the one who’s reviewing and fixing potentially and so on.
[00:25:42:27 - 00:26:07:18] Federico Ramallo Makes sense. Yeah, because when you’re building an application, it’s not only your code isolated, you’re trying to build a domain expert API internally, right? So you have the main knowledge of everything that happens on the application, right? So you’re trying to use that to build higher complex features, right? Exactly. Exactly, yeah.
[00:26:08:25 - 00:26:30:06] Federico Ramallo So being able to have a GPT component that can actually understand that context, it’s much more important than just go to chat GPT and type, writing a genetic code for whatever, right?
[00:26:31:20 - 00:26:38:32] Federico Ramallo It’s a much more complex situation. And what language is this? GPT pilot?
[00:26:38:32 - 00:27:28:13] Zvonimir Sabljić So it’s completely language agnostic. Just yesterday or two days ago, one of the members from our Discord actually posted that they created this kind of 3D game in C sharp where you have some kind of a bowl, like a sphere that is changing color and something like that. I never did any game or with the Unity or whatever, so I was super impressed with what they were able to do. And GPT pilot walks you through the entire process, it installs Unity, it installs all the dependencies, creates first, I don’t know how it went, but first, let’s say a layout, then it creates a sphere, then it creates some functionality, then some more functionality and so on and so on. So yeah, it’s completely language agnostic, so it can create any kind of application.
[00:27:30:35 - 00:27:40:34] Federico Ramallo And how does GPT pilot and Pythagoras work together? Are they part of the same application or two independent projects?
[00:27:40:34 - 00:27:46:35] Zvonimir Sabljić They’re completely two independent projects. They actually don’t have anything together.
[00:27:46:35 - 00:27:51:54] Federico Ramallo Okay, okay. So with Pythagoras, you’re trying to do the,
[00:27:53:02 - 00:28:00:22] Federico Ramallo you know, write the code from the test point of view, right? And with GPT pilot, you’re trying to build the code,
[00:28:01:40 - 00:28:03:28] Federico Ramallo help to build the code, is that correct?
[00:28:03:28 - 00:29:10:16] Zvonimir Sabljić Yeah, so Pythagoras actually how it works, it creates tests from the code, it analyzes tests code, but it also analyzes like server activity, like variables propagating and all these different things. And from that, it generates tests, and GPT pilot just does everything. And you know, this was also one of my threads during Pythagora, as I was, you know, thinking, and one, I believe, it was like one investor, who asked me, like, what do you think is the biggest threat for Pythagora becoming something huge? And my answer was that if an AI evolves too quick, automated tests will be relatively unimportant. Because if an AI can create the code, it can create the tests as well. I mean, tests are, you know, the smallest part. By the way, you know, Pythagora, we created it in like a month or so, whatever, and it works perfectly. Like, for example, on low dash, you run one command, and on low dash, you know, about low dash, the package.
[00:29:11:20 - 00:29:18:37] Zvonimir Sabljić Yeah, so low dash with one command, it created like 1600 tests, and it got 91% code coverage.
[00:29:19:39 - 00:29:50:50] Zvonimir Sabljić Right? Which is, you know, and it just shows how LLMs actually work good with tests. And I was thinking, man, if they do this so easily, what will happen in a year when we get, you know, GPT-5 or whatever, right? And this was my biggest, biggest threat. And we realized with after building, so as we were building GPT pilot, we didn’t know how it will end up. We didn’t know if it’s going to get traction, but also if it’s going to work. And we realized that even with GPT-4,
[00:29:52:55 - 00:30:04:16] Zvonimir Sabljić GPT pilot can do wonders, can work amazingly well, even with today’s LLMs. And can you imagine, like in a year, two or three, when we get, you know, even better, better LLMs?
[00:30:06:58 - 00:30:19:39] Federico Ramallo Yes, the LLMs are evolving very quickly. So it’s very interesting to see how is that happening. And the people that I talked that are coding with LLM help,
[00:30:20:44 - 00:30:35:39] Federico Ramallo they tell me that the code that they get is based on their style, on their application. And, you know, it’s exceeding the expectations in ways that nobody imagined it, right?
[00:30:35:39 - 00:31:30:29] Zvonimir Sabljić Literally every single thing since Chet GPT came out, every single thing that I try, I get amazed how good it works. You know, my first experience with Chet GPT was that, you know, I had to do for Patagra, we were doing this kind of interception. So we were intercepting packets on TCP level, right? And, you know, to be honest, I worked last time with TCP packets in college, that was like 10 years ago, whatever. And I didn’t, you know, know this kind of a low level programming, especially in C and whatever. And I spent two days trying to, you know, to find some package and change it a little bit and make it work. And I couldn’t do it in two days. So I said, okay, let’s, you know, get someone from Upwork to just create this script. It’s a standalone script. I’ll just get someone who knows TCP.
[00:31:31:30 - 00:31:37:40] Zvonimir Sabljić And I got proposals and the price was around like a thousand or fifteen hundred dollars for this script.
[00:31:38:48 - 00:31:44:11] Zvonimir Sabljić And it would take, you know, on average like 20 hours or so, or some 10 to 20 hours.
[00:31:45:29 - 00:32:19:36] Zvonimir Sabljić And literally before I went to hire a person, I talked to a couple of guys and they were like, okay, we can do it. And literally before I went to do it, I said, let me try this chat GPT thing. It was like in January or February, whatever it was pretty early. Let me try it with that. I took two hours to talk to chat GPT and this was GPT 3.5. And it created perfectly working Redis interceptor that I would usually the developers, developers who know how to do it would take 10 to 20 hours.
[00:32:20:43 - 00:32:58:38] Zvonimir Sabljić So, you know, when you look at it, it’s also amazed at that point. And even even that I created it. So Pythagoras built in Node.js and I created this kind of a with chat GPT in Python, this interceptor. And I thought like Node.js likely doesn’t have these kind of low level things like TCP packets. Let me try it out. I’ll just ask, you know, chat GPT. Hey, can you convert this to Node.js? And it did it. It converted Node.js and I used it and we literally implemented that thing into Pythagoras. So anyway, it’s definitely amazing. Yeah.
[00:33:00:21 - 00:33:10:59] Federico Ramallo Do you think that this type of applications would make the boot camp, you know, obsolete? The boot camps or training camps like that?
[00:33:12:00 - 00:33:46:06] Zvonimir Sabljić That’s a super hard question because at this point and what also GPT pallet stands for is that we’re not replacing the developer. We think that LLMs won’t be able to replace developers for the foreseeable amount of time. And this is basically what the premise behind GPT pallet is that it can offload, you know, 95% of coding tasks. But those 5% developers still needs to be here to review, to maybe fix some things to direct, to point LLM in the right direction.
[00:33:48:36 - 00:34:12:20] Zvonimir Sabljić But also, so that is on the one side. So on that side, I would say, no, you need to know how to code. But on the other hand, I’m thinking, you know, what will happen after five to 10 years after people use GPT pilot? You know, less and less people will actually know how to code and how to go deep into issues and all these things.
[00:34:13:45 - 00:34:50:38] Zvonimir Sabljić So I don’t know. I really don’t know. It’s a pretty hard question. I don’t know what will happen to social development industry. But I know that in like five years, it will be completely different than today. I think it will be just unrecognizably different than how software development works today. I’m not sure what will happen to bootcamps in that sense or what kind of developers will there be like most need for or something like that. But it will definitely be different a lot.
[00:34:51:50 - 00:34:52:02] Zvonimir Sabljić Right.
[00:34:52:02 - 00:34:56:37] Federico Ramallo Right. Yeah, I think that they will need to adapt to this new approach.
[00:34:56:37 - 00:34:59:24] Zvonimir Sabljić Oh, for sure. Oh, for sure.
[00:34:59:24 - 00:35:03:39] Federico Ramallo The other thing that I don’t know how to.
[00:35:05:46 - 00:35:39:22] Federico Ramallo So one thing you mentioned the low level TCP problem that you needed to solve. Right. So I believe that the GPT pilot would not replace that knowledge, you know, being able to understand how everything works internally. And, you know, being able to based on that design, you know, a solution or, you know, be saying, OK, this is the solution that I want or what I want to build and then have GPT pilot to help building that. Right.
[00:35:40:44 - 00:35:46:53] Federico Ramallo But I’m wondering how can a junior will be able to get that knowledge. Right.
[00:35:46:53 - 00:35:55:56] Zvonimir Sabljić Yeah. I mean, that is the right question. Right. And I think that at this point, no one knows knows the answer.
[00:35:58:21 - 00:35:58:33] Zvonimir Sabljić Yeah.
[00:35:59:40 - 00:36:07:25] Zvonimir Sabljić Yeah, I don’t know. But, you know, starting like like a commercial development experience today, you know.
[00:36:10:15 - 00:36:28:43] Zvonimir Sabljić Yeah, I don’t know. But likely, you know, developers in the future, especially junior developers today, I think that their biggest assets should be just using the same. They should be just using as many tools as they can, because it just you know, one thing why software development is really, from my perspective,
[00:36:29:46 - 00:37:03:47] Zvonimir Sabljić you know, one of the first in line for disruption for complete disruption is because, you know, LMS have one one big problem. Right. LMS doesn’t know if it’s it’s response is correct or not or how correct it is. Right. It just tells you some text and, you know, hopefully it’s good. So if let’s say you’re a lawyer and you ask GPT to create your document after it creates a document, you need to go through it yourself and make sure that it’s correct. Right. And then you can tell it, hey, it’s correct. This is not correct.
[00:37:04:47 - 00:37:21:43] Zvonimir Sabljić Why software is in amazing position is because when LLM creates code, we can automatically or programmatically, we can run the code, we can take the output, send it back to the LLM and as the LLM. Okay, is this correct?
[00:37:23:05 - 00:37:37:51] Zvonimir Sabljić And this way we enable LLM to actually know how good their response was and actually to iterate on it and automatically debug. This is what GPT pilot does. But, you know, it can do this automatically.
[00:37:40:15 - 00:37:42:17] Zvonimir Sabljić So, yeah, right. Right.
[00:37:42:17 - 00:37:51:48] Federico Ramallo Yeah, I mean, GPT pilot will give you the direction and then you need to make corrections. So it goes exactly where you want to go.
[00:37:51:48 - 00:37:55:51] Zvonimir Sabljić Yeah, you know, I always use the reference like a team lead.
[00:37:57:09 - 00:38:18:22] Zvonimir Sabljić So whatever team lead does with coders in their in their team, this is what developer will do with GPT pilot, you know, GPT pilot writes code, runs it, debugs it, does all these things. That it can, you know, everything is automatic. But whenever you get stuck, it tells you like, hey, I tried this and this, it’s not working, can help me out.
[00:38:19:42 - 00:38:20:25] Federico Ramallo Right.
[00:38:21:52 - 00:38:24:46] Federico Ramallo Right. Amazing, amazing technology.
[00:38:26:42 - 00:38:27:17] Zvonimir Sabljić Oh, for sure.
[00:38:27:17 - 00:38:46:35] Federico Ramallo Yeah, I haven’t had the time to I saw the demo on the demo you gave us on TechCrunch, right? But I haven’t had time to actually sit down and, you know, start coding again. And when I saw your your your prototype, I was like, well, I’m eager to start coding again. Right.
[00:38:48:40 - 00:38:54:20] Zvonimir Sabljić Yeah. Yeah, you were at the booth, right? So you saw my co-founders presentation.
[00:38:54:20 - 00:39:07:22] Federico Ramallo Yes, yes. And it was very interesting. And there were, you know, it’s pretty crowded, you know, pretty noisy. So it was difficult to, you know, to actually, you know, get into a lot of details, right?
[00:39:08:40 - 00:39:14:26] Federico Ramallo Yeah. Plus a lot of people asking more questions. So so yeah, but I find, too. Yeah, go ahead.
[00:39:14:26 - 00:39:52:06] Zvonimir Sabljić But yeah, one to ask you, like, as you were going through disrupt, like, was our booth like you said, like, there were a lot of people asking questions, all these things. And that’s what I really so I was mostly away from the booth. I was walking around the whole conference trying to find investors and so on. So I didn’t really see. But from what I understood is that people were around my co-founder alone quite, quite a lot. Would you agree that our booth was, you know, had like more people than than the regular booth? Or do you think like all booths had these kind of like packs of people, you know?
[00:39:52:06 - 00:39:59:31] Federico Ramallo I stopped by your booth like two or three times because I was walking the floor as well.
[00:40:01:08 - 00:40:29:25] Federico Ramallo And it was two out of the three times was crowded. So the third time it was when I was able to just talk with your co-founder and he was able to explain me in more detail how it worked, right? It was, you know, for a few minutes before, you know, somebody else showed up, interrupted, right? So, you know, my takeaway was like, wow, this technology looks amazing. I have so many more questions that I want to ask.
[00:40:30:46 - 00:40:48:58] Federico Ramallo But yeah, but it was a lot of people interrupting, right? So we didn’t have enough time to go into the details, right? But a few questions that I asked, I found it very interesting. Yeah. And I asked him the same question about the language, right?
[00:40:50:32 - 00:41:04:51] Federico Ramallo Yeah. But I was with a friend and he was confused because he saw Python, right? So he was like, well, this is Python, right? And so I had to explain to him after we left the booth, well, it is written in Python, but it can support any language, right?
[00:41:06:14 - 00:41:08:33] Federico Ramallo Yeah. Or that’s what I understood, right?
[00:41:08:33 - 00:41:16:24] Zvonimir Sabljić Oh, yeah, yeah, yeah, for sure. I mean, when I’m testing GPT power, I’m building Node.js applications.
[00:41:17:30 - 00:41:31:01] Zvonimir Sabljić Right. Actually, Node.js is more native for me than Python. So I just, you know, create Node.js applications. But obviously, you know, you can even the C sharp 3D game, you can do that as well.
[00:41:31:01 - 00:41:42:28] Federico Ramallo That’s amazing. Yeah. I’ve been using chat GPT and asked questions in English and Spanish, you know, and sometimes they just switch, you know, half-way the chat, they switch.
[00:41:43:29 - 00:41:47:55] Federico Ramallo Chat GPT just… And it still works. It still works. Yeah. It’s like… Yeah.
[00:41:47:55 - 00:41:49:16] Zvonimir Sabljić Yeah. Yeah. It’s amazing.
[00:41:49:16 - 00:42:11:08] Federico Ramallo Yeah. The best way that I describe in particular chat GPT, right, is a creative, you know, a creative mind, right? It can give you a lot of answers or ideas, but you have to choose which one would make sense to pursue and which one won’t, right? And in those ideas, some are going to be, you know, crazy or, you know,
[00:42:13:16 - 00:42:15:35] Federico Ramallo you need to discard those, right?
[00:42:16:50 - 00:42:17:51] Federico Ramallo But… Yeah. Yeah.
[00:42:19:15 - 00:42:19:24] Federico Ramallo Great.
[00:42:19:24 - 00:42:20:04] Zvonimir Sabljić Yeah.
[00:42:20:04 - 00:42:25:26] Federico Ramallo So what are your next goals with GPT Pilot and Pythagora?
[00:42:27:15 - 00:43:14:51] Zvonimir Sabljić So we’re focused on GPT Pilot, and the basic idea right now is that we need to completely build out the technology, you know. Currently, GPT Pilot works on like smaller applications, you know, and these things, but we want to make it work at any scale. So the idea here, and these are all the mechanisms we already created, is that it can work at any scale. It can create an application with like hundreds of thousands of lines of code, and you will still be able to get in and say like, “Hey, can you implement this feature?” And it will continue and implement a feature. So, you know, these are all the things that we still need to implement at that test out and all these things. So the basic, you know, timeline is likely we will take like six to 12 months to do so.
[00:43:16:12 - 00:44:01:20] Zvonimir Sabljić You know, then we’re hoping we will have this really amazing tool that, you know, whenever I think about GPT Pilot, I think that it’s going to be just so obvious for people to use rather than not use. Because the idea is that it really takes on the app development, and developers here, you know, it’s basically like a superpower, right? You come in and you don’t need like six months to create a product. You need three days, let’s say. So anyway, our main focus is just building it out likely six or 12 months to do so. Soon we will start fundraising, so we will always need funds to do so. And yeah, that’s pretty much it.
[00:44:02:21 - 00:44:09:51] Federico Ramallo Great. Great. You’re still looking for investors and VCs or you mentioned that before, right?
[00:44:11:08 - 00:44:15:29] Zvonimir Sabljić Yeah, yeah. So we’ll go into this process of fundraising soon.
[00:44:16:29 - 00:44:30:01] Federico Ramallo So a shout out to our audience if somebody knows an investor that can recommend you guys, then they should contact you. And we’ll leave the link below on the episode so they can reach out to you.
[00:44:31:19 - 00:44:52:17] Zvonimir Sabljić Awesome. No, no, perfect. I mean, obviously, you know, warm intros is something that we’re always looking for, especially from founders who have raised before, you know, for VCs to get an intro from an existing founder. That’s really a great thing to do. So if you’re a founder, please reach out to us. I’ll be happy to also pitch to you and then maybe you can connect us for a little longer.
[00:44:52:17 - 00:44:59:37] Federico Ramallo Great. Great. I appreciate very much for, you know, for joining us for this episode.
[00:45:00:39 - 00:45:08:07] Federico Ramallo And I’m looking forward to see where GPT Pilot, you know, what do you guys take GPT Pilot next?
[00:45:09:46 - 00:45:28:12] Zvonimir Sabljić Awesome. No, no, I’m also excited. Yeah, we have an exciting year in front of us. And obviously, we’re making changes all the time and launching new features and new things. So yeah, go check out our GitHub page and see how it works. And we’re always happy to hear any feedback.
[00:45:29:19 - 00:45:32:24] Zvonimir Sabljić So whatever you have in mind, we’re super curious to hear.
[00:45:32:24 - 00:45:36:45] Federico Ramallo Great. Great. We’ll leave the links below so people can check it out.
[00:45:38:08 - 00:45:39:13] Federico Ramallo Awesome. Thank you.