4 ms·

> Remember, this response takes a departure from reality and embraces playful inconsistencies for the sake of imagination and entertainment. The actual Hacker N

• junofan 3y ago
  Yeah, can’t wait until this garbage becomes unfashionable.
  • b33j0r 3y ago
    Have you used it? At worst it’s a way better intellisense. I am invested in it being a little more, but not like my robot wife lol. I’m genuinely curious, why does reading about transformer models and chatbot news upset you, and some others?
    • ckolkey 3y ago
      Its just so repetitive, with the same types of cargo-cult responses to any criticism. Very similar to bitcoin/tesla in how it gets shilled
      • dontupvoteme 3y ago
        It's basically the reddit hivemind with built in wikipedia quoting. I wonder if it is better in languages that never really used wikipedia. Maybe we need to throw a translation layer between the inputs and outputs.
    • ask_b123 3y ago
      Wasn't peyton bothered by the disclosure messages at the end rather than at the language model?
      • b33j0r 3y ago
        Yep, I made a categorization and sentiment error. As a non-large language model, my loss function is brief internet shame. :p good call
    • 1123581321 3y ago
      He’s clearly talking about the lectures tacked on at the end of output (or sometimes the entire output) for “safety,” not LLMs in general.
• ineedasername 3y ago
  I tell it to stop that, but it still does that sort of thing every few prompts when it loses the immediate token context window. Sometimes it’s wider continuity [1] will keep it off that pedantic path longer, but I still have to occasionally put into prompts things like “don’t remind me about blah blah blah every time” [1] “continuity” is what it called the ability when, after about a half dozen prompts of interrogating it, I finally got it to acknowledge that it was still aware of something outside of its immediate context window (without me giving the direct knowledge with a leading question). I don’t know if it hallucinated the term “continuity” to account for its ability to look back further than the token context window or if that’s actually what the openai folks call it, but ChatGPT said it was limited, not everything was available, and there was some type of pruning based on perceived importance. I’ll try to dig up the precise wording and post in a comment to myself here, it’s late and I’m on a mobile device though and too tired and lazy to touchscreen my way through tons of chats histories. We need a global search of prior chat sessions.
  • oidar 3y ago
    Did you find it? I'm interested to see if it's testable. I think ChatGPT just drops context that doesn't fit in the context window anymore - and then based on context it continues the conversation. within the 8k context window, it works really well because ppl keep revisiting important bits, so there isn't much lost there. I haven't seen OpenAI talk about condensing prior context, and I don't think ChatGPT has any other knowledge about itself other than what is in it's system part of the prompt. It doesn't really know what happens to the prior context, it just wants to finish the document that is being created.