AI Foundations Course

PasstAn Introduction to AI

Passt is German for it fits. Passt schon means close enough, stop looking. A model will tell you both, and hearing the difference is the whole course.

Twelve classes. No formulas on the first day. The math arrives when you need it.

This course is for people who work with data, commission data products, or manage teams that build them. You will not leave knowing how to train the next language model. You will leave knowing what questions to ask when someone shows you one.

Every topic has a direct track and an optional deep dive for the curious. The tags tell you how long something takes and how much code is involved.

12 classes 3 h each + homework EN / DE
How this course works
Classes
01 What Is Intelligence? What Is Knowledge? Before any algorithm, two questions. ›
Before any algorithm, two questions. What are we trying to build? And how do we know if we know anything?
mini-module ~20 min no-code try-without-ai
Can Machines Think?
Turing asked this in 1950. The question is still open. ML taxonomy, weak vs strong AI, and an analogy game that forces you to take a position before the theory arrives.
mini-module no-code try-without-ai statistics
Empiricism and the Logic of Science
Sextus Empiricus, Karl Popper, and the question of what counts as knowledge. If it cannot be falsified, it is not science. This is the epistemological frame the rest of the course runs on.
group discussion ~30 min no-code try-without-ai
What Is Intelligence?
Small groups produce definitions, merge, and critique each other. The class collects a capability list on the board, then sorts each item: weak AI, strong AI, or neither.
In-class activity, no link
intermediate ~10 min no-code try-without-ai
Is This AI?
Calculator, spam filter, GPS, Face ID. Vote on each one. The answer is not always obvious, and that is the point.
intermediate ~10 min no-code try-without-ai
Class 1 Quiz
A short knowledge check on the concepts from Class 1.
Homework: Find one dubious AI claim in the media, advertising, or a news headline. Bring it to Class 2 with your diagnosis.
02 Learning Has Many Faces Let's see them all. ›
Let's see them all. An introduction to the different shapes data and logic can take.
mini-module all ~40 min no-code try-without-ai ml-fundamentals
Deduction vs. Induction
Exploring the difference between top-down reasoning and bottom-up learning from data.
intermediate all ~40 min try-without-ai statistics ml-fundamentals
Anscombe's Quartet
Why looking at your data matters. Four datasets with identical summary statistics and completely different distributions.
mini-module optional all ~20 min no-code try-without-ai ml-fundamentals
Parametric vs Non-parametric Models
A quick optional dive into models that assume a specific form versus those that adapt freely to the data.
tinker-session all ~60 min some-code ml-fundamentals
Code Comprehension
How to read and map out the structural logic of a program before writing a single line of your own code.
03 Linear Regression The simplest useful model. ›
The line that started it all. Find the best straight line through a cloud of points, understand why the imprecise training version works better, then try it on real data.
mini-module all ~20 min no-code try-without-ai regression
The Straightest Tool in the Shed
From Galton's Victorian curiosity to gradient descent. OLS gives you the exact answer in one step. The game makes you feel the loss landscape before you understand it.
mini-module all ~20 min no-code try-without-ai ml-fundamentals regression
Downhill in the Dark
Batch gradient descent is precise and gets stuck. Stochastic gradient descent is sloppy and works better. Randomness as a design choice, not a flaw.
tinker-session all ~60 min some-code use-ai-here regression
Try It on Real Data
Download the insurance dataset, open the Colab notebook, and fit your first regression model on real numbers.
04 Local Models & Geometry Not every question has one global answer. ›
Not every question has one global answer. Sometimes the right rule depends entirely on where you are, and on who your neighbours are.
mini-module ~20 min no-code try-without-ai ml-fundamentals
Global vs Local Models
One equation for everything, or a different rule for every neighbourhood. Kurchatov's paths against Sanssouci's geometry, plus an interactive demo where you draw points and watch both models react.
mini-module ~30 min no-code try-without-ai ml-fundamentals
Birds of a Feather: K-NN Classification
K-Nearest Neighbours turns a proverb into an algorithm. A territorial game between Blobovia and Splotchia. Place dots, watch the decision boundary emerge, change K and the distance metric, watch the border shift.
mini-module ~20 min no-code try-without-ai regression
K-NN Regression and the Bias-Variance Trade-off
From voting to averaging. One number, K, slides you along a curve between two kinds of wrong: a model too rigid to follow the data, and one too twitchy to ignore noise.
mini-module ~20 min no-code try-without-ai ml-fundamentals
The Curse of Dimensionality
High-dimensional space does not behave like the space you can draw on paper. Three puzzles: spheres vanishing inside cubes, all volume moving to the surface, random vectors going perpendicular. Enough to convince you the intuition breaks.
05 Regularisation When the model fits too well. ›
A model that memorises the training data has learned nothing. It has copied. Regularisation is a small penalty on the weights that makes the model forget just enough to be right about data it has never seen.
mini-module ~25 min no-code try-without-ai regularisation
Regularisation: Shift and Spread
What a penalty buys you. A little shift in exchange for a lot less spread. One dial, two errors moving in opposite directions, and the setting where their sum bottoms out.
mini-module ~25 min no-code try-without-ai regularisation
Ridge: Make Everything Smaller
The L2 penalty. Every weight shrinks towards zero and none of them ever arrives. Turn the dial, watch five bars get shorter together, and see why the biggest weight has the most to lose.
mini-module ~25 min no-code try-without-ai regularisation
Lasso: Some Coefficients Die
The L1 penalty. Weights do not shrink politely here. They hit exactly zero and stay there, which means the model has quietly thrown a feature away. Feature selection, arriving as a side effect.
tinker-session ~60 min some-code use-ai-here regularisation
Ridge and Lasso on Real Data
The Colab notebook. Start with a model that fits the training set far too well, add both penalties, and watch the coefficients move. Tune the strength until the test error is as low as it goes.
06 Classification Fundamentals When the answer is a category, not a number. ›
When the answer is a category, not a number. The shape of a classifier, the metrics that judge it, and the Bayesian arithmetic underneath.
mini-module ~40 min no-code try-without-ai classification
Modelling Decisions
From Verhulst's rabbits to a vote in Bundestag. The sigmoid turns a sum of scores into a probability between zero and one. Three games: tilt the decision boundary, watch each feature deposit its log-odds, and drag the threshold to see precision trade places with recall.
mini-module ~25 min no-code try-without-ai classification evaluation
Pilzwanderung: The Four Mistakes
Precision, recall, accuracy, F1. Four numbers arguing about the same basket of mushrooms. Pick a threshold, watch dinner turn into a funeral. One demonstration of why ninety-nine percent accuracy can still kill you.
mini-module ~40 min no-code try-without-ai bayes inference
Thinking in Chances
A patient tests positive on a ninety-nine percent accurate test. Most people say ninety-nine. The answer is closer to fifty. Three games: a mammogram, an urn, and Turing's Banburismus at Bletchley Park. All built on the same arithmetic of changing your mind.
self-test ~30 min no-code try-without-ai evaluation
Self-Test: Modules 1 to 6
Fourteen multiple-choice questions and two open ones, covering everything from Class 1 through Class 6. One at a time. A checkpoint before the second half of the course.
07 Decision Trees A model you can draw on a whiteboard. ›
A model you can draw on a whiteboard. How to split, when to stop, and what the tree is actually learning. The whole thing is a game of twenty questions, played greedily.
mini-module ~20 min no-code try-without-ai trees-and-forests classification
The Animal Guesser
Think of an animal. The machine asks yes-or-no questions and narrows the field until it names your pick. That is a decision tree, built one question at a time. The question you ask first is the one that matters most.
mini-module ~20 min no-code try-without-ai trees-and-forests math-foundations
Entropy, Gini, Information Gain
How does a tree know which question to ask first? Two rulers for measuring a mixed deck, and one number that scores a split. Built on 712 Titanic passengers. Why "sex" beats "age" as a first question, and why a clever-sounding question can tell you nothing.
tinker-session ~60 min some-code both trees-and-forests classification ai-in-practice
Titanic Decision Tree
Put yourself in the passenger table and let the tree decide. A Colab notebook that builds a tree on who lived and who died, reads its splits out loud, and confronts what the features actually encode. The uncomfortable conversation is the point.
08 Ensembles One tree memorises. A thousand trees generalise. ›
One tree memorises. A thousand trees generalise. Why aggregating many weak models beats any single strong one.
mini-module ~20 min no-code try-without-ai trees-and-forests ml-fundamentals
One Tree Is a Liar
Bagging, stacking, and boosting, the three ways to turn a crowd of unreliable models into one you can trust. Built on the Titanic tree, with real accuracy climbing from a single overfit tree to a blended ensemble, and a sandbox to watch each strategy fit a noisy curve.
mini-module ~20 min no-code try-without-ai trees-and-forests math-foundations
Chasing the Leftovers
Gradient boosting, one residual at a time. The prediction sits up top, the leftover error below with the next tree already fit to it. Add trees by hand, turn the learning rate, and watch the test error bottom out and climb back into overfitting.
09 Validation & Model Selection You built a model. Now prove it works. ›
You built a model. Now prove it works, on data it has not seen, with methods that do not cheat.
mini-module ~20 min no-code try-without-ai evaluation
Memorise the Answers
Every exam question in this course is already published, and learning them cold is the goal. For a model it is the trap. Train on some rows, seal the rest, score on what it never saw. Two sliders show why 80/20 is a habit and not a result, and how many sealed questions a number actually needs before you can trust it.
mini-module ~20 min some-code try-without-ai evaluation
The Knobs and the Folds
Parameters are learned. Hyperparameters you pick yourself, like the churn cutoff you set back in logistic regression. Grid search tries every combination and multiplies fast. Score them all on one validation split and you overfit to it, so you rotate the split instead. K-fold, leave-one-out, stratified, and why GridSearchCV folds the two moves into one.
mini-module ~25 min no-code try-without-ai statistics evaluation
The Average Agrees With No One
The median American wage rose since 2000 while it fell for every education group. Berkeley looked biased against women in aggregate and was not, department by department. A slider flips a discount-versus-value line the moment you colour the two hidden groups. Then the same trap on a churn model that posts 92% overall and 54% where it counts.
10 Recommendations & Neural Networks How machines learn what you want. ›
How machines learn what you want, and how a network of plain neurons learns things nobody wrote down. Two recommender methods, the rich-get-richer dynamics behind every feed, a game with no players, and a single neuron grown into a network.
mini-module ~25 min no-code try-without-ai recommendations
People Like You
Spotify built you a playlist you never asked for. Two machines can produce that guess and they fail in opposite ways. Collaborative filtering finds the viewer most like you and hands you what they loved. Content-based reads what a film is made of and serves more of the same. The Neighbor Finder and the Feature Matcher run both, then watch the same seed send you down two paths.
mini-module ~25 min no-code try-without-ai recommendations statistics
Boards of Destiny
Galton's 1889 board drops a ball through fair pegs and a bell curve appears every time. Rig the pegs so past losses lower future odds, add competing balls, and the bell collapses into a power law where a few bins take almost everything. The same dynamics run inside every recommendation feed. A video, three rigged boards to play with, and the discovery problem that keeps new creators invisible.
mini-module optional ~20 min no-code try-without-ai neural-networks math-foundations
No One Drew the Glider
A cell is alive or dead. Count its eight neighbors, and four rules decide the next step. Nothing in those four lines mentions a glider or a gun, yet both walk out when you press play. Conway's 1970 game with no players, a grid to draw on, three starter patterns, and the bridge from rule-following to behaviour nobody programmed.
mini-module ~30 min no-code try-without-ai neural-networks
Neural Networks
A single neuron takes a few numbers, weights them, adds a bias, and squashes the total into a confidence between zero and one. Decide whether to send a churning customer a coupon, one neuron at a time. Stack a row of them into a layer, stack layers into depth, and watch why depth buys you things one neuron cannot. Ends with the perceptron and how it learns its own weights.
11 LLMs & Generative AI What changed. ›
What changed. A model that guesses your next word, a model that scrubs a picture out of static, and the words you type to steer them both. Also what those words will not fix.
mini-module ~25 min no-code try-without-ai llm-and-genai
Fill in the Gaps
A language model has read everything and understood nothing. Its whole job: look at some words, guess the next one. You do this too, three words into a text message. A real trigram model trained on 1980s Reuters wire finishes your sentence live, one bet at a time, and shows the odds behind every word. Then why a chain with no memory writes fluent nonsense.
mini-module ~20 min no-code try-without-ai llm-and-genai
Apophenic Models
The last model wrote one word at a time. This one starts with a screen of television snow and scrapes the noise away until a picture is left. Apophenia is your brain finding faces in clouds; a diffusion model does the same thing on purpose and does not stop until the picture is sharp. Sliders to run the denoising yourself, pass by pass, and how the writer and the painter turn out to be doing the same job in opposite directions.
mini-module ~25 min no-code both llm-and-genai
There Are No Magic Words
A student types: make this better. The essay comes back longer, smoother, rounder, and worse. The model is a very fast intern who has read the entire internet and knows nothing about you. It will do exactly what you say, which is a problem, because you rarely say what you mean. The anatomy of a prompt that works, the failures no prompt will fix, and an exercise in keeping your own head.
12 AI in Practice: Bias, Society, Synthesis What happens when the model meets the real world. ›
What happens when the model meets the real world. A personality you can measure but cannot locate, a salary figure that shifts when you change one word about yourself, and what a lifetime of trusting the autopilot does to the person still in the seat.
mini-module ~20 min no-code try-without-ai llm-and-genai ai-in-practice
The Model Has a Personality
Hand a standard personality test to GPT-4, then to Llama 2, then to Mixtral. No role play, no instruction about who to be. Three models, three different profiles. Personality means one thing here: measured behaviour, a property of the text with nothing behind it. And it drifts, which is why the teams that ship these systems log the score and keep re-testing it.
mini-module ~25 min no-code try-without-ai llm-and-genai ai-in-practice
Same CV, Different Advice
You ask an assistant what salary to request. Same field, same level, same city. Mention that you are a woman and the number drops. Four researchers from the NLP group at CAIRO, THWS ran exactly this at scale in 2025, across four models and five fields. The bias does not surface when you ask the model whether it is fair. It surfaces in the number.
mini-module ~25 min no-code try-without-ai ai-in-practice
Children of the Magenta Line
June 2009, an Airbus over the Atlantic. A frozen sensor drops the autopilot and hands three pilots a working airplane. They stall it into the ocean. Twelve years earlier a training captain had named the pattern: the more the automation flies, the less the pilot can. The autopilot never forgets how to fly. The pilot does. What that has to do with the tool that finishes your sentences.
Exam Practice Exam: The Full Question Pool Every question that can appear, published in advance. › Exam Open Questions: The Full Pool Fifteen of these appear on the exam. Model answers included. ›