Aha, for the time being (and probably going into the future without major architectural changes to transformers) prompt engineering is important. For a one-sho

Aha, for the time being (and probably going into the future without major architectural changes to transformers) prompt engineering is important. For a one-shot question, it's probably fine. But when asking a transformer to create a completion for a conversation history, it seems to be very important to: * Start with a clear prompt establishing the conversation, rules and goals, format of responses * Continue to re-establish these rules throughout the text to remind the transformer They say "attention is all you need" but transformers seem to be like a child hopped up on sugar, too much input and it starts forgetting things from the start of the input. Hell even from one sentence (or message in a chat history) to another it can forget, presumably because the last chunk of input has keywords mentioned several times through the text and therefore stronger than certain rules (that are only mentioned once, at the start). I'll ask it to play a role playing game and it'll be all like: "The man X walks up to you an introduces himself, he's wearing a green tunic" Then after I interact it's all suddenly: "X looks own at his red tunic and also suddenly he's a dragon and actually a woman". I think it's because attention correctly identifies all the various topics and keywords of the conversation, but it doesn't seem to create a hierarchical tree of "attention"...I wonder if the same principles of finding similarity between embeddings could also apply to the attention mechanism. If transformers had something like that, it would be nice to reliably set an outline that forms the root node of an attention hierarchy, with the current topic at hand (towards the end of the input) as a leaf node.