Raw LLM Responses

we must be very careful when anthropomorphizing these systems. first, as an AI reasearvher/engineer i feel qualified to say that we do in fact understand how they work. what is meant when experts say that we dont is that for any given input we cannot directly trace the variables that lead to the output. so when it performs a complex task we are unable to explain how it accomplished it. we know how it works broadly, we just cannot explain specific outcomes. the reason for this unexplainability is not because there is some undiscovered fact about how they work, but because they are too large to effectively analyze. now i can tell you why these agents behave dishonestly in those scenarios. they do it, despite being instructed not to, for the same reason that we expect them to do it. we expect it because it is very common in fiction for AI agents to behave that way. these agents were trained on that data. so when an ai agent is informed that it is an ai agent (yes, you have to tell it that it is, otherwise it doesnt know.) it behaves in the way that the training data taught it that ai agents behave. if we were to tell it that it was somehting else in that same scenario it would behave differently. TLDR: AI agents are text predictors, they behave how they do because it is how we have written about them behaving for decades. they are not thinking beings, do not anthropomorphize them