←back to thread

Hermes 4

(hermes4.nousresearch.com)
202 points sibellavia | 1 comments | 27 Aug 25 08:58 UTC | HN request time: 0.267s | source
Show context
mapontosevenths ◴[29 Aug 25 20:44 UTC] No.45069190[source]▶
I appreciate the effort they put into providing a neutral tool that hasn't been generically forced to behave like "Sue from HR".
replies(3): >>45070045 #>>45070200 #>>45076115 #
bckr ◴[30 Aug 25 16:57 UTC] No.45076115[source]▶
I’m having a hard time not being sarcastic here.

The most recent news about chatbots is that ChatGPT coached a kid on how to commit suicide.

Two arguments come to mind. 1) it’s the sycophancy! Nous and its ilk should be considered safer. 2) it’s the poor alignment. A better trained model like Claude wouldn’t have done that.

I lean #2

replies(2): >>45077511 #>>45080008 #
1. karan4d ◴[30 Aug 25 19:53 UTC] No.45077511[source]▶
the sycophancy is due to poor alignment. the instruct based mode collapse results in this mode collapse induced sycophancy. constitutional alignment is better than the straight torture OAI does to the model, but issues remain