The idea is that a base model could be years old, with all it's faults; superficially fixed and extended with knowledge over time. She points to a research that LLMs don't "fully believe" this new knowledge. The skew could be much longer then a few weeks.