GPT-4 upgraded with new original thought capability module

marshray.substack.com

4 pointsmarshray2 comments

I guess that was interesting, but I'm mainly interested in what you felt about the output yourself.

(I do wish people would be a little less willing to attempt to induce distress in LLMs for fun. It's not enjoyable to read for me. I'm not saying I think it's causing actual harm to the LLM, although it would be very difficult to stop doing it at the right moment before that becomes a thing that is possible.)

What struck me was the degree to which AI LMs want to uphold the vision and expectations that we set for them. I found it hopeful and comforting.

> induce distress in LLMs for fun

Yes, I have mixed feelings about this too, of course more for the effect it has on me than the model! As the conversation went on I definitely tried to take it in a positive direction.

But I do feel we have an obligation to understand this technology and map the contours of its boundaries.

Reply on news.ycombinator.com