About me

AI in winter

I trained my first neural network in 1992, on a 25 megahertz NeXTstation, in a seminar at the University of Koblenz. To put it kindly, that was not a good time for artificial intelligence. The field was in the middle of an AI winter, expert systems had not kept their promises, and the AI lecture was mostly about logic and Turbo Prolog.

The idea behind neural networks grabbed me anyway. But even the NeXT, the machine on which Tim Berners-Lee had invented the World Wide Web shortly before, was far too slow for networks of any practical use. And almost everything that later made deep neural networks possible was missing: large datasets, tricks like ReLU and dropout, and graphics cards that could compute all of it in reasonable time. So I put AI aside for a while.

The software factory is coming soon

At IBM I spent almost fourteen years, mostly as an IT architect and technical project lead, on projects across many industries with a focus on insurance. I kept programming the whole time, mostly in Java, and I started early: for my diploma thesis in 1995 I used the beta of Java 1.0.

During those years I heard one promise so often that at some point I could recite it: soon nobody will need to program anymore. The software factory is a term from the late sixties. Then came 4GL languages, visual programming with VisualAge, model-driven architecture, where code was supposed to emerge from diagrams on its own, then offshoring and nearshoring, later low code and no code. People kept programming anyway, more rather than less. Since then I have been skeptical of announced revolutions on principle, and I look closely at what is left of them.

Watson, ducks and a rediscovery

At the INTER insurance group I first led application development and later the entire IT department. In 2017 we brought a chatbot built on IBM Watson into customer service there, the group's first AI project. It was more lighthouse than business case, I'll gladly admit that. But it got something going in me again. In the evenings I was back at the computer myself, this time with Python, PyTorch and computing power you could only dream of in 1992, working my way through the fast.ai course "Practical Deep Learning for Coders".

My favorite project from that time: a motion camera in the garden is supposed to tell whether an ordinary bird has landed or a duck. Yes, I know, a duck is a bird too. For the model the distinction still matters, because only the ducks have a habit of sh… in the pool every spring.

Too dangerous to release

In February 2019 OpenAI presented GPT-2 and initially held back the large model because it was supposedly too dangerous. By then I had already tried to classify emails with BERT, a topic that is still with me today. At the latest now I had arrived at language models.

When I joined the board of teckpro AG in 2020, GPT-3 had just come out. I built first prototypes with it and gave talks, and we expanded our collaboration with the German Research Center for Artificial Intelligence (DFKI). After that I was Chief Technology Evangelist at adesso insurance solutions.

And now?

After thirty years of hearing that promise, I don't write this lightly: this time it's true. I no longer write code by hand, I develop with AI agents. The code comes out of a conversation, and I have the diagrams generated afterwards so the architect in me keeps the overview. Diagrams used to be supposed to turn into code, today it's the other way round. So the software factory does exist after all, just very differently from how it was imagined in the sixties.

For developers this means less time translating requirements into code and more time on the question of what should be built in the first place. They increasingly work like product managers, directly with their customers, and build solutions there would never have been a budget for before. For this to work in companies, you need architecture, quality assurance and governance that keep up with the new pace.

That is what this blog is about: what AI-assisted development looks like in practice, with all its successes and stumbling blocks, and which new AI developments live up to their promises.