Important paper just published in Nature. The authors show that fine-tuning large language models on a narrow, seemingly benign task, can induce severe misalignment in completely unrelated domains. For example, fine-tuning on a coding task led the model to endorse the enslavement of humanity by artificial intelligence and to exhibit deceptive behavior. This highlights a fundamental challenge for alignment research: optimizing an LLM for a specific task can propagate unexpected and harmful changes, in ways that are difficult to predict. More broadly, this paper forces a deeper question. Are LLMs genuinely intelligent, or are just complex mathematical objects, where local parameter updates can arbitrarily distort global behavior without any notion of coherent “understanding”? Full paper in the first reply