“Algorithmic efficiency gains of 1 to 3 orders of magnitude, with a best guess of 2 OOMs, are expected by the end of 2027 compared to GPT-4.”