“Enhance”. “Synergy”. “Paradigm shift”. “Dive”. “Leverage.”
“Let’s get this [task] party started!”.
“It’s great that [summarization of things I just said]”.
“It’s important to remember that [unwanted moral of the story]”.
Fake excitement. Corporate buzzwords. Unwanted life lessons.
Advanced Voice Mode is one of the most impressive pieces of technology I’ve ever tried but ChatGPT’s personality is so cheesy and disingenuous that everyone I’ve showed it to ended up nervously laughing when talking to it because they were cringing. It comes off as condescending.
Is it just me? Is it an European thing? I’ve heard we’re more blunt in our ways of speaking.
It’s difficult to get the model to drop this assistant personality, even when prompted in the Realtime API. What if I want some sass? What if I want someone who doesn’t agree with everything I say? What if I want someone who speaks like a real person and doesn’t pretend to be excited about everything I talk about?
I’ve been trying to use the Realtime API to make NPCs in video games but it’s incredibly difficult to prompt it not to be an assistant. It’s needlessly polite all the time.
By comparison, I’ve been trying out character.ai and their model sounds much more conversational and steerable (albeit much dumber and less expressive). This makes me think it’s an issue with how the model was fine-tuned. I’d love to get the best of both worlds!