Post

eevee
@eev.ee

afterthought 1 — "Finally, art is democratized," proclaims area programmer man making $200k per year. "No longer is it held solely in the hands of the elites: furry porn artists struggling to pay rent"

eevee
@eev.ee

afterthought 2 — clear bridge between "you can't customize LLMs like you can tools" and the almost comical unpredictability of image generation, where getting the same character twice is a crapshoot. but i didn't want to get into image stuff too much or the post would've been twice as long

eevee
@eev.ee

afterthought 3 — i forgot i muted "AI" and some related words on bluesky so i'm sorry if i don't see your reply lmao

eevee
@eev.ee

afterthought 5 — you could probably add a bridge of """the metaverse""" in here, but every time i think about it i just remember World Baby and i can't ever take it any more seriously than that

eevee
@eev.ee

afterthought 6 — i once heard A Hater was telling people not to even visit my website, so i wouldn't get the ad revenue. but there were no ads. i'm guessing they used adblock and just took for granted that there would be ads they weren't seeing, because what website doesn't have ads? (mine)

eevee
@eev.ee

afterthought 7 — if sama gets what he wants and everyone just has chatbots write code for them, i wonder: ⓐ how will the next generation develop enough skill to review the generated code? ⓑ when a new thing comes out, who's supposed to write the new code using it that the chatbots get trained on?

1 repost25 likes

Thorsten S.
@thesiteabout.me

I've been asking this myself from the early days of ChatGPT. The concept has an ever-worsening spiral of content degradation at the very core. Shit-in/shit-out, and the more shit AI models generate, the worse future models get. And nobody has been able to explain to me how this will be prevented.

Alpha
@alphatheskunk.bsky.social

I've though about this one already: 1. Companies will get divided in senior-churning machines (college 2.0) or senior-absorbing machines (the ones that use the thing). 2. Agents trained on byte-level tokenization can read the manuals and code the thing. Using RL a compiler agent would train it.