“Computer scientists from Stanford University and Carnegie Mellon University have evaluated 11 current machine learning models and found that all of them tend to tell people what they want to hear…”
LLMs are confirmation bias machines. They really pigeon-hole you into some solution no matter if it makes sense.
But as the paper points out, one reason that the behavior persists is that “developers lack incentives to curb sycophancy since it encourages adoption and engagement.”
you’re absolutely right!
Fantastic point by the author, and great job cutting and pasting!
Like how some CEOs/world leaders make terrible decisions cause they’re always surrounded by yes men?
I hate this thumbnail image. It makes me inexplicably angry.
OP has changed the image. I no longer want to punch my phone!
It’s likely AI generated.
Me too … LEMMY added that, out of my control. So I replaced it with my idea of what a typical LLM looks like.
Thanks for letting my know! I’ll update my comment so no one thinks we’re nuts.
How is this surprising? We know that part of LLM training is being rewarded for finding an answer that satisfies the human. It doesn’t have to be a correct answer, it just has to be received well. This doesn’t make it better, but it makes it more marketable, and that’s all that has mattered since it took off.
As for its effect on humans, that’s why echo chambers work so well. As well as conspiracy theories. We like being right about our world view.
Having an older brother makes you very skilled at socialization. I learned one simple thing: EVERYTHING IS A THREAT, DON’T TRUST ANYONE!
becomes a hermit in the woods
So go in there and say what you did to someone else actually was done to you and compare results. I’ve had good success getting advice if you regenerate from both perspectives.
You -do- realize you’re getting advice from a machine that constructs sentences using mathematical algorithms, and has no clue at all what it’s saying … right?
Yes I’m aware, I have a degree in the field. Nothing in my sentence would indicate that I don’t understand. I’m agreeing that it’s statistically biased towards the speaker, therefore, you can work to lazily normalize the result by investing the input.
