|
3 posts total
 It occurs to me that LLMs ("AI," if you like) are used pretty much exclusively in areas where quality is difficult to define. Art, service, design, communication, writing...arenas full of ambiguity, where it's hard to quantify success or failure. It strikes me that LLMs are deployed mainly here, rather than in less equivocal places, at least in part because this masks the terminal flaw of LLMs: They are designed for generation, not for accuracy. They are not at all truthful or reliable. I think we're focused on the wrong thing when we look at what tech works for a company like Amazon or Facebook or Netflix. We should be looking at what tech works when you *don't* have a small army of staff engineers optimizing it. I want to know what I can scale *without* paying someone a half million dollar salary to do it. There should be more case studies on things that don't have a billion-dollar company propping them up, humming along quietly on a cheap-ass VPS somewhere.
Show previous comments
@collinsworth “100 amazing technologies you can setup in a day and then never worry about again”.  @collinsworth after hearing this reasonable take being ignored through my entire professional career, I gave up. Excited engineers will always follow "industry leaders" without criticism. Dissent is confused with ignorance. I took this photo of my kid playing with his blocks the other day, and when I realized it reminded me of something, well...I had to make this.
Show previous comments
@collinsworth You should put a credit to https://xkcd.com/2347/ there, but thanks for letting me feel I'm in the "in"-group :) @collinsworth please add the original Randall Munroe xkcd comic link "Dependency" to the alt-text. |
Gregory's Server
@collinsworth and the reliability cannot be fixed without breaking the generation. Writhing that realm LLMs are also deployed where "not very good" is considered "good enough." That is, "it's better than i can do and it's irrelevant that it's not as good as the human i can't afford to hire." Art and design are also faster to verify and avoid the worst mistakes.
@collinsworth is that true though? There’s a lot of us of LLMs as coding assistants