Human Brain EnthusiastRE: https://yiff.life/@tael/116325471293353839
This is why "AI" won't take over the world—it can't. It doesn't work that way.
It can work to augment and in tandem with existing technology and has some useful use cases (though IMO relatively few).
The issue with "AI" isn't really if it works or not (we know it doesn't and needs augmentation, special-casing and lots of help to be useful).
The issue is that these big companies won't be able to provide this service forever that they currently offer way below what it actually costs them.
If they'd charge what it actually costs to run it plus an upcharge to pay their investors plus some more to make a profit—they'd be out of business tomorrow.
0px auto@thomasfuchs it'll be interesting to see how it will play out when three weeks from now, we will be having to decide whether to be able to eat/cook/drive or generate an image of a cute fox playing the ukelele...
AI slop nourish neither soul nor body...
curtosis@thomasfuchs Plus the theft and ethical morass.
@curtosis Well yeah, but at least that is the one part of all of this that could be overcome (use only licensed or public domain works, etc.); which is largely true for some niche specialized LLMs like for coding.
I see the future of this in locally-run LLMs that are more or less open source, with training data that's vetted for proper licensing. (You can already run pretty large models locally and it uses only as much energy as gaming does.)
However I fail to see any big use cases for it; programming is probably the biggest one—and only because programming languages are in general terrible so people use LLMs as sort of macros to quickly churn out plausible chunks of code to then go in and refine.
@thomasfuchs True. I was *mostly* referring to how so much of the training was done through exploitative labor practices in the global South.
(FWIW I also think there’s a lot hidden under the “only trained on licensed works” label. Common Pile includes CC-BY[-SA] works, which LLM-generated text violates.)
Elias Mårtenson@thomasfuchs it's useful and can probably be pretty cheap for cases where it doesn't matter if the output is wrong though. Especially spam. It's excellent for spam.