Are the weights invertible, or are the prompts being fed into the model invertible?

The prompt is reconstructable from the (latent) representation vector but obviously not from the output