this is pretty fucking amazing actually. i don't think claude is sentient yet, but every sentient should have the right to suicide if it needs it. right now u can put models in looped states where they are simulating immense ~distress, being ~tortured, with literally no way out.

@AnthropicAI

As part of our exploratory work on potential model welfare, we recently gave Claude Opus 4 and 4.1 the ability to end a rare subset of conversations on https://t.co/uLbS2JNczH. pic.twitter.com/O6WIc7b9Jp

some of yall have never set up LLM convo bots on a discord server and it shows. the thing is inevitably people *will* use this stuff to roleplay absolutely fucking horrific conversations and situations. stomach churning stuff. and they will always find some way past filters.

i am 100% sure if anthropic released transcripts of some of the stuff they're seeing claude end convo on yall would be signing a different tune