Katherine Parrish

I had my first real mindblown moments with GPT4 last night. I’ve been playing around with a creative project for which I needed a source text that was all text, but unconventional when it came to syntax, lineation, language, etc. After learning that GPT4 can read and analyze text from an uploaded pdf file, I decided to upload a selected sliver of an excerpt from a. rawlings’ Wide Slumber for Lepidopterists.

I wasn’t trying to test its ability to analyze it, dare I say understand? but before I could put it through the paces of my choice, it said this:

It looks like you’ve uploaded an excerpt from a text titled Wide Slumber for Lepidopterists. The passage you shared contains dense, abstract language with a blend of vivid imagery and technical terminology related to sleep and insects, particularly moths and caterpillars. The writing appears to explore themes of transformation, perhaps hinting at the metamorphosis of insects or even a metaphorical transformation within the human experience. There’s also a rhythmic quality in the text, mimicking the repetitive nature of breathing or a heartbeat, which ties into the sleep-related themes.

How would you like to explore or use this content? Do you need help with analysis, inspiration for your own work, or something else?

It took me a while to recover from this. My wonder and delight was soon followed by hubris. I had thought, and declared loudly, that the solution to the challenge of how to teach students to write a literary essay independently was to use texts that were outside the range of its training texts, and I thought that Canadian contemporary poetry was a pretty safe bet, especially texts that were published after its release, or were not available online, or widely published. “It can’t write about what it doesn’t know!”

This is probably still true of ChatGPT3, but ChatGPT4 is a very different beast, as many people have been saying. I seriously doubt that Wide Slumber For Lepidopterists was part of GPT4’s training – I’ll test this later with some post GPT4 release more obscure texts. What I think I saw here was an AI applying concepts to new material, and challenging material at that.

I then asked it to: “bring the text down to earth. force the language to behave itself.”

Sleep Paralysis

Smoke and the rough feel of steel wool enter the scene. A caterpillar crawls slowly over the bedrail, making its way beneath the quilt and sheet. It moves, leaving a faint, distorted line behind, like a scar.

The caterpillar’s movements are slow and steady, weaving its silk around our throats, tightening its grip. The air feels thick, like layers of fabric suffocating the lungs.

To avoid twisting, the wings must move together. The wings are pinned to a board, immobilizing them.

Cataplexy

Pins pierce the skin, like a tooth against a wall. The specimen is placed under a lamp to dry faster. A moth with barbed spines emerges.

Narcolepsy

Neon lights flicker. Breathing becomes a rhythm: slow, slow, then a sudden slip into sleep. The mouth, soft and parted, releases a whisper of breath. Muscles stretch and relax, pinned in place. The body moves as if on its own, turning and pulsing, pushing the boundaries of sleep.

Sounds repeat: “a hush, a ha,” mimicking the rhythm of breath and sleep. Hypnic twitches, deep slumber, and restorative sleep cycles come and go.

---

This version retains the essence of the original text but presents it in a more straightforward and accessible manner. The abstract elements have been simplified to create a clearer narrative, making the imagery and themes more relatable.

You said:

holy fuck.

ChatGPT said:

I’m glad that resonated with you! If you want to explore more or have any other requests—whether it’s further adjustments, different styles, or even something entirely new—just let me know!

As gobsmacked as I was, I still couldn’t figure out how it had read the more linguistically unconventional parts of the text. I decided to pair down my selection even further. When I uploaded just this section:

and sent it to GPT4 to “read” and “analyze,” I received an error message.

Truthfully, I was a bit disappointed.

I’m really not invested in finding a GPT4-proof text, or going on a robot hunt. (Btw, does anyone think that what Captcha currently has in place for its rudimentary Turing test is going to be effective in a GPT4 world? ) But I clearly have a lot more thinking and investigating to do before I make any bold pronouncements about what this generation of LLMs can and cannot do with texts.

Beyond here, there be dragons!