I think that this "paradox" for AI training and capabilities is predictable from first principles, and has very limited implications.