re: dissolving into compilation, I think the machine/human separation has been at work for some time. Modern languages (e.g., rust, swift) are already pioneering tracking aspects like effects, lifetimes, regions, etc. and then using whole-program optimization at compile- and link-time, largely based on intermediate languages like LLM IR/SIL, which are surfaced as user-visible features when they compose well with other user-visible language features. LLM training on these languages makes them suitable for generative AI; I doubt LLM's could pick up some new language, particularly if it weren't analogous to existing ones.
For what it's worth, the last time a language exploded on to the scene so quickly that people were picking it up commercially at a large scale before there was any sort of code base that LLMs could train on was Java. (Yes, there were no LLMs at the time, but I can back-project that as a measure OK.) And that was a money-infused attempt by Sun to buy their way into dominance on the Internet. It ultimately worked for Java, but not so much for Sun.
Any new language, of any kind, optimized for LLMs or not, is going to intrinsically take multiple LLM training cycles to grow anyhow. Net-net I would still expect AI adoption to accelerate those things and at least make it easier for hobbyiests to play with at some useful scale and get more feedback faster. Of course they'll also face increased competition from the other languages riding the same waves, which is possibly the bigger problem for anyone thinking of doing this.