Anton Osika on Lovable, Creative Software, and the Changing Work of Building

Anton Osika imagines a future in which people have more freedom to choose which problems they want to solve. Joel Hellermark introduces him as Sana's first employee and asks how his education shaped his thinking.

Osika says he has long been curious about how things work. A school environment that encouraged curiosity led him toward physics; at university he took courses in physics, computer science and neuroscience as deep learning and reinforcement learning began to show promise.

His interest in neuroscience led to a question: could machines reproduce something like the human brain's intelligence? In Sana's early days, he and colleagues researched student learning with neural networks. Alongside that research, he enjoyed taking an idea and making it real.

That interest in building connects to Lovable, which Osika describes as a new interface for people who have not been able to create software. He wants founders and employees with ideas and domain knowledge to solve business problems or start companies without needing to be developers.

Hellermark asks how the platform evolved from GPT Engineer. Osika says new models' more general reasoning convinced him software engineering would change quickly. To demonstrate the point, he built an open-source tool that could generate and run a snake game from a terminal instruction.

The decision to start a company came from seeing that developer tools served a very small share of the world's population. Osika wanted an interface for others. He cites a friend building ShiftNext as an example of a non-coder using the product like a technical co-founder.

Osika says the product has changed as its team has developed the toolkit in which an agent operates, chosen among language models and benefited from improvements in those models. Hellermark then asks why Lovable generates an initial application quickly instead of collecting extensive requirements first.

The team tried several approaches, including gathering requirements and letting agents propose plans. It found that people contribute more effectively when they can react to something concrete: a first version may show them what they do not want or help them recognize what they were trying to describe.

Reducing the time from an imperfect idea to a visible result was central to the launch, Osika says. He reports that 40 million people had built on the platform. Hellermark sees a possible return of the polymath, with people able to build, launch and market a product across formerly specialized roles.

Asked which skills matter, Osika says implementation and execution are being compressed. The more valuable work becomes identifying needs worth addressing and using creative judgment to shape a solution. Both can be learned by attempting solutions to business, scientific or other real-world problems.

Osika also wants builders to think about the emotions their products create and what unique contribution they can make alongside advanced systems. He believes people may gain more freedom to choose their problems. Hellermark asks whether software creation will remain concentrated among a small minority, as in many creator economies.

Osika says several futures are possible, but he wants Lovable to help people consume software and content more deliberately. He describes a mission of human self-actualization and creativity, and says he would avoid monetizing addictive attention even if it were profitable.

He believes people may be more creative when friends, family or a wider audience can try what they build. Hellermark asks for a historical analogy to a resource becoming abundant.

Osika compares access to cognitive labor with the industrial revolution's expansion of access to energy. In his view, cheaper capability can allow people to tackle problems that previously required too much money or time, and may also enable new ways for people and institutions to organize.

The conversation turns to the so-called SaaS apocalypse and whether companies will build their own software. Osika says chief executives, finance leaders and CIOs want control over security, data and the code they run, as well as the ability to change software quickly and connect information across systems.

He describes a global real-estate business with hundreds of websites and 80,000 agents. According to Osika, a marketing leader used Lovable to help replatform the legacy website stack and replace chatbots, saving $2 million and reducing the time needed to learn something from about a year to days or a week.

Osika says CIOs are also replacing disconnected internal tools with systems whose actual users help develop and continually change them. He sees opportunities beyond Lovable for businesses that connect customer data across applications and address newly solvable problems.

Asked which software companies he would back, Osika says he favors leaders who innovate, stay close to customers and understand their problems. His emphasis is on the people running a company and their willingness to keep adapting, rather than a fixed category of software.

Hellermark asks whether future software will be individually generated, custom-built for organizations or based on templates. Osika expects all three. People like adapting their surroundings, he says, and businesses often want software fitted to workflows they repeat frequently.

He argues that the emotional experience of using software becomes more important as capabilities are easier to create. Templates will still help people start quickly, but they may later want to change a proven foundation. The balance will depend on the user, business and problem.

The biggest change, Osika says, is the shrinking time required to customize software: teams can iterate with customers to agree on the right solution, then build it while addressing functionality, compliance and security requirements.

Asked what he would do as a Fortune 500 CIO, Osika says every employee can now contribute to improving technology systems. He recommends giving employees access while retaining visibility into what they build and connect, then trying a hackathon or another shared experiment.

Hellermark asks whether the future will bring one dominant agent or many specialized agents. Osika expects experimentation with multiple agents and models. At the time of the conversation, he says Lovable used four models in production for different situations, including models the company had trained and others from leading labs.

Asked who owns Lovable's prompt, Osika says there is not one prompt: different information is supplied to an agent in different situations. The goal is to give it useful context without overwhelming it. A human developer ultimately decides how the system should communicate and behave.

That developer uses data to compare versions, including how many users succeed or become frustrated and where they get stuck. The team also measures narrower tasks, such as how quickly and in how many attempts a requested login can be added, so it can learn from fewer examples.

Osika distinguishes two engineering challenges. Starting a product from scratch and launching it is becoming fast, he says; changing a complex product with many users is harder, like replacing part of a car while its engine is running. Engineers must systematically manage trade-offs and risks.

For complex existing systems, he says an AI needs enough context about what can go wrong. The model improvement he most wants is better replanning: instead of getting stuck in a loop, a system should step back and reassess. Better ways to supply and retrieve context would reinforce that capability.

In his predictions, Osika hopes politicians will talk more seriously about AI and strengthen diplomatic relationships. He imagines abundance and cooperation if humanity gets the transition right. Asked for a favorite recent memory, he chooses working shoulder to shoulder with colleagues on launches and production incidents.