“The Time Horizons benchmark was designed to avoid adversarially selecting tasks that current models fail at, in order to produce a more stable and predictable trend of AI progress.”