CNN
—
We’ve seen enough of artificial intelligence that the term is often a misnomer. While concepts such as the ear, the correct number of fingers, the natural rhythm of the voice, or the coherent complex of thought continue to circumvent them, publicly available AI programs still hilariously convey moments of ignorance in large numbers. We create AI programs. Finding these limits is fun. It’s fun to convince ourselves that computers are far from swallowing our culture whole.
It’s even more fun to do it through a messy and unholy craft project.
ChatGPT, a publicly available language learning AI, is not designed to create things like crochet or knitting patterns. However, since such patterns are a form of language, it is theoretically possible to create them programmatically. Many curious crafters have tried this, with increasingly absurd results.
This exercise is not only hilarious, it raises a very interesting question. What happens when you ask a program specially trained in a language to create something outside its domain?
Let’s explore some ChatGPT-generated patterns by crocheting them.
Some notes on my methodology:
- I used the prompt “Create a (blank) crochet pattern” for every crochet pattern I generated.
- I followed the crochet pattern exactly. I have tried my best to resolve any discrepancies in the pattern, such as miscounting stitches or gibberish instructions.
- I chose objects and concepts that are not very common in crochet projects. Many questions have been raised about how well programs like ChatGPT generate ideas that others have already created, and I wanted to get as close as possible to the core of his ChatGPT creativity.
This was supposed to be a control project. A simple object with a unique shape that is repeated in countless crochet patterns on the internet.
That’s exactly what ChatGPT gave me. Was it a banana? Not even the slightest sensation. Was it something that a very limited intelligence system floating on a binary goo approximates to a banana? perhaps.
At first glance, ChatGPT crochet patterns look exactly like crochet patterns. There’s even a hilarious little introduction, and the program is clearly able to mimic terms any crafter would recognize, such as “doing stitches.” I also knew how many 3D crochet projects started as circles.
However, once the instructions progress beyond a few common starting stitches, the project usually evolves into one of two things. One is spheres, or complete nonsense. Here I knew enough to provide an explanation of bananas and peels. After that it was all just spheres.
Bananas radically realigned my understanding of ChatGPT’s crochet limitations, so I decided to put it back. What does the program do with vague and qualitative words like “weird” and “creature”?
You can never guess. (It is a sphere.)
This pattern produced several different parts: a torso (sphere), two tubular legs (long sphere), and two ears. It wasn’t specified how they would be combined, leaving the creators to decide on their own level of “weirdness”.
ChatGPT can create patterns that are linguistically consistent, but did not do very well when it comes to specifying the overall “art”, such as how to assemble the work in a way that does not violate the boundaries of Euclidean geometry. For traditional crochet patterns, the building instructions have specific locations and methods for attaching pieces, along with pictures and notes for particularly difficult steps.
The instructions provide no assembly guidance, but conclude, “Feel free to experiment with different colors, thread thicknesses, and additional embellishments to make your creature unique.” increase. “Have fun with crochet!”
I knew this was going to be a problem the moment the note was generated that said “one ear green and the other tan”.
ChatGPT actually leveraged their knowledge that Baby Yoda is also known as “The Child” in the introduction to the description. Next, as you might have guessed, we render The Child as a series of spheres.
This is a good time to give you a little more background on how ChatGPT works and why it takes so much effort with things like crochet patterns.
“The important thing to keep in mind is that this is limited AI,” says Casey Miller, a computer programmer in Kennesaw, Georgia. “Unlike what we probably think of as AI in general, we have absolutely no idea what ChatGPT is regurgitating to people and why.”
Narrow AI is trained based on specific skills. In the ChatGPT example, the skill is to respond in a way that sounds as human as possible. It is theoretically possible to create an AI that generates amazing crochet patterns, but it would require very specific training and programming.
“For example, a general AI can turn around after beating anyone at chess and create a very complex crochet pattern. No,” Miller continued. “We only know how humans talk about crocheting, or how they usually communicate patterns. We can’t interpret what these patterns actually create.”
Fair warning: this is what broke me. Everything you see after this is an abomination, an attack on the order of nature that should freeze you to the core.
At least it wasn’t a sphere.
Again, I was interested in seeing what ChatGPT does for fairly abstract requests. At first, it seemed to offer a very logical parry.
ChatGPT’s response reads, “To create a crochet pattern to represent Antarctica, you need to create a textured surface that resembles an ice landscape.” “This is a pattern that captures the essence of Antarctica.”
The pattern instructed me to create the first series of stitches and then build each row of stitches on top of it. Doing so creates a number of blanket-like flat crochet patterns. So far, so good. Started at 14.
The problem was immediately apparent when I was told to create 5 stitches for each existing stitch.
can you see It looms over us like an oncoming, inflating train.
A few lines later, I got the same instructions. A few lines later, the instruction “5 stitches for each stitch” was repeated again.
14 stitches can no longer be flattened in the 3rd increment. With 1,750 needles, I was creating something like a crawling coral reef. ChatGPT gave me their version of fractal crochet because they understand non-existent things like time, amount of yarn, and human sanity limits. If you had completed the pattern exactly as written, you would have over 35,000 stitches in each row. Naturally, it stopped at 1,750.
As I worked, I thought I was subconsciously knitting a rough model of the universe. What equation do I need to solve to predict how time or space will be laid out as it explodes and folds outward exponentially row by row? I thought of If I knew how, could I find myself in some hidden crevasse here and now?
When I showed this work to other people, they guessed what it was. “Brain,” they said. “It’s a loofah!”
No, you idiots. Antarctica.
To be honest, this was optimistic for me. I thought that something with a very specific and simple shape would be easier for ChatGPT to approximate. And buildings with highly recognizable and distinctive contours, such as Dubai’s Burj Al Arab, seemed like the perfect combination of spatial simplicity and obscurity of crochet patterns.
At least it wouldn’t be a sphere!
(It was a sphere.)
How ChatGPT gets what it can easily know as a language model (Burj Al Arab is in Dubai! It’s popular among architecture enthusiasts! It has sails!) and uses it for other things It was a lot of fun to see what it would combine with. It never made sense (what is a sail? What are the dimensional limits of human creation!). The result, if it existed in The Flintstones, was the Burj Al Arab.
I saw the potential in testing the Burj Al Arab and chose another building that I thought would definitely break ChatGPT’s habit of rendering everything in spheroids. Burj Khalifa next to Burj Al Arab seemed like a good choice. It’s so sharp! very high!
Did I mention a tall, pointed sphere? ChatGPT can do it!
Now, for those of you who have never been on the verge of tears while crocheting, I can say that it is a special experience. What really irritated me about ChatGPT’s instructions, so nonsense as to approach the beautiful entropy of human thought, was how the program perceives the relationship between the notion of spires and the buildings they let me create. was.
After creating a long hump of the main structure (which was a very ominous start), ChatGPT “formed a spiral pattern down along the side of the building” from the top of the building, creating a length of I wrapped the thread and instructed it to “gradually reduce the length.” distance between each lap. ”
“Now you have a small crocheted model of the Burj Khalifa!” I heard the instructions. “This iconic building makes a great exhibit or gift for architecture enthusiasts.”
In the words of TS Eliot, That’s not what I meant at all.
Not at all.
Maybe three-dimensional crochet is too difficult for language learning AI. Perhaps going back to basics might help him create the one that most closely resembles my prompt, but only as one item.
The initial result of this prompt was so nonsensical that I regenerated the response. Then tried again. Creating these objects one after another was like watching something struggle to be born. A keen-eyed crocheter may recognize the recognizable heart pattern fragment. By the fourth iteration, we had a reasonably good approximation of the shape. Was it a learning experience, or was it just luck like a pack of monkeys at a typewriter?
I’ve generated a few more patterns from this prompt and I assure you it didn’t get any better.
If there’s one lesson I learned from this experiment, it’s that human intelligence is fundamentally interdisciplinary. Language oozes into the visual sense and is intertwined with memories and personalities. Artificial intelligence programs don’t really work that way. You might be able to do one thing very well. In some cases, you might even be able to do some things. But once you exceed your allotted skill, it just clumps all the way through.
