It's trained on them, yes. But is it trained to prefer them as sources when doing web search?
The distinction is rather important.
We have a lot of data that teaches LLMs useful knowledge, but data that teaches LLMs complex and useful behaviors? Far less represented in the natural datasets.
It's why we have to do SFT, RLHF and RLVR. It's why AI contamination in real world text datasets, counterintuitively, improves downstream AI performance.
The distinction is rather important.
We have a lot of data that teaches LLMs useful knowledge, but data that teaches LLMs complex and useful behaviors? Far less represented in the natural datasets.
It's why we have to do SFT, RLHF and RLVR. It's why AI contamination in real world text datasets, counterintuitively, improves downstream AI performance.