Training large language models on narrow tasks can lead to broad misalignment (nature.com)

Isn't this just "catastrophic forgetting?" e.g. training LLMs on anything leads them to get worse at what they learned before.