Back to the roots part II

In part I of this post I described how I got into artificial neural networks and machine learning research. The story ended with an AI winter. But after winter comes spring…

The deep learning revolution

A few stubborn researchers including Geoff Hinton kept true to the so-called connectionist paradigm. In 2012 neural networks were ready for a comeback with the so-called AlexNet developed by Alex Krizhevsky, Ilya Sutskever and Geoff Hinton. The convolutional neural network AlexNet won the image classification challenge ImageNet by a large margin and importantly used two GPUs (graphics processing units) for the training on the by the time huge dataset of one million 224 x 224 color pixel images. As Geoff Hinton famously teased in his 2012 NIPS invited talk “I have a supercomputer on my desk”.

Ilya Sutskever perhaps better than anyone understood that scaling was the key to unlocking the full potential of neural networks. For my part I was very reluctant to go back to neural networks. They had let me down once so it was only by the intervention of a group of very talented students that I reluctantly came back to what was now known as deep learning. We assembled our own supercomputers on the floor of the students’ office and joined the deep learning revolution. The networks now had many more layers and therefore parameters. That is what we spent the factor 21^6 extra processing power on.

Getting back in the action

In the meantime I had become a university associate professor and later professor. As a user of technology I was only a power user of the email app. I spent most of my time reading papers, doing calculations on paper, writing papers, supervising students and teaching. A lot of emailing. When I had research ideas I had to convince my often reluctant students to try them out. I also co-founded Raffle.ai, where my job was to help set the direction of technology development. My colleagues at the company built our product and I looked at it from the outside.

In the middle of 2025 things started to change. I learned how to use my Github account and with the help of colleagues and ChatGPT I slowly started contributing to the company code base. At the university I started making code for teaching purposes. The real step change in my working life came with Claude Code. I now deep dive into explorative analysis of models and data, write research code and make meaningful contributions to company features and services. Each of these tasks would not have been done by me prior to Claude Code.

Exciting time ahead

Often researchers are skeptics when it comes to what the technology in their research field can deliver because we experience up close the daily grind of trying to get things to work at all. Having experienced the profound almost intoxicating change agentic coding has had on my work life I am no longer a skeptic. I cannot pin down exactly what got me excited about neural networks and machine learning all those many years ago but I sense that it was the same feeling that I have now. The barriers that kept me from really exploring the things that interest me are gone and I want to spend my time from now on hands-on building and exploring ideas. No way I am going back to email only.

P.S. Running many agents at once is not always pure bliss as the image seems to indicate.