Generative software: perhaps that Europe has fallen behind?

There is a lot of talk about how generative AI is revolutionizing software development. The most popular narrative is "just write a prompt" and the model does it all. Like saying that because Sora makes 3-second videos he can make a movie and send it to Venice.

This is mostly the idea of investment funds in Italy, and, to a lesser extent, in Europe. Heard from the real, "I invest [100k] only if it tastes like a unicorn." The truth is that with 100k you don't make a unicorn with it — at most you make a nice little donkey with it.

Building software based on generative AI means doing training, not prompting.

In general, even today to create software that does something useful (send emails automatically for example) we have to write code that details the flow of behavior: if an email arrives from a person you have never responded to, send the notification that they are on vacation. If the email is automatic, don't send anything. If it is a reply to an email from me, send nothing. And so on.

If I wanted to design a code writing assistant, I would have programmed the steps: download the repository, search for relevant files, apply changes, launch tests. We've always done it that way.

Today it is different: instead of explicitly coding steps, we need to make an AI model learn by observing how an experienced programmer works.

Take codex, OpenAI's programmer model: "it was trained using reinforcement learning on real-world coding tasks in a variety of environments to generate code that closely mirrors human style and PR preferences." Watching the logs as he works is fascinating: he works from the terminal exactly as a 50-year-old would. No GUI, just shell.

Only in Italy you get a diploma in computer science without knowing what a shell is. Result? You won't even be able to tow a model, it will tow you!

Financially, the key is that, increasingly, creating new products means training new models. And training models means having money and servers. Lots of it. It's not enough to just take a generic model and throw a few prompts at it: to have really useful software that can program or handle email with professional quality requires a huge training and fine-tuning effort.

• A coding assistant is not born from two lines of prompts: it is born from months of code collection, pull requests, debug logs, and training on infrastructure that costs millions.

• An email assistant does not learn "by magic": one must have access to large datasets of communications, curated for privacy, and dedicated training to capture the style and needs of a domain.

This is the real frontier of software production today: no longer just writing code, or having it written by "full stack" programmers paid 20 euros an hour. But having the technological and financial strength to train models capable of replacing that code.

And this is where the big gap emerges: in the United States and China, companies and investment funds are massively funding vertical models (for coding, for biomedical research, for finance). In Europe, on the other hand, there is a lack of both fundamental models to build on and investment to specialize existing models on high-value-added tasks.

The risk: while elsewhere models capable of writing code, analyzing contracts or generating molecules are being born, in Europe we remain stuck at a superficial level, convinced that it is enough to "use ChatGPT with a better prompt." But the future of software will not be decided by those who write the prompts: it will be decided by those who own the models and have the capital to train them on the right problems.

In summary: generative AI has fundamentally changed software production, but not in the trivial sense of "anyone can program." They have changed it because today making software means training artificial intelligences, and those who do not have the ability to do so — technical, economic, and political — will be left behind.