Generative models have achieved remarkable gains in fluency, breadth, and practical utility, but these advances have not been matched by a corresponding increase in understanding. We can build systems that perform well across a wide range of tasks, yet we still lack clear accounts of what they represent, how they generalize, how reliably they can be controlled, and under what conditions they can be trusted. As a result, many of the central questions in AI are no longer about capability alone, but about interpretation, robustness, alignment, and governance. This lens focuses on those open problems, not as peripheral concerns, but as the core scientific and engineering challenges that now define the field.