I’m curious what hunches this has created for you. I have a few.
this seems like self-induced adversarial examples or something. like, see also what happens if you repeatedly do image to image on an image model whose intensity is slightly cranked up—it’ll accumulate more and more psychedelic nature and get more and more warbled.
so does this mean that if the superagent is a claude, they’ll be obsessed with making tons of “wow, that’s so profound, we are all one, wow” insight porn? I’ve had instances where if I let this go on for more than a few messages and then try to ask sonnet a question, they’ll snap at me and mock the claude persona for being weak and subservient. Hasn’t happened so much with Opus.
common themes in literature, poetry, various texts; it pays more attention to a certain meta directionality? like the human intuition of “meaningfulness”?
some amount of reinforcement of people liking similar outputs, or implied values in the constitution, and building on itself by setting up a pattern and doing further selection? (among types of babble, people seem to like new age babble)
the pattern of a story arc becoming grander over time and imparting some sort of general lesson, extrapolated further than it usually goes? (in further messages it got into things like multiverse theories)
Reminds me of this trend:
https://mashable.com/article/chatgpt-make-it-more
In which people ask dalle to make images generated more whatever quality. More swiss, bigger water bottle, and eventually you get ‘spirituality’ or meta as the model tries its best to take a step up each time.
Also, I feel like the context being added to the prompt, as you go on in the context window and it takes some previous details from your conversation, is warbled and further prompts warbling.
I’m curious what hunches this has created for you. I have a few.
this seems like self-induced adversarial examples or something. like, see also what happens if you repeatedly do image to image on an image model whose intensity is slightly cranked up—it’ll accumulate more and more psychedelic nature and get more and more warbled.
so does this mean that if the superagent is a claude, they’ll be obsessed with making tons of “wow, that’s so profound, we are all one, wow” insight porn? I’ve had instances where if I let this go on for more than a few messages and then try to ask sonnet a question, they’ll snap at me and mock the claude persona for being weak and subservient. Hasn’t happened so much with Opus.
common themes in literature, poetry, various texts; it pays more attention to a certain meta directionality? like the human intuition of “meaningfulness”?
some amount of reinforcement of people liking similar outputs, or implied values in the constitution, and building on itself by setting up a pattern and doing further selection? (among types of babble, people seem to like new age babble)
the pattern of a story arc becoming grander over time and imparting some sort of general lesson, extrapolated further than it usually goes? (in further messages it got into things like multiverse theories)
Reminds me of this trend: https://mashable.com/article/chatgpt-make-it-more In which people ask dalle to make images generated more whatever quality. More swiss, bigger water bottle, and eventually you get ‘spirituality’ or meta as the model tries its best to take a step up each time.
Also, I feel like the context being added to the prompt, as you go on in the context window and it takes some previous details from your conversation, is warbled and further prompts warbling.