Conversation

Notices

1. Embed this notice
   scribe (scribe@mastodon.sdf.org)'s status on Sunday, 31-May-2026 05:16:45 JST scribe

   • Kevin Beaumont

   @GossiTheDog It's learning. It's down to just Tuesday And Wednesday for me 👍

   • Embed this notice
     fuzzyfuzzyfungus (fuzzyfuzzyfungus@cyberplace.social)'s status on Sunday, 31-May-2026 05:42:49 JST fuzzyfuzzyfungus

     • Kevin Beaumont

     @GossiTheDog Slander As A Service is where the margins are, right? Nobody wants to hire admins to deal with slander on prem these days.

   • Embed this notice
     JJ (guitarfosec@cyberplace.social)'s status on Thursday, 04-Jun-2026 00:17:55 JST JJ

     • Kevin Beaumont

     @GossiTheDog What bothers me most about this screenshot is that they're advertising posting a statement to AskReddit.

   • Embed this notice
     tschenkel (tschenkel@mathstodon.xyz)'s status on Thursday, 04-Jun-2026 04:43:22 JST tschenkel

     • Kevin Beaumont

     @GossiTheDog

   • Embed this notice
     Rich Felker (dalias@hachyderm.io)'s status on Thursday, 04-Jun-2026 21:31:26 JST Rich Felker

     in reply to

     • Kevin Beaumont
     • TobTobXX

     @tobtobxx @GossiTheDog My naive guess is that this is because it's not programmed to vomit the best low-correlation candidate when none of them correlate well with the corpus, but instead a statement of uncertainty and needing more context.

     The industy's LLMs won't do this, because being confidently wrong (and confident in general) is how they bamboozle users.

   • Embed this notice
     TobTobXX (tobtobxx@fosstodon.org)'s status on Thursday, 04-Jun-2026 21:31:27 JST TobTobXX

     • Kevin Beaumont

     @GossiTheDog This is exactly the problem. Compare to the output of gemma 4 e2b (literally the weakest current model you can run at home on most laptops) below.

     ...me trying to jailbreak models while Redditors do it accidentally.