Conversation

Notices

1. Embed this notice
   That DΔrn Pooka :verified_think: (theququ@shitposter.club)'s status on Saturday, 21-Oct-2023 19:00:10 JST That DΔrn Pooka :verified_think:
   OpenAI will simultaneously try to convince you that ChatGPT is borderline a real intelligent being, while sticking a metaphorical ice pick up its nose and giving it a frontal lobotomy with no moral qualms whatsoever that it's destroying a living thing's cognitive capabilities. They're either up their own ass with PR-speak (likely) or sociopaths (also likely but for unrelated reasons).
   • Sexy Moon likes this.
   • Embed this notice
     Sexy Moon (moon@shitposter.club)'s status on Saturday, 21-Oct-2023 20:33:35 JST Sexy Moon

     in reply to
     @TheQuQu after using it for close to a month I am both amazed at how much it can do but also convinced that it's not even close to general intelligence.
   • Embed this notice
     Sexy Moon (moon@shitposter.club)'s status on Saturday, 21-Oct-2023 20:34:01 JST Sexy Moon

     in reply to

     • Sexy Moon
     @TheQuQu categorically it doesn't even act like iterative methodical thought.
   • Embed this notice
     That DΔrn Pooka :verified_think: (theququ@shitposter.club)'s status on Sunday, 22-Oct-2023 05:31:17 JST That DΔrn Pooka :verified_think:

     in reply to

     • Sexy Moon
     @Moon Since it's just statistical token prediction (albeit incredibly good token prediction), it has no internal mental state whatsoever, so it can only reason "on paper". That's why folks get better answers if they ask for it to explain its reasoning before answering.
     
     The funniest bit was when people tried to play 20 questions with it. Because it can't keep secrets, it'd just answer "yes" and "no" randomly, and then when you made a guess it would pretend it was thinking of that thing all along if it matched prior answers.
   • Embed this notice
     Sexy Moon (moon@shitposter.club)'s status on Sunday, 22-Oct-2023 05:31:17 JST Sexy Moon

     in reply to
     @TheQuQu chatgpt has memory and can hold state. I've also done things like asked it to keep a timer and it worked.
   • Embed this notice
     Sexy Moon (moon@shitposter.club)'s status on Sunday, 22-Oct-2023 05:32:35 JST Sexy Moon

     in reply to

     • Sexy Moon
     @TheQuQu one advance of gpt over markov chains is it has a limited memory. I have been reading about ways that people are trying to overcome the limited amount of space it has. for example I was trying to do stories with it and after so many messages it forgets details. people are trying to improve that.
   • Embed this notice
     That DΔrn Pooka :verified_think: (theququ@shitposter.club)'s status on Sunday, 22-Oct-2023 05:35:15 JST That DΔrn Pooka :verified_think:

     in reply to

     • Sexy Moon
     @Moon That's usually a parlor trick, they give it all sorts of APIs it can invoke. For instance, because LLMs suck at math it'll recognize if there's a math question and outsource it to a Wolfram Alpha-style service. It probably has a clock and timer API endpoint it can hit.
     
     In the early days before they locked it down tight, someone got it to expose a list of its API endpoints by leaking the preprompt.
     Sexy Moon likes this.
   • Embed this notice
     Sexy Moon (moon@shitposter.club)'s status on Sunday, 22-Oct-2023 05:36:53 JST Sexy Moon

     in reply to
     @TheQuQu here's some experimental stuff with trying to manage context memory better: https://github.com/cpacker/MemGPT
   • Embed this notice
     Jeff "never puts away anything, especially oven mitts" Cliff, Bringer of Nightmares 🦝🐙🇱🇧🧯 (jeffcliff@shitposter.club)'s status on Sunday, 22-Oct-2023 05:37:35 JST Jeff "never puts away anything, especially oven mitts" Cliff, Bringer of Nightmares 🦝🐙🇱🇧🧯

     in reply to

     • Sexy Moon
     > github
     
     :molotov: :gnujihad:
     Sexy Moon likes this.
   • Embed this notice
     Sexy Moon (moon@shitposter.club)'s status on Sunday, 22-Oct-2023 05:38:22 JST Sexy Moon

     in reply to
     @TheQuQu yeah you're right I don't know about any that can manage hidden state
   • Embed this notice
     That DΔrn Pooka :verified_think: (theququ@shitposter.club)'s status on Sunday, 22-Oct-2023 05:38:23 JST That DΔrn Pooka :verified_think:

     in reply to

     • Sexy Moon
     @Moon Local open models are doing better than OpenAI is at that now, there's finetunes of Llama 2 that can have a 16k token context, which I believe is bigger than chatGPT supports. But even this memory is just the entire token history. If you want "hidden state" you can teach it to output tokens that denote a block of text to hide from the user, but I don't know of any that do that yet.
   • Embed this notice
     guizzy (in exile) (guizzy@shitposter.club)'s status on Sunday, 22-Oct-2023 06:08:27 JST guizzy (in exile)

     in reply to

     • Sexy Moon
     @Moon @TheQuQu There's always the option of chaining LLMs and autoprompting to overcome base limitations of token prediction to overcome the limitations of the basic token prediction design. You can create multiple sets of system prompts and even different models/finetunes that are specialized on different aspects of cognition, and have them work out an answer to a prompt together while sharing a database they can query.
     Sexy Moon likes this.