The natural threshold to make a book (not a good book, just something which can be called “book”) is roughly “lots of text organized in readable sentences which are somehow related”, exactly what LLMs are good at.

The natural threshold to make a computer program is roughly “runs and does what it says it does”, and beyond very basic computer programs (Hello World level), LLMs by themselves and with no human adjustment of their output can’t even compile, much less run and work as expected.

One can use an LLM to make millions of lines of “code” in a very short time and it still won’t be a “computer program” if it doesn’t compile or doesn’t do what it’s supposed to do when it runs.

So since that threshold to be a “computer program” filters out things at a level that’s well beyond the mere textual arrangement (which is what LLMs are good at) that “book” or “story” require, it makes total sense that “book” and “story” submissions are exploding in the LLM age but not computer programs.

Then on top of this and as pointed out in the article, it turns out that even domain experts (who do know how to adjust the LLM output to make it into something that compiles and runs doing roughly what it’s supposed to do) aren’t in fact any faster at making programs that way than they are at just doing the whole thing themselves without using LLMs (probably because the “checking and adjusting” time eats up all the gains from “faster code production”, IMHO).

unreadable garbage you couldn’t publish doesn’t count, genies aren’t real