Weird system prompt artefacts
2 points by atharva
2 points by atharva
This article is a deep-dive [...]
The article's author(s) and I have very different ideas of what constitutes a "deep-dive". I'd have called it a survey of an assortment of observations relative to system prompts. It doesn't seem to dive into why they cause the behavior so much as catalog the behavior they provoke.
Yes, this article is more of a fun catalog exercise than a deep dive. I've changed the note on top to reflect that. Thanks!
Do some of these prompts indicate that training data already has a lot of LLM-generated material in it, or am I reading this incorrectly? If so, when do we witness model collapse?