While there is a large consensus about model projections for global trends such as increasing global mean surface temperature under various emissions scenarios, the reliability of more local (and long term) projections is far weaker. But impact assessment often rely on local, high-resolution climate projections. In many ways, climate modelling and climate decision-making are now at a turning point, facing the tension between, on the one hand, the current focus on more detailed, complex climate models and on increasing computational resources and, on the other hand, fundamental epistemic constraints (such as structural biases) and uncertainties linked to high-resolution (local, long term) projections. We consider the possibility that these latter may point towards some irreducible epistemic limitation for climate modelling and may require some sort of 'paradigm shift' in the epistemology of climate science, where the notions of scientific understanding and explanation play a more central role.