Hamel Husain argues that a 100% pass rate on AI evaluations is a warning sign, usually meaning the tests are too easy or the model is overfitted to them. Teams should keep some tests deliberately failing to get real feedback on weaknesses and ensure progress is genuine, not just gaming the metrics.
When you see 100% pass rate on evals, this is usually a really bad thing. What it means is your evals aren't telling you anything. If you're just passing every single test, you're not getting any signal on what your AI is not doing well on, and you're just patting yourself on the back. Usually when you see 100% pass rate, what it means is that you've overfit on your evals, meaning kind of train the prompt to overfit on those test cases. And that's really not good because you don't know if your AI is generalizing. It's either that or you've made your evals so easy that they're trivial to pass. Either way, this This is not a good idea. What you should aim for is having some evals that fail. You want some hard evals that are even reaches that you fail on so that you have something to iterate against.
