D&D General Deep Thoughts on AI- The Rise of DM 9000

[Hussar]

I think there are certainly worse places to train an AI.

I can certainly foresee the day when someone does take the time to train an AI dm. Heck even if it was just running npc’s alongside a live dm, that would be fantastic.

There really are a lot of possibilities here.

 

[Fanaelialae]

I bet if we feed the AI Enworld, it would become the best DM ever! All of our collective knowledge poured into it would level it up to twenty instantly. It would be like giving a fighter a +6 sword!

No, by and large all ENWorld would do is give the AI the ability to dispense advice on how to game. The AI lacks comprehension though, and can't actually put any of it into practice in the way that a human might. Even Actual Plays wouldn't really help, since an AI trained on those would probably just try to tell you a story (rather than run a game for you).

 

[Cadence]

I bet if we feed the AI Enworld, it would become the best DM ever! All of our collective knowledge poured into it would level it up to twenty instantly. It would be like giving a fighter a +6 sword!

I bet we could train it really well to argue that whatever we seemed to like in a game wasn't really what we liked....

 

[Dausuul]

If I were tasked with using a large language model to build an AI DM, this would be my approach:

1. Create a database which can hold "facts about the game world." These would be in some standardized format (I think there are such formats for knowledge bases). The important thing is that each fact is tagged with one or more topics.

2. Train a version of the model -- call it the "fact-finder" -- to identify relevant topics from an ongoing session. Each time a player says something, the fact-finder determines what topics are relevant now.

3. Train a second version of the model -- the actual "DM" -- which should emit two things: New/updated facts, and text to the players. As input, it takes a) the transcript of the session so far, and b) all of the facts identified by the fact-finder on the relevant topics.

4. To keep the list of facts from growing out of control, it might be necessary to extract most of the game mechanics into a separate engine, a purely programmatic one. In that case the "DM" model will also need to emit orders to this engine ("Have Bob make a greatsword attack against Roper #2"). The engine would then update the fact database when mechanical events result in "visible" effects, like Roper #2 dying.

If this approach worked, the fact database would keep the DM-bot's improvisations grounded in a consistent fictional reality. It might work. I can also see a ton of ways it could fail, though, and it certainly would not be easy to implement.

To be honest, I’m not super interested in metaphysical arguments based on opinions about what is feasible. I’m more interested in actual experiments.

It's perfectly reasonable to insist on experimental results... but then you can't just wave off GPT's failures and glitches during those experiments. You have to demonstrate that they can be fixed in a programmatic way (i.e., one that doesn't rely on a human devising ad hoc solutions to nudge the bot back on track).

 

[Stalker0]

No, by and large all ENWorld would do is give the AI the ability to dispense advice on how to game. The AI lacks comprehension though, and can't actually put any of it into practice in the way that a human might. Even Actual Plays wouldn't really help, since an AI trained on those would probably just try to tell you a story (rather than run a game for you).

An enworld trained AI would start off sharp as a tack, and then dive so far off track you’ll never find the plot again

 

[Clint_L]

It's perfectly reasonable to insist on experimental results... but then you can't just wave off GPT's failures and glitches during those experiments. You have to demonstrate that they can be fixed in a programmatic way (i.e., one that doesn't rely on a human devising ad hoc solutions to nudge the bot back on track).

No, I don't have to demonstrate that at all. Because I am not interested in whether or not ChatGPT is a true AI or can perfectly emulate a human being as a DM. I am interested in what it can potentially be used to do. In fact, I have specified in this thread (and in real life to my students) that it should be looked at as an assistant to human beings, not a replacement.

So I am not waving off failures. I am interested in them, and interested in which ones are easily solvable and which ones are not. I don't have an agenda other than finding out how this thing can be used.

Also, I note that human DMs often have glitches and require another human being to get them back on track.

 

[AdmundfortGeographer]

You know, I might be more interested in running a group of AI players than playing under an AI DM. Tuning the kinds of players would be a fun experiment.

 

[Guest 7030100]

This story is tangentially related, I think:

Sci-fi becomes real as renowned magazine closes submissions due to AI writers

Clarkesworld wrestles with flood of machine-made submissions—over 500 in Feb. alone.

arstechnica.com

The sci-fi magazine, Clarke's World, is temprarily suspending its submissions process until it can get a handle on all the AI spam. The article also mentions a rise in the number of AI generated titles on Amazon books, etc, as well as "get rich quick" youtube clickbait on the topic.

Makes me wonder if &/or how many mostly completely AI generated publications are popping up on DTRPG and the like. As some here have noted, there are quite a few areas where AI content would be helpful (or maybe even preferable) to a human-written publication, things like the ubiquitous books of encounter tables, d100 lists, fantasy market inventories, and the like.

I imagine we'll see a massive glut of this sort of uncreative content in the near future, at least until these AIs are routinely incorporated into apps aimed specifically at RPGing.

 

[Hussar]

So, I finally got around to fiddling with ChatGPT today. WOW.

Here's an example:
---------------

My input:create a unique monster for dungeons and dragons, complete with all game statistics, that is a CR 7 Aberration. I also need a history for this monster.

ChatGPT:
Sure, here's a unique monster for Dungeons and Dragons that is a CR 7 aberration, along with game statistics and a history.

Name: Gloomwraith

Description: The Gloomwraith is a creature of shadow and mist, with a twisted, humanoid form that seems to defy the laws of physics. Its body is insubstantial, made of a dark, smoky substance that can seep through solid objects. It has glowing, red eyes that seem to burn with malevolent intelligence. Gloomwraiths are often encountered in dark, abandoned places such as tombs, ruins, and forgotten cities.

Game Statistics:

AC: 16 (natural armor)

Hit Points: 120 (16d10 + 32)

Speed: 30 ft., fly 60 ft.

Abilities:

• Str: 12 (+1)
• Dex: 18 (+4)
• Con: 14 (+2)
• Int: 14 (+2)
• Wis: 16 (+3)
• Cha: 20 (+5)

Skills: Stealth +8, Perception +7

Damage Resistances: Cold, Fire, Lightning

Damage Immunities: Poison

Condition Immunities: Poisoned

Senses: Darkvision 120 ft., passive Perception 17

Languages: Deep Speech, Undercommon

Special Abilities:

• Shadowmeld: As an action, the Gloomwraith can transform into a shadowy mist and move through solid objects such as walls and doors. While in this form, it is immune to all damage and effects.
• Miasmic Aura: Any creature within 20 feet of the Gloomwraith at the start of its turn must make a DC 16 Constitution saving throw or take 2d8 poison damage and be poisoned for 1 minute.

Actions:

• Shadow Claws: Melee Weapon Attack: +8 to hit, reach 5 ft., one target. Hit: 11 (2d6 + 4) slashing damage plus 10 (3d6) necrotic damage.
• Mist Blast: Ranged Spell Attack: +8 to hit, range 60 ft., one target. Hit: 18 (4d8) cold damage and the target's speed is reduced by 10 feet until the end of the Gloomwraith's next turn.
• Mist Form: The Gloomwraith transforms into a cloud of mist and moves up to 60 feet. While in mist form, it is immune to all damage and effects.

Challenge Rating: 7 (2,900 XP)

History:

According to legend, the Gloomwraith was once a powerful sorcerer who became obsessed with the idea of immortality. In his quest for eternal life, he delved into dark, forbidden magic and made a terrible bargain with a powerful entity from the Shadowfell. The sorcerer's body was transformed into a being of shadow and mist, and he became the first Gloomwraith.

Now, the Gloomwraith roams the world, seeking knowledge and power to increase its already considerable abilities. It is said that the Gloomwraith has the power to control shadows and bend them to its will, making it a feared and dangerous opponent. Few who have encountered the Gloomwraith have lived to tell the tale, and even fewer have managed to defeat it.

------

Just eyeballing that, that looks like a pretty damn good monster. Holy crap. It repeated the Mist Form effect, but, other than that, I've seen far worse designed monsters.

 

[Hussar]

Noodling around with ChatGPT, I could easily see how this would MASSIVELY help when adventure writing. Banging out bespoke monsters, room descriptions, that sort of thing. Sure, yes, you have to go back and fix it, fair enough. But, wed this to an art generator and a map generator, and you'd be able to bang out a fully fledged adventure, complete with stat blocks REALLY quickly.

Heck, even as a random encounter generator, it would be fantastic. There's a lot of grunt work that the AI can do quickly.