Numerator: 1

The passage around this formula

The method is a curve fit, not a lookup table, and stating it plainly exposes the assumption it rests on. For a task of human-expert duration t , METR fits a logistic curve to the model’s observed success rate across many tasks of varying length: $p(t) = \frac{1}{1 + \exp\left(a + b \log t\right)}$. with b > 0 , so predicted success falls as human-equivalent task length grows. The reported time horizon $T_x$ at success level x is then the duration at which the fitted curve crosses that threshold, i.e. the t solving p(t) = x . Two things follow directly from this formulation that a single reported number obscures: the curve, not the crossing point, is the actual result, and a different chosen threshold x produces a different…

Read this part in the article →