AI commentator Prakash claims models are doubling performance every 40 days on the METR task benchmark · Digg