Learn with an AI assistant
AI assistance is optional. You can complete the core path using the lesson, documentation, scripts, practice, and public feedback without a paid service or AI account. Your instructor's local rules govern assessed work.
A repeatable workflow
- Attempt a small step and write what you expect.
- If permitted, ask for an explanation, a hint, or a critique of your reasoning.
- Inspect proposed code and source claims before running it.
- Test a bounded change and compare the actual output with your prediction.
- Explain the decision independently and record useful assistance briefly.
A useful prompt is: “Explain why fitting this scaler on all rows can leak information. Give me a question to check my understanding before suggesting code.” Another is: “Here are the metric names and confusion counts, without private records. Identify an unsupported conclusion in my memo.”
Verify the output
Check imports and API names against the tested environment. Inspect paths and commands, especially deletion, installation, network calls, and credential requests. Use a small local case and relevant behavior checks. Validate numerical claims against saved results. An assistant's confident explanation is not a source citation or proof that code ran.
Do not let an assistant inspect final-test results repeatedly and then optimize the method while calling that test untouched. Keep errors and invalid model output visible; do not silently repair an answer and credit the original system.
Protect data and authorship
Do not submit student records, credentials, restricted assessments, private logs, or identifiable messages to an external service. Public datasets can still contain personal content; use local processing and summarized evidence when possible. Record tool/date, the bounded help received, what changed, and what you checked. A full private conversation transcript is unnecessary.
Learn from failure
The small NLP and language-model examples deliberately preserve poor observed outcomes. Explain whether the failure concerns learning, label imbalance, schema compliance, or an unjustified claim before choosing an intervention. Tiny LoRA updates demonstrate mechanics, not useful deployment behavior.
If you choose no assistant, use the same prediction/run/check/explain cycle and the public feedback. No additional work is required solely because you did not use GenAI.