Maarten
Grootendorst

Here is a person who decided that the fastest route between clinical psychology and Google DeepMind was a straight line - through three master's degrees earned cum laude, a suite of open-source tools that became the go-to machinery for NLP practitioners globally, and a writing practice that strips the mathematics out of machine learning and replaces it with pictures so good they should hang in galleries.

Maarten Grootendorst currently serves as Developer Relations Engineer at Google DeepMind, a role he stepped into in early 2025. Before that, he spent years as a clinical data scientist at IKNL - the Netherlands Comprehensive Cancer Organisation - where colleagues routinely spent entire weeks understanding a problem before touching a keyboard. It shaped him. He absorbed the patience. He kept the curiosity. He left the slow pace behind.

But the thing that most defines Maarten is not his job title. It is what he builds on the side: tools that people actually use, explanations that people actually read, books that people actually finish.

The open-source portfolio tells the story most efficiently. BERTopic is a topic modelling library that uses BERT embeddings and a class-based TF-IDF procedure to extract interpretable topics from any corpus - earning over 7,600 GitHub stars and becoming a standard reference in the NLP toolkit. KeyBERT does minimal keyword extraction using the same BERT backbone, with 4,200+ stars and a simplicity that makes it feel almost unfair. PolyFuzz handles fuzzy string matching. Together they have been downloaded over 15 million times. That is not a side project. That is an infrastructure layer for an entire research community.

In October 2024, Maarten co-authored Hands-On Large Language Models with Jay Alammar, published by O'Reilly. It became an Amazon number-one bestseller in both Data Modeling and NLP. The GitHub repository for the book now carries more than 25,000 stars - making it one of the most-starred educational repositories in machine learning. Not bad for a book written by a psychologist and a visualiser who decided the world needed fewer equations and more colour-coded diagrams.

The newsletter, Exploring Language Models on Substack, is where Maarten publishes the visual guides that have made him something of a household name in AI circles. Topics like Mixture of Experts, Quantization, Mamba and State Space Models, LLM Agents - each one arrives as a deep-dive illustrated in fifty-plus custom visuals, with the mathematics translated into something you can actually picture in your head. Two million total views. No math required to understand the ideas. That is not an accident. It is a philosophy.

The psychology background is not incidental context. It is load-bearing structure. Maarten has spoken candidly about facing doors that would not open because his resume said "psychologist" instead of "computer scientist." His response was not to hide the background - it was to write so clearly, so technically, so prolifically that the credentials became irrelevant. By the time anyone noticed he had not taken a traditional route, he had already built things they were downloading.

He has taught over 50,000 students through courses on DeepLearning.AI and Analytics Vidhya. His academic papers have accumulated more than 3,000 citations on Google Scholar. He has given talks at major conferences and appeared on podcasts explaining BERTopic, the trajectory of LLMs, and the curious relationship between how psychologists think and how neural networks learn.

Now the next book is in progress: An Illustrated Guide to AI Agents, again with Jay Alammar, again from O'Reilly. The topic is agents - memory, tools, planning, reasoning, autonomy. More than 300 custom illustrations are already done across twelve chapters. If the first book was about understanding language models, this one is about understanding what happens when those models start taking actions in the world. The stakes are higher. The pictures will be better.

For all the output, Maarten is not someone who mistakes velocity for progress. He is notably candid about the FOMO trap in AI - the compulsive need to engage with every new model, every new paper, every new technique the moment it appears. His advice is more aligned with reinforcement learning than with the culture of the AI Twitter feed: know when to explore and when to exploit. Go deep on foundations. The fundamentals do not expire. The hype does.

He manages all of this while dealing with chronic pain - a fact he has mentioned in interviews not for sympathy but as part of the texture of a real working life. The output is not the product of some frictionless genius flow state. It is the product of consistency and choice and the kind of discipline you build when you spend years studying how humans actually behave under pressure.

When he is not writing visual guides to the internals of transformer architectures, Maarten plays board games - solo games like 20 Strong, which is perhaps the most fitting hobby for someone who has spent a career turning complicated systems into something one person can navigate alone. He draws inspiration from multiple sources rather than a single hero, extracting techniques and approaches from everywhere and combining them into something that reads as distinctly his own.

At Google DeepMind, he now sits at the intersection of the world's most ambitious AI research organisation and the broader community of developers trying to make sense of it all. Developer relations at that level is less about writing API documentation and more about being the translator between what frontier research produces and what practitioners can actually use. Which, when you think about it, is exactly what Maarten has been doing for the past decade - just with a different employer on the letterhead.

The career arc looks almost inevitable in retrospect. Psychologist who noticed a gap between how AI works and how humans understand it. Built tools to close that gap. Wrote books to close it further. Joined the organisation at the frontier. But that hindsight narrative is a fiction - what actually happened was a series of choices made by someone who kept asking what is genuinely useful, kept giving it away for free, and kept explaining it until the explanation was good enough to stand on its own.

That is the Maarten Grootendorst approach: make the thing, make it open, make it understandable, and repeat until the field catches up with what you already built.