AI Fundamentals
How language models actually behave, assuming no prior AI knowledge. The prerequisite for everything else in the track.
6 lessons
What you will be able to do
- Explain next-token prediction without hand-waving
- Budget a prompt against a context window and measure it from C#
- Choose between an embedding call and a completion call for a given problem
- Set sampling options as a policy, and say what temperature 0 does not buy you
- Pick a model and an SDK, and keep both cheap to change
- Name the failure modes a production .NET system has to contain
Lessons
Foundations
- How language models workbeginner · 12 min
Next-token prediction, what training does and does not give a model, and why that shapes everything you build on top of it.
- Tokens and context windowsbeginner · 14 min
Tokenisation, why the context window is a budget rather than a memory, and how to measure both from C#.
- Temperature, sampling and determinismbeginner · 18 min
The step that turns a distribution into a token: what each sampling option does, which values are policy, and what "deterministic" can honestly mean.
Working with models
- Embeddings vs completionsbeginner · 14 min
Two different model calls with two different jobs — when you want a vector, when you want text, and what each costs.
- Choosing a model and an SDK in .NETbeginner · 15 min
Two decisions, not one: which model answers the question, which library carries the call, and how to keep either one cheap to change.
- Capabilities and limitsbeginner · 17 min
Hallucination, staleness and non-determinism as engineering constraints — and the .NET patterns that contain them.
Capstone exercise
One exercise across the whole course. You write it locally, in your own editor.
A console app that counts, calls and costs
One tool that exercises the whole course. Point it at a text file and a question. It tokenises the file with the real vocabulary, fits the prompt to a stated window with the output reserved first, sends one call against a pinned dated model version at temperature 0 asking for structured output, checks FinishReason before parsing, validates the parsed result, and prints a report: input tokens, output tokens, the cost of this call from a price table you supply in configuration, and the model id the provider actually served. Persist a CallRecord per run so a second run can be compared with the first. It calls a real model against your own key, so it costs real money — a few pence at most, and the report is what tells you exactly how much.
- In the repository
exercises/ai-fundamentals/capstone-token-cost-report- Verify with
dotnet test exercises/ai-fundamentals/capstone-token-cost-report
Self-reported and not verified. Nothing runs your code or reads your repository — this records what you say you did, in this browser.