• Smorty [she/her]
      link
      fedilink
      arrow-up
      3
      ·
      20 days ago

      It indeed is one of these cases. Individual letter recognition doesn’t work well, since they tend to be clumped up into larger “tokens”, which then can’t be differentiated anymore. That’s also the same reason why it can’t count the letters in words, It sees words as just a single, or at most three “tokens”

      • Swedneck@discuss.tchncs.de
        link
        fedilink
        arrow-up
        1
        ·
        14 days ago

        what i don’t get is that surely it must be perfectly feasible to just implement letter tracking as a separate function that the LLMs can somehow interact with? they can clearly hook up the LLMs to stuff like wikipedia for up to date information…

        • Smorty [she/her]
          link
          fedilink
          arrow-up
          1
          ·
          14 days ago

          And right again, ai enthusiasts have already hooked up LLMs to the ability to write their own code. It’s infact super easy to let a model write a script which then countanthe letters.
          Then again, it’s also the case, that you could just use a normal program for that in the first place.

          This whole "how many Rs are there in ‘Strawberry’ " thing is nice to show people that LLMs can’t do everything. But it’s also not a reasonable usecase.

          You don’t hire people to count words in a document, that’s what computers are for.

          The reason for my LLMs don’t read every character individually is because it’s way more efficient to let it detect entire words, rather than letters.

          Building a “letter tracker” into an LLM would only be useful to specifically make these people happy, who make fun of LLMs not counting the Rs, and literally no one else…