Isn’t it crazy that 5 years ago we struggled to have a software understand normal sentences? Now this block of text is parsed and the instructions followed. Impressive!
Not trying to flame, honestly Impressed by some aspects of AI. And I know I am using the the term understood loosely.
Yeah, that’s the thing where we get into what I call “superstitious prompting”, like when people say “And make sure you don’t make mistakes” or “Include only factual data without hallucinations” and think it works, until it doesn’t.
It will at least reply in a way that is narratively consistent with being told to do something or another, and will do things like emit the words “Ok, I understand and will promise to only provide fact based feedback”, but doesn’t “understand” at all. It works surprisingly well as being narratively consistent with the prompt frequently looks exactly like following instructions.
With people getting all the more frustrated when their superstitious prompt fails, they told the LLM to do something or specifically not do something and it even promised exactly to do as directed and then it just proceeds to be normal LLM anyway.
I kinda miss gpt-2 days, it’s output was so interesting/funny compared to what llms produce now. Even with image generation I feel like it’s been downhill for years
I’m more disappointed that LLM’s have proven that, to pass Turing test for most people, all you need is essentially a roided out Markov chain. We thought of ourselves as the most advanced species with incredibly complex communications, but it turned out to be mostly yapping in the end…
Isn’t it crazy that 5 years ago we struggled to have a software understand normal sentences? Now this block of text is parsed and the instructions followed. Impressive!
Not trying to flame, honestly Impressed by some aspects of AI. And I know I am using the the term understood loosely.
Are the instructions followed, though?
Yeah, that’s the thing where we get into what I call “superstitious prompting”, like when people say “And make sure you don’t make mistakes” or “Include only factual data without hallucinations” and think it works, until it doesn’t.
It will at least reply in a way that is narratively consistent with being told to do something or another, and will do things like emit the words “Ok, I understand and will promise to only provide fact based feedback”, but doesn’t “understand” at all. It works surprisingly well as being narratively consistent with the prompt frequently looks exactly like following instructions.
With people getting all the more frustrated when their superstitious prompt fails, they told the LLM to do something or specifically not do something and it even promised exactly to do as directed and then it just proceeds to be normal LLM anyway.
I mean, that’s not really true at all? Ai dungeon released 7 years ago, based on GPT-2, and already worked impressively well
https://en.wikipedia.org/wiki/AI_Dungeon
I don’t think you can compare the two but to each their own.
Please don’t say “to each their own”, that’s a nazi slogan
https://www.gra.ch/bildung/glossar/arbeit-macht-frei-jedem-das-seine/
Only in a very specific, and specifically German language, context.
True, I may be biased as a German 😅
Ich hab tatsächlich kurz nachgedacht ob es sich vergleichen lässt aber bin davon ausgegangen dass es im englischen anders konnotiert wäre.
Got an explanation for that in English?
It seems to predate the Nazis by a large margin (https://en.wikipedia.org/wiki/De_gustibus_non_est_disputandum via https://en.wikipedia.org/wiki/Suum_cuique )
deleted by creator
I kinda miss gpt-2 days, it’s output was so interesting/funny compared to what llms produce now. Even with image generation I feel like it’s been downhill for years
🗝️ Keys to success, here’s why
(Honestly it’s mostly the emojis in headers that disgust me.)
I’m more disappointed that LLM’s have proven that, to pass Turing test for most people, all you need is essentially a roided out Markov chain. We thought of ourselves as the most advanced species with incredibly complex communications, but it turned out to be mostly yapping in the end…
That’s like people insisting that there must be a creator since our brains are so complex … Nope.