This Session

We have 2 hours! Here’s how we’ll spend them:

Why This Course?

• High school students are already using AI, but rarely understand it
• Most AI education focuses on one domain; this course covers four
• Students leave able to build things with AI, not just consume it
• Real datasets, real hardware, real APIs: no toy examples

What Students Will Leave With

• A working mental model of how machines learn from data
• Experience training a real model on hardware
• Confidence running computer vision algorithms on live video
• The ability to call a language model from Python and make it do something useful
• A capstone project they built themselves

Who Is This Course For?

• High school students. Basic Python is a plus, but also suitable for students with less experience
• Students won’t feel frustrated if they don’t know Python, but if they do, they’ll be able to do more
• Students who are curious about AI but haven’t seen inside it
• The course is intentionally broad over deep. Four modules in two weeks means we prioritize exposure and excitement over exhaustive mastery

At a Glance

How the Modules Connect

Each module builds on the last:

• Module 1: students train a regression model in Python
• Module 2: students train a gesture model on hardware (same idea, different form)
• Module 3: students use pre-trained neural networks (understanding what’s under the hood)
• Module 4: students call powerful pre-trained models via API (applying everything)
• (The progression is: build intuition → train something → use the trained thing → build with it)

The Mongolian Thread

Week 1 uses datasets from Mongolia, by design:

• Day 2: Ulaanbaatar air quality: PM2.5 levels 30–40× the WHO safe limit
• Day 3: Dzud livestock mortality: can we predict a catastrophic winter event?

The Mongolian Thread

Why Mongolia?

• Striking, unfamiliar data that isn’t the usual US/EU context
• Real stakes. These datasets describe real harm
• Students can relate to the context before they’re asked to analyze

What Is a Jupyter Notebook?

• Every activity in this course runs in a notebook.
• No installation required: Students open a link in Chrome.
• A notebook is a document that mixes code and explanation in a single file.
• Code cells: write and run Python; output appears immediately below
• Markdown cells: text, headings, images, and formatted explanations

Why Notebooks for AI Education?

• Students see code and output together. No switching between files
• Instructors can demo live and students follow along cell by cell
• Partial completion is meaningful: A student who runs half a notebook has learned half the lesson
• Output can be text, charts, images, audio, 3d scenes, or live video: all in the same document

Activity: Introduce Notebooks!

Let’s walk through the same notebook students open on Day 1:

1. Open labs/01/notebooks/hello_world.ipynb in Jupyter K-12
2. Run the first few code cells: notice variables persist across cells
3. Find the hidden cell and use the eye icon to reveal it
4. Try the #@param form fields: change NAME and GREETING and re-run
5. Scroll to the Check Your Understanding section at the end

Jupyter K-12: Debrief

• How is this different from tools you’ve used before (Word, Google Docs, PowerPoint)?
• How might the eye icon change how you deliver a demo?

Module 1: Days 2 & 3

Day 2: Working with Data

Students load a real Ulaanbaatar air quality dataset and use pandas and matplotlib to investigate why the city has some of the worst air pollution on the planet.

Day 3: How Machines Learn

Students load a dzud disaster dataset, find which climate variables correlate with livestock mortality, then train a linear regression model using scikit-learn to predict future events.

By the end: students have trained their first ML model in Python.

Activity: Air Quality Investigation

What students do on Day 2 (you’ll try a version now):

1. Open the air quality notebook in Jupyter K-12
2. Load the Ulaanbaatar AQI dataset with pandas (.head(), .describe(), .info())
3. Plot PM2.5 levels over time with matplotlib
4. Identify the seasonal pattern: what time of year is worst, and why?
5. Calculate a correlation coefficient between temperature and PM2.5

Module 1: Debrief

• What did you notice in the data?
• Did the correlation coefficient match your intuition?
• How would you frame this investigation for a student who has never used pandas?

Module 2: Days 4 & 5

Day 4: AI in the Physical World

Students connect a BBC micro:bit wearable via radio to CreateAI, record 4–5 gesture actions, and train a gesture recognition model. Then they swap with another pair to test whether the model generalises.

Day 5: Coding the Response

Students write MakeCode Python event handlers that respond to each detected gesture, flash the model to the wearable, and build a mini-project (fitness tracker, clapper, safe driving app, or original idea).

Activity: Gesture Recording

What students do on Day 4:

1. Pair up: one wearable micro:bit, one radio receiver per pair
2. Connect to CreateAI via the radio option
3. Connect the wearable first, then the receiver
4. Record 4–5 gesture actions (e.g. waving, writing, shaking)
5. Click “Train model” and test each action
6. Swap with another pair: does their model recognize your gestures?

Module 3: Days 6 & 7

Day 6: How Computers See

Students explore how neural networks work (interactive browser demo), then run five pre-trained computer vision algorithms: face detection, object detection, pose estimation, and a 3D pose avatar.

Day 7: CV Mini-Project

Students add gesture recognition and image segmentation, then combine any of the five CV algorithms with their own Python logic to build an original mini-app.

Five CV Algorithms

Each algorithm has its own notebook. Students run them on live webcam video.

Activity: Pose Estimation

What students do on Day 6 (you’ll try a version now):

1. Open the pose estimation notebook in Jupyter K-12
2. Run the notebook: your webcam will activate
3. Move around and observe the 33 landmarks being tracked
4. Then open the 3D pose avatar notebook: the camera view disappears, replaced by joint positions in a live 3D scene
5. Reflect: what could you build with this?

Module 3: Debrief

• What surprised you about how accurate the landmark tracking was?
• What did it get wrong?
• What mini-app would you have built in the 25-minute project time?

Module 4: Days 8 & 9

Day 8: Language Models

Students explore word embeddings (King − Man + Woman ≈ Queen), then call a language model from Python using the OpenAI SDK. They apply streaming and structured output to build more useful programs: including generating a piano melody and a 3D scene.

Day 9: Images and Vision Language Models

Students generate images from text prompts using a diffusion model, restyle photos with image-to-image, and query a Vision Language Model (VLM) with their own images: ending with a three-step seeing assistant.

Activity: OpenAI SDK + Structured Output

What students do on Day 8 (you’ll try a version now):

1. Open the OpenAI SDK notebook
2. Send a basic chat message to a language model
3. Modify the system prompt: observe how it changes the model’s behavior
4. Open the structured output notebook
5. Use Pydantic to ask the model to return a structured response (e.g. a list of items with fields)
6. Try the piano notebook: ask the model to compose a short melody as structured data

Module 4: Debrief

• How did changing the system prompt change the model’s behaviour?
• What did structured output unlock that plain text couldn’t?
• How would you explain a diffusion model to a 16-year-old in one sentence?

Day 10: Capstone

• Students choose what to build:
  • Explore more datasets with pandas and matplotlib
  • Re-train micro:bits for a more complex project
  • Create a new game or app using CV algorithms
  • Use a generative model to create text, images, or structured data
  • Combine two or more module types
• The capstone is open-ended by design: the point is student agency, not a prescribed deliverable.

Key Takeaways

• The four modules are a coherent progression, not four separate topics
• Real data and real stakes do the motivational work
• Students don’t need to master each tool; they need to see what’s possible
• The hands-on activities are the lesson (slides are the scaffolding for the class)
• Failure is a teaching moment: the gesture model that doesn’t generalize, the VLM that hallucinates, the prompt that produces nonsense

Resources

• Syllabus: labs/Syllabus.qmd
• All notebooks: labs/01/ through labs/09/
• Jupyter K-12 platform: jupyter-k12.org
• createai.microbit.org
• OpenRouter API keys: contact the course coordinator

Questions?