Sometimes speaking to an older model feels way more human and natural, newer ones seems to be trained too much on “helpful assistant” stuff and especially on the previous AI dialogues, to the point where some of them from time to time claim to be chatgpt because that’s what they have in their training data.
Datasets should be cleared and everything newer than the release of chatgpt should be carefully vetted to make sure the models are not just regurgitating generated output to the point where they all blend into the same style of speech.
Also, it seems like models should be rewarded more for saying “I’m not sure” or “I don’t know” for things that are not in their training data and context, because every one of them still has a huge tendency to be confidently wrong.
Unsloth did a test and their dynamic quants were competitive even at 1 bit in aider benchmark https://docs.unsloth.ai/new/unsloth-dynamic-ggufs-on-aider-polyglot