Conversation

Notices

1. Embed this notice
   Paul Cantrell (inthehands@hachyderm.io)'s status on Tuesday, 02-Dec-2025 14:00:41 JST Paul Cantrell

   “Would it be particularly helpful to have some text that would be •typical• in a given context, even if the text is possibly •incorrect•?”

   ↑ vet potential applications of LLMs with this one weird question!

   • Embed this notice
     Paul Cantrell (inthehands@hachyderm.io)'s status on Tuesday, 02-Dec-2025 14:05:06 JST Paul Cantrell

     in reply to

     I’m not saying the answer will always be “no,” but people are wildly, wildly overselling the number of situations in which the answer will be “yes.”

     Sometimes the answer is “no” because incorrectness is a deal-breaker — and that scenario’s been getting a lot of attention (fake legal briefs, vibe-coding faceplants, etc).

     Sometimes, however, the answer is “no” because making something •typical• isn’t really helpful. Where humanity needs to come to the foreground, homogeneity is positively unhelpful. Here’s a fascinating example:

     https://hci.social/@bwaber/115647856415818492

   • Embed this notice
     ben (benjamineskola@hachyderm.io)'s status on Tuesday, 02-Dec-2025 20:17:05 JST ben

     in reply to

     @inthehands is an aspect of this also when correctness is hard to verify?

     I can think of examples where the verification process would basically involve doing most of the same work over again. (e.g., summarisation, where if the summary needs to be correct, then a human has to read over all the same inputs and ensure that nothing has been missed or invented.)