I'm not so sure. A decision tree could do this, it would just need to traverse the syntax and make a decision at each token. I think what has changed is that the decision tree would not have been seen as impressive because we knew how they worked. A lot of people seem to think LLMs are magic so that makes it more impressive to them.
LOL, you really have no clue, do you :DDDD decision tree writing a working compiler, sure :DDDDDDD I wonder why none of AI researchers ever demonstrated anything like that, a decision tree writing a working C compiler, I mean it'd easily be a Nature paper :DDDDD But of course I bet YOU could easily do that over weekend if you just put your mind to it, you just have more important things on you, right? :DDDDDDDD
Even if all that did was "just transpile GCC code it was trained on into Rust" its hugely impressive, I mean just look at actual existing cutting edge transpilers, and the code they produce. Hint: they suck, and the code they make is usually unreadable even if it works.
The "let's rewrite everything into Rust for the hell of it" crowd I bet is already eyeing this thing. uutils people are doing exactly this scenario.
Also, how quick those goalposts move, eh? Seems only last year you were like "LLMs SUUUUUCK, they can't even figure out simple 10 line coding assignments reliably" today it's "LLMs SUUUUUUUCK a compiler they wrote from scratch generates less efficient code than a decades old project", wonder what your next year's goalposts will be?