“Developing AI control techniques seems tractable with current technology, but does not add much value at present because there is not yet much misalignment risk.”