That seems like a slightly different aspect of this to me – we have seen tons of examples now of models being fine-tuned on carefully created artificial data, but to me that doesn't speak to the risk of unintentional artificial data affecting the models at the pre-training stage

→ View original post on X — @simonw,