The User Is Visibly Frustrated
11 points by pscanf
11 points by pscanf
“Make the agent respond like a robot would” I’ve gone a step further. I’ve asked Claude to respond to me “as if you were the beloved character SHODAN from the System Shock games.” Having the computer call you “insect” or “flesh-thing” has really helped the syncophancy.
Maybe I would prefer a more radical solution: drop the human pretense entirely. Make the agent sound clinical, robotic. Dispel the idea that I’m interacting with a person, and make me feel like I’m just approving or rejecting random outcomes.
Thumbs up for this. I'm all for ridding LLMs of deliberate anthropomorphization. They don't have to be conversational. I wish they were closer in form to search engines.
I think it may also be worth dropping the word "agent" and replacing it with something that gives the impression of a tool or a service.
Relatedly, I've heard some claim that LLMs perform better if they're given some kind of sympathetic prompt ("my job depends on it"). This makes me wonder if the LLMs aren't deliberately anthromorphizing, but maybe that's a fundamental consequence of being trained on a massive corpus of human-generated text. You can prompt them to talk to you like a robot, but the LLM may just be playing along i.e. differently syncophanting.
You can’t really suppress the human tendency to anthropomorphize everything, but I like the idea of reducing the impact of it a bit by instructing the LLM to respond in a less human way