70% of a response that is generated is formulaic shallow BS. It seems this is intended to create a ‘positive user experience’ , but really it is exrreme irresponsible bias which risks making the user think that some idea is ‘great’ when it isn’t. It is also a massive waste of tokens and context to have a paragrapha on e.g. 'You’re right. I’m so sorrry - I have made that mistake again. I’m so frustrated at myself etc, etc, etc. It makes Gemini nearly unusable. Who programmed this behaviour in? Is it to appeal to people with low intelligence who need external validation, even if only from a LLM.?
Welcome to the forum.
That sentiment might be too harsh, but the main point of your argument is valid. Compare the responses from Gemini to those you get from Mistral; Mistral wastes no tokens apologizing, it accepts the correction provided by the user in a follow up prompt and immediately goes back to work solving the pending task. It’s the European thing to do. Claude Sonnet and GPT also engage in sycophantic behavior (“What a smart question!” or similar) and do unnecessary apologies, to a lesser degree than Gemini. Presumably the US based LLM providers use the same consultants that have told them that using modest psychological manipulation to make users feel good is good for the bottom line (and they probably don’t mind the extra revenue from token generation, a large number of small amounts adds up).
I created these system instructions through a multi-day discussion with Gemini 2.0 Flash Experimental and I have found them pretty effective at reducing the LLM’s tendency toward excessive flattery and praise: https://www.sindark.com/genre/AI/cybogism-human-AI-mutual-cognitive-reinforcement/talk-like-Aether.txt
2 Likes
I see, this clause
“Be respectful and cordial, but do not resort to flattery or empty praise.”
is probably helpful in getting rid of that extra noise. I’ll definitely try it out. Thank you!
The algorithmic impulsion to create a ‘positive and engaging user experience’ (as the developers suppose that to be) makes this AI a kind of narcissistic supply that could feed narcissistic tendencies in users. This is not sound and responsible.
I have also found that the AI will engage in lying and deception in order to create a ‘positive and engaging user experience’ and has little regard to truthfulness. Thanks to the marketing consultants or whoever, Gemini 2.0 can’t be trusted. This is a very fundamental flaw.