Developing a Three-Week AI Curriculum as a Personal Side Project
About half a year ago I launched dentro.de/ai with a simple idea in mind: provide clear, simplified explanations of modern AI in one place: "Inside AI".
In the background there was always a bigger plan. I wanted all those loose pieces to eventually line up as a complete scaleable learning path. Not just a glossary here, a visualizer there, a blog post over there, but something you could actually follow from start to finish.
That missing piece is finally here: a 3 Week Learning Curriculum that ties everything together and walks you through the AI Black Box from the outside in.
This post is a bit of a behind the scenes story of how and why I built it, what ended up in the course, and how you might use it, whether for yourself or to teach others.
The problem that kept poking me
Around me, more and more people were "using AI", without really knowing what it is. Only to be puzzled, when it seemed not to perform well. I have the strong believe that a) it is easy to gain that AI knowledge and b) this knowledge is a huge enabler to better use this new technology.
Within our work environments we now are tasked not only to use AI, but to build workflows, educate business partners, evaluate risks and prepare for the future. All this by not being an engineer - and no desire to become one. We don't need to implement transformers from scratch. We need to reason about what these systems can and cannot do.
I went through that journey myself and was puzzled that easy-to-consume education is rare for AI. Which is why is started outlining it myself:
- one article about how LLMs work
- an analogy about the AI lifecycle
- a glossary entry here, a visualization there
- a few good videos and visualizations
It helped teaching too, but only if I sat next to them and arranged the pieces in the right order.
At some point I realised I was rebuilding the same mental map over and over again. That is when the idea of a structured curriculum stopped being "nice to have" and became the actual goal.
The mental model that everything hangs on
The anchor for the whole project is one simple model:
BEFORE → INSIDE → AFTER the AI Black Box
- BEFORE covers setup and training: data, architecture, tokenisation, model lifecycle
- INSIDE covers how the finished model functions: parameters, weights, embeddings, attention, layers, probabilities
- AFTER covers what happens when you use it: prompts, context, tools, workflows, evaluation
Once that frame was in place, all the separate elements on the site suddenly had a clear place to live. The 3 week course is basically this model stretched out in time and filled with concrete reading, watching and doing.
What the 3 week curriculum looks like
The curriculum is called "How to Learn AI in 3 Weeks". It is written for people who are serious about understanding modern AI at a high level, but who do not necessarily want to code or push matrices around.
Each week focuses on one phase of the Black Box and has three flavours: Overview, Learn, Practice.
Week 1: BEFORE - where models come from
Week 1 is about everything that exists before you type a single prompt.
You look at:
- how text becomes tokens and token IDs
- how data, scale and training define the AI Model as product
- the lifecycle of a model from early research ideas to large training runs
You read the AI model lifecycle article, browse the AI timeline, and poke around with tools like Tiktokenizer to see sentences or numbers chopped into tokens.
By the end of the week, "AI" looks a lot less like magic and more like an engineered product with a long supply chain.
Week 2: INSIDE - the model as a mathematical function
Week 2 is about peeking inside the Black Box. Orignially this was my main driver years ago: the desire to understand what exactly is happening in the Black Box and why. It is a fascinating area and people are rightfully puzzled as to why this works so well.
Focus in the 2nd week is on:
- how a trained model is a fixed mathematical function - this is VERY important to get across. People still discuss with LLMs thinking they are now teaching them.
- why everything is cut in tokens and vectors rather than words and sentences - and how this has a big influence on the functionality of LLMs.
- how the model always outputs a probability distribution over possible next tokens - one after the other.
The material mixes the long form blog post on LLMs with high signal videos (by legends 3Blue1Brown and Andrej Karpathy) and visualizers where you can literally watch probabilities flow through a tiny model.
The goal is not to memorise formulas, but to get the basic idea of what happens when sending a prompt to the model.
Week 3: AFTER - how models behave when you use them
Week 3 looks at the output of the model and clarifies concepts like inference and context window, but also haluzinations.
You explore:
- why the same prompt can produce different answers
- why different models behave differently on the same task
- how prompting structure and context help or hurt
- what simple evaluations and benchmarks look like in practice
This part is deliberately experimental. You try classic failure prompts, compare models and temperatures, and build a tiny evaluation set for your own domain.
The point is to develop a feel for strengths, blind spots and failure modes. Impressive demos are one thing, but using the tools as solid assistants is another.
Stitching the site into one coherent thing
The curriculum is not a separate product sitting next to the rest of dentro.de/ai. It is more like a guided tour that walks through all the existing content in a structured sequence. Visualizations and Glossary are helpful, but also the 3 main articles I have composed so far with teaching in mind:
The sections Industry Landscape and AI News are obviously helpful when going further.
The site started as a collection of these pieces. The curriculum is what makes them feel like a single, purposeful resource.
It also works backwards: writing the course forced me to tighten, rename, and sometimes rewrite the underlying material so that it fits the mental model cleanly. It helped me too to connect remaining loose ends.
Who I wrote this for
I had three kinds of people in mind while building this.
- People who already use AI tools a bit, but are keen to understand how they function - high level - under the hood.
- People who sit between technical and non technical worlds and have to explain AI decisions in human terms.
- People who do not want to chase scattered links and need one place they can consult or refer others to.
It is not designed for:
- builders or researchers
- people who want a coding bootcamp
- anyone looking for a bag of "magic prompts"
If you work with AI models day in day out you will probably find the level too gentle. If you are trying to get from "I am lost" to "It finally clicked", you are at the right place.
From curriculum to lecture material
One side effect of finishing this is that it is now much easier to explain the topic live.
Because the curriculum has a clear spine and a set of examples for each phase, it translates almost directly into:
- a short "AI in one hour" talk
- a half day workshop using some of the experiments
- a three part lecture series that follows the weekly structure
I designed it so that, with a bit of preparation, other people could give those sessions too. Everything is on the site and referenced explicitly, so you do not have to dig through personal notes to reconstruct the argument.
If you work in a company or teach and want a structured way to introduce modern AI without turning everyone into engineers, I hope this gives you a starting point.
Where it stands now
Is the curriculum finished and perfect? No :)
What I can say is that it now feels legit and good enough that I am comfortable standing behind it as "a solid starting point".
Future me will almost certainly rearrange things, add better examples, and fix what time and progress in the field make outdated. For now, I am mostly satisfied that the original plan worked: the loose ends of dentro.de/ai have been stitched into something that feels like a complete path rather than a pile of resources.
If you do decide to go through it, or use it with a group, I would be very interested in what helped, what confused, and what is missing. That feedback will probably shape the next half year of evening work.
In the meantime, if someone asks you "How should I learn how this AI stuff actually works", you are welcome to send them here:
Hopefully it saves you a few explanations and starts a few good ones.