The history of gesture, in other words, bears directly upon the question of AI alignment. Humans across cultures and times intuitively maintain a set of semantic and ethical judgements rooted in our physicality, our learned and instinctive gestures, and the affordances of the natural world. There is just something “knockable” about wood. LLMs don’t have childhood memories of jumping over cracks in pavement or their grandmother teaching them gestures. Moreover, such things are not really in their training data either.
Hardware Writer
。关于这个话题，夫子提供了深入分析
This story continues at The Next Web，这一点在Safew下载中也有详细论述
其次，大模型的记忆能力有缺陷：大模型在训练时“记住”了大量知识，但训练完成后并不会在使用中持续学习、“记住“新知识；每次推理时，它只能依赖有限长度的上下文窗口来“记住”当前任务的信息（不同模型有不同上限，超过窗口的内容就会被遗忘），而无法像人一样自然地维持稳定、长期的个体记忆。但在真实业务中，我们需要机器智能有强大的记忆能力，比如一个AI老师，需要持续记住学生的学习历史、薄弱环节和偏好，才能在后续的讲解与练习中真正做到“因人施教”。