One nesting map
AI, ML, deep learning, generative AI and LLMs, in their place.
What you’ll learn
Four ideas, one module.
Model vs product
Same model, different products, different failures.
Training vs inference
Why your chat doesn’t teach the model.
The system map
Eight boxes you’ll use for the rest of the course.
The system map
Eight boxes around one model.
Context
What the model sees right now
Knowledge
Documents and retrieval
Memory and state
What persists between turns
Tools and actions
What it can do in the world
Model
Generates the next tokens
Orchestration
Steps and loops around calls
Guardrails
What it may see and do
Evaluation
How we know it works
Interface
Where people meet it
Every later module zooms into one box.
Applied exercise
Dissect three real products
Dissect ChatGPT, a coding assistant and a support bot into model, context, tools, state and external systems.
- ChatGPT
- A coding assistant
- A customer-support bot
Cheat sheet · for your projects
Most of what users call “the AI” is the system you build around the model.
Ask on every project
- Which ring is this really: hand-written rules, classic ML or a generative model?
- What exactly is sent to the model on each call, and who decides that?
- Where do facts the model can’t know come from: private data, recent events?
- Which actions can it take, and who or what approves them before they run?
- How will we know it got worse after a model, prompt or data change?
Rules of thumb
- Name the box before you fix anything: model, context, knowledge, tools or another.
- Rule out context, knowledge and tools before you blame the model.
- Use a plain rule instead of a model when a fixed rule does the job reliably.
- Judge quality on a set of test cases you can rerun, not on one good demo.
Checklist
System map in five minutes
- Pick one real user request and trace it from the interface to the answer, step by step.
- Mark the model call. Every other step belongs to the system.
- Sort each step into the eight boxes; one step can touch two boxes.
- For each empty box, ask: missing by design, or missing by accident?
- Note what each box adds per request in cost and delay.
- Circle the box most likely to fail first, and name who owns it.
Leave at the door
Five myths this module breaks.
| Myth | What’s actually true |
|---|---|
| “The model learns from my chats.” | Weights don’t change during inference. |
| “It looks things up to answer.” | Only if the system gives it a search tool. |
| “Better prompts are the main lever.” | Context, data and tools usually matter more. |
| “The demo worked, so it works.” | A demo is one sample, not an evaluation. |
| “The AI decided.” | Someone configured that behavior. |
Next · Module 01