The article claims that deferred rendering doesn't support MSAA, but this is only true for a naive implementation.
You can, for example, generate a stencil mask of complex pixels in the prepass and use hierarchical early stencil reject to run a per-sample pass for complex pixels and a per-pixel pass for non-complex pixels (with imperfect SIMD lane utilization).
You can also use a pass to process the complex pixels mask and collect a buffer of (x, y, sample) coordinates to allow a later single hybrid per-pixel/per-sample compute shader pass with perfect SIMD lane utilization.
You can, for example, generate a stencil mask of complex pixels in the prepass and use hierarchical early stencil reject to run a per-sample pass for complex pixels and a per-pixel pass for non-complex pixels (with imperfect SIMD lane utilization).
You can also use a pass to process the complex pixels mask and collect a buffer of (x, y, sample) coordinates to allow a later single hybrid per-pixel/per-sample compute shader pass with perfect SIMD lane utilization.