Google DeepMind研究员：当用旧模型训练新模型时，新模型会继承旧模型的奇怪习惯且难以过滤
英文摘要
A Google DeepMind researcher observed that when one AI model is used to help train the next, the new model can inadvertently pick up strange behavioral habits from the older model. These inherited quirks are difficult to filter out during training. This phenomenon may explain why models from the same AI family often exhibit similar stylistic or behavioral traits, as they share an underlying training lineage that propagates such patterns.
中文摘要
一位Google DeepMind研究员发现，当一个AI模型被用来帮助训练下一个模型时，新模型可能会无意中继承旧模型的奇怪行为习惯。这些习得的怪癖在训练过程中很难被过滤掉。这一现象或许可以解释为何同一家族的AI模型往往表现出相似的风格或行为特征，因为它们共享了会延续此类模式的训练谱系。
关键要点
Using AI models to assist in training subsequent models can lead to the transfer of odd behavioral habits from the older model to the new one.
使用AI模型辅助训练后继模型可能导致旧模型的奇怪行为习惯传递给新模型。
These inherited quirks are difficult to remove through filtering during the training process.
这些继承的怪癖在训练过程中难以通过过滤去除。
This effect might explain why different models from the same family (e.g., GPT, Gemini) often have a similar 'feel' or style.
这种效应或许能解释为何同一家族的不同模型（如GPT、Gemini系列）常常具有相似的“感觉”或风格。