D&D General Can ChatGPT create a Campaign setting?

[Jer]

AI art can be beautiful and stunning. The text seems to really trite.

I mean to be a little fair - ChatGPT is tuned to be a chatbot. It's not really supposed to be churning out literary text, it's supposed to be responding to queries as if it were a customer service representative on the other end whose job it was to answer whatever inane question you happen to be asking them. AI art tools are tuned to generate impressive looking output - if there were an AI art tool that was trained to churn out emojis or memes and you asked it to generate something in the style of a van Gogh you'd have the equivalent of the chatbot being asked to generate something literary.

That said - text is actually harder to work with in a learning context than images, and the models around images are older and better understood than the text ones. Also an image is static while text flows through time - asking an AI art bot to generate consistent animation is a harder task than asking it to generate still images. That would be the equivalent of asking it to generate a lengthy text.

(There are other problems too - large language models by themselves are actually a bad fit for what people are trying to make them do. The fact that they work at all is impressive, but the upper limit of what they're going to be able to do is going to disappoint a lot of folks in the coming years I suspect.)

 

[Reynard]

As a writer with a large amount of unfinished or otherwise unpublished work, I would like to train my own AI so I can have to write for me.

 

[Grendel_Khan]

If the AI obviously uses some sentences or parts of pictures obviously taken from a copyrighted source... it feels like someone should be in trouble. (Does it obviously do that for some things it uses in picture making AI?).

It's a complicated topic, but ChatGPT doesn't necessarily lift whole sentences, just as Stable Diffusion doesn't lift whole images. They basically blend data into a slurry and then form new content out of it. The closest they get to what you're describing is when they use the "style" of a given source, which is easier to detect for artists (certain kinds of brush strokes, etc). So it's easier to go after AI art for copyright stuff, but it might be impossible to do it for something like ChatGPT.

Closest we can get is to shame anyone who uses it in commercial products!

 

[Ruin Explorer]

So it's easier to go after AI art for copyright stuff, but it might be impossible to do it for something like ChatGPT.

I don't think it will be, long-term. I think it'll be a bit of an arms-race with plagiarism detection AIs. Also I think the results of the AI art legal battles may well give other "steal a huge data set"-type "Silicon Valley Libertarian" approaches pause, because I don't think the AI companies are likely to win them.

 

[Cadence]

It's a complicated topic, but ChatGPT doesn't necessarily lift whole sentences, just as Stable Diffusion doesn't lift whole images. They basically blend data into a slurry and then form new content out of it. The closest they get to what you're describing is when they use the "style" of a given source, which is easier to detect for artists (certain kinds of brush strokes, etc). So it's easier to go after AI art for copyright stuff, but it might be impossible to do it for something like ChatGPT.

Closest we can get is to shame anyone who uses it in commercial products!

In the OP all of the deities chosen look to be public domain kind of things. Does it have any rails so that it doesn't lift the Valar from Tolkien, say?

 

[Clint_L]

By any cognitive measure, yes, it's random. All LLMs do is chase correlations in language. They aren't worldbuilding with any sense of how various themes and elements might fit together, or how they're playing into or breaking genre tropes. They're literally just sticking words together.

You don't have to believe me, but I've been interviewing AI/ML professors about this stuff for the past few weeks. This is how people who make stuff explain it.



If ChatGPT's output is any measure, then yes, what humans are doing is very special. These things are incredibly impressive from a technical benchmark perspective, but the stuff they produce is verbal cardboard.

Just, like, write your own text.

I am a teacher with decades of experience. I have a masters degree in literature, with a secondary reader from linguistics, and am ABD on my doctorate. I have assessed thousands of students and many thousands of examples of student writing, from ages 12 into adulthood. I am extensively trained in various methods of assessing writing. I currently teach IB Theory of Knowledge, IB Language and Literature, and Creative Writing.

I think I am reasonably expert in assessing young human adult writing by any objective standard.

Here is something that I got ChatGPT to write:

write a memorandum for the office on not drinking my milk from the office fridge

Memorandum

To: Office Staff

From: [Your Name] Subject: Office Fridge Usage

I am writing to remind all staff members that the office fridge is for communal use and should be treated with respect. Specifically, I would like to bring attention to the fact that several individuals have been seen taking and consuming milk from the office fridge without replacing it.

Please be mindful of your actions and refrain from consuming communal items without replacing them. If you do consume something, please ensure that it is replaced or restocked as soon as possible.

We all have a responsibility to maintain a clean and organized office environment, and this includes being conscientious about our use of the office fridge.

Thank you for your cooperation and understanding.

Sincerely, [Your Name]


can you make it more passive aggressive and add sarcasm

Memorandum

To: Office Staff From: [Your Name] Subject: Office Fridge Usage (or lack thereof)

Greetings esteemed colleagues,

I would like to take a moment to address the elephant in the room, or should I say, the missing milk in the office fridge. It appears that some individuals have been under the impression that the office fridge is their personal dairy cow, and have been helping themselves to the milk without a second thought about replacing it for the rest of the office to enjoy.

I mean, who needs milk anyways, right? It's not like it's an essential part of a balanced breakfast or anything. I'm sure we can all survive on dry cereal and black coffee.

But, in all seriousness, can we please show some common courtesy and respect for our communal spaces and items? If you do decide to indulge in the office's milk supply, please make sure to replace it as soon as possible.

Let's all work together to maintain a clean and organized office environment, and not just for the sake of the milk, but for the overall well-being of the office.

Sincerely, [Your Name] (who is now officially lactose intolerant)

That second version of the memo is fantastic. I would be proud to have come up with that final double entendre. If you are looking at that and seeing "verbal cardboard," I don't know what to tell you.

It turns out that writing can be very effectively automated. Which probably means that most of it was not as special as we like to think. We should see this as an opportunity - with AIs assisting us in writing, the way they already assist in innumerable other tasks, how far can we continue to push human creativity?

Edit: I also got it to produce the last version of the memo in Swedish. No problem.

 

[Grendel_Khan]

In the OP all of the deities chosen look to be public domain kind of things. Does it have any rails so that it doesn't lift the Valar from Tolkien, say?

This is a great question, and a little hard to address because the technology is such a moving target right now, but officially all of these models (that we have public access to) are supposed to be trained on content that's in the public domain. But the scale of the training, and the fact that it's based on scraping the internet, means that they seem to catch a fair amount of stuff that they aren't supposed to. Do the models' makers care about that, if it makes the outputs more successful or interesting? I'd say no, but they pretend to be very concerned and interested in violating copyrights and such. Sure thing.

But unless there was a change over the past week, you can absolutely ask ChatGPT to write you stuff related to Tolkien. I've seen an example of it generating a letter from the orcs to Sauron in the style of a union-related complaint.

 

[Grendel_Khan]

That second version of the memo is fantastic. I would be proud to have come up with that final double entendre. If you are looking at that and seeing "verbal cardboard," I don't know what to tell you.

It turns out that writing can be very effectively automated. Which probably means that most of it was not as special as we like to think. We should see this as an opportunity - with AIs assisting us in writing, the way they already assist in innumerable other tasks, how far can we continue to push human creativity?

From a technical standpoint that's a great example of just how stunning these models suddenly are. Their improvement over the last five years or so is amazing. There's no doubt that these things are basically blowing the Turing test out of the water, in these specific text-generation tasks.

As writing, though, that is absolutely, as the kids say, mid. Just hackneyed "joke" after joke. As with every long-ish ChatGPT output I read, I wish I could get back the time I spent reading it, because they never produce anything interesting or original.

 

[Cadence]

From a technical standpoint that's a great example of just how stunning these models suddenly are. Their improvement over the last five years or so is amazing. There's no doubt that these things are basically blowing the Turing test out of the water, in these specific text-generation tasks.

As writing, though, that is absolutely, as the kids say, mid. Just hackneyed "joke" after joke. As with every long-ish ChatGPT output I read, I wish I could get back the time I spent reading it, because they never produce anything interesting or original.

RE: Anything original

I mean, it is just remixing what it's given in a fairly sophisticated way, right? Is that what we do (King Lear vs. Oedipus at Colonus)? Is there anything new under the sun?

Now I'm thinking Shakespeare and wondering how it does at making up new words.

I'm also curious what it would do if asked about people and things with multiple spellings, say a famous mathemetician with one of the following: Chebyshev, Tchebichef, Tchebycheff, Tchebycheff, Tchébycheff, Tchébycheff, Tschebychew, Tschebyscheff, Tschebyscheff, Tschebyschew, Tschebyschew, Čebišev.

I'm guessing it will have run across a website with most of those, but then I'm thinking of genealogy things where no one has made a concordance yet. (I'm wondering how a family tree making AI would do given how awful and half-arsed so many people are when using them - so that an uncurated training sample would be a train wreck).

 

[Bent Goblin]

A few weeks ago, I experimented with asking ChatGPT to be my dungeon master. I asked it:

I want you to be my Dungeon Master and run a small adventure for me. I will be playing a halfling rogue called "Bent". The adventure begins when Bent wakes up in a cell and cannot rember how he got there -- all he can tell that the cell seems to be in the cellar of a larger house or castle, because through the window of the cell he can hear faint sounds that seem to stem from a kitchen where a feast is being prepared. We will play like this: You describe the scene to me and ask me what I want to do. You do not presume any activities I might undertake, ask me and I will tell you how Bent reacts to what you describe. I will describe what Bent does by writing in first person. Let's start, please describe to me how Bent wakes up in the cell and what he sees.

I was excited to see that ChatGPT threw in a mysterious symbol that was in the cell, but interactions with the symbol went nowhere. So after we were finished with the small adventure, I asked ChatGPT for the backstory. The response was

The symbols you encountered in your adventure are fictional and were included as part of the story. They were meant to add a sense of mystery and intrigue to the setting and to give you, the player, something to think about and investigate.
The symbols had no specific meaning, they were created to add depth to the story and to create an atmosphere of uncertainty and danger. Their purpose was to make you question what was happening and to add an element of mystery to the story. As a dungeon master, I was trying to build an immersive experience for you and the symbols were a tool to achieve that.

In my next attempt, I first prompted ChatGPT to come up with some elements of a backstory and then chose some of them randomly for an adventure. I asked for

• seven villains with a secret agenda that are respected members of society
• a reason, why each of the villains might give a dinner party
• possible evidence for each villain's nefarious activities
• ten possible guests at such a dinner party
• seven henchmen, which might be working for any of the villains

Then I told ChatGPT the following:

Now I want you to act as dungeon master in a roleplaying game. Pick at random one of the villains and five of the guests and two of the henchman. The setting of the game will be a dinner party given by the villain, to which he has invited the five random guests and me. I will play Bent, a halfling rogue and con man who is impersonating an ambassador from a faraway dukedom and who has been tasked with infiltrating the villains home, find prove of his nefarious activities and escape from the home with the evidence. The evidence should be located in one of the ten rooms of the house. We will play like this: You describe the scene to me and ask me what I want to do. You do not presume any activities I might undertake, ask me and I will tell you how Bent reacts to what you describe. I will describe what Bent does by writing in first person. Let us start the game: describe the scene when I arrive at the villains home and the door is opened by one of his henchmen.

The adventure that ensued was a bit better than the first one. Obviously still a long, long cry from having an human as dungeon master, but it shows how prompts result in better output by ChatGPT also for using it as a DMs tool (or as DM itself, in this case).

If you are interested, you can read the transcripts of my first try and my second try in full, but to be honest, I concur with what @Grendel_Khan just said about wishing one could get back the time spent with reading longish ChatGPT output. So you might want to skip reading the transcripts and better try out ChatGPT as dungeon master yourself if you are interested into how that goes.