x
An embedding decoder model, trained with a different objective on a different dataset, can decode another model's embeddings surprisingly accurately — AI Alignment Forum