it's interesting to think about. When farming took 99% of time, all people did was farm. Maybe they complained that the machines weren't as good and yield and quality would drop?

What can the computer chuds start doing when finally not every little stupid thing takes 2 weeks? Where to spend effort?

There is this relationship in the real world, where very professional people tell you "make it work first, make it good later" / do a vomit draft / just tape a cardboard tube to a widget as a handle to see if you like it, etc.but with software, common wisdom was if you don't get the fundamentals right, you will end up in trouble. there never was an identical draft to polish path for coding. reimplementing your entire UI in a better-suited lib took ages, for example.
AI might align effort more with the Pareto curve

Of course maybe those kinds of people were always wrong and you should "just get better at writing code" and then a rewrite wouldn't have been that big of a deal, but even if so, maybe AI will help people realize that.

RT: https://shitposter.world/objects/2cccb614-758f-4263-a10b-feb95cc9d8e0

@WandererUber @sun one big issue used to be that 'doing it right' took exactly as long as becoming wise about your problem domain, so total rewrites were mostly impractical. that's not the case now, a total rewrite of a system that took 10 years to build can be done in a few weeks.

karpathy said the LLMs show that there were always two kinds of programmers, 'builders' and 'coders'. the builders now use the new tool, they usually didn't dislike programming, but if they can make superpleroma 2 without writing a single line, they'll gladly do it. the coders really care about the coding itself. i sort of agree with that split.

'just get better' really doesn't work here, because a deep understanding of the problem space only comes from actually engaging with the issue. there's no 'just plan ahead moron', humans can't handle 'make no mistakes' either.

i think that's also something you see with projects like opencode: before, the issue would have been to write the code for it. now, the problem is to curate the feature set and to figure out where you actually want to be in that space. so you might have 1000 PRs open, each with a good feature, but still not want to merge a single one.