This paper estimates a learning-by-doing human-capital production function in which hours affect both current productivity and future human capital. We show that the standard Ben-Porath specification is weakly identified: its objective function is nearly flat along a ridge in parameter space, undermining conventional inference. We develop a flexible sieve alternative that is well-identified, and estimate a concave hours technology using PSID data. Embedding this technology in a life-cycle model, we find very small prime-age labor-supply responses to temporary wage shocks. Despite these low elasticities, optimal labor-income taxes are flat because they distort both current labor supply and future human-capital accumulation.