I will have my last original story of the year for you next week, and then either a year end wrap-up post or possibly just a rest week on New Year’s Eve. Hard to believe that 2022 is almost over. What a crazy year it has been!
This week, however, I bring you my attempt to develop an original short story concept entirely using ChatGPT. In doing this, I tried to have ChatGPT come up with as many of the ideas as possible, so that I was guiding and shooting things down, but I was always coming back to the A.I. doing the creative work where possible.
So it begins:
What we’re doing here is batting around concepts. GPT—as I would have guessed—starts out with pretty generic concepts. Remember, it’s scraping the internet for ideas, not thinking laterally like a human would do.
In other words: Most of my ideas are hearing something that’s not obviously related to sci fi in any way and going “I wonder if that could be a story?” GPT gets ideas by associating things with the concept of “sci fi story ideas” directly. This gives it a huge library of possibilities, but a lot of them are going to be generic.
This is a crucial fact for understanding how to use GPT to best advantage. So I continue to prompt it to differentiate:
That last idea is very clever: Appeal to the control system’s sense of self-preservation. That’s a real “aha” moment as far as a story goes. It makes sense when you say it, but is also not obvious before you do. Now we’re getting somewhere.
This is just really good stuff. This level of thinking is better than many development executives currently pulling down a healthy six figures in Hollywood. Where does it come up with the name “Ava” out of all possible names? I don’t know but it seems perfect somehow.
I also love how enthusiastic and unflappable it is. Besides being funny, that’s actually an incredibly important and helpful thing for doing story development. I could tell it “no, not that, do another idea” a thousand times in a row and it’s still going to respond with the same “Sure, no problem!” attitude, and even though I know on one level that its enthusiasm isn’t “real”, it still provides helpful encouragement.
One of the really hardest things about creativity is keeping your energy and enthusiasm up. What you’re seeing in these chats is a huge part of the process—sitting around and going “what else could it be”, “what else does that imply”, etc etc. That process can be really deflating. Having a conversational partner whose enthusiasm is unshakable is genuinely helpful no matter if it’s human or not.
This is a killer app for GPT as far as I’m concerned. I didn’t take a note on what we were talking about, I just concentrated on prompting well and my creativity, and then the moment I was ready, it spit out everything we had been talking about in a helpfully bullet-pointed format.
I have sat through many development meetings where I took notes on what a bunch of people were saying, and then went back afterwards and spent longer than the meeting lasted trying to reconstruct the meeting from the notes to summarize it, and in some cases having to email the participants and ask them to reiterate details I was worried I’d missed. The time savings this feature represents are *huge*.
I then prompted GPT to turn those four bullet points into something more resembling “story beats”, and this is what it came up with:
It’s interesting to me what and how much it added to the story here. It has clearly absorbed some classical story archetypes from the internet, because it’s trying to add a “bad guy comes back after you think it’s vanquished” moment to the story, a la every horror movie ever. This is good development—it’s a bit generic, but it’s extrapolating the generic beats to the specifics of our story well, and honestly, when you’re trying to break a story, putting some generic stuff in there and seeing how it looks is a standard move that often works. The prompting continues:
This is honestly a really good—if still too generic—story outline, and at this point I would consider the initial concept development phase to be complete. We understand what the story is, who the main character is, and what the “arc” of the story is going to be. Plus, we have an ending—Ava is going to convince the A.I. to enter the digital realm and then say an emotional goodbye to it. That’s a story concept.
Just taking a moment to appreciate this, because now we’re going to get into some stuff that GPT doesn’t do as well, but I don’t want to undersell what it’s done here. I hardly came up with *any* of this story. GPT suggested almost everything about it, and with only a couple of small exceptions, all I did was go “I like that part” and “I don’t like that, do something else.” That is incredible!
That said, it’s not a complete development process. you can’t write a story based on this outline without making a bunch more decisions about what *specifically* the story is going to be. Still too much generic stuff in there. So I keep pushing and we see how well GPT can do in this next phase:
So far so good. We’re starting to flesh out some of the specific beats, and GPT is performing well so far. But now we need to add a whole other level of detail. This is really hard stuff, and it’s something that most humans really struggle to do. It requires a very high level of conceptual thinking, and a combination of abstraction with specificity that trips up most brains. Observe:
Uh oh. Now I don’t know if this was a problem with the model, or if it’s a meta-problem of resource limitation at OpenAI itself. Meaning, ChatGPT was going viral as I was doing this, thousands upon thousands of people were using it, and it’s possible there just wasn’t the computing power available to complete its answer to my prompt. Intuitively it does seem like the result of a computing power shortage would be a straight denial of service, rather than sort of trailing off mid-sentence (which is itself a very human response!), which would seem to lean in the direction of the model itself being unable to handle the complexity. But, it’s impossible to say for sure without being privy to functionality data from OpenAI that they are not likely to provide.
My thinking here is to try to give it the task in smaller chunks. Rather than “add subbeats to each beat” where it has 7+ things to do before responding, will it have more success if I prompt it individually for each beat with an example?
For some reason (I don’t have a good theory here) it’s struggling to only add one sub-beat instead of two until I got very specific about it. And, to be fair, some of these sub beats don’t add detail or surprise to the scene the way I asked. They remain quite generic. However, it feels like we’re getting back on track.
My intention with this piece was to push GPT as far as it would go, all the way to having it write the story paragraph-by-paragraph from my prompts if that was possible. At this point I started to have doubts it was going to be able to handle the complexity though. It’s certainly the case that getting GPT to write the story was going to take *way* more time and effort than just writing it myself from the outline we concocted would have.
I would not have guessed this, but GPT seems at its best when its coming up with “whys” and “hows” for things. Like all those options could use more specificity to them, but none of them are stupid or unworkable! And trust me, when you sit around in development meetings trying to break stories, most of the ideas are stupid or unworkable, so this is really quite impressive.
This is where I quit in the development process for this attempt. It really seems to get confused with the reprinting and collation of the outline, even though intuitively that seems like it should be the easiest step for a model whose entire thing is scraping text and regurgitating it in legible form.
For me, this is only a partial success, but still very much a success, insofar as I started with nothing and now have a story concept that I could go and write and feel good about. I will definitely be using ChatGPT more to discuss story ideas and get inspiration. It’s strength with generating plausible reasons for things to happen or ways they could happen is helpful, because for me those are the hardest parts of the process!
I will be writing this story and releasing it as an official part of OGWiseman’s Stories at some point, for sure. Not only will it be cool to write a story from an A.I.-generated idea, but it’s a cool idea and I think it will make a genuinely cool story! It has real literary pathos (Ava wants the Climate A.I. to reach a Heaven since she herself cannot believe she will), and those aren’t easy to come by.
I also plan to return to this process in the future. There are many, many iterations of this that could yield really interesting results. (e.g. Can GPT come up with a story concept that’s funny? Is it capable of making jokes?) Besides, at some point they’re going to release a version of ChatGPT that’s paid, I *will* be buying access, and at that point there should be no compute limitations to blame for any model failures.
Whether ChatGPT is capable of writing an entire story from its own outline remains to be seen, but if it can’t, a later version will. Seeing that things have gotten to this point has totally convinced me. Maybe it won’t just be a single prompt—”write me a sci fi story”—and then it spits out a finished story, but at some point it will take my coaching, not lose track of where it is in the outline, and paragraph-by-paragraph it will crank one out.
Frankly at this point, I’m just trying to stay one step ahead of the job curve by getting good at writing prompts and coaching A.I., because some version of this is going to be what most white-collar jobs become, and before I retire at that. We’d better get ready!
END
Thanks for reading! If you enjoyed this post, please help me out by liking, commenting, and sharing with friends. Have a great week, and I’ll be back next weekend with another original! :)
Merry Christmas!
Once again, I am both impressed and terrified by what ChatGPT can do.
As I was reading this, it occurred to me that ChatGPT could be used to help with writer’s block. For example, if you have 50% of a story done but then are at a loss as to what to do next, you could always explain to the AI the general premise of your story and then ask it to continue where you left off. Even if you used none of the suggestions, it could still help to get the creative juices flowing again.