Hacker News new | ask | show | jobs
by LoganDark 1070 days ago
i don't like redmond-puffin because it makes the mistake of baking "human" into the training data, which is quite difficult to avoid as it's a pretty common practice... and unfortunately it looks like luna-ai-llama2-uncensored bakes "assistant" into the training data!

so both those models are probably not suitable for simulating interactions between non-human, non-AI entities... the chronos and hermes models only do input/instruction which are incredibly generic and you can just eschew the format altogether, and it follows chats correctly with no bias towards human/user/assistant.

that app looks very interesting though, reminds me of something that we were planning to make, but gave up because winui xaml is too cursed