This is a mirror of my article at https://stroke.so/blog/ai-and-creativity written for Stroke blog.
Have you ever tried using AI for tasks that require creativity, initiative, or vision? When formulating such a task, you don’t have a clear understanding of what you really want to see. But once you see it, you instantly know — this is it, or it’s not.
If you’ve ever tried, you know: you can ask an AI to write a text for you—you’ll get something no one wants to read. You can ask it to slightly refine your text — you’ll get something no one wants to read. Weird, huh? It almost never satisfies your sense of what’s okay.
At the same time, AI is getting better and better at working with codebase and logic operations, and it handles such requests quite well. In these tasks, AI is effective due to its ability to follow structured rules.
Every programming language has syntax rules, every popular framework has extensive documentation, and there are widely accepted best practices for naming files, functions, classes, and so on. Combining these provides clear boundaries of what we expect from AI as response (predicted output). Moreover, the presence of clear criteria for what constitutes a predicted output makes it easy to train models through backpropagation in such fields, bringing their performance to the top level.
Definition of creativity
Things are entirely different when it comes to creative tasks with AI. Creative tasks don’t have structured rules, documentation, or widely accepted best practices. There’s no concept of normality or a right answer for artworks. Each time, both form and essence are entirely a reflection of the creator’s worldview. And the form is no less important and unique than the essence.
Do you see any patterns or best practices in these paintings?
In creativity, the author’s inner worldview is reflected. This worldview is a perception, created and shaped by a cocktail of numerous events that have occurred in their life. A work of art is an imprint of this perception at the moment of the time when the work was created.
Anil R. Doshi, Assistant Professor at UCL School of Management, and Oliver Hauser, an economist at the University of Exeter, made excellent research on this:
Generative AI enhances individual creativity but reduces the collective diversity of novel content
Here are some quotes:
Having access to generative AI effectively equalizes the evaluations of stories, removing any disadvantage or advantage based on the writers’ inherent creativity.
If the publishing (and self-publishing) industry were to embrace more generative AI-inspired stories, our findings suggest that the produced stories would become less unique in aggregate and more similar to each other.
In other words, the chatbot made low-creative people more creative, but it made the whole group that had AI help less creative because it doesn’t have any diversity.
Okay, now we understand where the problem lies. The lack of a unique perception filter and criteria of predicted output makes AI practically useless for creative tasks. This is true, of course, if your goal is to create a result above the average.
As a result, we see that most people use AI solely for spell-checking, translation, or minor editing.
But can AI be integrated into creativity in a way that produces not mediocre results, but something truly worthy and unique? Let’s dive in and find the answer.
AI Spirits
What if we had a tool that let us to create our digital reflection? Our worldview, our thoughts, our reasoning, our perfection, our writing style — all of this would be embedded in the very core of the AI model (its weights), and you wouldn’t need to struggle with creating an input layer workaround just to get the AI to somewhat understand and adopt your style.
The value of such an LLM model might be low for a broad audience. But for you and your work its value would be immense. When asking such an LLM to help create something, you would recognize yourself in the output. Any such material would be, if not the final version, an excellent starting point for creating something outstanding.
Steve Jobs, the founder of Apple, thought about such a tool 40 years ago. He dreamed of the possibility of creating a digital imprint of a person and their worldview, allowing interaction with it as if that person were sitting across from you. Here’s a video where he talks about it:
I came across this video after writing the initial draft of this article. This is it, I thought.
And today, we have all the resources needed to create and implement such tools on a large scale. We have access to many great pre-trained large models, like Llama-3, MPT-7B, MosaicGPT, and others, that could be serve as a foundation for the fine-tuning process on your thoughts, works, artworks, or other personal materials and data.
The cost to train a fine-tuned LLM varies, but it’s not super expensive. It could range from $10 to $5,000, depending on the amount of data, personality level, and other factors.
This model would retain the encyclopedic knowledge of general large models, like ChatGPT / Grok, but its style, values, and decision-making evaluation system would reflect you more than any some average, politically correct consensus.
Thus, the issue of a unique perception filter would be resolved, and the output would be far more personal and yours than some generic average.
This LLM model could easily reside locally on your device, encrypted, and use cloud computing power to process your requests when needed.
The ability to create such personalized LLM models is one of the features of Stroke. And this is what really excites me, and that’s one of the reasons I work on Stroke.
Imagine that even if you can’t clearly formalize your request, you’ll still get workable and usable output. And it will be unique — just as creativity and our world should be. That sounds pretty great and fun.
Context Windows. Creativity as a system tree
Every AI model has a maximum context window size (the maximum size of input it can process). When this limit is reached, it starts truncating incoming data and losing track of the conversation.
As shown in the image, during a long conversation with AI, the context window can fill up rather quickly, since the input layer includes the entire conversation history every time.
But in practice, problems arise even before reaching the context window’s limit (200k tokens in the image). The fuller the context window, the more likely the AI model is to hallucinate, often getting confused by instructions that may contradict each other from message to message. You can read more about context windows on Anthropic’s website.
From a creative perspective i’m a writer. And it’s easier for me to reason through the example of writing, as I understand its creative process better.
Any large book is essentially a complex, well-thought-out system. This system consists of various interconnected components. For example, we can identify the following components in any large book:
- Main characters
- Their backstory
- Their personality
- Their behavior and speech style
- Their true goals
- …
- Supporting characters
- Their backstory
- Their personality
- Their behavior and speech style
- Their true goals
- …
- The book’s scene
- The book’s time period
- Events of the period described in the book
- And so on
If we go to write chapter after chapter with AI straightforward, we’re guaranteed to get a result no one will want to read.
But if we are consistent and work with AI initially on each component separately, and then weave those prepared components into the plot, we can create decent material.
What makes a book unique is how these components are interconnected and their interplay. If we have a general picture in our mind, the end-result of such consistent work can be a book with a complex and coherent world.
I’ve personally read examples of such works, and they are excellent, engaging books.
Conclusion
It’s our personal and unique worldview perception and inner reflection make a text interesting to read and convert painting into true artwork. This imprint, which we leave on everything we touch, creates true value for us and other people.
Now we have open and accessible large LLM models as a foundation. Until recently, the costs of creating a personalized LLM model was too high for practical implementation of the described idea. Now, those costs are no longer prohibitive. And we have the opportunity to leverage all the capabilities of LLMs without losing the personality and diversity essential to the creativity field.