We introduce LCDB 1.1, a large database of learning curves, and show that learning curves are less well-behaved than previously thought, impacting downstream tasks, such as model selection.