Mechanistic theory proposes data drives exponentially increasing AI time horizons
a new mechanistic theory suggests that the metr graph's trend of exponentially increasing ai time horizons is driven primarily by data accumulation rather than time.
responding to the famous metr graph showing consistent ai capability gains, this theory introduces a mechanistic explanation linking task completion to data scaling via wright's law. it argues that the 'time horizon' metric functions as a noisy estimate for the number of distinct subtasks required, implying that the exponential trend relies on the availability of training data. this deeper understanding allows for better predictions about when the trend might saturate or break, offering a sanity check on simple extrapolations.