Over the past two decades, studies into GNSS (GPS) time series have revealed that subduction zones are not in a steady convergence state between large earthquakes. Sometimes these faults release stress with slow-slip events lasting weeks-to-months (e.g. Rogers & Dragert, 2003). Other times it appears that background creep and locking rates have changed (e.g. Heki & Mitsui, 2013), although the physical explanation for such subduction accelerations remains unclear.  Isolating interseismic transients has been a longstanding objective of the GPS research community, since it is believed that a more complete characterization of interseismic transients will improve our mechanical understanding of how large asperities eventually fail.  However, it has been challenging to isolate these transients from the full signal due to noise, steps, and the lack of a-priori knowledge on the location of transients in the time series.  In this presentation I will show a solution to this problem using the GrAtSiD algorithm (Bedford & Bevis, 2018) applied to GPS time series in Japan and Chile.  Using GrAtSiD, we are able to automatically and robustly estimate the accelerations of GPS stations via a linear regression (least squares optimization) of a modified trajectory model.  Accordingly, we observe that velocities in both the Chilean and Japanese subduction zones are highly transient on the GPS observable timescales. Before the Maule (Mw 8.8) and Tohoku-oki (Mw 9.0) earthquakes, we capture a large reversing signal acting over the 1,000 km scale and lasting several months that we suggest is related to the interaction between deeper and shallower subduction processes.  I will present these signals and discuss the interpretation from our recent paper "Months-long thousand-kilometre-scale wobbling before great subduction earthquakes