aplassard··on Evals in 2025: going beyond simple benchmarks to build models people can useI think cost should also be a direct consideration. Model performance varies wildly on benchmarks when given a budget. https://substack.com/@andrewplassard/note/p-173487568?r=2fqo...