Mixtral has a weird tendency to
Here’s just one example of that but it has happened MANY times to me already:
It made me laugh. 
Mixtral seems to have some issues and until now it is much worse than the premium llama model it replaced, which is weird because every test I find on the internet says it should be better. It is worse at fixing up text, leaving mistakes way more often in it, or randomly deciding to completely rewrite the text instead of fixing it like I instruct it, ending up with text that has completely different length.
And then this funny thing happened:
and it started giving me these “Confidence: X%” at the end of every reply for that discussion.
I don’t know if the LLM is wrapped wrongly or if it expects a differently structured prompt but it seems very unstable in the sense that it keeps surprising me with unexpected behavior. It’s strange cause in theory it’s better than claude-instant-v1 (though Claude’s larger context window is extremely useful for summarizing since most youtube videos and articles don’t fit in Mixtral’s window).
Still, 
you Brave for trying to host more models yourselves. I’ll play around with Mixtral more but I’m switching to claude for my primary no-nonsense one for now.
It’s weird as responses vary. Most of the things I have used the Mixtral for has been just direct answers. There was nothing about confidence levels, numbering, etc. So I guess it will vary based on what it’s being used for.
I’m going to assume that the grammar one had been labeled so as you went through different variations, it would be easier to refer to a suggestion it made.
As to confidence, no idea why that showed except perhaps it was just letting you know it wasn’t really confident about the answer. Then again, when you consider him being a Linux blogger before and then everything that happened after, I could understand the hesitancy.
