Comment by andai
1 day ago
There was a paper a while ago, where they found a strategy for GPT beating humans on the Turing test 2/3rds of the time.
They just prompted it to talk the way a 13 year old girl texts.
(Also they prompted it to pretend not to know anything about anything.)
Most of the LLM tells we usually recognise are due to their heavy attunement to role playing a "helpful assistant" with a knack for correct typography.
If you used a smaller corpus, followed a different kind of fine tuning routine with rather normal conversation patterns, and wrote a different base prompt, I don’t see why it shouldn’t be possible to come up with a very particular personality, so to speak.
> where they found a strategy for GPT beating humans on the Turing test 2/3rds of the time.
IIRC, PARRY https://simulation-and-society.org/parry/ got there in 1972 by pretending to be a paranoid schizophrenic person
>while ELIZA has a comforting effect that many crave, PARRY often actively makes users uncomfortable and afraid
PARRY is my spirit animal.
> In empirical evaluations conducted during the 1970s, experienced psychiatrists proved unable to reliably distinguish PARRY's responses from those of actual patients diagnosed with paranoid disorders, achieving only 48 percent accuracy, a figure statistically indistinguishable from random guessing.
https://ai.miraheze.org/wiki/PARRY
Also I found the source code!
https://www.cs.cmu.edu/afs/cs/project/ai-repository/ai/areas...
I feel like pretending to be a child is cheating somewhat. Might as well print random key presses and claim it's a toddler.
[dead]
https://xkcd.com/2635/