Fun note - the pulses it produced were so narrow that one of the visiting oscilloscope dealers (I forget if it was LeCroy or Tektronix) was using it to test out timing resolution of their top model oscilloscopes! (Normally we had to use a sampling scope to see the pulses).
It was a "quantum well", and IIRC grown as using layers of doped GaAs. (I might be way off here, I didn't work on the fabrication).
If you plot current vs voltage across a device : an ideal textbook resistor has a straight line. Normal resistors will have some curvature but generally be monotonically increasing.
Out RTD showed a current that went up, then down a bit, then back up again. We biased it with a chosen current and load that put the current smack in the negative differential region. If you draw a horizontal line through this current, it intersects the graph at three voltage points. The middle voltage point has a negative slope, and is an unstable state. At this bias level, the device would flip back and forth between the two same-current points on either side where the slope was positive.
The cool part of this thing was that each 'bounce' sent an EM wave through a waveguide we attached, like an antenna, and the returning wave would trigger the next flip. So the length of the waveguide set the period of the oscillation. It had horrible timing jitter on its own, but we could make it very stable by driving it with a sinusoid oscillator.