The Future is Open Source

AI is redefining software development day after day, pushing us toward a breaking point where closed-source programs become obsolete.

It has been less than two years since modern agentic coding tools first appeared. At the beginning, the underlying models were merely halfway capable of creating simple code and performing basic debugging. Soon they became sophisticated enough to read through long codebases and generate functioning, more complex code. Before last week, frontier models were alredy able to reproduce entire applications by observing program functionality through screenshots, without ever having seen their source code. They did this by reconstructing the behaviors through their own knowledge and eventual software-based documentation. Today we arrived at a point where AI can recreate software by analyzing its binary files!

While the technology is not yet perfect, and these new programs are not bug-free, the trajectory points toward a future where the software industry stands on the brink of transformation, since the AI models we use today are the worst models we will ever have to deploy.

How did it start?

Last week, the AI community started analyzing games and swapping characters or game logic with each other, creating all sorts of modded games, like Mario in World of Warcraft or Minecraft characters and tools in Call of Duty and so on.

Some of them went even further and started recreating closed-source programs such as the complete Adobe Suite. For example, now we have a free and open source version of Photoshop, called Photocraft. It runs on Windows, Linux and macOS. The Linux version is one of the most interesting releases, since Adobe did not support it natively.

How does it work

Agentic AI, when coupled with tools that can read executable binaries, like the open‑source reverse‑engineering framework Ghidra (released by the U.S. National Security Agency) or community projects such as “Reverse Engineer Anything” (REA), allows the model to begin with raw machine code, infer algorithms, and produce a high‑level representation that can be recompiled. However, decompiling proprietary software to obtain its source code is generally illegal.

This brings us to the behavioral imitation approach. By treating a program as a black box, an AI can interact with its graphical interface, observe user interactions, record inputs and outputs, consult publicly available APIs or documentation, and synthesize a functional replica. Since only the functionality is reproduced, not the original source or binary, this method sidesteps many copyright restrictions, which protect expressions (the concrete code) rather than underlying ideas or methods.

Near Future Forcast

Contrary to the popular belief that “everyone will become a coder,” most people will remain users rather than developers. The barrier to interacting with a terminal, managing API keys, or debugging Docker containers remains high for the average person, at least for now.

Small developers might find new niches in crafting specialized prompts, fine-tuning models for vertical domains, or providing the “glue” that connects AI-generated components to legacy systems. In effect, the software industry could migrate from a product-centric economy to a service-centric one, where the primary commodity is the ability to translate intent into executable code.

For professional developers, the role will shift toward prompt engineering, model validation, and integration architecture. Rather than writing boilerplate CRUD endpoints, they will focus on designing robust interfaces between AI‑generated modules, ensuring security, performance, and maintainability. Knowledge of software fundamentals, algorithms, data structures, system design, will remain essential, but the daily grind of syntax memorization will fade.

What will come soon

However, soon we will have polished interfaces with “plug-and-play” setups, and since the rise of natural‑language interfaces lowers the threshold for specifying what software should do, within five years, software becomes “invisible”. Instead of manually installing distinct programs like Photoshop, Excel, or whatever… users will talk with an AI assistant that dynamically assembles the necessary functionality on the fly. (By the way, we already have this, it’s in its infancy.) The assistant might retain a personal library of “mini‑apps” built over time, like a personal “App Store” each tailored to the individual’s workflow and refined through repeated interaction. (It’s a bit like Tony Stark with JARVIS.) These micro-applications could live entirely within the chat environment or in a dedicated folder on your PC, therefore there will be no explicit installation step, no visible executable, and no traditional “program” to manage. The AI will handle it all.

So in a few years, at work you might simply say, “I need a dashboard that pulls sales data from Amazon, compares it with our data, shows daily totals and lets me filter by region… plus setup a cloud workspace, invite the XY Team with read-write rights and upload the results there as well” and the AI would spin up a secure web app with authentication, database integration and all the required functions without the manager ever writing a line of code.

Challenges to solve

Despite the optimism, several hurdles remain. The reliability and safety of AI‑generated code are still concerns. Hallucinated logic, subtle bugs, or security vulnerabilities could propagate if users blindly trust the output.

Although remember: the AI models we use today are among the worst models we’ll ever deploy!