Appearance
How It Actually Works: The Context Completion Engine
Workshop 1 of the introductory sequence.
A large language model does one thing: it predicts the next token from the text it's given, then does it again, and again. That text is its context. Almost everything an AI product does, from its tone to what it knows, is decided by what goes into that context, and by a few settings around it. In this workshop, you play with interactive demos that each reveal one piece of what happens behind the scenes. Then you build a small agent of your own and change its context one piece at a time, to feel where the design lives.
This workshop opens the series "Mini-Me: Building System Behavior".
Before the Demos: Where Are You Now?
Take a few minutes to answer these for yourself. Write your answers down; you'll look back at them later in your process journal.
- What is AI?
- What's been exciting or troubling you about AI?
- What have you been doing about it?
- How would you describe your hands-on use of AI right now?
- What gaps in your understanding of AI do you want to fill?
What You'll Do
- Guess the next word a language model will write, and see the loop it runs.
- Change the "randomness" setting, and learn that it's a design choice.
- See how a model breaks text into tokens, and how vocabularies differ between models.
- Name the five design surfaces of an AI application.
- Build and test a small agent in Agent Studio, one design surface at a time.
Before You Start
You'll need:
- A computer with a web browser and an internet connection.
- Time to explore. You can stop after any guide and pick up later.
- Your own material to test with. You'll give it to your agent in the last guide, so prepare it before you start:
- your résumé as a PDF;
- three paragraphs about your goals that you'd be happy to share with a classmate;
- a paragraph about your superpower, something you're good at;
- a paragraph about a hobby you enjoy, such as cooking, music or sports, with details like specific meals or instruments;
- five questions you'd ask someone you were getting to know at a party, on a hike or over dinner.
Save the paragraphs as a PDF too, so you can upload them. Don't include anything you wouldn't show a classmate.
Earlier workshops: none. This is the first workshop in the series, and you don't need to know how to program.
Alone or with others: the demos work either way. In a group, the first demo becomes a game, and comparing notes in the last guide is half the fun.
Two Ways Through
- For a quick run-through, follow the TLDR. It's one page of short activities, with a link or a prompt for each.
- For a deeper understanding, follow the guides below, in order. They explain what each demo shows and why it matters for design.
| TLDR steps | Guide | What it covers |
|---|---|---|
| 1 | The Next-Token Loop | The next-word game, and the loop every language model runs. |
| 2 | Temperature and Sampling | Why the output varies, and who decides how much. |
| 3 | Tokens and Vocabularies | What a model actually reads, and why it differs between models. |
| 4 | The Design Surfaces | The five places where an AI product's behavior is designed. |
| 5–6 | Playing in Agent Studio | Building a small agent, changing one surface at a time, and reflecting on what changed. |
Then do the Practice before the next workshop.
Next
Workshop 2: AI as a Material: Modeling Behavior. You'll learn how a model is made, then design an agent's behavior by acting it out.
Examples
See prototypes made in this workshop on the Examples page.