Hamel Husain argues that automated evaluation platforms can't understand what your code is supposed to do without you telling them. AI needs human guidance, detailed notes, and error analysis to catch real problems. Build your own evaluations instead of trusting a black box.
Many engineers believe that you can completely automate your evals. After all, we live in the age of AI. Can't you just press a magic button and it just does the evals for you? Who wants to even do these evals? You want to just build your product. You should just buy an automated eval solution. That simply doesn't work. Don't do that. That's a huge trap. And the reason it doesn't work is because AI can't read your mind. Until we get neural brain implants and and an AI can see what you're thinking, you have to go through the process of iterating and explaining to an AI what you want. And the only way to really do that effectively is to go through this iterative process of evals, where you write down what went wrong and sort of refine the prompt as you go. So here's the takeaway. If someone promises you that they will just do your evals for you automatically, that you will just summon a new AI to just do it, run as fast as you can in the other direction. What you need to do is look at your data, you need to take notes, and you need to do error analysis and write some automated evaluations that are very specific to your application.
