D&D General Would you buy an AI-generated Castle Greyhawk "by" Gary Gygax?" Should you?

[Maggan]

I wonder if Chatgpt has already been trained on Gary's text as it might be in the corpus of text that has been scraped from the Web.

Google has partly trained its AI on posts made on EN World, so I wouldn't be surprised if other AIs have done so too. So I assume Gygax's online style is in there, although to "summon it" you might have to invoke "in the style of Col.Pladoh".

https://www.enworld.org/threads/en-world-ranked-11-248-out-of-15-million-sites-used-to-train-ai.697275/

 

[Whizbang Dustyboots]

I wonder if Chatgpt has already been trained on Gary's text as it might be in the corpus of text that has been scraped from the Web.

The fact that people on this thread could get it talk convincingly like Gary suggests it was.

 

[Clint_L]

Current AI models can never reproduce human creativity, because they are statistical models on steroids. So no, I am not interested in novels or D&D adventures that are produced based on empirical probability.

What if they are good? Also, how do you know that human creativity is not essentially "statistical models on steroids"?

A short side rant: We don't even know how human brains work exactly, we are not even close to build AIs that are coming close to human brains.

Both these statements are true. You might add that we don't fully understand how generative AI works, as well, and that considerable ongoing research is showing that generative AI exceeds what were thought to be its parameters in a number of surprising ways.

Although we don't understand how human cognition works, it's not magical. It's a physical process happening in a bioelectrical machine. An incredibly complex one; much, much more complicated than anything else in the known universe.

I teach Theory of Knowledge, Creative Writing, and Language&Literature, with an MA (and ABD on my doctorate in the latter). I have decades of classroom experience. One reason generative AI writing has shaken me, and many others, is that it has exposed potential problems with how we have approached human writing. Because, like it or not, generative AI is very good at certain writing tasks; far more so than even competent humans. This is not my opinion; generative AI has done much better on a great number of standardized written tests than most humans can. Much as a calculator can perform mathematical computations much better and, in particular, faster than humans, generative AI can already do basic writing tasks much better and, in particular, faster than humans.

As it turns out, there is tremendous value in being able to do mathematical computations accurately and quickly, and there is tremendous value in being able to do basic writing tasks accurately (more or less; writing is generally more subjective than math) and quickly. For example, my BFF's law firm now employs far less office staff because a great deal of the rote writing jobs are now prepared by AI and then reviewed by a lawyer, rather than being written by clerks and reviewed by a lawyer.

On top of that, this suggests that much of the writing that humans do might be basically repetition. This is what generative AI is forcing teachers to confront. I spend years teaching students how to write essays. I am spending a good chunk of my winter break marking ToK Essays. These are from IB Diploma candidates in their final year, and the average marks I am giving on the first draft are D/C- (they'll get there, but plenty of revision is needed!). And a huge amount of my time is spent on helping students express their ideas more clearly; basically, teaching them to use grammar and syntax to convey a complicated thought.

Generative AI doesn't have the complicated thought. But it does have a very good handle on the grammar and syntax! So in a way, I feel like math teachers probably felt in the 1960s, when they spent most of their time trying to teach fundamental rules just so that a few students could one day go on to do interesting things in the field.

What I am getting at is that, from my perspective, generative AI offers an opportunity to accelerate human learning in the arts similar to what calculators did for mathematics. Right now, the education system strongly rewards brains that are very good at quickly (by human standards) picking up the repetitive aspects of reading and writing, because that's the model of the mind that was prioritized a few centuries ago when public education was being invented. And this has led to many, many people not getting much from education, being left feeling stupid and useless, and often being shunted away into lower socio-economic strata. I think it has created a vast waste of human potential. Generative AI poses an existential challenge to the fundamental premises of modern education, and I am hoping it will be the catalyst we need to fundamentally shake up a model of the human mind that is ever more outdated.

 

[Whizbang Dustyboots]

Going along with what @Clint_L said, I write professionally. (That may not always be clear in my typo-ridden posts here.) I could not tell you what a gerund is, beyond knowing it's one part of speech playing the role of another part, or many other grammatical rules, as my despairing AP English teacher could attest.

But on the job, I write famously clean copy because I read a great deal and write a great deal. In other words, the machine in my skull was trained the same way as generative AI is.

Now, I don't write trying to guess what the next word in a sentence is most likely to be, based on other writing on the same subject, as generative AI does. But my grammar and vocabulary are largely the product of having read widely and having my writing challenged by adversarial AI -- my editors, some of whom have been extremely adversarial over the years -- rejecting my work, forcing me to improve my work so as to pass the next challenge. (This is the same system that's been used to make deep fakes get so good: One AI produces an image, the second one tries to detect if it's a fake, and the first one keeps refining until it can fake out the second AI.)

The route may be somewhat different -- I have intention that current AI models don't -- but at the end of the day, I think it's reaching a number of the same landmarks nonetheless. And I think it's worth remembering that it's only been a year since most of us have had the ability to access AI as members of the public and it's made huge strides in that time alone.

Statements declaring that AI "can never" or "will never" do something are pretty bold. We don't have to be excited about it -- I am confident AI is going to eliminate a lot of jobs in my industry, maybe even mine -- but counting it out is denialism, IMO.

 

[Cergorach]

I'm pretty sure you could get a LLM to produce better adventures the 90% of the garbage that's on DTRPG. But what people seem to forget is that it also depends heavily on the one providing the prompts. Just like with googling, the quality of the search results depends heavily on the google-fu skill of the user. And if people have thought me anything the last 25+ years, people's google-fu skill suck! It's a skill like any other, not everyone has the skill and most that have it have it at a very low level. Prompting is relatively new and with the rate of advancement of LLM, things change so often, so people have to spend a lot of time keeping their prompting skills even relevant.

LLM is automation, just like we've been doing for decades. Some jobs might disappear, but it's mostly resulting in fewer people doing more, cheaper. In writing we have the equivalent of dime store novels, cheap romance novels vs. what we call literature. Not everyone reads 'literature' all the time, there's a huge market for literary 'crap' out there. And if we have to believe things closer at home, getting back to the mountains of crap RPG supplements and adventures on a DTRPG, people still make very low quality, shoddy products that actually sell. Why complain if with the right person at the helm a LLM can make better product then the oodles of low quality stuff that we currently have? And I've seen way better descriptions, texts, etc. from LLM then from (apparently) way too many writers.

"Would I buy an AI generated Castle Greyhawk?" What others have said, it depends, if it's any good. I've seen fan projects from decades ago that I found a better read then most of the official Greyhawk products that I had access to at the time... Greyhawk is also a WotC IP/trademark, so it would probably would have to be a product made by WotC and I've not bought many D&D books in the 5E era (only PHB/DMG/MM)...

Also with current LLM (and I don't see that changing), it lies, it lies extremely well. So you would still need a very knowledgable person (on the IP) prompting a LLM for creating a Castle Greyhawk. ChatGPT could swear up and down that Elminster is the resident royal wizard at Castle Greyhawk and that it's located in Monaco...

 

[Retros_x]

What if they are good? Also, how do you know that human creativity is not essentially "statistical models on steroids"?

They aren't good. At least not for me. Because they are not innovative, they lack smart ideas. They are normalized, generic. Because human creativity doesn't work on mathematical functions and can't be mapped to any mathematical functions. Which brings me to the second point.

How do I know that human creativity is not essentially statistical models on steroids? Because humans invented them. They all have a target function underlying them. They have predictable outcomes and are designed for very specific goals and cannot be used outside of them. They use tons of training data. We did not designed human creativity. We cannot find underlying mathematical functions in the human brain, because we are not even close to the complexity any approximation would take. We have moments of sudden inspiration when we stand under the shower. We can change the rules, think outside the box. We don't need a tiny fraction of the training data AIs need.

Here is a famous blog article by Andrej Karpathy (computer scientist, specialised on deep learning AI) in which he explains what information the human brain can get from looking for a few seconds at a simple picture: https://karpathy.github.io/2012/10/22/state-of-computer-vision/
This article is from 2012, but it still holds up. Image recognition AIs still can't do most of this and we are still really, really far away.

That lack of creative thinking in AIs is even true for mathematic itself! Because, yes computers are much better at calculating than humans, but if you would give a modern AI all the base rules and axioms of finite maths and basic arithmetic, there is not a single chance it would invent modern mathematics itself, because it is not able to abstract from these rules to invent new ones like humans did. The same is true for any creative thinking, because it is not calculating and not based on calculating.

We know for a fact that computers can calculate much faster than any human being. If human creativity and thinking would be a statistical model, modern computers would've already "calculated" it. But they haven't and never will, unless we find much, much, much more complex new models that we can't even imagine yet. The models modern AIs today are using were invented in the 1950's so I wouldn't hope on fast iterations on that.

Both these statements are true. You might add that we don't fully understand how generative AI works, as well, and that considerable ongoing research is showing that generative AI exceeds what were thought to be its parameters in a number of surprising ways.

Its different. Because we understand the principles of generative AI much more precise than we understand the principles of human creativity. We understand it enough to actually design and build them. We know the mathematical functions and statistical methodologies that are used in them. We know the precise formulas and equations to do so, because we wrote them ourselves (again, in the 1950s). Human thinking in comparision - we aren't even close. Yes we know a lot of the surrounding "framework" of the biochemical reactions that are needed for it, but we don't know anything about the logic behind them, we can't write any equation or function that maps human thinking.

Although we don't understand how human cognition works, it's not magical. It's a physical process happening in a bioelectrical machine. An incredibly complex one; much, much more complicated than anything else in the known universe.

Thats like saying a computer is a physical process in a electrical machine. Its like really, really, really basic. Its also not true. Humans tend to compare themselves to the current dominant technology. Its the body-as-a-machine metaphor, and before computers telegram networks, or steam machines were used for example as metaphors for the human body. But they are nothing more than a metaphor. That doesn't mean they don't have uses - simplifications like that can help us understand certain processes. But that doesn't mean the lung is actual a smith's bellows or the brain is actual an electrical machine.
https://www.theguardian.com/science...er-neuroscience-neural-networks-consciousness

https://www.infoq.com/articles/brain-not-computer/#:~:text=Key Takeaways,not to a malfunctioning brain.

Because, like it or not, generative AI is very good at certain writing tasks; far more so competent humans.

I never doubted that. But its really bad at all the other tasks outside of that certainty and writing creative texts is one of those. Writers who write soulless ad texts made for search engine optimization or artists who make stock art - they can be easily replaced by generative AI. But I think we can all agree that these are not quite the pinnacle of human creativity.

But my grammar and vocabulary are largely the product of having read widely and having my writing challenged by adversarial AI -- my editors, some of whom have been extremely adversarial over the years -- rejecting my work, forcing me to improve my work so as to pass the next challenge.

Believe me, the extensiveness of all the books you read and the effort your editiors put to give you a hard time? Its nothing compared to the extensiveness of data AI processes. And yet, if you are not completely untalented, I bet your work will be more evocative than any text ChatGPT could generate.

Statements declaring that AI "can never" or "will never" do something are pretty bold. We don't have to be excited about it -- I am confident AI is going to mean a lot of jobs in my industry, maybe even mine -- but counting it out is denialism, IMO.

I wouldn't say never, but I would say "the chance it will happen in our lifetime is so close to 0, you could almost say never". And I am not saying it out of denialism, I say it because I am a computer scientist, who actually learned a little bit about what AI can do, how it works and more importantly its limitations. I would never act like I am an expert, but I learned the foundational mathematical theory behind most modern ML algorithms and implemented some basic implementations. I've also read some opinions of actual experts on this. A very short and sweet example by the AI team of Google:

"While AI systems are nearing or outperforming humanbeings at increasingly complex tasks [...], they remain narrow and brittle, and lack true agency or creativity. [...], machines with human intelligence remain a long way off." Bolded by me to come back to the topic at hand a bit.

https://ai.google/static/documents/exploring-6-myths.pdf

If some want to read an excellent laymans introduction in to AI where the principles, chances, limitations, dangers etc. are described without dabbling too much into technical details and mathematics, I can highly recommend the book "Artificial Intelligence" by Melanie Mitchell.

 

[Whizbang Dustyboots]

They aren't good. At least not for me. Because they are not innovative, they lack smart ideas. They are normalized, generic. Because human creativity doesn't work on mathematical functions and can't be mapped to any mathematical functions.

If we choose to program a hypothetical dungeon creator in the most lazy way possible (like saying "hey, ChatGPT, create a 5E D&D tier 1 dungeon for me"), sure.

But if we did something like feed it the Tome of Adventure Design, which famously has some extremely wild and crazy prompts that, at first glance, don't have anything to do with one another, or have it use something like the Iron DM system here at ENWorld, which does something similar, I think the end results of having to incorporate extremely random elements would be so close to true human creativity as to be indistinguishable. Because what is "creativity," really, other than applying unexpected approaches to conventional problems?

And it's a mistake to assume no one else has talked to AI experts about this stuff, for the record.

 

[Clint_L]

They aren't good. At least not for me. Because they are not innovative, they lack smart ideas. They are normalized, generic. Because human creativity doesn't work on mathematical functions and can't be mapped to any mathematical functions. Which brings me to the second point.

As you previously pointed out, we don't know exactly how the human brain works. As I pointed out, it isn't magical. So the claim that it can't be mapped to mathematical functions is pretty confident.

How do I know that human creativity is not essentially statistical models on steroids? Because humans invented them. They all have a target function underlying them. They have predictable outcomes and are designed for very specific goals and cannot be used outside of them. They use tons of training data. We did not designed human creativity. We cannot find underlying mathematical functions in the human brain, because we are not even close to the complexity any approximation would take. We have moments of sudden inspiration when we stand under the shower. We can change the rules, think outside the box. We don't need a tiny fraction of the training data AIs need.

AI hasn't needed the billions of years of evolution that we've had. And statements like "we can change the rules, think outside the box," those are aphorisms. What do they mean, really?

We don't know how the human brain and human mind work, though we are developing increasingly complex theories. We don't entirely know how generative AI works, though we understand how it was built. We do know that it appears to be evolving unexpected abilities. So I'm going to be cautious about declaring what it can or can't do...or what humans can or can't do.

That lack of creative thinking in AIs is even true for mathematic itself! Because, yes computers are much better at calculating than humans, but if you would give a modern AI all the base rules and axioms of finite maths and basic arithmetic, there is not a single chance it would invent modern mathematics itself, because it is not able to abstract from these rules to invent new ones like humans did. The same is true for any creative thinking, because it is not calculating and not based on calculating.

AIs have been discovering new proofs and axioms in mathematics. Famously.

We know for a fact that computers can calculate much faster than any human being. If human creativity and thinking would be a statistical model, modern computers would've already "calculated" it.

I think that human creativity probably involves a great deal of statistical modelling at an unconscious level, but it would be very reductionist to claim that there is nothing more. However, I don't exactly follow your point here.

But they haven't and never will, unless we find much, much, much more complex new models that we can't even imagine yet. The models modern AIs today are using were invented in the 1950's so I wouldn't hope on fast iterations on that.

Its different. Because we understand the principles of generative AI much more precise than we understand the principles of human creativity. We understand it enough to actually design and build them. We know the mathematical functions and statistical methodologies that are used in them. We know the precise formulas and equations to do so, because we wrote them ourselves (again, in the 1950s). Human thinking in comparision - we aren't even close. Yes we know a lot of the surrounding "framework" of the biochemical reactions that are needed for it, but we don't know anything about the logic behind them, we can't write any equation or function that maps human thinking.

I'm not sure what your point is. I already agreed that the human brain is astoundingly complex. I am also aware that AIs function profoundly differently. It does not then follow that AIs cannot potentially exceed the human brain at various tasks. In fact, we already know they can. You seem to think that one particular task, "creativity," cannot be achieved, but you haven't defined what that task is, beyond vague aphorisms.

Thats like saying a computer is a physical process in an electrical machine. It's like really, really, really basic. It's also not true. Humans tend to compare themselves to the current dominant technology. It's the body-as-a-machine metaphor, and before computers telegram networks, or steam machines were used for example as metaphors for the human body. But they are nothing more than a metaphor. That doesn't mean they don't have uses - simplifications like that can help us understand certain processes. But that doesn't mean the lung is actual a smith's bellows or the brain is actual an electrical machine.
https://www.theguardian.com/science...er-neuroscience-neural-networks-consciousness

Thank you for explaining what a metaphor is. I hadn't realized.

But its really bad at all the other tasks outside of that certainty and writing creative texts is one of those. Writers who write soulless ad texts made for search engine optimization or artists who make stock art - they can be easily replaced by generative AI. But I think we can all agree that these are not quite the pinnacle of human creativity.

"Bad" is a qualitative judgment. I have gotten ChatGPT to write things that were quite good, IMO. The Dall-E thread is full of images that I consider quite good. What does "soulless" mean? Your argument keeps coming down to undefinable terms.

I wouldn't say never, but I would say "the chance it will happen in our lifetime is so close to 0, you could almost say never". And I am not saying it out of denialism, I say it because I am a computer scientist, who actually learned a little bit about what AI can do, how it works and more importantly its limitations. I would never act like I am an expert, but I learned the foundational mathematical theory behind most modern ML algorithms and implemented some basic implementations. I've also read some opinions of actual experts on this. A very short and sweet example by the AI team of Google:

"While AI systems are nearing or outperforming humanbeings at increasingly complex tasks [...], they remain narrow and brittle, and lack true agency or creativity. [...], machines with human intelligence remain a long way off." Bolded by me to come back to the topic at hand a bit.

https://ai.google/static/documents/exploring-6-myths.pdf

If some want to read an excellent laymans introduction in to AI where the principles, chances, limitations, dangers etc. are described without dabbling too much into technical details and mathematics, I can highly recommend the book "Artificial Intelligence" by Melanie Mitchell.

This is...pretty condescending.

 

[M.T. Black]

To get back to @Whizbang Dustyboots original question, we actually have the map and key to level 1 of Castle Greyhawk (see below, sourced from Delta's D&D hotspot).

So, I fed the key into ChatGPT 4 and asked it to write it up, Gygax style, using this prompt:

the table above is a dungeon key for a Dungeon called Castle Greyhawk Level 1. I want you to pretend to be Gary Gygax, and you are turning this dungeon key into a full adventure. Please write out rooms 1 to 5 fully in the voice and style of Gary Gygax. Each room should have a number, a name, some read-aloud text, and then a description of the room and contents, including monsters (use the average), what they are doing, treasure and value, and other objects. bold the monster name.

Below are the rooms that ChatGPT created using Gygax's key. I think the prompt could be tuned far better, but the output seems ok to me. I've seen worse by humans! It is a bit bare-bones. It would be useful to feed the tool a model example of how Gary wrote a room.



Room 1: The Scowling Den

Your torchlight flickers upon the uneven walls of this dank chamber, revealing scurrying shapes that chitter and jeer in the gloom. They brandish spears no taller than a child, but with a menace that belies their stature.

Within this lair dwell Kobolds, a pack of 12 by average count, each as craven as they are cruel. These diminutive dragonkin are currently honing their weapons and setting snares for the unwary. Their eyes gleam with greed and malice, for they guard a relic of old: a dust-colored silver mirror, set in strange metal and etched with forgotten runes. This mirror, if hauled back to the realms of sunlight, could fetch a sum of 750 gold pieces.

Room 2: The Goblin Warren

The stench of refuse and the sounds of bickering greet you as you enter. In the flickering shadow, green-skinned figures squabble over a pile of pilfered loot.

This chamber is infested with Goblins, nine in number, their skin the color of moldy leaves. They are engaged in the division of spoils from their latest raid, oblivious to the world beyond their squabbles. Among their trove is a sum of 10 gold pieces — not a princely sum, but a start. And goblins, though easily scattered, can be vicious when cornered.

Room 3: Lair of the Giant Vermin

The crunch of bone underfoot betrays the chamber ahead. Shadows skitter across the floor, and the air is filled with the musk of feral hunger.

This place is claimed by Giant Rats, seven in all, their fur matted and eyes alight with hunger. They feast upon the remnants of less fortunate adventurers and guard a treasured potion of strength with the ferocity of beasts twice their size. This potion, lying amidst the rubble, could prove invaluable and is valued at 50 gold pieces.

Room 4: The Brigand's Holdfast

The clink of coin and the rustle of maps can be heard as you peer into a room that bears the mark of organized crime. A band of armed men and women look up from their plotting, hands inching towards weapons.

The room is the domain of the Bandits and their warrior leader, a formidable group of 13. These rogues are in the midst of dividing their ill-gotten gains and planning their next foray. Among their cache is a shield of fine make, a potion of undetected control, and a sum of 1,000 gold pieces. The bandits are cunning and will fight with strategy and guile.

Room 5: The Centipede's Nest

The sound of your steps is drowned out by the chittering and clattering of countless legs on stone. The room before you writhes with long, segmented bodies coiling in and out of the rock.

Occupying this room are Giant Centipedes, six in their number, a writhing mass of venom and hunger. They are currently feeding on a patch of fungus growing in the damp corner of their lair. There is no treasure to be found here, save for the venom that coats their mandibles — a dangerous prize for those who know its worth.

 

[M.T. Black]

If anyone wants to try a prompt to turn the key into some rooms, here is the raw key data:

Design Key: Castle Greyhawk, Level 1

1,-,Kobolds,9-16,5,63,-,"Dust colored silver mirror, 750 gp"
2,-,Goblins,7-12,8,71,-,10 gp
3,-,Giant Rats,3-12,5,38,-,Potion of strength
4,-,"Bandits, warrior leader",10-16,150,-,-,"+1 shield, potion of undetected control, 1000 gp"
5,-,Giant Centipedes,5-8,6,39,-,-
6,-,Hobgoblins,2-3,15,38,-,"+1 battle axe, gem 400-800 gp"
7,-,"Giant black widow, skeleton in plate",2,106,212,-,50 gp
8,"Iron chest, empty (false bottom)",Needle trap,-,-,-,"6 gems @200",1200 gp
9,-,Orcs,3-7,50,50,"Chest w/8 flasks oil",30 sp, 8 gp
10,"Water seepage",-,-,-,-,-,(drink!)
11,-,Giant Constrictor,1,650,650,"Gems inside (1-4) or nothing 50/50",-
12,"Rustling corridor (no bats)",-,0,-,-,-,-
13,-,Skeletons,2-5,10,35,"One piece of jewelry, in pile, d10 x 30",-
14,Rustling,Giant Bats,1-2,175,263,-,-
15,-,"Giant Stink Beetle (2 attack in 5')",1,275,275,"Nothing/10-100 sp & 1-10 gp",-
16,-,Berserkers,3,19,57,"Ring of protection, 20-40 gp each, Idol with curse (-1/3d)",90 gp
17,-,"Seer (chaotic, light, charm person)",1,25,25,"+1 dagger, spellbooks, 1-20 gp",-
18,-,"Evil adept (cause light wounds) & gnoll",2,44,54,"Staff of healing (6 ch.)",-