I've read whole "stories" written by AI (in the low thousands of words).
They're usually pretty terrible. I can't imagine a novel's length of it.
That said, having played with the LLMs - they are super helpful for doing things like bouncing conversations off a "character", or describing a scene for me, or just being whacky idea generators.
I think it's similar to the stable-diffusion side of things, it's going to make some authors more productive, but it won't necessarily replace the need for an author. (whether that leads to more or less authors long term is a related, but tangential, question and influenced by a lot of non-ai society factors).
Beyond a few thousand words (32k tokens for GPT-4), it seems the beginning of the story will be outside of the range that can influence the output, though perhaps there are techniques to retain some memory, maybe you could first generate an outline and then fill in the chapters, scenes and so forth, so the limited memory becomes less of an issue.
The problem is less that the context runs out (although for a full novel, it does become a challenge, since 32k tokens is only about half a book).
The problem is more than the AI has no concept of actually telling a coherent story. It doesn't tell anything exciting or novel or cohesive - it just spits out similar literary sentences, and the more it spits out sentences that look like random novel content, the less cohesive the story gets as it matches more an more strongly on previous content that looks like random novel.
Honestly, it's less like a real novel, and more like a panorama of small chunks of all sorts of different stories - Mostly unrelated and uninteresting.
That's the impression I've been getting. AI lowered the barrier to entry of just generating a large body of text and self-publishing, but there are more subtle market dynamics at play here than people seem to realize.
AI lowered the barrier to entry of just generating a large body of text and self-publishing
Honestly, most of the generated stuff I've seen at length (as in, more than 1 or 2 page lengths of content) just devolves to gibberish. The words and sentences are syntactically correct, but mostly irrelevant or meaningless in the context of the story as a whole. It's like reading a collage of stories jammed together with no real artistry.
You can self publish that (people do) but it's not really much improved on the old scam of just publishing copy-pasta bullshit.
I think there can be space for LLMs to do things here, but right now they are absolutely not a replacement for an author with a story they want to tell.
They are more like a calculator: If you have a clear picture of a scene/character/setting, the LLM can flesh out a lot of options for you to pick and choose from to best match your intent. It does mundane work in support of your goal - it has no goal itself.
Comments
I've read whole "stories" written by AI (in the low thousands of words).
They're usually pretty terrible. I can't imagine a novel's length of it.
That said, having played with the LLMs - they are super helpful for doing things like bouncing conversations off a "character", or describing a scene for me, or just being whacky idea generators.
I think it's similar to the stable-diffusion side of things, it's going to make some authors more productive, but it won't necessarily replace the need for an author. (whether that leads to more or less authors long term is a related, but tangential, question and influenced by a lot of non-ai society factors).
Beyond a few thousand words (32k tokens for GPT-4), it seems the beginning of the story will be outside of the range that can influence the output, though perhaps there are techniques to retain some memory, maybe you could first generate an outline and then fill in the chapters, scenes and so forth, so the limited memory becomes less of an issue.
The problem is less that the context runs out (although for a full novel, it does become a challenge, since 32k tokens is only about half a book).
The problem is more than the AI has no concept of actually telling a coherent story. It doesn't tell anything exciting or novel or cohesive - it just spits out similar literary sentences, and the more it spits out sentences that look like random novel content, the less cohesive the story gets as it matches more an more strongly on previous content that looks like random novel.
Honestly, it's less like a real novel, and more like a panorama of small chunks of all sorts of different stories - Mostly unrelated and uninteresting.
That's the impression I've been getting. AI lowered the barrier to entry of just generating a large body of text and self-publishing, but there are more subtle market dynamics at play here than people seem to realize.
Honestly, most of the generated stuff I've seen at length (as in, more than 1 or 2 page lengths of content) just devolves to gibberish. The words and sentences are syntactically correct, but mostly irrelevant or meaningless in the context of the story as a whole. It's like reading a collage of stories jammed together with no real artistry.
You can self publish that (people do) but it's not really much improved on the old scam of just publishing copy-pasta bullshit.
I think there can be space for LLMs to do things here, but right now they are absolutely not a replacement for an author with a story they want to tell.
They are more like a calculator: If you have a clear picture of a scene/character/setting, the LLM can flesh out a lot of options for you to pick and choose from to best match your intent. It does mundane work in support of your goal - it has no goal itself.