Apparently someone got into a debate with a bot in Reddit. Maybe it’s not a surprise that the bot was anti-mod, pro-Reddit admin. Make sure you scroll down in the comments to see the second screenshot (it’s a screenshot.of the same conversation but in dark mode).

  • geoffervescent@kbin.social
    link
    fedilink
    arrow-up
    5
    ·
    1 year ago

    The only thing they ‘always’ do is mimic human language conventions based on scraped data. It stands to reason that as a higher percentage of people talking about Chat GPT refer to it using different words, the model would take that into account when referring to itself.

    • soft_frog@kbin.social
      link
      fedilink
      arrow-up
      4
      ·
      1 year ago

      The models are only trained up to 2020, they’re not updated with new data but the internal prompt and moderation system are updated to reflect changes in use and moderation escapes.