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