Assistant output can look correct while being wrong, from invented names to subtle bugs. This part covers common failure modes, checks that catch them, and small evaluations for judging tools.
Topic | Lessons | Description | Time |
|---|---|---|---|
5 | How assistants fail on engineering tasks, and the checks that catch each failure. | 40 min | |
5 | A habit and checklist for checking assistant-written changes, with differential, property and diff checks. | 40 min | |
4 | A small test set, a fair baseline, and cost per accepted result for judging AI tools. | 30 min |
Assistants produce plausible output without grounding, including invented names, wrong numbers, and bugs that compile, so confidence carries no evidence. The Hallucination and Failure Modes topic covers each failure and its check.