Nvidia researchers found a simple linear math technique that swaps AI models mid-task up to 25x faster than recomputing from ...