> Well for one, programming actually sucks. Punching cards sucks. Copywriting sucks.

There's a significant difference between past software advancements and this one. When we previously reduced the manual work when developing software it was empowering the language we were defining our logic within so that each statement from a developer covered more conceptual ground and fewer statements were required to solve our problems. This meant that software was composed of fewer and more significant statements that individually carried more weight.

The LLM revolution has actually increased code bloat at the level humans are (probably, get to that in a moment) meant to interact with it. It is harder to comprehend code written today than code written in 2019 and that's an extremely dangerous direction to move in. To that earlier marker - it may be that we're thinking about code wrong now and that software, as we're meant to read it, exists at the prompt level. Maybe we shouldn't read or test the actual output but instead read and test the prompts used to generate that output - that'd be more in line with previous software advancements and it would present an astounding leap forward in clarity. My concern with that line of thinking is that LLMs (at least the ones we're using right now for software dev) are intentionally non-deterministic so a prompt evaluated multiple times won't resolve to the same output. If we pushed in this direction for deterministic prompt evaluation then I think we could really achieve a new safe level of programming - but that doesn't seem to be anyone's goal - and if we don't push in that direction then prompts are a way to efficiently generate large amounts of unmaintained, mysterious and untested software that won't cause problems immediately... but absolutely does cause problems in a year or two when we need to revise the logic.