Worldbuilding 202: AI and Worldbuilding
Greetings and sferics! ...I don’t want a storm now! NOOO!!! Welcome to my blog, I am your favourite alien Limax, Vivian! Today, I will be talking about AI; to be more precise, the text generating kind, and a wee bit on image generation. Please read my AI policy to see where I generally stand.
Modern text AI
AI has been around for a long time, and it has always been considered “the latest development” until the AI becomes so everyday that it loses the feeling of “AI” and is just another machine thing doing a crap job that makes lives easier. So I feel modern AI will fall into that category as well in due time, but how do ChatGPT and other language-based AIs, or large language modes (LLMs), inherently work today?
All modern AIs require enormous amounts of data to draw experience from. This is similar to how an actual brain works in that it takes lots of information and tries to construct a model of the world. Modern chat AIs have neural networks that are meant to mimic to a certain degree the behaviour of how nerves and brains work. The big difference is that a brain works to understand things, and AIs do not. I will return to this point.
On top of this, the system constructs a many-dimensional tensor. What’s a tensor?
A tensor is a mathematical object that can transform things mathematically.
The exact nature of the tensor is not relevant beyond that it mathematically works. The big thing with this tensor is that it has a number of words that it can look back on, including what you type in and what has come before, in order to create the illusion of it having an understanding of context, then it uses all that context to give a probability of what words come next. It either then picks the most probable or uses randomness to pick the next word. Then it runs that again, and again, and again, until it reaches a point where it is considered “finished.”
Some models pick at random as it generates, as I said above, and some are more deterministic, though they have a random number initially as an input in order to make certain that no two responses are identical. Generating this tensor used to determine probabilities of words is absolutely enormous in terms of data and even bigger in terms of how much training and data is required to build it. This has often been done through scrapping the internet for anything that is available, which today is backfiring as too much of what is collected is getting to be AI itself, but I will get back to scrapping and what it means in terms of theft.
The important thing to remember here is that it only creates statistical models of words and how they correlate with each other. An AI has no understanding of what anything means. AIs can further be trained on other factors based on humans and other things, but for now, we are ignoring those.
AI Creativity
Say after me, literally say after me:
AI HAS NO CREATIVITY
You need to understand this as a fundamental thing when it comes to using AI as a tool: it has no creativity. If you thought about what I said earlier, AIs build statistical models based on “context” to create the illusion of someone capable of understanding context and having prior knowledge.
This means it cannot come up with actual ideas or anything creative. Anyone who has used it, and I have used ChatGPT extensively, sees that it has its own very unique style. This style arises because the aforementioned tensor will only output the same probabilities and words over and over again, and it has, when scrapped from the internet, a very generic model of how the language is used.
The same comes in that it cannot make anything really new. After countless tests of various things and stuff I have been working on, when ChatGPT is asked to render something, it takes the most likely things that fit the prompt, and it ranges from the most boring and mundane to things so outrageous (due to me prompting it to be more out there) that I am flabbergasted at what it suggested. I am not kidding, it once suggested cannibalism for a festival…
AI hallucination
It all keeps coming back to that wretched tensor I spoke off before! While Wikipedia and many other information storages contain, well, information and knowledge, a modern AI does not contain any form of knowledge. Sure, the neural network and the tensor have been trained on a lot of information, but the tensor does not contain the actual information about, for example, the General Theory of Relativity. Yes, it has been trained on it, but it doesn’t know anything about it because it doesn’t contain knowledge.
Some are now trying, with online searches and linking to knowledge bases we have, such as Wikipedia, to train the chatbot to access knowledge, but that is still being run through the probabilistic tensor which means it can still run wild with what it says and be completely wrong.
The phenomena of an AI stating factually incorrect things with a level of certainty that is entirely unwarranted and almost no human or alien would do is called “hallucination.” The AI has learned that when it comes to prompts asking for facts and knowledge, real authorities speak in an authoritative manner, so it puts on the air of authority in the text and conjures up whatever its tensor is deeming a probable and likely outcome from it. That is not knowledge, that is called making shit up when a human does it.
AIs forget context
Another factor with AIs that is a huge nuisance and diminishes their usefulness is the fact that context is forgotten. You might sit there talking to it and having it do this and that and then want to go back to something that is 1100 words earlier and ask it about stuff there and… the AI says it has no knowledge of that stuff.
This is because AI has no memory. As I stated before, it has no repository for knowledge beyond the probabilistic tensor for words. So every time it is about to answer, it has to use the previous text to get the context. But unlike humans who have a short term memory where the brain stores relevant context bits for when you are 800 words later, modern LLMs don’t and can’t do that. They only have a finite number of words back to look for context which means it is fundamentally limited. Combine that with the fact that it hallucinates, and you can get some hilariously wrong answers.
Scrapping for AI
Of course, when dealing with AI, we cannot avoid the elephant in the mainframe: the acquisition of data for the AI. It has been scrapped from the internet and in virtually all cases, if not all cases, without people's consent. Consent is a major issue, especially in what I will call “less well-trained” AIs. With that, I mean that they are not trained to actively avoid their own source data as to not infringe on copyright. We have seen several images where the copyright text on the image is still present.
That is obviously bad, but when it comes to generated content that is from well-trained AI, it is, as far as I am aware, legally questionable in many ways. I have heard from some that text and such generated from AIs cannot themselves be copyrighted because it is derived, or whatever the argument is. I am not a legal expert.
While laws are trying to figure out where things land and people settle on the morality of using sewing machines, I mean AI, this does cause an issue when using AI. Namely, it is wholly dependent on what it has acquired. For text generation, it is generally not an issue as people write a lot of things all over the place. For image generation, the issue is much clearer. I had it try to generate a 5-eyed creature because I like toying around with it… but animals are bilateral, and no one alive has a middle eye (that is clearly visible), so it went 4 eyes, 6 eyes, 8 eyes, 2 eyes…1 eye (it got the one, I think, because of cyclops), but it just couldn’t do 5 eyes as it had nothing in its training set that bore any resemblance to it, so it panicked, threw whatever it had at the problem, and started sobbing. You can see the result here.
Very fine 6 legs and 5 eyes, don’t you agree?
How to use it
Now one thing I disagreed about with Anne, and then realised I am a teacher, and she is right, is that people lock onto ideas and focus too much on them once they have heard them. And she is absolutely right, people do that; some people can avoid this, but if you are new to creativity, you are likely not one of those.
With this in mind, the first thing to start with AI is to have your own ideas or a framework for what you likely want. For example, when I use it in some instances to get ideas for my species, I have general rules for how I want my alien species: what kingdom they belong to, some ideas for the basic look, and so on.
Then, the next recommendation I give is to not have it tell you exactly what something should be. Instead say things like “Give me 5 ideas for X”. This partially helps the aforementioned locking-on-ideas problem because you now have 5 to consider. To continue my example, for the Raixhe, I had the appearance of the vines for primales and ultimales done, and it felt right to me, but the midmales I was a bit uncertain on. So I fed ChatGPT what I did have and asked for a bunch of suggestions for midmales’ vines. I got a bunch of suggestions that I didn’t like, and after a few tries, it suggested changing colour with the seasons, and I liked that. I had it then suggest colours for the seasons, and I seriously liked nothing there, so I took the suggestions I got and fused them together into one that I liked.
Notice what I do? I let it use its ability to generate ideas, and then I take suggestions and modify them to fit what I want. I don’t let it mindlessly generate things and then walk with it. This is why having some initial ideas and feelings for what you want is important. It allows you to sort out those ideas that don't sit right with you. Sometimes you can even take a random off hand remark and run with it. I did it with the Raixher again; I can’t remember what ChatGPT said, but it was something about the nose and smelling that made me go “What the?” and then I let my imagination run with it, and they got their different olfactory system!
I want to place emphasis once again on how the process I have used is about using ChatGPT, or whatever LLM strikes your fancy, to inspire me, and then use my own imagination on top of it. If you do not keep this in mind and just run with the things it spits out, or worse, you don’t give your mind options before and during usage, you will fall victim to having your mind focused on one thing and be out of creativity.
As a final note on this, after a lot of trying, do not use image generating AIs. They are incredibly boring and fairly predictable. Beyond showy images, it is not that useful. You can give the most detailed descriptions possible, and it will fail spectacularly there. That is why when I use it, I use it for generic images where details currently do not matter.
AI & the future
This is my personal opinion and nothing else, but as I see it, AI usage is inevitable for both good and bad. We will see, and already are, people using it in academics to cheat and have entire “books” generated. Of course, that is all bad and should be discouraged. Same goes for images. I have seen some use it for book covers, and they are atrocious because AI isn’t that good yet, but they will continue to get better. As AI evolves, we have to just keep ahead of the curve and find ways to promote the useful elements and discourage the negative ones.
Summa Summarum
In summary, AI is useful and will become more so, but it takes thought in how you use it and learning how they work in order to get the most use out of it. It can also be used for a lot of bad, and as with every time when there is a new technology wild west, the bad uses are the first and easiest uses people go for in order to make money.
But I feel that just because some people use kitchen knives to stab people to death, that doesn't mean we should hate on what will become a great tool. At the time of this post, AI and all these neural network systems that are used are still in their infancy, and you are trying to figure out how to properly train them to do their job. But like the first homo (early lineage of humans) who had to figure out not to gouge out their eye with their new knife, we have to figure out how to use this tool correctly and where all the lines go. But like a knife, it will become more present, more used, and a strong tool to use to enhance human work, and not replace it.
And I feel that two of the most important aspects of this are honesty and transparency. If you use it, be honest.
Here is a song in which Anne fed this blogpost into ChatGPT 3.5 and asked it to create lyrics, then put the lyrics into Suno in order to create a song with this associated video.
Blogpost image was also AI generated with ChatGPT 4.0 and edited by humans.
Want to dive into a discussion about Stellima or the art of writing on Discord? We’d love to have you! And if you have any topics you struggle with or that you would like to suggest for a future blogpost, we’re open to suggestions!
Interested in supporting our work? Join our Patreon and become a part of Stellima as a citizen of Mjatreonn! Or would you like to give us some caffeine to fuel our writing? Consider buying us a coffee at Ko-fi! Every contribution inspires our creativity and keeps us going. Thank you for your support!
Copyright ©️ 2024 Vivian Sayan. Original ideas belong to the respective authors. Generic concepts such as large language models and ChatGPT are copyrighted under Creative Commons with attribution, and any derivatives must also be Creative Commons. However, specific ideas such as the Raixher and all language or exact phrasing are individually copyrighted by the respective authors. Contact them for information on usage and questions if uncertain what falls under Creative Commons. We’re almost always happy to give permission. Please contact the authors through this website’s contact page.
We at Stellima value human creativity but are exploring ways AI can be ethically used. Please read our policy on AI and know that every word in the blog is written and edited by humans or aliens.
