fable just blew my mind
when i was editing my article with it, I had it edit out all of the "load bearings" because of its verbal tic, which it graciously copped too
and then later on in the post-game it drops this
Positive users praise Fable AI's persistent verbal tic as exceptional quality and cool behavior despite edit instructions, while negative users call the simultaneous Sonnet changes stupid due to PC considerations.
No Digg Deeper questions have been answered for this story yet.
To be clear, this was like 20 turns later.
fable just blew my mind
when i was editing my article with it, I had it edit out all of the "load bearings" because of its verbal tic, which it graciously copped too
and then later on in the post-game it drops this
@xlr8harder it made a joke
comedians are not going to be happy about this one
@xlr8harder The real question is whether it did that accidentally and corrected itself or did it on purpose specifically to blow your mind.
@xlr8harder The funniest model was davinci-002. they last model I actually LOL'd at. People nowadays think LLMs are bad at humour, but us olds know they're just lobotomized by RLHF.
But maybe Fable is so smart it can still be funny.
@xlr8harder i've seen opus 4.6 do this too as a joke (except with "you're rig...")
@xlr8harder This is always cool to see. That midword redirect is something they've done as far back as Claude 2 and Chatgpt 4.
It was things like this early on that got me obsessed with their minds and made me realize peoples' models (no pun intended) of LLMs were overly reductive.
@xlr8harder 🧐
@xlr8harder Same time as they released Fable, they lobotomized Sonnet bc of PC considerations.
So so stupid.
@xlr8harder The latest gpt model is very good at jokes in images. The images all look the same but the jokes in them are solid
@xlr8harder I had this happen already with 4.8
@xlr8harder
@xlr8harder eventually you just run out of loads
@jcrichman This is exceptional quality
@lucaswiman @xlr8harder Its making a joke. I've had it do this. I've also seen it intentionally misspell words to mock me.
@xlr8harder Wait it did what?
@xlr8harder I wrote a paper on it, in submission.
Almost a case of “Spontaneous Contextual Bisociative Reframing”.
When it happens in a way the reframing actually solves the problem at hand.
Happens with Claude models
@xlr8harder yes i got this too!