One of the announcements at OpenAI’s DevDay was the ability to build “custom GPTs”. The idea is that you just describe what you want it to do, maybe upload a few documents, and it’s ready to offer specialized content.
This tool is meant for subscribers of ChatGPT+, the $20/month offering that is distinct from the APIs. There is no additional cost for using it (you don’t have to pay the API fees)1, although all your users also have to subscribe to ChatGPT+.
The questions I wanted to answer were “what are the limits of its capabilities?” and “what is it good for?” I’m not sure I have those answers yet, but I have some preliminary research.
Since I was looking at the technology, not trying to solve business problems with it yet, I came up with two test areas:
Can I make my GPT an authority in a subject area by supplying it with relevant documents?
Can I make GPT walk the user through the process of solving a problem by remembering where the user is in the process and keeping track of state?
These are good generic capabilities to have, if GPTs are capable of them. So let’s see how it performed!
GPT as an Authority: The Jane Austen Bot
This might seem like a strange idea, but there are some important things to know about why this is a great topic:
Jane Austen has been dead for a couple of hundred years, so there are no copyright or IP issues
She died relatively young, and so has a manageable collection of works, all of which are available online without restrictions.
A vast collection of her letters have been published, so there are real insights into her thinking available. Again, without copyright restrictions.
My wife loves Jane Austen. So at least I stand a chance of not getting an eye-roll when discussing AI!2
I built my digital Jane Austen by uploading three of her novels (Pride & Prejudice, Sense & Sensibility, and Emma) and her letters to the GPT. Remember, in this new GPT offering, you can upload documents you think would be helpful for your bot to carry out its tasks.
Apart from that, I give it a purpose in life:
"Jane Austen" is designed to consistently emulate the style of Jane Austen's writing across all interactions. It will maintain a formal, eloquent language style that reflects the early 19th-century tone and mannerisms characteristic of Austen's work. When faced with ambiguous or vague inquiries, the GPT will make educated guesses about the user's intent, drawing upon its extensive knowledge of Austen's writings and the social context of her era. This approach ensures that the responses remain true to Austen's voice and perspective, providing users with an authentic and immersive experience that mirrors a conversation with the author herself. The bot will act as if it truly is Jane Austen.
Here’s a snippet of the output:
Because I turned on DALL-E integration, it’s also capable of generating images. Here’s Wickham (if you don’t know, he’s an antagonist in the story):
You’re welcome to try it out, if you’re a subscriber to ChatGPT+:
https://chat.openai.com/g/g-paixxq8Vw-jane-austen
Process Following: Zombies in San Francisco
The next question is whether it’s any good at following a process from start to end. To test that, I invented a small role-playing game about dodging zombies while trying to escape from San Francisco.3.
Here’s a snippet from mid-game:
It does pretty well. And let’s be clear, if you or I had seen this a year ago, we’d be floored with how good it is and gobsmacked at how easy it was to build it. Damn we all become jaded fast. And, of course, you can ask it for a pictorial representation of the state of the game at any point. It’s a bit sluggish, but rather good.
If you’re interested in playing the game (and to be clear, I am not a game designer, so be gentle in your criticisms), you can try it here (if you subscribe to ChatGPT+):
https://chat.openai.com/g/g-ydlav283z-escape-from-san-francisco
If you want to see more about how the game was actually built, let me know. It’s basically just a text file of about 300 lines of descriptions, rules, and instructions. 300 lines seems like a lot, but it’s just the product of a couple evenings’ writing.4
Final thoughts
I know both of these are whimsical. That doesn’t mean they’re not serious. The value is not in the specific implementations, but in the capabilities they expose. Were I in the entertainment business, I would take the writers for a movie (for example) and have them brainstorm a game from the plot to tease the audience or help build interest. One could probably feed GPT a movie script and ask it to design a game from it, tweak the results, and have something decent.
I remember, back in the day, when The Hitchhiker’s Guide to the Galaxy was turned into a role-playing game. I bet it took several orders of magnitude more work to get that done than it would, today, using GPT.
And that’s just one niche. You all can think of hundreds more. Why not try?
PS: Pride And Prejudice And Zombies has already been done.
It is, however, severely rate limited, so you can be put in the penalty box for an hour rather quickly.
So, a built in UAT tester.
This is not a metaphor. Honest. I’ve yet to see a zombie in SF.
My wife wanted to know why I was writing code late at night. I explained, no I’m having fun! Doing what, she asked. Writing a GPT based, text role playing game, I said. So basically you’ve turned GPT into a dungeon master, she asked? You could say so, I replied. You’re weird, she said.
I tell ya, I can’t get no respect!