Whenever people complain about the energy usage of LLM training runs I wonder how this stacks up against the energy we waste by pointlessly redownloading/recompiling things (even large things) all the time in CI runs.