…Man created AI and it was good. Very good. Very very very good. By now, you’ve probably used ChatGPT, Gemini (aka Bard), Claude or some other variant to help with a cover letter, a resignation, or resume.
You can use some AIs to generate images from a textual prompt and / or other images. Much like computing back in the 70s and 80s, we are at the forefront of a new iteration of technology - a culmination of everything that has come before. Anything that can be translated to 1s and 0s can be used by AI to discern something that hasn’t existed before.
In that vain, I am curious to see AI does with scripture. While this may be the beginning of a bigger project, for now I just wanted to see how Claude would imagine in SVG the first chapter of Genesis. My initial prompt was to render to an image the text of Genesis Chapter 1 which I supplied into the prompt. Claude politely told me it can’t render images but it could do vector based graphics, effectively SVG.
The first SVG kind of surprised me. Super simple.
Again, super simple. Yet it is well organized and easy to interpret. If you showed the rendering to someone who had some familiarity with Biblical stories, I am fairly certain they would relate it to the story of Genesis.
I inquired to Claude if it did a second rendering, would it be the same? Claude said it would not. As someone who develops applications for a living, I was somewhat surprised. In my natural state as a developer, I expect code to work the same way each time it’s run. Generally speaking, software needs to have a predictable outcome for it to have a predictable use.
Claude said there were a few reasons why the renders are different each time (verbatim from Claude):
-
Non-deterministic choices: As an AI, I make many small decisions when creating these images. These decisions are somewhat random within certain parameters. For example:
- Exact coordinates for elements
- Specific sizes of shapes
- Color choices within a reasonable palette
- Which artistic elements to include or exclude
-
No visual memory: I don’t actually “remember” or reference the previous image when creating a new one. Each time, I start fresh, interpreting the prompt anew.
-
Multiple valid interpretations: The creation story can be visualized in many different ways. Sometimes I might focus on a linear timeline, other times on a circular layout. One version might emphasize the separation of elements, another might focus on the emergence of life.
-
Creative problem-solving: Each time I create an SVG, I’m solving the problem of “how to represent this story visually” from scratch. This leads to different solutions each time.
Being new to AI, these are concepts I had no idea about until this little experiment. Claude has no visual memory so on it’s own it can’t “think” to use what it has done as input to what it is doing. Claude is randomizing certain things like colors and positioning of objects within what it calls reasonable parameters. If I had to akin this to a human behavior, I’d consider it a mood. Then there is the “creative problem-solving” and “multiple valid interpretations” - is this what we would consider “thinking?”
For the moment, I am just going to call it fascinating.
When I asked Claude if the rendering would be different, it gave me the response above and then proceeded to create another version to show me how it would be different.
We can see that there are some similarities with the first SVG, but this one feels different and seems to be more understandable. That said the waters are above the land, and the lights are detached from Day 4. I think the first rendering is a better telling of the story, but the second render is more interesting to look at.
As part of the second rendering, Claude added:
This variability is actually similar to how human artists work. If you asked two artists to illustrate the same story, or even the same artist to draw it twice, you’d get different results each time. The core elements would be the same, but the specific artistic choices would vary.
This idea is like going to an art class where everyone is tasked with painting the same vase of flowers. Every painting will be different because each person is making decisions based on their experiences and interpretions. The main difference is that a person would likely be able to provide a reason for their decision whereas Claude has been designed to emulate decision by randomization. That is something we should be wary of when looking at AI results.
Claude offered a third version. I found this one to be a little less inspired and it’s here where you can see the randonness leaking into the overall effectness of the outcome. Claude stated it made the following design choices:
- Layout: Uses a curved dividing line with elements arranged along it, rather than the vertical or circular layouts of the previous versions
- Background: Incorporates a hexagonal grid pattern for a more technical/geometric feel
- Style: More minimalist representations of each day’s creations
- Color Palette: Uses a different range of colors, though still keeping with the theme of creation
- Element Organization: Arranges the days in a flowing, cascading manner
- Decorative Elements: Adds horizontal lines for visual interest and structure
The third version has all the parts, but it feels closer to a textual list than the other two images. Reviewing the list of design choices, the SVG is not in the spirit of the Creation story. The first two kind of feel right…but perhaps that’s just luck. In Claude’s defense, a textual list is what the Bible or any other volume work is: lists of books that contain lists of chapters that contain lists of verses that contain lists of words.
For someone who is just getting started in the AI space, I find this simple experiment very intersting. The exact same prompt was used for the three versions of Genesis 1: Chapter 1, yet Claude rendered each one differently based on - in simplistic terms - random design choices. However, the reason for those choices isn’t derived internally like an artist would; but instead they are made from external ideas as if to ask what would a different artist do. The benefit here is that you can get different options, but it might take a number of tries to get the one you like. I would imagine in the future to yield specific stylized results there would be distinct versions of Claude - daresay promptless personalities?
Personally, I am looking forward to continuing my trek through the AI-verse. Just like text editors, IDEs, and compilers, AI is an extension of things we know. This difference here is that AI may yield us outcomes we don’t know or even understand.