There's actually quite a few AA terms thrown about...
■SSAA - Super-sample AA. Renders the entire scene with an increased resolution, then uses filtering to scale it down to the output resolution. Technically this method yields the most perfect result; the principle of scaling that way is the same to how, say, a digital camera takes an image at its own resolution, but with the equivalent of infinite SSAA. The downisde is that this is an absolute performance hog.
■MSAA - Multi-sample AA. Begins each render pass with a quick stencil check, making generally 2, 3, 4, 6, or 8 (though other numbers are possible) checks per pixel, at slightly different locations "within" the pixel. If it finds that all of them match the same polygon, then it renders that pixel normally without any form of AA. If It does find multiple polygons present in the pixel, it then it takes a sample of color from each point it'd checked earlier, and blends them together. This technically may not be as precise as SSAA, but the visual effect is almost indiscernible for most, and the performance drain is markedly reduced.
■CSAA - (something)-sample AA. This is a new type of AA, used by nVidia on their GeForce 8 cards. Technically, at heart, it's a form of MSAA, as it works in the exact same manner. HOWEVER, even though it will perform the initial stencil pass with 8 or 16 samples per pixel, when it needs to blend, it will only draw 4 of those samples, and then blend them together according to the ratio that the stencil pass found; a 16x stencil pass that found polygon A with 1 part, B with 3 parts, C with 5 parts, and D with 7 parts, would make one sample from each polygon, then blend them together at the aforementioned ratio.
■FSAA - Full-scene AA. This technically is not a "form" of AA; it refers to AA methods applied to the whole scene, as opposed to object- or line-based AA, which is not found at all in consumer graphics cards and commercial PC games.