Diederik Kingma

Some scientists change a field, and then there are scientists who become the invisible infrastructure of that field. Diederik Kingma, known as "Durk" in the AI community, belongs to the second group. This Dutch researcher published two papers within the same period (2013-2014) that are now part of the DNA of virtually every generative AI system in existence: VAEs and the Adam optimizer. He published them while still a doctoral student at the University of Amsterdam, and his ideas spread so fast that the field never trained neural networks the same way again.

In December 2013, together with professor Max Welling, he presented "Auto-Encoding Variational Bayes": the Variational Autoencoders (VAEs). It was an elegant way to teach a neural network not just to recognize things, but to generate them, to imagine them from within. A year later, in 2014, he published the Adam optimizer alongside Jimmy Ba: a training algorithm that automatically adapts the learning rate for each parameter. The result was that training deep neural networks became much easier, faster, and more stable. Everyone uses it today, from first-year students on their laptops to the teams training the world's largest models. His two papers combined have accumulated more than 380,000 citations in the scientific literature.

His career maps almost perfectly onto where AI has been over the past decade. He earned his PhD cum laude from the University of Amsterdam in 2017. He was part of OpenAI's founding team in 2015. He worked as a researcher at Google Brain and later Google DeepMind from 2018 to 2024, leading research on generative models for text, images, and video, including Glow, an invertible flow model published in 2018. In May 2024, the ICLR conference awarded him its inaugural Test of Time Award for the VAE paper, recognizing its lasting impact on the field. That same year he joined Anthropic, the company behind Claude, to continue his work on large-scale machine learning.

What inspires me most about Kingma's story is that his contributions were not isolated flashes of brilliance: they were practical solutions to real problems researchers faced every day. Adam did not arrive to impress anyone; it arrived to make training work better. That logic, of building tools that empower others, is exactly what makes AI advance for everyone. Every time you create something with a generative model, whether it is an image, text, or audio, there is a very high probability that Adam or VAEs are working underneath.