Given our modern understanding of how LLMs work (like the recent Anthropic work), I wonder if that insight can be used to quantize better. For example, we know

Given our modern understanding of how LLMs work (like the recent Anthropic work), I wonder if that insight can be used to quantize better. For example, we know that LLMs encode concepts through rotations (but not magnitude) of several neurons. Bringing this up because the abstract (and the mention of rotations) reminded me of recent LLM interpretability posts.