AI course for school students in Germany

Source: Gesellschaft für Informatik, Informatik-Monitor 2026/27, map of the gymnasiale Oberstufe, each Land marked "Stand 2024/25". Read on 3 November 2026. Bremen is left out because the same page gives two figures for it that do not agree.

Read the table the other way round and it says that in most Länder, eight or nine pupils in every ten leave the Gymnasium without a single Informatik course in their final two years. Many of them are strong in maths. They simply never had the place where maths turns into a program that learns. Our AI course is meant for exactly those pupils, and also for the ones who do take Informatik and want to go further than the school course has time for. It runs alongside school; the school course is the one that counts for grades and the Abitur.

# Celsius readings and the Fahrenheit values that go with them (our own example data)
celsius = [-10, -3, 0, 4, 9, 15, 21, 28, 34]
fahrenheit = [c * 9 / 5 + 32 for c in celsius]
n = len(celsius)

def loss(w, b):
    return sum((w * c + b - f) ** 2 for c, f in zip(celsius, fahrenheit)) / n

def step(w, b, rate):
    # the two partial derivatives of the loss, straight from Oberstufe Analysis
    dw = sum(2 * (w * c + b - f) * c for c, f in zip(celsius, fahrenheit)) / n
    db = sum(2 * (w * c + b - f) for c, f in zip(celsius, fahrenheit)) / n
    return w - rate * dw, b - rate * db

for rate in [0.0001, 0.001, 0.003]:
    w, b = 0.0, 0.0
    for _ in range(5000):
        w, b = step(w, b, rate)
    print("rate", rate, " w =", round(w, 3), " b =", round(b, 3), " loss =", f"{loss(w, b):.3g}")

The true rule is Fahrenheit = 1.8 times Celsius + 32, and nobody typed it in. With the smallest step the model is still far off after 5,000 steps: it has found a slope near 2.4 and an offset near 15, a line that is roughly right in the middle of the data and wrong at the ends. With a step ten times larger it is almost there. With 0.003 it has found 1.8 and 32 to the last digit shown. So is a bigger step always better? Version 2 answers that.

for rate in [0.003, 0.005]:
    w, b = 0.0, 0.0
    history = []
    for _ in range(6):
        w, b = step(w, b, rate)
        history.append(f"{loss(w, b):.0f}")
    print("rate", rate, " loss after steps 1 to 6:", ", ".join(history))

The data, the code and every number in this section are our own, run with Python 3 on 3 November 2026. The temperatures are a teaching example, not measurements.

Between 0.003 and 0.005 lies a cliff. Our calculation puts the edge at about 0.0032: beyond that, every step jumps across the valley of the error and lands higher up the other side, and the next step jumps further. The edge is set by how steeply the loss curves, and the Celsius values up to 34 make it curve steeply in the w direction. That is why engineers rescale their inputs before training, and why a pupil who understands a second derivative understands a real training problem. We work through that argument with Oberstufe pupils on paper after they have watched it happen on screen.

1. Predict first. Before running Version 1, write down which of the three rates you expect to win. Then run it.
2. Find the cliff. Try 0.0031, 0.0032 and 0.0033 in Version 2 and describe what changes in the six losses.
3. Rescale. Divide every Celsius value by 10 before training, and adjust the final w. How large can the rate be now?
4. Going further. Add a little random noise to the Fahrenheit values. The model can no longer reach zero loss. What does it reach instead, and why?

Anthropic's consumer terms say, in their own words: "You must be at least 18 years old or the minimum age required to consent to use the Services in your location, whichever is higher." So Claude is for adults only, and we never use it with a pupil under 18, in class or as homework. Other tools have different rules; OpenAI, for example, allows teenagers with a parent's permission. We keep every provider's quoted rule, with the date it was read, in our AI tools age guide.

The age rules matter less in our course than parents expect, because the course does not run on chatbots. Pupils write the model themselves, in Python, on their own computer, as in the example above. No account with an AI company is needed for any of it. Where a pupil wants to try a chatbot for a project, the mentor checks the provider's rule first and the family decides.