Meet the Experts: Roy van Rijn - Making AI Understandable From the Inside Out
How a Java Champion turned a question about language models into an interactive way to understand them, one token at a time.
Series introduction: In this “Meet the Experts” blog series, OpenValue developers share who they are, what drives them, and what makes working at OpenValue different. After Simone de Gijt, we continue with Roy van Rijn, Founder and Director of OpenValue Rotterdam, Java Champion, and international conference speaker.
It started with a simple question. During a discussion among OpenValue’s AI champions, a group of developers pioneering AI usage in software development, someone asked Roy van Rijn if he could explain how modern language models actually work.
He could have made a slide deck. He built an interactive visual explainer instead.
The result is The Anatomy of an LLM. It is a guided journey through the machinery inside a large language model. It walks you from raw text, to tokens, to vectors, to attention, all the way to the next predicted token. And you do not just read about it. You play with it.
“I wanted to create something more intuitive,” Roy says. “Something you can play around with. Something that helps you really get an idea of how it all works.”
LLMs are everywhere now. They write code. They summarize documents. They answer questions and generate ideas. They show up inside the tools developers use every day. But for many people, even technical people, they still feel like a black box.
That was exactly the itch Roy wanted to scratch.
“For me, sharing is the highest form of knowledge. I only truly understand something when I can explain it clearly to someone else.”
Not a big database
One misconception Roy hears a lot is that AI is “just a big database” full of stored facts.
“That’s not what modern LLMs do,” he explains. “They do memorize knowledge, but not in any human-interpretable way.”
For Roy, that distinction matters. If you think of an LLM as a lookup table, you will misjudge its strengths, its limits and its risks. A model does not find an answer in a row of data. It transforms tokens through layers of learned mathematical structure. Then it predicts what should come next.
That sounds technical. It is. But Roy believes the basics can still be explained clearly. The explainer at royvanrijn.com/anatomy-of-an-llm is built around that belief.
It is aimed primarily at developers. But Roy deliberately made it approachable for people with little or no programming background too.
“I’m a visual learner,” he says. “Some concepts just click better when you can change parameters and immediately see what happens.”
From curiosity to understanding
Roy has been following AI for years. He was already experimenting with large language models before ChatGPT pushed them into the mainstream. He understood the concepts well before he started the project. But building the explainer forced him one level deeper.
That is a familiar pattern for him.
Roy has a reputation for diving into hard subjects and making them understandable. The One Billion Row Challenge. Quantum computing. SAT solving. Algorithms. Low-level Java performance. His approach is always the same. Open the black box. Figure out what is really happening. Then find a way to explain it.
That philosophy shaped The Anatomy of an LLM. It is not just a technical demo. It is a teaching tool.
“I think AI is a very powerful tool in the right hands. And I think expert knowledge will only become more valuable with AI.”
Making the invisible visible
The hardest part was not understanding the concepts. It was finding simple interactive examples that were both intuitive and honest.
Some parts of an LLM are easy to visualize. Tokenization is a good example. Type a sentence into the tokenization chapter and watch it split into tokens. You quickly notice that words are not always the units the model sees. A token can be a whole word, part of a word, a space, or something stranger.
Other concepts are far more subtle.
How do you explain attention without drowning people in matrices? How do you show the purpose of Query, Key and Value vectors without pretending they are simpler than they are? How do you make something like RoPE or the KV cache visual enough to be useful?
“The hardest chapters were the ones where a simple interactive example was hard to come up with,” Roy says.
That tension became the most important design rule of the whole project. Simplify, but do not lie.
Building with AI, guided by expertise
Roy built the project as a static website. Easy to generate, change and upload. He used IntelliJ and local OpenAI Codex extensively throughout the process.
The workflow was structured. First he generated a chapter-by-chapter outline. Then he created a progress file with the sections and implementation steps for each chapter. This file became the guide for Codex, helping implement the explainer step by step.
After each chapter, Roy inspected the result himself. He iterated on the design. He improved the interactions. He only moved on once the visual quality felt right.
It is a good example of how Roy sees AI in software development. Not as a replacement for expertise. As a powerful accelerator in the hands of someone who knows what they are doing.
At the same time, he sees AI literacy becoming a basic skill across professions.
“It is becoming something people need to learn to use correctly,” Roy says. “Not just developers, but people in all kinds of work.”
OpenValue and the culture of curiosity
Projects like this fit naturally at OpenValue. According to Roy, OpenValue is home to curious, inquisitive developers. People who want to understand how things work, not just use them.
That culture matters.
“OpenValue encourages sharing knowledge,” Roy says. “And projects like this come from that.”
It connects directly to the role developers now have in the age of AI. As AI tools become part of everyday software development, developers need more than prompt tricks. They need intuition. They need to know where the model is powerful, where it is fragile, and why it sometimes behaves in surprising ways.
That is exactly what The Anatomy of an LLM sets out to give. Not a PhD-level treatment of transformer architecture. Not a hype-driven sales story. A guided, visual way to build useful intuition.
“I want to remove that black box feeling a lot of people have.”
Removing the black box feeling
The response has been strong. People reached out through LinkedIn, X, Slack and even WhatsApp to thank Roy for making the subject easier to grasp.
That was the whole point.
“I hope people get a better understanding of what is happening inside modern LLMs,” Roy says.
He believes understanding the basics is important. Not only to appreciate what LLMs can do, but also to understand their risks and shortcomings. Because the better you understand the tool, the better you can use it.
You can follow the full path yourself at royvanrijn.com/anatomy-of-an-llm. Text becomes tokens. Tokens become vectors. Attention moves information between tokens. Layers rewrite each representation. The final state becomes a next-token distribution. The black box is still big, but it is no longer sealed.
What’s next?
For now, Roy has no fixed sequel planned. But turning The Anatomy of an LLM into a broader series is on his mind.
There are plenty of candidates. Text search. Compression. Error correction. Algorithms. Inference engines. Technical subjects that get used every day but are rarely understood deeply.
So maybe the real question is not whether there will be another anatomy. The question is what Roy should research next.
Explore Roy’s interactive explainer here: The Anatomy of an LLM.
Curious about working with developers like Roy? Visit openvalue.eu.
