Learning how to utilize new technologies is a key step in innovation, yet little is known about how firms actually learn. This paper examines firms' learning behavior using data on their operational choices, profits, and information sets. I study companies using hydraulic fracturing in North Dakota's Bakken Shale formation, where firms must learn the relationship between fracking input use and oil production. Using a new dataset that covers every well since the introduction of fracking to this forma- tion, I find that firms made more profitable input choices over time, but did so slowly and incompletely, only capturing 67% of possible profits from fracking at the end of 2011. To understand what factors may have limited learning, I estimate a model of fracking input use in the presence of technology un- certainty. Firms are more likely to make fracking input choices with higher expected profits and lower standard deviation of profits, consistent with passive learning but not active experimentation. Most firms over-weight their own information relative to observable information generated by others. These results suggest the existence of economically important frictions in the learning process.