Deep Black Box - How AI actually works, in 21 hands-on problems

What you will learn

Ch.	Title	What is inside	Problems
1	Machine learning	What does it actually mean for a machine to learn? You work the smallest version of it by hand: nudge numbers until the prediction gets closer to the answer	1-5
2	Deep learning	What exactly is deep about deep learning? You stack the layers yourself, and see both what stacking buys you and why stacking alone does not work	6-11
3	LLMs	What is ChatGPT doing inside? You run a working language model, watch it pick the next word by probability, and see why the same question can come back answered differently	12-14
4	How an LLM is built	Why is AI polite, strong at math, and still confidently wrong? You run the three stages separately: practicing next-word prediction, tuning on choices people made, and training on problems that can be marked right or wrong	15-17
5	What decides accuracy on real data	From the records of 891 passengers on the Titanic, you predict who survived using age, sex and cabin class. It is the most widely used beginner task on Kaggle, and the tool here is a decision tree	18-21

There are formulas, but you never have to do the arithmetic yourself. Move a slider and the page recalculates on the spot.

A thirteen-year-old can get through it. If you have already met derivatives and logarithms, you can push on into why it comes out that way.

Chapter 1 Machine learning

Function (a rule that turns an input into an output) — we start from y = ax + b. a and b are the parameters, the knobs you turn
Error (how far off you are) — the gap between prediction and answer, collapsed into a single number. Making that number small is the goal
Gradient descent (walking downhill) — the height of the hill is the error. You read the slope and move the parameters a little in the direction that lowers it. This loop is all there is to a machine choosing its own parameters
Learning rate (step size) — too large, and you stride straight past the bottom; the error keeps growing and you cannot get back
Limits of the form — with a straight line there are problems no setting of the parameters can reach. Add parameters so the line can bend, and you get there
Softmax (turning scores into probabilities) — adding and subtracting raw scores squeezes the gaps, so we use exponentials to turn them into probabilities that add up to 1. The order from largest to smallest stays the same

Chapter 2 Deep learning

Features (the columns you feed in) — you go from one column to several. Adding a column that carries no information about the answer does not lower the error
Normalization (putting columns on the same scale) — mix height in centimeters with weight in kilograms and one column takes over. The step size is shared by every parameter, so without rescaling, learning breaks
Limits of a straight line — some arrangements of points cannot be separated by a single line
Hidden layer (a stage in between) — you insert a stage, which gives you more parameters
Activation function (the bend) — without a bend, any number of stacked layers collapses into one straight line. So we bend between the stages
Vanishing gradients (why depth breaks) — add stages and the error signal stops reaching the early ones, so the early stages stop learning. A path that skips ahead, a residual connection, fixes it
Overfitting — fitted too closely to the training data alone (you run into this for real in Chapter 5)

Chapter 3 LLMs

Vocabulary (the exits) — the choices on the output side. There are some 50,000 candidates for the next word, each with a probability attached
Probability distribution — that lineup of 50,000 probabilities. Keep taking the highest one and you get safe, unremarkable phrasing
Temperature — the setting for how hard to lean toward the high-probability words. Raise it and low-probability words start turning up. The lineup itself does not change, so when you want to change the lineup, you add conditions to the question
Self-attention — the weights that decide which words in the sentence to use as cues. The weights add up to 1, so there is a ceiling on looking hard at several words at once
Multi-head — which is why the weights are split into several sets, so that different words can be looked at in parallel

Chapter 4 How an LLM is built

Pre-training — on a large amount of text, the model practices one thing only: predicting the next word. Grammar and facts arrive as a by-product of that practice
RLHF (tuning on human preference) — a good answer cannot be written down as a formula. So people are shown two answers, asked which is better, and that record is used to tune the model. Who does the choosing changes what you get
RLVR (training on right and wrong) — only problems a machine can mark, such as math and code. That is why math improves, and why anything without a correct answer, being interesting for instance, is out of reach at this stage
Why it is wrong with such confidence — what the training selects for is not what is true, but what is likely to come next

Chapter 5 What decides accuracy on real data

The data — the records of 891 passengers on the Titanic, the liner that sank in 1912. From age, sex, cabin class and the rest, you predict who survived. It is the most widely used beginner task on Kaggle
Feature engineering (building the columns you feed in) — when a person builds a new column, the number of correct predictions goes up. Pulling the title out of a passenger's name, for example
The training score is not a ruler — add columns and the score on the training data rises, but that is not ability. Taken far enough, it is memorization
Decision tree — a different mechanism from gradient descent. It tries every split point and keeps the best one
Cross-validation — measure once and the number wobbles. Split the data, measure several times, and take the average
Line up models built in quite different ways and they end up at much the same level. Which tells you there are situations where the input matters more than the model

What this is

You read a conversation, then work the same thing yourself

Each problem runs as an exchange between a professor and a high-school student. Rather than reading an explanation, you follow a conversation until it clicks.

Conversational introductions to AI, visual tools you can play with, and drill sites that pose one problem at a time and mark your answer all exist already.

What we could not find was one that did all three. So we built it.

	Reads as a conversation	You work it yourself	Poses problems and marks them
Conversational primers (the Math Girls series, for example)	Yes	No	No
Visual explainers (Transformer Explainer, for example)	No	Yes	No
Drill sites with automatic marking (Deep-ML, for example)	No	Yes	Yes
Deep Black Box	Yes	Yes	Yes

In all 21 problems, you are the one moving things

Not one problem is read-only. Most of them put knobs in front of you, sliders you drag left and right, and what you are moving is the numbers inside the model itself, its parameters. The problems without sliders switch things with buttons, or ask you to commit to one of three predictions first.

Every number on the screen was measured

AI is hard, so it has to be written plainly; write it too plainly and it stops being accurate. So the mechanisms here are tied to 32 original papers, the ones covered in Generative AI, Read Through the Original Papers by Yohei Kikuta (in Japanese). And wherever something was simplified for this course, the page says so on the spot.

Nothing to set up

No installation, no environment to build. Open it in a browser. It works on a phone.

Why we built it

If you want more people building AI, there has to be a way in to the principles.

That is how Yuto Yamai came to see it. He took a bronze medal at the International Olympiad in Artificial Intelligence (IOAI) 2025, and he is in his final year at Kaisei High School in Tokyo. He first ran into this field in the winter of his first year of high school, by chance. Less than a year later he had the medal. What had been missing was not ability. It was the chance to find out the field existed.

How to use AI gets talked about constantly. What is going on inside it rarely gets taught.

At the first session of Agora, run by Roppongi Venture Capital, five people spent two hours in discussion with an AI. This came out of that session, built to hand one way in to the students who light up at math and structure.

Agora is a program where exceptional young people go deep in discussion with AI and produce something that does not exist yet.