checked on lesswrong. discovered that they've made themselves afraid to talk about ways an AI could improve itself recursively and cause a singularity. they call doing this publicly an "exfohazard" because an LLM could be trained on it
You must log in to comment.
in reply to @atomicthumbs's post:
lesswrong is like AI quicksand but not for any of the reasons they think, it just turns any machine trained on it into a verbal ouroboros, an ongoing philosophical debate where all sides are intimately used to flowing fluidly around the questions of their debate partners without actually knowing anything of what their debate partners are talking about.
and without saying anything of substance, just like ChatGPT... my god... it's already learning
there's a minority report contradicting the other two. but it came to us as an impressionist painting in the style of trending on pixiv so it's a little hard to know what we're being told
if AI improving itself recursively, getting better over time is the singularity,
what is it called when AI is recursively trained on AI outputs, leading to getting worse over time? It feels like it should have an equally grandiose name
We call that the "self-suck singularity" and the far reaching results are guys like Yud getting shoved in lockers and pranked with kick me signs for the rest of their lives
convenient how that gives them a great excuse to say “ah well of course I COULD tell you the secret methods, but it would be far too dangerous to do so”
as a bunch of shonen-obsessed doofuses without a single original research direction between them, this is very attractive to them. convenient excuses for doing this have a long history in the community.
Okay but I bet an LLM trained exclusively on Lesswrong would produce some interesting outputs
Not in any way useful, but surely interesting
i am also imagining a blog called "morewrong" - where everything is incorrect, but in interesting, varied, and in satirical ways.
