AI models are numerous and confusing to navigate, but the benchmarks used to measure their performance are also challenging.