Angles of attack for continual learning safety

See this post on LessWrong here.