I also wonder whether brainfuck (ie turing machine like) coding would be a more efficient interface to LLMs

For those who want to try it, there’s always the https://raku.org module…

  use Polyglot::Brainfuck;
    
    bf hi { 
        ++++++++[>++++[>++>+++>+++>+<<<<-]>
        +>+>->>+[<]<-]>>.>---.+++++++..+++.
        >>.<-.<.+++.------.--------.>>+.>++. 
    }
    
    say hi.decode; # Hello World!
    
    bf plus-two {
        ,++.
    }
    
    say plus-two(buf8.new: 40).head; # 42