GPD for can compress and decompress prompts into non human readable forms. It will generate this kind of interesting output of a mix of compressed kind of keywords and emojis. And then in some cases, when you feed it back to GPD for separate, so there's no memory of how it compressed it, GPD for can then decompress and tell you what that original text was. Yeah, I mean, we've seen that already since last year, we have been multiple papers like Seiken where you don't need to retrain it. If you just kind of prompt it, it can tell you go here do this to that. The challenge then becomes, well, okay