MLB 14 The Show Livestream (PS4)

Hi, video game artist from Guerrilla Games here! You may know us from the Killzone series.
While the picture OP posted is accurate in the sense of downres/upres methodologies, there is a very important point to be made here.
It's true that subdividing a 2k poly model will not result in a more detailed model. And that downscaling a 20k model will result in a more accurate 2k representation.
But this is not what game graphics are about. We have almost stopped caring about polygon counts in the last couple of years. Increasing the poly count beyond 5k for the bust that is shown in the example picture will add almost nothing. The contour may look a little better, but that's it.
What actually happens in current AND next gen games is that the detail contained in the hipoly model is baked into a "normal map". It's a texture that defines the surface angle for every pixel. That way, even though the model may be relatively low poly, the lighting on it looks like the full high res model.
The guys over at Epic games did a great demonstration of this back in 2006, but it looks like they removed that part of their website by now. It was a great demonstration of how a 4 million poly character model was downressed to 9000 polies and effectively looked the same because of normal maps.
So anyways - about the actual issue at hand; there really is a diminishing returns issue in games. Every time you double the amount of polygons, the subject will only look marginally better than the previous generation did. The difference between PS1 and PS2 was enormous. The difference between PS2 and PS3 was smaller, although still very significant. The difference between PS3 and PS4 is clearly noticeable, but it's not as big a leap as previous generations were. Future generations will no doubt offer smaller changes in graphical fidelity, and put more focus on added features.
#edit: Wow! So many responses. I'd love to answer them all but don't have the time. Some clarification:

1. Poly counts still matter! When I said "stopped caring" I meant that we don't design objects with the intent of saving on polygons. We still do crazy amounts of optimization once the object is made. And of course we use Level of Detail models to reduce poly counts of objects that are further away from the camera.
2. There are still great improvements to be made in graphics technology. We're moving away from static pre-baked lighting and that means we want to simulate light bouncing around in real time. That's ridiculously demanding on the hardware. And you wouldn't see the difference in a screenshot, but you will see it when objects and lights move around.

Game graphics can still get loads better, but the improvements won't come in the form of higher poly counts. At least not by much. Look forward to stuff like realtime global illumination, particles, dynamics, fluids, cloth and hair, etc. It's not necessarily stuff that will make screenshots look better, but it will definitely make game worlds feel more alive.
Also; THANKS FOR THE GOLD! Feel free to call on me if you need me in the future. Cheers!