AI Engineering from Scratch: choose a route through the curriculum
Measure learning cost and progress with runnable evidence
Compare study routes using your time, tool spend and demonstrated understanding.
What you will learn
- Define a result before the session
- Account for paid dependencies
- Review mistakes as data
Before you start
- Basic command-line access
- A bounded learning goal and disposable checkout
Turn lesson commands and code changes into a reviewable record of what you can do.
Key takeaways
- Lesson count is not a duration estimate.
- Provider cost depends on the exercises chosen.
- Mistakes should remain visible in the learning record.
Define a result before the session
For a math lesson, a result might be explaining a matrix-vector operation and changing an input without breaking the script. For an agent lesson, it might be tracing a tool call and an error case. Decide this before reading, so the quiz is not the only measure.
Track minutes spent reading, running and debugging separately. A short route can be harder than a long one when it hides prerequisites. The repository’s lesson count does not predict your completion time.
Account for paid dependencies
Early examples can run locally without model API calls. Later LLM and agent exercises may use providers or heavier hardware. Record provider, model, token usage or device time only when an actual run produces those values.
Do not copy cost figures from another person’s hardware. Keep a budget cap for practice and a stop condition when a lesson would require a paid service you have not chosen.
Review mistakes as data
Keep failed commands and wrong quiz answers with the corresponding lesson path. Re-run one case after a correction and explain what changed. This is more informative than a completion percentage alone.
No benchmark was run for this series. The worksheet below is a suggested reader artifact, not a feature shipped by the repository.
Decision guide
| Criterion | Option A | Option B |
|---|---|---|
| Best when | You need predictable behavior and easy auditing | You need adaptive optimization and have reliable telemetry |
| Main risk | May leave performance on the table | Can become difficult to explain or debug |
Implementation steps
- 1
Set one observable outcome for a chosen lesson.
- 2
Record time and any paid usage from an actual run.
- 3
Revisit one failed case and explain the correction.
Copy-ready example
lesson,goal,minutes_reading,minutes_running,paid_usage,result,unresolved
01-dev-environment,explain preflight,,,,not-run,Frequently asked questions
How long does the whole curriculum take?
The repository does not establish a universal duration; measure your own route and prerequisites.
Is every lesson free to execute?
The material is open source, but a selected exercise may require paid compute or model access.
Sources
- AI Engineering from Scratch / README.mdSource checked 2026-09-29
- AI Engineering from Scratch / phases/00-setup-and-tooling/01-dev-environment/docs/en.mdSource checked 2026-09-29