Re: thoughts on future regulation of encryption
"Otherwise, it'll simply improve over time."
By what metric? The benefit of a computer playing Go, Chess, etc is that while the game is famously complex with incredible numbers of permutations, the rules are clearly defined, the objective that must be achieved to win is also fairly clearly defined (take/trap the King, etc). All a learning machine needs is to have a sufficient number of games to figure out the best way to get from a known starting position to a desired winning end when playing an opponent whose exact moves cannot be entirely predicted in advance. That's the skill, responding to a "random" behaviour of the human player to keep the best advantage (which probably requires tracking player moves to work out what the human is attempting to do).
Now let's turn our attention to a learning system for spotting terrorists. There's no defined starting situation. There's no defined end game. There's no defined list of behaviours that may occur in a message in order to indicate terrorist activity. "Don't forget to put the cake in to bake at three o'clock this evening" posted on Mumsnet could be a message from one scumbag to another - they're hardly going to say "Westminster, 3pm, bang" are they? So... We don't exactly have a start or an end or even a middle. Good luck getting a machine to "learn" that.