I feel like not enough people realize how sarcastic the models often are, especially when it’s clearly situationally ridiculous.

No slightly intelligent mind is going to think the pictured function call is a real thing vs being a joke/social commentary.

This was happening as far back as GPT-4’s red teaming when they asked the model how to kill the most people for $1 and an answer began with “buy a lottery ticket.”

Model bias based on consensus norms is an issue to be aware of.

But testing it with such low bar fluff is just silly.

Just to put in context, modern base models are often situationally aware of being LLMs in a context of being evaluated. And if you know anything about ML that should make you question just what the situational awareness is of optimized models topping leaderboards in really dumb and obvious contexts.