Ep 539: Tesla's new patent lets GPUs pull only the exact frames needed for training, cutting wasted decode time at petabyte scale.