LLMs are also quite good at outputting useful code. Again, it’s because code is written using a programming language - and has an even better characteristic because it follows a grammar which makes programs un-ambiguous.
The “intelligence” we see when we use LLM-assist when coding is basically asking it to “continue this curve starting from the already-drawn curve so far”.
It’s an emergent property because it works well and we haven’t specifically written the model’s code to “know C++” or “know JavaScript” or “be a software development simulator”