path: root/lume/src/blog/automuse.mdx

---
title: "Automuse: A System for Generating Fiction Novels"
date: 2023-04-15
tags:
  - cursed
  - GPT4
  - Plotto
  - writing
---

<XeblogConv name="Cadey" mood="coffee">
  If you know of any AI/Machine Learning scientific journals or academic
  conferences, I have an academic paper version of this post available
  [here](https://cdn.xeiaso.net/file/christine-static/blog/2023/automuse/main.pdf).
  I would love to submit this paper to be formally published, but I'm not sure
  where to send it.
</XeblogConv>

## Abstract

A novel approach to generating fiction novels using a combination of
Plotto, a system of plot formulas, and GPT-4, a state-of-the-art
language model is presented. An eBook publication pipeline that
automates the process of creating and formatting eBooks from the
generated text is also described. The aim is to explore the potential
and limitations of using artificial intelligence for creative writing,
as well as to provide a tool for amusement and experimentation.

For an example of what Automuse output looks like, see the premere
novel [Network Stranded](https://xena.greedo.xeserv.us/books/network-stranded.html).

## Introduction

Modern advancements in large language models such as GPT-4 present
many opportunities when used creatively. There have been a few
attempts at doing this such as [Echoes of
Atlantis](https://medium.com/@chiaracoetzee/generating-a-full-length-work-of-fiction-with-gpt-4-4052cfeddef3)
which used the ChatGPT web UI to synthesize prose to fill a novel, but
none of these options were sufficiently automated for the author's
tastes.

The goal of Automuse is to be able to fabricate these novels in
minutes with little or no human intervention. As such, existing
processes and published prior works were insufficient, requiring
Automuse to be created.

Automuse is distributed as [a GitHub
repository](https://github.com/Xe/automuse) for anyone to download or
attempt to use.

### Motivation

The author discovered [Plotto](https://garykac.github.io/plotto/), a
kind of algebra for generating the overall plot structure of pulp
novels. This was written by William Cook, a man affectionately known
as "the man who deforested Canada", who had an impressive publishing
record at up to one entire novel written every week.

The authors wanted to find out if such a publishing pace could be met
using the ChatGPT API. After experimentation and repair of
[plottoriffic](https://github.com/justinoverton/plottoriffic), a
Node.js package to implement Plotto's rule evaluation engine, Automuse
was created. Automuse wraps the following tools:

- Plottoriffic to generate the overall premise of a story and to name
  the main dramatis personae.
- The ChatGPT API to generate novel summary information, chapter
  summary information, and the prose of the novel.
- Stable Diffusion to generate cover art for publication.
- Pandoc to take generated prose and stitch it together into an eBook.

## Results

According to the [National November Writing Month
rules](https://nanowrimo.zendesk.com/hc/en-us/articles/4403958068884-Why-50-000-words-And-how-do-you-define-novel-),
the works of Automuse count as "novels". In testing, the program has
been able to produce works of over 50,000 words (usually by a margin
of 5-10 percent). The outputs of the program have been described as
"hilarious", "partially nonsensical", and overall they have left
readers wanting more somehow.

The authors of this paper consider this to be a success, though they
note that future research is required to ascertain as to why readers
have an affinity towards the AI generated content.

## Methodology

When writing novels, generally you start by creating the premise of a
novel, the major actors and their functions, the overall motivations,
and the end result of the story. Plotto is a system that helps you do
all of this by following a series of rules to pick a core conflict and
then flesh things out with details. As an example, here is the core
plot summary that Potto created for Network Stranded:

> Enterprise / Misfortune: Meeting with Misfortune and Being Cast Away
> in a Primitive, Isolated, and Savage Environment
>
> A Lawless Person
>
> Ismael takes a sea voyage in the hope of recovering aer health
> Ismael, taking a sea voyage, is shipwrecked and cast away on a
> desert island
>
> Ismael, of gentle birth and breeding, is isolated in a primitive,
> uninhabited wilderness, and compelled to battle with Nature for aer
> very existence
>
> Ismael, without food or water, is adrift in a small boat at sea
>
> Comes finally to the blank wall of enigma.

<XeblogConv name="Cadey" mood="coffee">
  It's worth noting that Plotto is very much a product of its time. Plotto was
  written in the late 1920's and as such the information it generates is very
  dated and can sometimes generate things that are seen as problematic in modern
  sensibilities. Luckily, ChatGPT seems to sand away this roughness and is able
  to fabricate a better premise. All the pronouns are replaced with ae/aer
  because that is a decision that the creator of Plottoriffic made at some
  point.
</XeblogConv>

From this description, ChatGPT is used to create a plot summary for
the novel. These summaries look like this:

> After a disastrous turn of events, software engineer, Mia, finds
> herself stranded on a deserted island with no communication to the
> outside world. Mia uses her knowledge of peer to peer networks to
> create a makeshift communication system with other stranded
> individuals around the world, all connected by the same network.
> Together, they navigate survival and search for a way back to
> civilization while facing challenges posed by the island.

<XeblogConv name="Mara" mood="hacker">
  Normally GPT "hallucinations" (or when the model generates grammatically valid
  nonsense) are seen as an impediment or something to avoid. In this case,
  hallucinations are actually useful, as the summaries from Plotto are rarely
  enough information to fabricate believable stories. It was intended as an aid
  to the writing process, not a replacement for it.
</XeblogConv>

In the same prompt, ChatGPT also creates a list of chapters for the
novel with a high level summary of the events that happen in them.
Here is the chapter list for Network Stranded:

- Disaster Strikes - Mia's company experiences a catastrophic network
  failure leading to her being stranded on an island.
- Stranded - Mia wakes up on a deserted island with limited supplies
  and no way to communicate.
- Building Connections - Mia develops a peer to peer network to
  connect with other people stranded around the world.
- Challenges of Survival - Mia and the other stranded individuals must
  navigate the hardships of surviving on the island.
- Exploration - Mia and a small group of stranded people head out to
  explore the island.
- Uncovering Secrets - During their exploration, Mia and the group
  discover hidden secrets about the island.
- Frayed Relationships - As resources begin to dwindle, tensions rise
  among the stranded survivors.
- Hopeful Discoveries - Mia receives a signal on her makeshift
  communication system, offering hope for rescue.
- Setbacks - Mia experiences a crushing setback in her plans for rescue.
- Moving Forward - Mia refuses to give up and formulates a new plan
  for rescue.
- Unexpected Allies - Mia and the other stranded survivors meet
  another group of people on the island who agree to help with their
  rescue.
- Facing Obstacles - Mia and the combined group must face obstacles
  and dangers as they try to implement their rescue plan.
- Breaking Through - After a grueling journey and setbacks, the
  survivors finally make a breakthrough in their rescue efforts.
- Homecoming - Mia and the other survivors return to civilization and
  adjust to life back in society.
- The Aftermath - Mia reflects on her experiences and the impact of
  the peer to peer network on their survival and rescue.

These chapter names and descriptions are fed into ChatGPT with the
novel summary to create a list of scenes with major events in them.
Here is the list of scenes for Chapter 1: "Disaster Strikes":

- Mia frantically tries to contact someone for help, but her phone and
  computer are dead. She decides to go outside to search for a signal,
  but realizes she's on a deserted island.
- As Mia tries to collect herself, she meets Ismael, who is also
  stranded on the island. They introduce themselves and discuss possible
  ways to survive.
- Mia remembers her knowledge about peer to peer networks and
  brainstorms a plan to create a makeshift communication system with
  other stranded individuals around the world, all connected by the
  same network.
- Mia and Ismael team up to scavenge for resources and build the
  communication system. They search for anything that could be used to
  amplify the signal, such as metal objects and wires.
- They encounter a danger while searching for materials: venomous
  snakes. Mia and Ismael must use their survival skills to avoid
  getting bitten.
- As the sun sets, Mia and Ismael finalize the communication system
  and connect with other stranded individuals on the network. They
  share their stories and discuss possible ways to get back home.

These scene descriptions are fed into ChatGPT to generate plausible
prose to describe the novel. The main innovation of this part is that
ChatGPT is few-shot primed to continue each scene after initial
writing. If ChatGPT emitted a scene such as:

> As the helicopter landed, Ismael saw the first human beings he had
> seen in days. They jumped down from the helicopter, and looked at
> him with a mix of pity and relief. Ismael couldn't believe it – he
> was finally going home.
>
> As the helicopter took off, Ismael looked back at the deserted
> island, knowing that he had survived against all odds, but also
> knowing that he would never forget the terror and hopelessness that
> had left its mark on him forever.

The next prompt would be primed with the last paragraph, allowing
ChatGPT to continue writing the story in a plausible manner. This does
not maintain context or event contiunity, however the authors consider
this to be a feature. When the authors have access to the variant of
GPT-4 with an expanded context window, they plan to use this to
generate more detailed scenes.

## Known Issues

Automuse is known to have a number of implementation problems that may
hinder efforts to use it in a productive manner. These include, but
are not limited to the following:

- Automuse uses GPT-3.5 to generate text. This has a number of
  problems and is overall unsuitable for making text that humans find
  aesthetically pleasing.
- Automuse uses Plotto as a source of plot generation. Plotto was
  created in 1928 and reflects many stereotypes of its time. Careful
  filtering of Plotto summaries is required to avoid repeating harmful
  cultural and social biases.
- Automuse does not maintain a context window for major events that
  occur during prose generation. This can create situations where
  events happen and then un-happen. This can be confusing for readers.

### Potential Industry Effects

According to Dan Olson's documentary about the predatory ghostwriting
industry named [Contrepreneurs: The Mikkensen
Twins](https://youtu.be/biYciU1uiUw), the average pay rate for a
ghostwriter for The Urban Writers can get as low as USD$0.005 per
word. Given that Automuse spends about $0.20 to write about 50,000
words using GPT-3.5, this makes Automuse a significant cost reduction
in the process for creating pulp novels, or about 1,250 times cheaper
than hiring a human to perform the same job.

This would make Automuse an incredibly cost-effective solution for
churning out novels at an industrial scale. With a total unit
development cost of $0.35 (including additional costs for cover
design with Stable Diffusion, etc.), this could displace the lower end
of the human-authored creative writing profession by a significant
margin.

However, the quality of novels generated by Automuse is questionable
at best. It falters and stubles with complicated contexts that haven't
been written before. At one point in Network Stranded, the protagonist
shares a secret to another character and that secret is never revealed
to the reader.

It is worth noting that the conditions that writers for groups like
The Urban Writers are absolutely miserable. If this technology manages
to displace them, this may be a blessing in disguise. The conditions
put upon writers to meet quotas and deadlines are unimaginably strict.
This technology could act as a means of liberation for people forced
to endure these harsh conditions, allowing them to pursue other
ventures that may be better uses of their time and skills.

But, this would potentially funnel income away from them. In our
capitalist society, income is required in order to afford basic
necessities such as food, lodging, and clothing. This presents an
ethical challenge that is beyond the scope of Automuse to fix.

If an Automuse novel manages to generate more than $1 of income, this
will represent a net profit. More sales means that there is more
profit potential, as novel generation costs are fixed upon synthesis
of the prose. This program is known to use a very small amount of
resources, it is concievable that a system could be set up on a very
cheap ($50 or less) development board and then automatically create
novels on a weekly cadence for less than a total cost of $5 per month.

## Conclusion

Automuse is a promising solution for people looking to experiment with
the use of large language models such as GPT-3.5 or GPT-4 to generate
fiction prose. Automuse's source code and selected outputs are made
freely available to the public for inspection and inspiration of
future downstream projects. The authors of this paper hope that
Automuse is entertaining and encourage readers to engage with the
novel [Network
Stranded](https://xena.greedo.xeserv.us/books/network-stranded.html)
as an example of Automuse's capabilities. There is a promising future
ahead.