I assume that only applies to natural language, not computer languages?
[–] needfish 1 hour ago
[–] 0x262d 5 hours ago
I think the key difference is that code is a tool used for a given purpose. It can be slop, or structured better or worse, but at the end of the day it exists to be executed, and its execution is either correct or not. Additionally, good code that does a given thing can only be expressed so many reasonable ways, probably. So LLM-generated code can be perfectly fine.
By contrast, human writing is about exchanging thoughts. Often those thoughts are about infinitely complicated and poorly understood systems, like other people, or emotions, or politics, so it's not a matter of exactly encoding something but attempting to describe it imperfectly. All the tiny little details encoded in how someone writes something can work to this end and are also a product of the writer's thought patterns and overall personality; this all conveys useful information to the reader. In other words, there is a much greater degree of fine detail that affects the quality of writing than of code, and LLMs suck at those fine details for most subjects that are interesting. On top of that, they don't have original thoughts.
It is impossible to reconstruct the intent/spec from the code. More often than not the intent is not clear in the code, and one has to track down the original author to check their context and decision at the time.
As the saying goes, naming things well is a hard problem. It's kinda mystical in a way, where only by knowing an entity's true name will one understand it.
However if it's communicated between agents one may care even less.