We develop a general model, with the exponential bandit as a special case, in which high-ability agents are more likely to achieve early success but also learn faster that their project is not promising. These counteracting effects give rise to a signaling model in which the single-crossing...