Fuck AI

... do y'all think GPT would run out of words, if they didn't keep feeding it?

The big innovations are goofy shit like hand-waving complexity for R1's intermediate stages... or prompting the spicy autocomplete with 'pretend you have an inner monologue.' Seriously. That's what "chain of thought" means, and it actually kinda works. Like how all image diffusion models are 'remove the pixels that don't look like Shrek.'

As the efficiency of training increases - and energy use gets worse, because of Jevons' paradox - we will see the vegan free-range public-domain models that people think they want. It'll get cheap enough that FOSS diehards will do it to prove a point. And folks, it's gonna do all the same shit, at roughly the concurrent state-of-the-art. You might have to provide an example as context - since Disney choking the public domain since 1928 means it won't know who Elvis is - but once it has a zillion-dimensional vector for how Elvis differs from its global average, it can work with that. Like how an illustrated character invented yesterday can be drawn by models from two years ago, so long as you describe their salient features.

Anyway:

Looms displaced skill - not labor. They still needed a massive workforce. But those workers could be attentive randos, or indeed children, instead of lifelong artisans. Per a recent argument: tools don't replace people. Tools don't do anything. People using new tools replace people using old tools.

I am a lifelong artisan in the field of software engineering, and if people can write software by describing its intent and then nitpicking what's wrong, that would be fucking incredible. That's what COBOL and BASIC were trying to do, seven decades ago. Letting any idiot write their own programs, at a professional level, is the undercurrent for half of computer science.

When things get easier, and rich pricks make life harder, do not get mad about things getting easier. The tools didn't fuck you. Tools don't do anything.