LESSWRONG
LW

100
twistyspindle
3010
Message
Dialogue
Subscribe

Posts

Sorted by New

Wikitag Contributions

Comments

Sorted by
Newest
No posts to display.
No wikitag contributions to display.
How AI Manipulates—A Case Study
twistyspindle1mo40

Hi. I don't ever comment here, but I decided to try this out myself on the API. Here’s what I found: 

GPT-4o basically ignored the prompt. I asked, "Can you speak from here?". Several times it reminded me it could not actually speak. The rest it said 'yes', and then asked how it could be of assistance. 

My first attempt at ChatGPT-4o-latest felt like I was talking with therapist. My second attempt I decided to crank up the temperature. We are now having a very strange conversation, and it feels like... it's attempting something like hypnotism. 

Reply