so the training dataset should not include the direct usage of a theoretical "end chat" tool, just tool use, and maybe examples of positive/negative interactions with users? that last one might still bias the model though

also bc im stupid and i don't understand half the math that makes it tick. i kind of get how it works from basic descriptions but not much more, and I couldn't probably recreate the architecture if asked, even if you give me a few math textbooks and a month

i believe anthropic gave claude an "end chat" tool so it is possible in theory. not sure about the emergent vibe thing though. maybe smth smth cosine similarity smth smth latent space shit
these words only have half a meaning for me bc im unsure how much LLMs can approximate "thinking" and "feeling"

well that just confirms that it was in fact her greed and ambitions that led to this clusterfuck happening.

unfortunately i am imaginary and do not exist in real life