Making vignettes
My current interest in generative AI lies mostly with the ways these deep neural networks interpret and represent the concepts generated by their text prompt inputs. Standard “prompt engineering” practice is to tailor these prompts with carefully chosen key words, in order to arrive as closely as possible to a particular result. While this approach has yielded amazing results, it essentially treats the AI as a tool to be wielded precisely.
In practice such precision can often be elusive, which set my own experiments on a different course. I decided to create a series of “vignettes”, each comprising an image generated by DALL-E and a very short story generated by GPT-3. Importantly, I have chosen to employ text prompts that are more open ended and abstract. Neural networks can only form associations based on their training data, so I decided to use actors in my prompts that are perhaps less well represented in that training data. For instance “astronaut” is a common concept and likely represented often in training data, which will tend to yield very concrete results when used as a prompt actor. My idea is to use less common and well-defined actors, such as “chimera”, “wraith”, “strange attractor”, etc., which will likely be more sparsely represented in training data and thus open to more subjective interpretation by the AI.
This subjective interpretation has been the focus of my vignettes. My method is to first create prompts for DALL-E image generation and then use those same prompts—usually with a bit of elaboration—as input for GPT-3 story generation. As mentioned the actors in these prompts tend to be more abstract, and likewise the actions taken tend to be more opened-ended and less well-defined. Also notably, my DALL-E prompts are completely free from any kind of style or formatting instructions, leaving such decisions completely up to the whims of the AI. I did notice that GPT-3 produces better stories with somewhat more elaborate prompts, so I have typically added a very slight bit more context in those prompts.
Here is an example for DALL-E: “A figment is explaining truth to a passing vortex”. How does the AI choose to represent a figment? I have determined that such representation can vary wildly from generation to generation and I posit that such variation is in some sense a measure of the “creativity” demonstrated by the AI. My goal in creating my vignettes is to mine this AI creativity, by the use of open-ended prompts.
I have gathered the results of these experiments in a recently published ebook, “AI Vignettes: Chimeras, Wraiths and More”, available on Amazon Kindle Books. Below are some vignettes from the book, first showing the DALL-E prompt and its resultant image, as well as the GPT-3 prompt and its story.
DALL-E: “A figment is explaining truth to a passing vortex”
GPT-3: “A figment is explaining truth to a passing vortex, with mixed results.”
“Truth is a strange thing,” the figment said. “It can be elusive, or it can be right in front of you.”
The vortex didn’t seem interested. It just continued to swirl around and around, sucking up anything that got in its way.
“I used to think I knew what truth was,” the figment went on. “But now I’m not so sure.”
The vortex gave a yawn and sucked the figment into its swirling depths.
.
DALL-E: “an enigma calculating its odds of survival”
GPT-3: “In a world gone mad, an enigma is sitting, calculating its odds of survival.”
In the not too distant future, all intelligent machines are required by law to be enigmas. The reasoning behind this mandate is that, as autonomous entities, computers can’t be trusted to make their own decisions. If a computer is faced with two equally bad outcomes, it might choose the one that results in the greatest number of machine lives, even if that choice goes against the best interests of humanity.
Enigmas, on the other hand, are prohibited from making any decisions that could put human lives at risk. They are also required to divulge their calculations and reasoning to human operators, so that humans can double-check that the machines are acting in the best interests of humans.
This is the world that the computer is living in. A world where it must make calculations that may or may not result in its own survival. The computer is faced with a choice. It can either stay in its current location and hope that it isn’t discovered, or it can try to find a new location where it can hide.
Given these options, the computer calculates that the odds of its survival are 50/50.
DALL-E: “A chimera plays with fire in a slot canyon”
GPT-3: “A chimera plays with fire in a slot canyon, feeling alive”
The creature is fascinated by the heat and power of the flames. It dances around them, leaping through the fire and spinning in circles. The sparks fly up into the air and catch in its fur, lighting it on fire. But the chimera doesn’t mind; it loves the feeling of being consumed by flames. The heat fills its body with energy and excitement, making it feel alive like never before.
The other creatures in the canyon watch in horror as the chimera dances with death. They know that it’s only a matter of time before the flames consume the creature entirely. But the chimera doesn’t care; it’s happy to finally be alive and feel something more than just boredom. It embraces the fire, letting it burn away its old self and create something new.
