I don't think this is an appropriate post for a Show HN (something you've made that others can tinker with) but I do agree that it's an impressive demonstration of apparent understanding. Have you tried this role play using any of the other popular LLMs?
I also wonder how ChatGPT would react if the roles were reversed, so that it becomes the buyer in this scenario, depending on if the seller was using anchoring bias or not.
Except there's a fatal flaw in the response, in that the $15 anchor was immediately implausible (in that it was way higher than the normally accepted price). In order for anchors work, they need to fall somewhere within the range of what might be plausible to the counterparty -- otherwise they fail.
For same reason that if for example you're in a salary negotiation, you might get somewhere by starting off asking for a 50 percent higher salary than what's being offered -- but if you start off asking for 2x, you'll be flagged as a primadonna and immediately shut down.
So if anything, this example (while certainly interesting) further showcases the fact that at the end of the day, LLMs have at best a very limited understanding of what's going on. And are in most cases are basically "satisficing" -- cranking out a stream of tokens which looks like a satisfying answer, but you when you break it down, you find it isn't, really.
3 comments
[ 0.23 ms ] story [ 19.7 ms ] threadI also wonder how ChatGPT would react if the roles were reversed, so that it becomes the buyer in this scenario, depending on if the seller was using anchoring bias or not.
For same reason that if for example you're in a salary negotiation, you might get somewhere by starting off asking for a 50 percent higher salary than what's being offered -- but if you start off asking for 2x, you'll be flagged as a primadonna and immediately shut down.
So if anything, this example (while certainly interesting) further showcases the fact that at the end of the day, LLMs have at best a very limited understanding of what's going on. And are in most cases are basically "satisficing" -- cranking out a stream of tokens which looks like a satisfying answer, but you when you break it down, you find it isn't, really.