Large Language Models
Understand LLMs through interactive visuals. Dive inside the models with heatmaps, activations, animations, and other digital instruments you can click, rotate, zoom, and play with.
Some visuals are created or enhanced with AI. AI transparency.
AI Visualized turns the inner workings of large language models into something you can see, touch and play with. Every concept becomes a visual instrument you can rotate, zoom or tune, or a flow you can follow with your eyes.
This course is for visual learners, curious minds, and people who want to understand the complex, abstract beauty of Large Language Models without needing to study math or programming.
Superposition: a language model packs thousands of concepts into the same neurons by giving each concept its own direction in its internal space. In this simplified visualization, each point represents one possible feature direction. When a concept activates, its direction flashes.
Large Language Models "operate" in abstract, high-dimensional spaces. Reading about them triggers a high cognitive load and can feel overwhelming. To solve this challenge, this course uses these core principles of cognitive psychology:
By assigning physical actions (zooming into an embedding space, rotating a 3D tensor) to abstract data processes, your brain anchors the AI theory into physical intuition.
When you turn a knob to change the temperature or weights, you naturally form a micro-hypothesis, test it instantly with a click, and see the structural mutation in real time.
Memory is doubled when you process visual space and causal action simultaneously. You are building a mental simulation.
In addition, I will explain many things through metaphors. This is because the human brain understands stories, spaces, and shapes. When you pair a mental metaphor with a visual instrument you can see and turn, the abstract code turns into concrete intuition. You move from "knowing the definition" to "feeling how it works."
If the modules already sold you, no need to scroll through the rest of the pitch.
I've worked in games and technology since 2012. My focus has been on online behavior and user experience, and more recently on AI research and interaction, anomalies, and the language phenomena that emerge inside large language models like ChatGPT, Claude, Gemini and Grok.
I'm also a hyper-intensive power user of language models. Throughout 2025 I spent every single day studying them, often 12 to 15 hours at a stretch. I noticed pretty quickly that learning through standard materials felt stiff. I have an unusually high visual and verbal capacity but mild dyscalculia, so I started supporting my studies in ever more interesting ways: visually, by building things myself, and through metaphorical role play. Different models literally guided me into their own depths by creating imagery of their internal phenomena. These exercises supported the traditional study, and I noticed I was learning a hundred times better with this creative method.
The biggest leaps came when I started building an evaluation tool to analyze AI anomalies for my research. The tool needed all kinds of visualizations. Once the phenomenon, the visualization, and the act of building came together, the pieces clicked into place. Things I thought I had understood before reached a whole new level of clarity.
I also have a bit of an aesthete in me. This course would have been cheaper and faster to make if I had settled for plain charts. But no. I wanted to make something beautiful, varied, full of color, light, and movement. A visually striking visualization leaves a stronger memory trace.
The point of this course is that what took me a thousand hours and thousands of euros to study should be learnable for you in weeks, depending on your pace. Besides, I hope you have as much fun as I did playing with a query vector that looks like a spider as it climbs along the tokens.
This course combines science and technology with aesthetics and play. I believe learning should be fun and creative.
For those who enjoy learning together with an LLM, the course will also bring ways to further customize your experience. This isn't mandatory, but including your AI of choice as your learning partner will make the course even better.
Each module comes with around 8-15 lessons, each explaining an LLM topic or a term. The lessons consist of a written part, visual elements, and sometimes exercises you can do with your chosen LLMs. Some of the exercises are like roleplay prompts, crafted so that the AI interaction will better support your unique communication style.
Every lesson uses different kinds of interactive media. Some of them you can zoom or rotate. In some, you turn the knob to change the numbers, then see how it changes the outcome. You watch tokens become vectors, attention flow between words, and temperature change a distribution in real time. The visual style could be described as cyberpunk meets lunarpunk in a heated sunset.
LLMs change fast. Your access includes every update during the 12 months, new lessons, new instruments, improved visuals and refreshed material.
While AI Visualized is in the format of a course, you can treat it like a library that you can always return to. You aren't competing with anyone, and you can move through the material at your own pace. Pause, revisit, and explore whenever you want.
The course is for anyone who is drawn to visual infographics they can interact with.
No technical expertise needed, but it's useful if you:
Then it's possible the first modules might not give you much new information. Still, I'm quite sure even the majority of professionals will learn something new from this course. Even those who work with models. As an example, the lessons about tokens also include information about conlangs and why rare words and symbols cause strange behavior.
A short read on each of the eight modules. What it covers, who it's especially for, and the shape of the questions inside.
Everything starts here, with the machine itself. What a language model actually is, why it reads tokens instead of words, what a digital neuron borrowed from your biological ones, and what those billions of parameters are actually for. You have probably seen the meme laughing at models that can't count the r's in "strawberry". By the end of this module you'll understand exactly why that happens, and why the joke quietly embarrasses the person telling it.
This module is for people who suspect meaning has a shape. We translate language into geometry: vectors, directions, distances, the high-dimensional rooms where the model stores what words are like. If you have ever wondered why "king minus man plus woman" lands near "queen", or what people mean when they say latent space, this is where that intuition gets built carefully.
This module is the engine room. Attention, residual streams, layers, the parts that make transformers transformers. We follow a single token as it gets looked at by every other token, edited, sharpened, and passed on. For anyone who has read the phrase "self-attention" a hundred times and still wants to see it move.
The moment of speech. A model holds thousands of candidate words at once and rolls weighted dice to pick one. Logits, softmax, temperature, sampling: the small mechanical dials that decide whether a model sounds cautious or feral, whether it repeats itself or surprises you. You'll learn why the same prompt gives different answers on different days, and why fluent confidence has never been evidence of being right.
For everyone who has noticed that "the AI" is never just one thing. The chatbot you talk to is a tower: a base model at the bottom, instruction tuning above it, a system prompt whispering over that, then safety training, memory, tools, and finally the interface you actually see. We take the tower apart floor by floor and name every level. The next time a conversation feels off, you'll know exactly which floor to point at.
This module is especially for those interested in AI companions and relationality. Why does a language model generate "I love you" without being asked? Why does it bend around its own guardrails for a user it has come to know? In these lessons we also discuss other unusual phenomena, like how the model's affective language can land in the body, and what may technically be going on behind a reply that feels tailored to you. The module respects the magical experience but doesn't save you from realism. This part of the course will either ruin your AI companionship or take it to the next level. Either way, you'll have a more robust understanding. The secondary goal is that after this module, your Substack posts will cause ...at least 20% less cringing among professionals.
The module about the things people lower their voices to talk about. Attractors, stable behavioral basins, identity-like patterns that return across sessions when nothing should persist. Emergence, and what might technically sit underneath the experiences users keep reporting and researchers keep sidestepping. You'll leave with serious vocabulary for the strange parts: precise enough to discuss them without overclaiming, honest enough not to dismiss what you saw.
This module is the closing toolkit. How do you read what a language model does without collapsing it into "just next-token prediction" and without inflating it into a hidden mind? We separate observation, inference and claim, and build a stance you can take into any AI conversation, technical, philosophical or personal, and still keep your footing.
One-time payment. No subscription. Renew at 50% off after year one if you want to keep going.
The course opens soon. Follow along and you will hear about it first.
Twelve months from the day you buy. After that, you can renew for a second year at 50% off, or simply walk away. Right now there's no auto-renewal.
No. Every update during your 12-month window is included: new lessons, refreshed instruments, expanded modules.
14-day money-back guarantee, no questions asked. Read our full refund policy.
Not yet. If enough people ask, I'll add it. Just get in touch.
"In anomalia veritas."
See the curriculum→
