J Bostock comments on 10x more training compute = 5x greater task length (kind of)