Teaching a Model a Language You Just Invented

Codebreakers have a word for the moment a cipher cracks open: the break . You stare at a wall of Wklv pdnhv qr vhqvh and count letters, and then all at once the shape of the language underneath swims up and the gibberish is just English wearing a mask. Human cryptanalysts describe it as a click. I wanted to know what that click looks like for a language model — and, more precisely, how fast it happens when the model has never been told the key. This is the second piece in a loose series on the…