@vitaut by the way, if you actually build such a system (maybe not with a human-language prompt of maximum word count in the middle, but say with a noise radio channel model as what's called a penalty), and train it so that what gets through is the closest possible to what went in, that's called and autoencoder, and it's a useful thing, because it allows machine learning to find a nice mapping from input data to storable/transmittable compressed form back to full data, for cases where theory so

@vitaut I hope that guy is called NLohmann and this code is used as a cautionary tale???

@vitaut please – don't. Don't give them ideas.

@vitaut I think that might be what is usually called "fraud" (narrow once more, and it becommes frauf).