Sana’s first engineer and Lovable founder Anton Osika joins Joel Hellermark to discuss how AI is turning software creation from a specialist skill into something anyone can do. He shares how Lovable lets non‑developers build and iterate on real products in minutes, why the future of work will hinge on choosing the right problems and crafting great experiences, and how this “SaaSpocalypse” could free people to focus on creativity over routine execution.
Watch the full episode above or read the transcript of their conversation below.
Opening remarks and interview introduction
Humans fundamentally will get more freedom to be able to select what type of problems do I want to try to solve for and I think that's a very exciting future.
We're so excited to have you here with us, Anton. For those that don't know, Anton was Sana's first ever employee. And I remember our first interactions and you were one of the most clever people I had ever met. But you also had a pretty interesting background and I'd love to start from there. So maybe we could just go back to where you went to school and how that affected how you think.
Early curiosity, schooling, and university
Sure. I've been obsessed by understanding how things work, similarly to you I think a bit. And that made me later go into physics, you know, and a lot of that curiosity, understanding how things work. I think it could have come from going to the school where you don't learn so much. They really want kids to not be burnt out by the teaching, but rather be curious, to want to learn. And so then when I got into university, I was like, wow, here I can select more and more courses.
So I took way too many courses in physics, computer science, and then towards the end, AI and deep learning started to become a thing. And through that, you know, that's how we met. When we started the meeting, there were starting to be like some glimpses of AI. It was, you know, traditionally had been more expert based systems. Deep learning started to work in certain domains. There was this sort of AlphaGo moment with reinforcement learning starting to work.
But what were sort of your initial curiosities around AI? Like why did you so early decide that would be sort of your life's work?
Neuroscience, intelligence, and early AI research
So before I went into university, which I think was very fun. I know some projects, they just skip that completely. But then I was thinking, should I go into something with neuroscience? Like how does intelligence work? So that's where the curiosity started for me. And I took some courses on top of the engineering courses in neuroscience. And then very naturally, if you see that there is a sign that you might be able to do what a human brain does, like the most interesting complex thing in the universe, we artificially, that's obviously a very attracting topic to think about and to practice in as a researcher, which is what we were doing a lot in the early days of SANA, right, as well.
Modeling student learning with neural networks. And since then, for me, the other big thing has always been how exciting it is to build something, to take an idea that you want to create and make that become reality. And I mean, that's what you've been doing, right? And what everyone who's building a startup or just a software engineer building a product is doing. And then this is, of course, very connected to starting Lovable three years ago, which is fundamentally a new interface where the 99% who has not been able to create in this domain can take their ideas and solve business problems, create entire new companies.
Lovable's mission and early architecture experiments
And that is growing really fast just because of what I thought was true before, which is that every human is born to create and has a lot of creativity that we want to enable them to get out into the real world. And if you look at the history of Lovable, there was first there was GPT Engineer, there was then also an early version of Lovable where you were using a lot of sub-agents and so on.
Could you talk through like from GPT Engineer to the initial version of Lovable to where we're at today, how the architecture of the platform has shifted?
Yes. So when we had new AI models that were able to actually reason more generally, it was obvious to me that this is going to completely change a lot of things, especially in software engineering, and it's going to happen quite fast. And since most people didn't understand what I was talking about, I wanted to prove a point. And that's how I went on to create this open source project that is just you type into your terminal, create me a snake game, and then it creates all the code and then it starts. starts the game, right? And that was just before I started the company, and I even wasn't sure if I would be building in this domain if I was starting a new company.
But the realization that made me do it is that just much less than 1% of the world are developers, which all the AI developer tooling companies were building for. And instead, you should be focusing on what is the next interface that is not what developers are used to. And by doing that, you can just empower so many more humans on this planet.
Focusing on non-developers and empowering new founders
And that's what we're seeing now that a lot of founders and people in large companies that have great domain expertise, they have drive and creativity in how to solve new problems, they can use our product and just work with it as a technical co-founder. And someone like Alan, a friend of mine who is building ShiftNext, he's building a company making a million dollars in just five months. And this is because he's an amazing individual who was not able to code, but now he's able to do that.
And during the one and a half years since we launched, the product has of course changed a lot. And it's driven by how we create this platform and the toolkit environment that an agent operates in, and how we optimize using different large language models, and of course, how those models are getting better and better every single month almost. And I think one of our dreams had always been that we would create sort of ubiquitous software and find this magic moment. And with Lovable, it felt like you truly hit that.
Instant app generation and fast human feedback
And as you built such a system, one approach could have been that you had a lot of upfront data gathering, and then you generated the app. But the magic of Lovable was that you sort of generate an app instantly. How do you think about that from a user experience perspective of generating something very quick that the user can then iterate on versus gathering a lot of that Intel upfront?
So we went through many different iterations before we launched the product, right? And some of them were like, okay, asking what are the requirements? And some of them were more like the agent going out on itself and proposing plans and spawning sub-agents. But what we realized is that if you want to get the most intelligence out from the human, you want them to as quickly as possible have something to react to. And then you trigger a lot of patterns in the human brain of like, no, that's not at all what I'm thinking. Or like, now I understand what I was actually looking for by just seeing the first example. So making that time from putting in your idea, even if it's not well articulated, and getting something back is something you kind of want to reduce. And when we did that, this is where we saw, okay, now the product is really good. And we went on to launch it.
The return of the polymath and changing skills
And since then, we've had 40 million people building on the platform. And the beauty of this is too that it's the return of the polymath in a sense. You know, over the last decade, we've seen increased specialization. Now, Lovable is empowering people to be very polymatic. They can do everything from generate their app to launch it, do the marketing campaign and so on.
What do you think this means in terms of what's really useful for skills? Do you think people should just develop exceptional taste and as a society will just be bottlenecked by people with a lot of taste or will taste also get commodified and something that comes after that? What do you think will be useful to learn?
It's a very good question. I think before the bottleneck was really this implementation execution that took a lot of time. Now that is being compressed. And what we're seeing being the more valuable thing is to understand what are the needs and problems that are valuable to where we're solving where you should be spending your time. And then some of this creative taste genius on how does the solution to those problems actually look like. And I think both of these things are things that you can learn by trying to create solutions, both for more kind of like business problems that have to be solved or science problems and some things out in the real world.
And of course, the problems that many companies and many of us care a lot about, which is to create the right emotions and positive emotions in users and humans that interact with what we create. And I think everyone should think a lot about what is my unique ad on top of and together in synergy with this very, very sophisticated systems here as we progress. And we humans fundamentally will get more freedom to be able to select what type of problems do I want to try to solve. For, and I think that's a very exciting future, and when you think about all other creator economies, it's often been a power law: you have one, two percent creating all of the content that the rest consume.
Creator economies, addiction, and healthier tech
Do you think software is different, or will we see a scenario where everyone is truly creating software, or will we still have this sort of one, two percent folks creating software but just using these tools?
I think there are many possible future scenarios here, and one of the things that I want our company to drive is to make it easier for us humans to be more deliberate in how we consume software and how we consume content, rather than just doing what our primal drives make us feel like.
So I hear that there's a lot of people who are addicted to technology in a way that, if they reflect on it, they are not very happy about the relationship with technology, and our mission is to empower human self-actualization and, specifically, creativity, and that means that if we get a chance to monetize, just addictive behavior, which is like human attention, can be quite valuable to monetize. I will actively avoid doing so and instead make sure that humans spend more time on creativity, and I think one of the important parts on that is that you will be more creative if you have friends or family or like a large part of a pool of users that will be interested in consuming and trying out what you have built. So I think that's an important ingredient in the future and I'm betting on that that is going to be the case in the future as well.
Industrial revolution analogy and re-organizing work
Yeah, that's really cool. What do you think are the best historical analogies for this? Like when you look back in history, something becoming abundant. What sort of historical analogy do you think can give us clues of how this is going to play out?
I mean, it's an industrial revolution, but in terms of getting access to cognitive labor, right, and it's in the industrial revolution. You got access to energy, which made it possible to build a lot of new things that were previously too expensive to build, and now the same thing is happening and that is fundamentally going to make all of us not work on this like existing companies that need a lot of manpower to keep these companies running, but go out and be able to solve problems that were too expensive to solve before, for they just need a longer time.
Does it really result in breakthroughs on the technology or breakthroughs in terms of how us humans organize ourselves to really make some new type of organization—maybe, say, the government organization or similar—come together, that fundamentally does something that us humans thinks is a good thing to happen, and there currently is a sort of sas-pocalypse ongoing that you and I are sort of on the two different sides of.
The 'SaaS-pocalypse' and companies owning their stacks
Yes, it's an exciting event I'd love to hear your thoughts on. You know, where do you actually see scenarios play out of companies just completely creating their own software soup to nuts, and in what areas do you think this is a bit overblown?
Yeah, so we're seeing a lot of people—CEOs, CFOs and CIOs—who are looking at how do we run this company on software, and they want to be in control both of that. There's no risk of security incidents. They want to know exactly where the data is and control of what code is being running, and they want to make sure that they can evolve this software stack quickly and easily and have that- the control over the data so you can connect it together with being able to evolve it. That's fundamentally unlocking a bottleneck for any business to run. So that's the reason we're seeing.
There's a fun example of a real estate agency that's global. They have dozens of countries, hundreds of websites and 80,000 real estate agents where their technology cycle is super slow to move around and the VP of marketing that really understands like you need to be able to open new brands in new countries all the time that he could use Lovable, and they were able to re-platform this legacy stack for hundreds of websites on to something that was running on Lovable and, as well, replace the AI chatbots on the websites, saving $2 million and, more importantly, going from a timeline of a year to days or a week to learn something new.
And we're seeing the same thing driven by CIOs that are using internal tools that are the data is not connected, and replacing that with something that the actual operator of that internal tool, who's the actual customer of it, is a big part of developing it and who can then continuously change it over time. And that is a huge unlock right. And I think there's this SaaS apocalypse. It opens up a lot of opportunities for many businesses, not just us, who can be good at understanding. Okay, what are the new capabilities that open up here? Fundamentally, buyers want data to be shared across all of their applications, and if you can do that really well and you can open up to solve new problems for your customers, I think there's a lot of opportunities for many different businesses in this super exciting age.
And if you take this SaaS bucket, what would you be short and what would you be long?
I would be long. Companies where leadership is really leaning in and innovating and they're close to their customers, which is fundamentally what I think is the most important value that the employees who work with me, that they put their customers first and really understand their problems and try to find new ways to solve them with less friction and with more new capabilities that did not exist before. So I would look at the humans running those companies to decide what to go along on, because everyone has to constantly disrupt themselves and that will continue to be the case.
Who will build software: employees, companies, and templates
So when you start looking at the types of software that is going to be generated, do you see every employee increasingly generating their own software? Are companies going to build bespoke software for their companies? Is it going to be more templatized? So you have a certain set of taste makers creating templates and you can adopt their software. How do you see that playing out?
I think we're going to see all three of those things happening. A lot of us humans who are creative, we like to adapt things around us. I want to decorate my house like this. I want my software and the technology part of my life to work in a certain way and in many cases that creates a lot of value. If you're a business as well, if you have a workflow that you're going through many, many times and you know you can optimize it, you will do that and you will want to have the perfect software for whatever you want to create. And in the end of the day, both for businesses and for individuals, the emotions that you feel when you interact with your software is becoming more of the important thing now that it's becoming quite easy to make sure it's number one secure and number two has all the capabilities that you want it to have.
So we're going to see both of those and then, in terms of things being templatized, it's definitely the case that if you want to move fast- you're just starting a business. You don't want to code or recreate an entire software stack from scratch—absolutely not—but you will want to once you have something that is battle-tested, many other humans have kind of proved that this is good, this is 100% secure, then you will want to do changes to it, and that's the mix that we're going to see, and it both depends on who the user is, exactly how this mix looks and what type of business or actually problems someone tries to solve there, and regardless, the biggest change we're seeing is that if you do want to change something, you do want something custom. The time that it takes is just rapidly compressing, both because of how much faster it is to align on what should we build, like what is the right type of solution for your customers, by going back and forth with them in real time with different versions. And once you know what you're going to build, making that become reality and making sure that it checks off all of your requirements in terms of compliance and security and functionality.
CIO strategy and empowering every employee
So if you took over as a CIO of a Fortune 500 company tomorrow, how would you think about your AI strategy and how would you be implementing Lovable?
There has never been a better time to be a CIO right now in how much you can impact your organization. I think you should be able to see every person at your company as someone who is contributing to creating better IT technology systems. And that's because they can finally build.
And what you can do if you're using Lovable is that you can have control over what are people building, what are people connecting, which is a huge part of being able to build an AI native company. Now anyone can go to Lovable and say, I want to connect to Sana, I want to connect to Workday or I want to connect to my data warehouse. And that opens up so much possibility for people doing their jobs more effectively. Humans having more context, which is what fundamentally humans also need to take better decisions and develop your business into do new things, serve your customers better. I just think you should get your hands on this technology as a CIO and you should give access to your employees and run a hackathon with Lovable or do something together with both of us at the same time.
Future agent patterns and multi-model ecosystems
And what do you think the patterns of Lovable will look like? Do you think there will be one global super genius agent or are you going to have a lot of individual agents fighting it out? Are you going to have a Johnny Ive agent fighting a Jeff Dean agent fighting a Demis Hassabis agent that are all sort of collaborating always on improving your software? Or will this sort of all consolidate into one sort of giant agent?
I mean, if you go and look at the very technical parts, I could imagine that already now in one large language model, there are multiple kind of personalities that are battling it out in a mixture of agents framework. So there's definitely going to be some version of that over the long term. And then since things are very much evolving and we have to experiment like we always have since like the beginning of life, things have been a big experiment.
I do think it's going to continue to be completely different agents and models battling it out or being used for different parts of products like Lovable, which is what we're doing now. I think we have four different models used in production that are used in different cases when you're talking to Lovable right now. Some of them we have trained them and many others are coming from the leading labs.
Prompt design, ownership, and data-driven iteration
And when you look at the Lovable prompt, how do you iterate on that? Who is the sort of master of the of the prompt? I remember seeing a picture of the sort of prompt printed out on the wall. Could you give a bit of color of like, how do you iterate on that prompt? Who owns it? What metrics are driving any changes to it and so on?
That's a fun question. First of all, it is not just one prompt, which is a bit misleading. Yeah, I would hope so. It's many parts of what that goes into the agent in different scenarios, you know, different parts actually goes into the agent in different scenarios, which is a big part of what has made our products so good. The people just feel a difference when they use it. It knows what to do, but it's also not over-prompted with too much information in one go, as we say.
But the main decider of how the prompt should be changed is a human developer who's trying out the changes and have this human feel for that. This is how we want it to talk to our users. And then at their disposal, that human, of course, is using data. And otherwise they can't make that decision. And looking at what changes fundamentally when you compare two versions of the prompt is what is the main source of that data. And when we say what changes, we're both looking at how many users are successful, how many users are frustrated if they get stuck somewhere, where they're trying to build something that was harder to build than they thought that it takes many iterations. And then we're looking at actually more specific things, like if the user asks for adding login, how fast is it? How many tries does that take specifically? And that makes it possible to take those decisions faster with fewer examples, because you're zooming in on this very, very specific thing.
So yeah, the data flywheel is the biggest decider. But in the end, it's a human that kind of takes everything into account, which us humans, I think, are much better than AI at doing.
Limits of software engineering automation and needed model breakthroughs
And how far away do you think we're from the limit of software engineering and what do you think are going to be the bottlenecks along the way?
So I think it's fundamentally two big parts of software engineering. If you're building from scratch, our product specifically makes it very fast to build almost anything and then be able to launch that and get users on the platform. Something that no one has solved yet, and we're not that close either, is that when you have a product that has a lot of users and it's a complicated product, then changing it is like changing the engine of a car that's kind of running or changing parts of that car when the engine is running. And that's a harder problem and that's where proper software engineers need to think a lot about different trade-offs in terms of how do we minimize risks that anything goes wrong and look at this from a very, very systematic approach. So the first one, I think we're very close. It's up to the human operator to use the AI in an efficient way and by just like trying and then learning from exactly how to use, how to instruct the AI, it's more about that.
For an existing complex product, I don't think we're so far either. You will, you kind of, depending on what the system is exactly, you need to make sure that the AI has context on all the things that can go wrong as you're changing the engine of this running car. What do you think, what model breakthroughs would make the biggest difference? Would it be context length, just increasing the intelligence of the models? What do you think would make the biggest difference for the quality of the software it generates?
I think the most stupid things we're seeing now is that the model tries out doing a few things and then it kind of gets stuck or just keeps looping, it doesn't take a step back and do a proper re-planning, re-assessing the situation of what it's trying to achieve. So I think more intelligence in that dimension is the biggest unlock. And then coupled with that, the context engineering of like how, what can you even put into the model or how does the model retrieve information from the outside world is the other important enabler that has a synergy with the first.
Predictions, politics, and a hopeful future
And back in the days, we always used to do this end of year predictions where you would force us to put stuff into your spreadsheet and then we would revisit them at the end of the year. What are your predictions for this year?
I think I remember us betting on SANA Learn being very successful. Yeah. And there was... that was very right there. Exactly.
So predictions for this year, I think fortunately we're going to have politicians starting to talk about AI for real, which I've been surprised by that it hasn't happened as much yet. And I think it's a very, very important topic, especially on how diplomatic relationships should have to be strengthened as we go through this as one species of humanity together. And talking about that, if we get this right, we're going to live in a world of abundance where we should be able to align our differences in value so that we have a world without wars and where we just colonize the future light of the universe in a way that every human is happy about. And that's something I really want for politicians. I think they will start to talk about this year.
Reflections and fondest memories at Lovable
So this last 18 months with Lovable feels like a lifetime. What are some of your fondest memories from the last year?
I think it's been the most fun when you work like shoulder to shoulder with other humans, my colleagues, and we solve a hard problem together. So going through everything that goes into making a launch really good, or even solving a production incident, those have always been my fondest memories