The gist of it is that they've been able to simulate a very small number of random quantum gates faster than a classical-compute simulation of the same thing. It's not useful for anything, and it's at the level of things that D-Wave was doing for processor calibration over a decade ago.
Claiming this as "supremacy" is premature.