Writing
Notes on building AI for long-form fiction.
Where we let the model drive, and where we refuse to
We build two AI systems for fiction with opposite philosophies. One is a free-roaming agent. The other refuses to let a model make its own decisions. Here is why the difference is the whole point.
The mistakes an LLM will never catch itself making
A language model translates fluently and confidently, even when it is wrong. Here are the errors it cannot catch on its own, and why we check its work with rules instead of more model.
Keeping a character consistent 500 chapters later
In a long serial, a fact from chapter 3 has to still hold in chapter 480, and it has to be the version that was true back then. That is much harder than it sounds.