Evaluating AI Systems
You cannot improve what you cannot measure — but measuring LLM quality is genuinely hard. Automated benchmarks measure proxies, not what users care about. Human eval is expensive and slow. LLM-as-judge is scalable but biased. Regression testing catches regressions. The answer is all four, used together.
Right yesterday
Your support bot said this today. Yesterday it said 14 days, which is correct.
Someone tweaked the prompt overnight. Nothing crashed, and nobody noticed.
Quiet failures
Normal code breaks loudly. A model breaks quietly, with a fluent wrong answer.
Reading a few replies won’t catch it. You need a check you can rerun after every change.
Spot the mistake
The policy: 14 days, receipt needed, no returns on opened items. Tap what the bot got wrong.
A test set
So write down questions where you already know the right answer.
Run them before every change to the prompt or model. If the score drops, look before you ship.
Good or weak
Which of these make a good test?
Tap the box where each one belongs.
A model as judge
Some answers have no single right wording. So people ask another model to grade them.
Give it a clear checklist and spot-check its grades. Model judges are known to favour longer answers.
Pick the check
Use the simplest check that works. Save the judge for what code can’t check.
Test it like software
Test a model like software: fixed questions, known answers, rerun after every change.
A set of test questions is called an eval. A drop after a change is a regression.