> There's just too much stuff you have to codify, and basing such a system on tokens is inherently limiting.

As a complete amateur who works in embedded: I imagine the restriction to a linear, ordered input stream is fundamentally limiting as well, even with the use of attention layers.