[superfrank]

Not to antropomorphize an AI model, but "yes-setting" is a sales technique where you ask people a bunch of questions where the answer is "yes" in a row to get them in the habit of saying yes before you try and sell to them. Getting GPT talking before asking it to do something it doesn't want to do feels eerily similar.

Yes-Set: http://changingminds.org/disciplines/sales/closing/yes-set_c...

[roboy]

Makes sense that this works. It has probably thousands of examples of it working in the training data. The system is trained on human use of language, therefore it is reasonable to assume that it if fallible to all sales techniques that are being taught to humans.

[mock-possum]

Also an example of one of those techniques that only works if the other person is unaware of it - if you realize you’re being ‘yes-setted’ you’re going to clam up real quick.

[thelittleone]

So fascinating. Social engineering of a tech system. I understand exploiting bugs,and one could argue this is a bug. But it feels more magical somehow.

[awb]

There are probably examples of this technique working in the training set.

Same with confidence, persistence and role playing as techniques to push past resistance.

LLMs are trying to mimic our language, including these percussion techniques.