I know that fragments are rasterized in a 2x2 quad. That's why the seam between two triangles can be rasterized twice. At least that article states this. I wonder is it still true today on modern hardware?
It is true to on today's GPUs, unless one or more vendors has implemented a technique such as this: http://graphics.stanford.edu/papers/fragmerging/shade_sig10.pdf
To clarify, rasterization is not being done multiple times, or redundantly - rather it is the shading work that is done redundantly. Each 2x2 quad will be shaded once for every triangle that covers at least one sub pixel sample in that 2x2 quad. So, for 4x MSAA, that could be as many as 16 times if every sample covers a different (very small) triangle.
That's why the seam between two triangles can be rasterized twice.
The seam won't be rasterized twice because of the OpenGL or D3D fill/rasterisation rules.
What I think Humus is alluding to is that a fragment shader may be executed multiple times on pixels spanning the shared edge, which just implies wasted effort.
The moral of the story is "avoid small or long/thin triangles if you want efficient rendering".