Hi @rhyolight as @floybix and @mrcslws say, this is because the RDSE is stateful. The way it works is that each bucket's encoding differs from its neighbours by exactly one bit. If you have a new input value greater than your current biggest bucket, you need to add buckets one by one (changing one bit "at random" each time) until you have a bucket for the new value. The same applies in the negative direction.
This means that a naive implementation would have an encoding which depends on the entire and exact sequence of inputs seen, and the which is a really bad idea. Even worse, if you use the global random number generator, your encoding will change depending on who else is pulling numbers out of the RNG. The only option is to ensure that the bucket creation process is deterministic, and that means that you have to use a private RNG and a starting "center", and you must build out your buckets in both positive and negative directions whenever you need to extend the range.
Using your example in the video, let's say you start off with a center of 500 and a resolution of 0.25. If the first input is 600, you'll have to grow 400 buckets up from 500 to 600, and 400 down from 500 to 400. So you'll end up with 800 (or 801) buckets, many of which might never be used (in your example all the non-integer buckets are waste).
NuPIC does this by using its
offset parameter to identify the center of the RDSE, and it holds a private stateful RNG in its
random field. But NuPIC appears to have a bug because it doesn't do the growing in both directions (see the createBucket() method. This means that the encoding will differ depending on whether you grow upwards or downwards when you need a new bucket.
For a very early study of how this works, see this page.