Earnings heterogeneity plays a crucial role in modern macroeconomics. We document that mean earnings and measures of earnings dispersion and skewness all increase in US data over most of the working life-cycle for a typical cohort as the cohort ages. We show that (i) a human capital model can replicate these properties from the right distribution of initial human capital and learning ability, (ii) differences in learning ability are essential to produce an increase in earnings dispersion over the life cycle and (iii) differences in learning ability account for the bulk of the variation in the present value of earnings across agents. These findings emphasize the need to further understand the role and origins of initial conditions.