Comment by SatvikBeri
3 days ago
I've noticed over and over again at various jobs that people underestimate the benefit of speed, because they imagine doing the same workflow faster rather than doing a different workflow.
For example, if you're running experiments in one big batch overnight, making that faster doesn't seem very helpful. But with a big enough improvement, you can now run several batches of experiments during the day, which is much more productive.
I think people also vastly underestimate the cost of context switching. They look at a command that takes 30 seconds and say "what's the point of making it take 3 seconds? you only run it 10 times in a day; it's only 5 minutes". But the cost is definitely way more than that.
Me, looking at multi-hour CI pipelines, thinking how many little lint warnings I'd fix up if CI could run in like 20 minutes
Whenever we make our code faster the users just run bigger models :P.