Post #2205672

What's ethnographically interesting about vibe coding is that while vibe code proponents liken LLMs to the next abstraction up from compilers, nobody is passing around prompts. Everyone is passing around the extrusion created by said LLMs. It's like inventing compilers, but exclusively passing around the assembly code, and then using decompilers to figure out what the assembly code does (bearing in mind that compilation is often a lossy transformation). A huge advantage of high level languages is how much easier they are to reason about compared to assembly or even (yikes!) binary files. Of course, you can still "compile" high-level languages to go from more readable -> less readable, see what happens whenever you minify JavaScript. There is also a reason why we don't work directly with minified JavaScript. Essentially, source code becomes a sort of "ground truth" where they can (deterministically) run a computer, and where programmers can co-reason about it, using the idioms and affordances of the high-level language and agreed-upon code style. Now LLMs don't exactly produce readable code. So my best guess if that the only way to work with them is to go from human idea -> LLM (sender's side) -> unreadable code -> LLM (receiver's side) -> human idea. They effectively create code where there is no shared idea of what it does. I struggle to imagine a world in which this does not cause disaster at some point. Perhaps we are already seeing such disasters