Well, there are roughly 15 million Comcast remotes that can control the RNG110 easily (Cable 1982).

Here's why the XMP protocol is difficult to learn, and probably why the Next Generation transmitter has difficulty.

Most IR signals are transmitted as binary code-- the signal is made up of 2 burst pairs which represent either zero and one. As an example, the NEC IR protocol uses a short burst of light (let's call the length 1 unit) followed by 1 unit of dark to represent 0. A one is represented by 1 unit of light and 3 units of dark.

The XMP protocol has uses 16 burst pairs which represent the 16 hexadecimal digits 0 through F. Zero is represented by 3 (much shorter) units of on, and 10 units of off. Each succeeding digit has 2 units more of off time, so digit F has 40 units of off time. This kind of coding ends up needing about 4 times less data, but it puts real stress on the timing accuracy of the remotes, especially in learning. The XMP receivers in the electronic components can correct for timing drift in a received signal, but the learning circuitry in a remote doesn't know which protocol it is getting. Also, in a learning situation, two remotes are used, and the timing inaccuracy of both contribute to errors.

There's another complication for learning remotes. Remotes typically try to characterize the IR signal as recognizable burst pairs, and they don't expect 16 genuinely different burst pairs. So even with a perfect time base, they may not be able to learn the signals.

I think the Nex Gen uses electrical noise from the remote's microprocessor to determine the temporal pattern of the RF signal it transmits. Probably it also need to classify the signals into recognizable burst pairs, and it may be programmed to reject signals which have 20 times more off time than on time.