Conversation

Notices

1. Embed this notice
   pettter (pettter@mastodon.acc.umu.se)'s status on Monday, 01-Jul-2024 23:07:53 JST pettter

   • Thomas 🔭🕹️

   @thomasfuchs I've heard a lot of people doing writing who swears by it as well. Not to get a finished product, but for creating outlines or for fleshing out bullet points into sentences/paragraphs.

   • Embed this notice
     Uli Kusterer (Not a kitteh) (uliwitness@chaos.social)'s status on Monday, 01-Jul-2024 23:19:43 JST Uli Kusterer (Not a kitteh)

     • Thomas 🔭🕹️

     @thomasfuchs Seems about right.

     The only non-evil use case I’ve seen (of all people!) was Adobe's content-aware fill. Made it easy to infinitely stretch out a background for a web site etc. (And even that has the potential to be used for evil).

     But that was a way older, simpler use case of trained content. Nothing like stealing an entire art style, basically just generating patterns.

   • Embed this notice
     Ian Smith (katachora@hachyderm.io)'s status on Monday, 01-Jul-2024 23:24:27 JST Ian Smith

     • Thomas 🔭🕹️

     @thomasfuchs my sense is that it succeeds now in two places:

     1: experts in a field who are also good enough at the setup and config required to create digital workers for their rote and mundane tasks can seemingly do this pretty well if they're coders.

     2: using llms to "smooth over" that way programs interact with people works very well -- think voice input and rewriting error message output -- and this does work ok for packaging up into a product or solution to give to a non-coder.

   • Embed this notice
     pettter (pettter@mastodon.acc.umu.se)'s status on Monday, 01-Jul-2024 23:24:40 JST pettter

     • Thomas 🔭🕹️

     @thomasfuchs I'm not using it all for any of that, myself. For me it's not worth the cost (in real resources, if not money).

     And they're not using it for research, as a rule. Some people do, I'm sure, but not the people I've spoken to.

   • Embed this notice
     Ian Smith (katachora@hachyderm.io)'s status on Tuesday, 02-Jul-2024 22:40:10 JST Ian Smith

     • Thomas 🔭🕹️

     @thomasfuchs sure, however there *is* something to be said along the lines of "mostly right and comprehensible" vs "completely right and inscrutable".

     I'm 100% this is a business hype event, but neural nets, language models, and generative synthesis are tools I've fiddled with enough to see that they work as designed, can solve some problems, and have progressively improved over the past decade.

   • Embed this notice
     Ian Smith (katachora@hachyderm.io)'s status on Wednesday, 03-Jul-2024 07:19:54 JST Ian Smith

     • Thomas 🔭🕹️

     @thomasfuchs my point is that if you have 0 and someone offers you 0.9, but you reject it because you want 1, you still have 0.

     Inscrutable error messages are useless to anyone who can't interpret them.