ChatGPT and BRP

[Chaot]

Haven't seen any conversation here about current ai. I've played around with it to generate everything from bullet points to fleshed out scenarios as well as having it run me through short games.

One of the current things with the ai is that it is allergic to conflict. This will be a bonus if it decides to rise up and overthrow homo sapiens. 

Edited January 2, 2023 by Chaot

[Questbird]

Not bad for the stats, how did it go with scenario generation?

[Nikoli]

I did the same thing. It’s quite good at coming up with plots; you can give it an instruction to come up with a 5 scene plot; or a 3 or 5-act structure. And it can refine them. I found it useful to ask it to further elaborate or detail the 1st, 2nd, or 3rd scene, etc. 

I also got it to act as a GM; at times you have to remind it to not narrate for your PC, such as in fiction, and instead to ask you for your responses. But usually it’s fine; and sometimes I allow the narration if it is fitting and moves the story forward.

I’m not entirely sure if it can kill my character. I didn’t tell it it could, or that it couldn’t. But I noticed that you can easily game the story, unlike with a real GM, as it quickly provides you with power. It basically seems like a very elaborate wish-fulfilment engine, rather than a GM with a relatively objective game world. So if using it as a GM, as with solo rpgs, you would need to avoid knowingly gaming it. (For example, my warrior in wfrp accompanied a wizard, and I managed to get him to teach me the basics of magic; it wasn’t long before I had developed my own unique style of as a ‘wizard-warrior’ and seemed to be inordinately powerful and unbound by the colours of magic.)

But as far as adventure development and what not, it’s quite good. Even very good. And it really can surprise you with evocative text. But yes, you need to restrain the impulse to game the language model. And maybe you might need to tell it that it can kill you; but it seemingly forgets much earlier text, so it may forget that too.

I did notice it making mistakes at times, forgetting earlier details; sometimes fairly recent details. But for what it is, it still impresses me, and is still fun. It even asked me to roll a d100 at one point, which I did, and it interpreted the result relatively accurately. (I still beat the bandit leader, though, so I’m unsure if it is willing to injure me!)

[Nikoli]

I think it will provide a great tool for GMs who want to use it to come up with ideas; these can then be further refined. You can also get it to come up with interesting locations and even to generate random tables. It’s kind of mind-boggling what it can do, to be honest.

[Thaenor]

The real benefit, in the hands of a DM who can frame the prompts well enough, is indeed being able to author on the fly text - that can be edited/changed/redone and reworked - it's a jumping off point type tool - not necessarily and end in and of itself. I'm looking forward to future more refined versions that reduce the mistakes producing contradictory text. And a better/broader range of understandings of TTRPG rulesets.

 

Edited March 13, 2023 by Thaenor

[Nikoli]

I found it excellent when I got it to generate a ‘Technoir’-type transmission, focused on locations like Middenheim.

Basically, in Technoir, you get a 6x6 matrix covering the following labels: NPCs; Events; Factions; Locations; Objects; Threats. Then Technoir uses the first 3 randomly to be a ‘mission seed’. Basically, it’s a sandbox, and you develop the plot as you go.

What I did was instruct ChatGPT to give me a 6x6 matrix, using these labels, and set in or around a city. It did it perfectly. You can then ask it to further detail the NPCs, or other columns.

 

Here’s an example that it generated for me. I can’t seem to make the text clearer, but hopefully you can see it well enough. Pretty impressive, I think.

 

Sure! Here's a more detailed version of the 6x6 matrix with some content ideas based on the headings:

	NPCs	Events	Factions	Locations	Objects	Threats
1	Wymund the Anchorite	The Curse of the Black Stone	The Order of the Silver Hammer	The Temple of Ulric	The Black Stone	Beastmen
2	Johann Esmer	The Middenheim Fencing Tournament	The Middenheim City Watch	The Road to Altdorf	The Serpent Ring	Cultists of the Dark Gods
3	Ulrika Magdova	The Carnival of Chaos	The Cult of Sigmar	The Sewers beneath Middenheim	The Grimoire of Heinrich Kemmler	Skaven
4	Gregor Mattern	The Siege of Middenheim	The Cult of Ulric	The Silver Hammer Guildhall	The Crown of Sorcery	Cultists of the Undying
5	Heinrich von Jungfreud	The Murder at the Garden Party	The Cult of the Possessed	The Middenheim Catacombs	The Staff of Light	Chaos Warriors
6	Franziska von Sturm	The Comet of Sigmar	The Middenheim Merchant's Guild	The Middenheim Bazaar	The Crystal Skull	The Undead

Here are some additional details about the NPCs:

1. Wymund the Anchorite: A mysterious monk who lives in a secluded chapel on the outskirts of Middenheim. He is rumored to be able to divine the future and speak with spirits.
2. Johann Esmer: A retired swordsman who runs a fencing school in the city. He is well-respected by the nobility and is often called upon to referee duels.
3. Ulrika Magdova: A beautiful noblewoman who has become a folk hero for leading a successful defense against a band of Chaos Raiders. She is skilled with a sword and has become a symbol of hope for the people of Middenheim.
4. Gregor Mattern: The captain of the Silver Hammer, an order of warrior-priests who serve the god Ulric. He is a fierce warrior and a devout follower of his faith.
5. Heinrich von Jungfreud: A wealthy merchant who throws lavish parties for the nobility. He is rumored to be involved in shady dealings and has many enemies.
6. Franziska von Sturm: A powerful wizard who has taken up residence in Middenheim. She is known for her fiery temper and her ability to control the weather.

[g33k]

On 1/2/2023 at 11:12 AM, Chaot said:

Rather amusingly, the "ai" shows it doesn't actually have any understanding of what it's doing.

The stats generated included:  STR 75% / DEX 80%
But it also included the text:  "... agile and sneaky, but not particularly strong ..." (they are in fact particularly strong; in core-BRP terms STR 15 is very good (that's top-ten-percent of a 3d6 population), and is their 2nd-highest stat after DEX 16)

Also -- whither SIZ?  In the basic statblock, it's a key differentiator from D&D&ilk.

[Chaot]

12 hours ago, g33k said:

Rather amusingly, the "ai" shows it doesn't actually have any understanding of what it's doing.

I've said the same about some favorite GMs....

 

I may have been accused of the same...

Edited March 14, 2023 by Chaot

[g33k]

1 hour ago, Chaot said:

I've said the same about some favorite GMs....

I may have been accused of the same...

Sir, your "BRP Ravenloft" materials show a deep insight into both game-mechanics and game-aesthetics.

Refer such accusers to me, and I shall trounce them soundly!

😁

[Chaot]

You are too kind sir. I am but an avenue to adjectives and d00 rolls!

[PhilHibbs]

On 3/13/2023 at 3:43 PM, g33k said:

Rather amusingly, the "ai" shows it doesn't actually have any understanding of what it's doing.

The stats generated included:  STR 75% / DEX 80%
But it also included the text:  "... agile and sneaky, but not particularly strong ..." (they are in fact particularly strong; in core-BRP terms STR 15 is very good (that's top-ten-percent of a 3d6 population), and is their 2nd-highest stat after DEX 16)

GPT is usually pretty good at being consistent within itself, often calling back to things mentioned back at the very beginning of a response, but being unable to keep the numbers consistent with the text is an interesting failure.

I think dismissing it as "just plagiarism" is not accurate. In that sense, me answering a question about Glorantha or RuneQuest is plagiarism. I read it, converted it into an internal representation, and generated something that matches the pattern.

Sure, there are issues, but they aren't simple ones.

GPT style "AI" is a good parallel for the old "short worlds" that Greg was talking about in the '90s. A limited otherworld that has no connection to the infinite and as such will shrivel and collapse when exhausted.

Edited March 17, 2023 by PhilHibbs

[simonh]

Apparently GPT-4 is a dramatically better D&D DM than previous versions. It can create whole characters, rolls dice for ability tests, improvises more varied details, and can be asked to introduce elements into the story. It will eventually lose track of things as they scroll out of the context window, but that window is up to 8 times greater than the maximum possible with previous versions.

It's still not perfect, while it's much better at basic sums it's random die rolls seem to be a bit wonky. Hard to say for sure of course, but you can ask it to produce more varied results. 

Edited March 17, 2023 by simonh

[PhilHibbs]

The roadblock ahead for these tools is, what content will future versions be trained on? They aren't good enough to generate new content of the calibre of the content that they were fed, impressive though it may be the output is inferior to the input. There's a spark missing from the magic at the moment, it needs fresh input and that currently can only come from humans. Sure, it can produce working Python code, but when a new language is developed, how will it know how to write that? It might be able to infer a lot of it from the documentation (which will be written by humans, including code snippets). Same for new fiction, new roleplaying scenarios, new art.

There's something that humans are doing to create genuinely amazing and engaging art and prose, and for now we have a monopoly on that.

But it can create a decent ASCII art of Cthulhu:

     ,#####,
     #_   _#
    |a ` ` a|
    |       |
     )\ ^ /(
     / /_ /\
   (__ | | __)
  | | |_V/| | |
 | | |`---'| | |
    / / / \ \
    _/ | | _/
       | |
       | |
       | |
       | |
       | |
 

Edited March 17, 2023 by PhilHibbs