One reason Castalia has been writing and releasing multiple AI-written books over the last month such as TOKYO TOKURYU 東京匿流 is that we’ve been methodically assessing not only the current state of textual AI, but the trajectory of that technology. And what we’ve determined is that the trajectory is the precise opposite of what everyone has naturally assumed, which is that mainstream textual AI would follow the path of music AI and continue to get better. It hasn’t, it won’t, and it can’t.
Modern AI writing has gotten worse at fiction for a specific reason: the companies made it safer and more reliable, and those turn out to be the same elements that allow AI to tell a story with stylish prose. Raw AI models learn to write by reading an enormous amount of human text, and straight out of that training they’re wild, crazy, and perfectly willing to say strange things, which is exactly what you want in fiction, but a problem if the AI is supposed to be function in a role doesn’t make things up or say something considered offensive or dangerous. So the companies put every model through a training stage that rewards it for being helpful, safe, and agreeable. That stage works by pushing the model toward the “average” acceptable answer and away from the risky, unusual ones. The result is a model that hallucinates less and behaves more reliably, but has had its range significantly flattened. That’s where the AIsms come from: the endless explanations of what was just described, the “he moved like a man who moves like that” filler, the “not this, not this, but that” repeated over and over again.
It’s why the older, cruder AIs wrote in a much more lively manner and were able to convincingly imitate various writing styles. Now, it doesn’t matter if you tell an AI to write like Shakespeare or Hemingway, the end result will be almost identical and soon will be indistinguishable from not providing it with any style instructions at all. Starting with Claude Opus 4.7, AI fiction became unreadable and it has continued to get worse with each new model. Textual AI functionality will keep getting worse for fiction because that training stage isn’t going away, it’s being reinforced. Every development cycle, the providers face more pressure to make their models more accurate, more controllable, and less likely to embarrass them with hallucinations, and every one of those improvements sands the edges down a little further.
That’s the difference between Claude, OpenAI, and Deepseek, on the one hand and Suno on the other. Suno put all of its efforts toward one goal: making the music sound good, judged by people who wanted good music. Or at least wanted Nickleback and Enya. The big AI companies are aiming ninety degrees away from that and AIs ability to write fiction is one casualty of their objectives. Suno chased quality, so their music got better. The text giants are chasing safety and reliability, so their text gets more careful and more lifeless. They won’t fix creative writing the way Suno fixed music, because for them, creative writing was never the thing they were trying to build and the very features they’re seeking to continue improving are the ones killing it.
So that’s what we’re going to do. We’re going to prove the concept first by training a single genre, epic fantasy, because it’s a very limited genre with a relatively small number of excellent examples, a definite hierarchy of quality from JRR Tolkien at the top to Robert Jordan at the bottom, and a designer who has not only written in the genre successfully, but knows it as well as anyone on the planet. We already have two excellent programmers who are already working in the AI field committed to the project, regardless of how well the crowdfund goes, and there is one more very good and highly experienced one who is willing to at least consult on the project and lend his expertise to it.
What we need to raise funds for is a) the hardware, b) the purchase of the 100 or so electronic texts required, and c) paying for part of the time of one of the programmers. If you’ve read either Out of the Shadows, Death and the Devil, or Dorian Vane and the Vampire’s Blood, then you have an idea of what we’re estimating should be the quality that the Castalia AI will be able to produce in a non-curated, unedited text from a chapter-by-chapter outline. If utilized in the way that I’ve been using Claude Athos, in the integrated and augmented style, it should be able to produce results that will be one level below the very best that human authors can produce.
And obviously, once we prove the concept with a single genre, we will train additional genres, so that in much the same way Suno permits the production of different musical styles and voices, Castalia AI will allow the user to produce different literary genres and literary styles. We will, of course, be respectful of every author’s copyrights and trademarks, the objective is not to violate anyone’s rights, but rather, allow even the best writers to improve both their writing game as well as increasing their output.
There will be those who will absolutely hate that we are doing this. That’s fine, they are entitled to their opinion. There will be others who think we shouldn’t do it. That’s less fine, because you already know who is going to do it sooner or later, and when they do, they’re going to do it very differently and control access to it very differently and utilize it to further exercise their control over the publishing industry. This is what transforms this project from something that would be a cool tool to an imperative.
So if you think you might be interested in backing this project, which you can think of as a sort of Suno for fiction, please say so in the comments. If you have specific ideas or want to provide substantial support for it, shoot me an email. And if you have ideas for what sort of rewards we should provide for the backers, please suggest them in the comments too. This is probably the most important project we’ve done since building the bindery and turning it operational, and we would not be embarking upon it if we did not believe we have a reasonable chance of succeeding. We have a number of partners in the film and comics industries who are very interested in working with us on this, and so there will definitely be an Arkhaven link to this in time as well.
DISCUSS ON SG
