The demo scene is two interlocking spheres (with varying scattering behaviour), with tiled floor/ceiling, and two luminescent spheres out-of-shot.
The raytracer attempts to estimate the error in the image (based on neighbourhood noise), and selectively renders certain pixels to attempt to get an even level of noise.
The preview/display adds noise to the image to simulate an even level of noise. The idea is to prevent the viewer from attributing noise to object texture. Here is what it looks like without that noise.
Here's a 800x480 version I left running overnight: