The breakthroughs would be mostly in optimizations for computing tokens, not as much in expanding parameter count (if anything we want to shrink that).