Hamel Husain explains why model benchmarks like SWE-Bench and GPQA Diamond don't tell you if your AI product actually works. You need product-specific evals instead, custom tests that measure whether your system does what it's supposed to do, like correctly canceling an order.
What are AI evals? AI evals are something that confuses lots of people, and there's 2 categories of evals. One is model benchmarks. Model benchmarks are things that you see when new models come out. So there are general-purpose benchmarks, things like SWE-Bench. SWE-Bench tells you how good models are across a battery of coding tests. GPQA Diamond is a benchmark that tests models against biology, chemistry, physics, and other advanced topics. But when you're building AI products, you need a different kind of eval. You need an eval that is specific to your product. And we call that product evals. Product evals measure whether your specific AI product does what you want it to do. The product eval for this agent would check to see if the correct order was selected and if the order was actually canceled. I'll show a link above where you can go to read a lot about evals, where to get started, how to implement them, and a lot of the traps people face when designing evals for their AI products.
