• vivendi
    link
    fedilink
    English
    arrow-up
    1
    arrow-down
    6
    ·
    edit-2
    19 days ago

    Not making these famous logical errors

    For example, how many Rs are in Strawberry? Or shit like that

    (Although that one is a bad example because token based models will fundamentally make such mistakes[1]. There is a new technique that lets LLMs process byte level information that fixes it, however)

    EIDT: [1] This sentence is badly written. I meant text based errors like number of letters of a word in this sentence. Token based LLMs operate on atomic units of tokens which may be part of a word, a complete word, or some sentence structure. Because of that they can’t interact with text the same way humans do, but a new paradigm that lets LLMs read their input as raw bytes will help with this.

    • froztbyte@awful.systems
      link
      fedilink
      English
      arrow-up
      6
      arrow-down
      1
      ·
      19 days ago

      oh, I get it, you personally choose not to make these structurally-repeatable-by-foundation errors? you personally choose to be a Unique And Correct Snowflake?

      wow shit damn, I sure want to read your eventual uni paper, see what kind of distinctly novel insight you’ve had to wrangle this domain!

      • vivendi
        link
        fedilink
        English
        arrow-up
        1
        ·
        19 days ago

        As I said new techniques really help with those problems, like selectively operating on raw data or tokens

        Technology isn’t standing still. If your neckbeard ass knows about it so do researchers