I've always felt that the ideal would be to somehow create many different tokenizations for different use cases. And allow them to sometimes build on each other. Like a lot of domain-specific languages.
Maybe there could be something like a mixture-of-experts but with a thousand experts and each has its own tokenization.
Fortunately I don't actually understand this stuff, so I am going to go ahead and congratulate myself on my brilliant ideas and let the geniuses work out the details. :P