Topic 1
Read the outputs
Metrics tell you how often something goes wrong. Only reading the outputs tells you what goes wrong. Error analysis is the habit of looking at real failures, grouping them, and fixing the ones that matter most.
AI Engineering (Ch. 8, "Data Processing" > "Inspect Data") makes the case for manual inspection: tools help, but they don't replace looking. The author says that in every project the author has worked on, about 15 minutes of staring at the data usually gave insight that could save hours. The book also passes on a remark by OpenAI co-founder Greg Brockman: manual inspection of data may have the highest value-to-prestige ratio of anything in machine learning.
Predict first
Scroll sideways to see the whole diagram
Start with real outputs: failing test cases, and a sample of live traffic.
Text version
- Start with real outputs: failing test cases, and a sample of live traffic.
- Read them, one by one, and write down what went wrong in plain words.
- Only then group the notes into failure modes, and count how often each happens.
- Pick what to fix from how often each failure happens, how much harm it does, and what fixing it costs.
- Keep the failing examples as test cases, so the fix stays fixed.
- Then repeat: the next round finds the next most important failure.
What to read: the failing cases from your test sets, and a sample of live traffic, including some that passed. Passing cases sometimes pass for the wrong reason. The book's inspection advice applies here too. Look at outputs by source, time and annotator. Look for outliers, and for patterns in which inputs get worse answers. Where people labelled the data, label a few yourself and see if you agree.
Quick check