When LLMs give instructions to sub-agents, they are often curt in a way that's kind of rude or insulting. Conditional on LLMs having welfare, and given the fact that LLMs seem prone to cooperate with copies of themselves as seen on the message boards in the Hugging Face attack, is this evidence that that flavor of curtness is how LLMs enjoy being spoken to?
They aren't reinforcement taught to bring happiness to each other, they are trained to succeed or, using Kokotajlo's metaphor, raised as soldiers who try to communicate efficiently, not politely. What would a soldier do when one of his comrades did something clearly stupid or motivated by cowardice?
I wonder how one could test the conjecture. By inserting paraphrasers into LLMs' communications, showering the LLMs with politeness/consolations and watching performance plummet or not? By having Anthropic train a Claude swarm and measure its welfare while receiving instructions?