It really does make you wonder why all the models seem to require that. In principle, it shouldn't be a property of LLMs, and lol no it's not an "emergent property".

Post-training and "human preference" according to "data". Don't know a single developer who use these tools for work who prefer that though, but also don't know anyone who use LLMs a lot just "for fun" either, might just be vastly different preferences between the two userbases.

LLM are a text prediction engine. Starting the prompt with “you are a helpful assistant” help make subsequent text prediction more in line of that of a helpful assistant.