Help Me Obi-Wan KenChatGPT, You’re My Only Hope
After having two shoulder dislocations in three months, my doctors told me I needed to have a fun procedure called arthroscopy labral repair shoulder. During the surgery, they put you under, cut three holes in your shoulder, open up the joint and poke around until they find the reason for your reoccurring dislocations. Then they helpfully drill some holes in your bones and sew everything back to where it should be and close you up.
All in all, a very fun procedure. After the surgery, they told me I had to wear my sling at all times for 5 weeks. 5 whole weeks!
But I had a bunch of astrophotography photos to post here and on the socials, I had work, I have to type!
My surgeon didn’t care, he granted me no mercy.
The AI Hype Train Only Needs One Arm
Just preceding the surgery, I had the good fortune of having two weeks in a row without rain and the majority of the nights clear enough to do astrophotography. This is pretty rare in Kentucky. We only have ~24% of our nights suitable for DSO (Deep Space Object) astrophotography, and rarely do those string together. Setting up the telescope takes an hour or so, so if I can string together nights where I only have to set it up once and then can leave it out, I can’t miss those opportunities.
So as soon as I came out of this surgery, I had around 5 photos to edit, put together and post. But how to do all of it with just one hand?
The ChatGPT release has taken the world by storm so I decided to give in to the hype and try it out.
Actually Getting Stuff Done
If you head over to my photo section, you’ll see my posts are pretty formulaic. Image is the leader, then I list some brief stats about the object, its size, distance, type, other names, etc. Then in the description, I give the object some context, maybe describing its discovery a bit and its characteristics in detail. Then I go through a few fun facts and try to use them to put into scope the grandness and awesomeness of the object. I try to phrase things in a way someone just looking at pretty pictures can understand. Hopefully, if they read the description, they’ll walk away knowing a bit more about space.
Normally with two hands, it takes about 30 minutes for me to research and write up a blurb about a photo. A lot of work goes into these! With one hand, I could barely do the research, much less type up paragraphs of content.
This, turns out, is ideally suited for AI writers. With something formulaic, I can tell ChatGPT my formula, my tone and what I want to impart. Then it can spit out a rough draft and I can edit it, check it for correctness and post it. All in all, with one hand, it took me just over an hour to figure out the prompt that worked best and then generate and post 4 new images.
Prompt Engineering
It was really quite intuitive to set up, I logged into ChatGPT and gave it this prompt:
Hello chatgpt. I am an astrophotography nerd. I take photos and then post them online with some context around
what is in the photo. Can you help me write those blurbs?
They are typically 2-4 paragraphs in length. The must include both the common name for the deep space object
as well as it's catalogue designations. It must include the following information, only if known, do not make
guesses: object size, object distance, object type (i.e. emissions nebula, dark nebula, galaxy, etc).
Optionally, include if possible, provide history of its discovery and why it is named what it is named. If it
has any connection with a historical event.
Please include one to three fun facts about the object that a common person can understand. One of those fun
facts should put some aspect of the object into perspective that a common person can understand
ChatGPT responded with a really good example for M42, the Orion nebula, however, we had some stylistic disagreements, and I had concerns about authenticity so I asked it for citations. So I asked ChatGPT for citations. This is the sample it provided:
In this stunning photo, we see the Orion Nebula (also known as Messier 42 or M42). It is a diffuse emission
nebula located in the constellation Orion, about 1,344 light-years away from us [1]. The nebula spans
approximately 24 light-years across and is estimated to be around 2 million years old [1]. It is one of the
most famous and easily recognizable deep sky objects, visible to the naked eye as a fuzzy patch in the Orion
constellation.
The Orion Nebula is a stellar nursery, where new stars are born from the gas and dust that make up the nebula.
In fact, it is one of the closest regions of massive star formation to Earth. The bright stars that are
visible in the center of the nebula are actually a cluster of young stars known as the Trapezium. These stars
are only a few hundred thousand years old and are some of the hottest and brightest stars in the Milky Way [2].
Fun facts:
1) The Orion Nebula is so bright that it can be seen even from urban areas with a pair of binoculars or a
small telescope [3].
2) The gas in the nebula is mostly hydrogen, which glows red due to the ionization caused by the Trapezium
stars [4].
3) If the Orion Nebula were as close to us as the nearest star (Proxima Centauri, 4.24 light-years away),
it would appear about as large as the full Moon in the sky [5].
Citations:
Messier 42 (Orion Nebula), NASA/IPAC Extragalactic Database (NED)
Orion Nebula, Universe Today
Orion Nebula, Sky & Telescope
The Orion Nebula (M42), Hubblesite
How Far Is the Orion Nebula?, Space.com
As you can see, its citations are a joke. Sky & Telescope is a magazine. It just cited the entire magazine. Pretty useless for checking its work So I figured I knew enough about the objects I was imaging I could weed fact from hallucination, and asked it to omit the citations. They weren’t helpful anyway.
One last thing I had to do, I hated the bulleted list, I wanted the fun facts given more conversationally. More as a description of the object and less as a list of facts.
I told ChatGPT:
Instead of listing the fun facts as a bulleted list, include them conversationally in a paragraph
And while that had an effect, it didn’t do what I wanted. I never could get it to properly pepper the fun facts throughout an explanation of the object like I would have done. But that’s easy enough to finesse in editing, even with one hand.
This was the end result of what it spits out for the fun facts paragraph:
Fun facts: The Cygnus Loop is so large that it spans an area of the sky about 36 times the size of the full
Moon. The Western Veil of the Cygnus Loop is also known as the Witch's Broom, due to its resemblance to a
broomstick. Finally, the Cygnus Loop was first discovered by the English astronomer William Herschel in 1784.
Not up to my quality, but easy enough to whip into shape with a bit of one hand editing.
Lessons
Would I do this again? Probably. I’m 7 weeks out of the surgery now and have my arm back. I might revisit using this especially since I have such a backlog of images to post. But the quality is lower than what I’d normally produce.
Now that I have both hands again, maybe it’ll be easier to edit it to my quality. Maybe I can play around with the prompt a bit more and get a better response.
End of the day, ChatGPT is quite adaptable and can do a lot of the grunt work for you, but it can’t replace you. Efficiency isn’t a replacement.
Tone
Tone was the biggest thing about my posts I have a specific tone I like to write my blurbs from, and I was able to impart MOST of that on ChatGPT. But ChatGPT has its own tone that it likes to use. It likes lists and it likes to start the ending paragraph with “in conclusion” which makes my writing sound like I was doing it for a high school essay.
If you work with ChatGPT a lot, you’ll notice it also has a distinct tone way of organizing its sentences. Every paragraph is almost exactly 3 sentences long, +/- 1 sentence, and the breaking into paragraphs is done by length. Not by idea. This makes it easy to spot AI work in the wild if you’re aware of it. It sticks out like a sore thumb.
Accuracy
I think it’s generous to call these things “hallucinations” when they’re just flat-out wrong. If your online bank balance was showing the wrong value, you wouldn’t let the bank tell you it was “hallucinating” it’s a bug.
That said, I didn’t encounter too many of them on these blurbs. I don’t know if it’s because I told ChatGPT to not guess and stick to well-known facts, or if its baseline is just decent. I did run into a few though, in particular, it told me that the pleiades cluster is gravitationally bound, meaning they’ll eventually all move apart, which is the complete opposite of what gravitationally bound means.