I was thinking of the same yesterday... shower thoughts only for now... but my motivation was token savings....

For example, Mandarin is anywhere from 2x to 6x more information dense compared to English.

And I guess a language built especially for llm's can be far more. Of course, you will have to train them in that first.

I see the purpose of the intent and the reasoning files/languages, but it appears to me like this has already been solved by the hundreds of MCP servers and skills for code intelligence and agent memory out there.

Elon has mentioned that future versions of Grok will write binary code directly. Once that happens, programming languages won't matter much to an AI.