> That graph looks like a hockey stick. Is it really accelerating or are we just getting better at detecting it (this makes more sense)?

I suspect it is a piecewise combination of two linear functions. There was probably a ~fixed sensitivity across Runs 1-2, with some upgrades before Run 3[0], which resulted in an improved sensitivity for that run. So in effect the slope of detections/time is steeper with the improved sensitivity.

[0] https://indico.cern.ch/event/577856/contributions/3422625/