I'm not sure I'd appreciate this feature. When I used LLMs I found conversations got more difficult the longer they went on - I'd ask one question, then another, and it'd try its best to answer the latter in the context of the former in a way I didn't intend. I used new chats as an explicit way to escape this. I also created new chats when I felt the LLM was going down the "wrong path" or generally interpreting things obtusely.

That is strange, I thought this had always been the case. Long ago, I had a chat about feature selection and dimension reduction, and I expressed my preference of not resorting to PCA because it doesn't actually remove features. After some time, I had a new session chatting about multicollinearity and what adverse effects it might cause, and among the suggestions for addressing that, it suggested dimension reduction, saying "Although you don't prefer PCA ...".

I wonder how this works? Are they just RAG-ing the chat history or something more clever?

i’ve told it not to change my contractions a thousand times, still can’t remember

[dead]