After reading more, it seems the novelty is in the two specific schemes that use these techniques to store data with high probability into tables of extremely high load factors with bounded pointer sizes, which are more complex than simple array indices due to having to resolve which fallback table is storing the key-value pair.
One scheme proves tiny pointer size bounds for fixed length tiny pointers. The other proves bounds for variable length pointers.
These are real RNNs, they still depend upon the prior hidden state, it’s just that the gating does not. The basic RNN equation can be parallelized with parallel prefix scan algorithms.
"Take a look at what happened", then three initial shots, "Get down! Get down! Get down!", followed by four rapid shots, all likely from the shooter. Trump ducks after the third initial shot. They sound like "crack---boom" where the "crack" is the bullet passing the microphone and the "boom" is the gunfire arriving later due to speed of sound at distance from the shooter.
Immediately after, there is a three-round burst of automatic fire, then several seconds later there is a single silenced shot ("kill confirmed") all likely secret service.
I'd have to look at the audio waveform to distinguish whether the four rapid shots were the same shooter or return fire from secret service.
It seems like the first 3 and then next 5 (not 4) shots were from the shooter, and there was a single sniper shot returned by secret service immediately after (some echo made it sound like a 3 round burst in the first video), with the final follow up confirmation shot several seconds later.
I think the best implementation (without eye-tracking) would not motion-blur the entire scene when the player moves themselves (ego-motion-blur) but would blur moving objects. However, if the player's view motion happens to coincide with the motion of objects in the scene, for instance if the player is tracking the object (maybe another player so they can shoot it in an FPS) then the object-motion-blur should be reduced according to the degree of tracking. I'm sure some game engines already do this.
The point is that ego-motion-blur is dumb. (It can also make players motion-sick.)
You say it yourself: "your brain compensates for saccades by censoring input for a while. And it has built-in motion stabilization which allows you to read a sign while walking down the street."