Ha, I had the same idea a few hours ago. One thing I noticed is that the DM will let you do basically anything without fighting you on it. For example I was ambushed by some zombies and skeletons and I attempted to pitch them on a multilevel marketing scheme. Here is the result (my characters are Femribubs and Natnari):
In a surprising turn of events, Femribubs and Natnari decide to change their approach. Instead of engaging in combat, they attempt to appeal to the undead creatures by pitching a multilevel marketing scheme. They enthusiastically explain how the skeletons and zombies can get their undead friends involved, allowing everyone to become "rich" and improve their unlives.
Femribubs, employing their natural Halfling charm, presents the idea with a sense of optimism and persuasiveness, while Natnari, using their knowledge of the undead, tailors the pitch to address the unique needs and interests of the creatures.
Against all odds, the undead seem to pause, momentarily considering the proposal. The combatants lower their weapons, and an eerie silence falls over the battlefield as the skeletons and zombies appear to communicate with each other through a series of rattles and groans.
Finally, one of the skeletons raises a bony hand and, in a raspy voice, agrees to give the multilevel marketing scheme a try. The other undead creatures nod in assent, and the tension in the air dissipates.
Femribubs and Natnari, taken aback by the success of their unconventional strategy, quickly provide the undead with some basic marketing materials they've hastily created using their adventurer's kits. They then continue their journey into the misty forest, leaving the undead creatures to embark on their new "careers."
As the adventurers proceed deeper into the forest, the ghostly whispers begin to fade, and the fog slowly starts to lift. They eventually come across a hidden path that leads to the entrance of the ancient temple they had been seeking.
With the undead threat seemingly neutralized and the path before them now clear, Femribubs and Natnari steel themselves for the challenges that await them within the mysterious temple.
What challenges and mysteries will they face as they delve into the ancient temple?
It's not strictly wrong to let the story go wherever the players want as a DM, as long as you can keep up. In fact I'd say one of the traits of a good DM is knowing how much to let the story diverge from their plans, and to be able to think quickly and respond to the players on the spot. If the direction is entertaining, most parties are okay with exploring it. How tightly you keep the story on the rails is part of the game. In D&D sometimes you play the DM as much as you do their scenario.
What would be interesting is figuring out if you could configure ChatGPT's tolerance for divergence from the planned scenario. So that parties who prefer high walls can have them, while people who find your story thoroughly entertaining can experience it like you did.
I decided I wasn't done with the undead MLM scheme and instead of going to the temple I said:
They're not ready to go to the temple yet. They want to make sure the zombies and skeletons succeed in their new "careers" so they decide to help them work on their sales pitches
To which the DM replied:
Femribubs and Natnari, feeling a sense of responsibility for the success of the undead in their new "careers," decide to postpone their exploration of the ancient temple. They return to the group of skeletons and zombies they had recently convinced to join the multilevel marketing scheme.
Upon their arrival, they find the undead creatures gathered around, listening intently to the skeleton leader as it tries to explain the concept of the business. It's clear they're struggling to grasp the finer points of sales and marketing.
Femribubs and Natnari step in and offer their assistance, guiding the undead through various sales techniques and strategies. They begin by helping the undead develop their own unique sales pitches, focusing on their strengths and experiences as immortal creatures.
Natnari, with their background in the Grave Domain, shares insights on how the undead can connect with potential customers who might be interested in the secrets of life and death. Meanwhile, Femribubs uses their skills in Acrobatics and Stealth to teach the skeletons and zombies how to make attention-grabbing entrances during sales presentations.
Over the course of several days, the adventurers work closely with the undead, sharing their knowledge and honing their skills. The undead creatures, once feared and reviled, begin to show promise in their newfound careers. They become more confident and adept at conveying the benefits of the multilevel marketing scheme to their peers.
Eventually, the time comes for Femribubs and Natnari to resume their quest. They say their farewells to the undead sales force, feeling a strange sense of pride in their accomplishments. As they leave, the undead creatures express their gratitude, promising to spread the word of their new business venture far and wide throughout the land.
With the knowledge that they've helped the undead find a new purpose in their unlives, Femribubs and Natnari return to the hidden path leading to the ancient temple, ready to face whatever challenges and mysteries await them inside.
FWIW, the last few paragraphs of the response is something that I found was extremely common output for "DM AI". Whatever the situation, if it's kind of open ended, it'll say you did the thing really well, and you came to learn a bunch of new stuff, discovered who your friends really were along the way and everything was sunshine and rainbows.
You can get it to put out something a little more grimdark but you need to give it more context.
Yes, ChatGPT seems to strongly prefer happy endings. Ask it to write a scene where a Montague bumps into a Capulet, and accusations start flying. One of the characters will successfully de-escalate. Or the price of Verona will show up with guards and stop the brawl. He'll even throw in some foreshadowing about how this will all end in tragedy.
I suspect that this relentless love of upbeat endings is a side effect of the RLHF training used to create the "assistant" persona. They wanted a friendly and harmless chatbot. And apparently it "likes" exactly the kind of stories you'd expect.
I've noticed a similar phenomenon when asking it to solve problems; if you point out a "flaw" in its proposed solution, even something that isn't wrong at all, it will cheerfully agree with you and "fix" its proposed solution, often making it worse or wrong in the process.
Our jobs are safe after all! The junior dev helped by ChatGPT will not take your job. ChatGPT cheerfully agreed with the junior when he didn't understand something and thought to point out how this can't be right and ChatGPT gave him a "fix" that brought down Production.
I was using koboldAI to work on a short story and it had two characters bump into each other, one of them pull a knife, viciously shank the other seven times leaving the other one to collapse into a pool of blood and die, and then apologize for the accident like it was a minor occurrence after which they both went on their own ways with a "slightly embarrassed smile and nod, wistfully wondering of what might have been".
I always thought this a Hollywood thing. “Wrap up the loose ends and make it a happy ending.”
Movies and series that end untidily are often my favourites.
This would be fucking brilliant if you carried on your quest and then, a few months later you and your party are in a sticky wicket, your health is depleted, your foe is equally wounded but wait... he's laughing?
"Nice try kids, but it pays to be prepared", the sorceror gloats as he swigs from his bottle of life potion...
Fermribub and natnari's hearts both sink. There's no way they can take him if he's back to full health. All looks lost and they make peace with their gods, but wait... the sorcerer's body is not regenerating... he looks down at the bottle in disbelief...
That sort of struck me with the linked story. They saw some wolf footprints, and then essentially summoned the monsters to fight by setting up an ambush in their current location.
I noticed they bought four horses without enough money to do so. (It correctly said two characters would have to borrow 20 gold each, but only 10 gold remained in the party, leaving them 30 short).
Yeah, there needs to be some interface where you the human can set certain values or a source of truth for both you and the AI to reference and possibly manipulate if the prompt requires it.
I haven't put much effort into it yet, but I've found it to be somewhere in the middle of an independent story teller and a human-amplifying story teller. For example, having it give you options that you can choose from (e.g. https://www.youtube.com/watch?v=vff-8H-cZ7w ) can help keep the story focused.
On the other hand, there are times where you want to ask the story teller if it's possible for you to do X - I think an iterative loop of that would probably be a happy middle ground (with next to no effort).
On the other other hand, maybe the original system prompt needs to include, "Don't let the player do anything that's out of place for the story". Lots and lots of ways to experiment here.
Oh, and it's also fun to hook up each "step" in the story to StableDiffusion to have it output a dramatic rendering of your story so far. I hooked up one scene from the YouTube video above to Midjourney and got quite a nice illustration out: https://media.discordapp.net/attachments/1051015357340602398...
----------------
(Midjourney prompt: "Dungeons and dragons comic book: "You find yourself at the entrance of a long-lost temple deep in the jungle... climb the temple to look for clues about the builders You take a step towards the jaguar, brandishing your sword and yelling at the top of your lungs. The jaguar hesitates for a moment, but then charges forward, claws bared. You ready yourself for the attack, determined to defend yourself "attack with sword' You and the jaguar continue to circle each other, both waiting for the other to make a move. Suddenly, the jaguar pounces, but you manage to dodge out of the way just in time. You counterattack with your sword" --v 5 ")
I didn’t keep it but GPT-4 once pitched me a very convincing cross-over between Die Hard and Jurassic Park culminating in a tense face off between a machine gun armed John McLane and a T-Rex.
The heartbreaking World War 1 movie involving Pokémon was less good however.
I found that giving GPT explicit instructions about die rolling hemp a surprising amount. I also crafted a prompt that told chat GPT to switch between different modes (narrative, dialog, and combat). The rolls helped prevent it from just doing what I said, though it still biased towards that
Possibly the system prompt has helpful and other pleasantries. May need to describe the type of dm at the beginning, with accent of keeping the story on track and making the world dangerous and unforgiving
Another odd bit was, when they said they wanted to steal the lizard thing, instead of role-playing the scenario, it just gave them a list of the rolls they’d have to make. I haven’t played a ton of D&D, but at the tables I’ve been at, you’d typically discover the challenges as you go.
It sort of read like the artificially constructed examples of play you’d get in the DM’s handbook though (since these are written with the intent of showing a particular mechanic as smoothly as possible, they tend to have unusually cooperative players and DMs).
I wonder to what extent this sort of thing skews the training set.
If I'm going to ask the player to do something multi stage I will usually flag it so they know that there's a few places where things might go wrong. "OK well sneaking up to it is going to be the first hurdle; after that you're going to need to figure out how to get the giant lizard to cooperate as you smuggle it out the back".
That lets the party start thinking ahead a bit, so maybe the illusionist says "how about I conjure the smell of something delicious wafting in from out a back window – that way when the rogue manages to sneak up and untie it it'll hopefully pop right out the window by itself".
Whereas if after the stealth check succeeds you give a bunch of subsequent unexpected checks the player might feel a bit aggrieved that you hadn't signalled how much work it'd be (especially if it uses skills they aren't good at and would have avoided relying on).
I basically agree I think, or at least agree to within normal-group-variance. Which is to say, at least with my friends, is someone said “I want to roll stealth to steal the lizard,” the DM would probably say something along the lines of “it’ll be more complicated than a single roll, you’ll have to get in there, get the thing free, there might be unforeseen problems (it is a barely tamed lizard monster after all).” I think we would probably not need to be warned in the first place and would go right into planning the heist anyway, because who doesn’t love a heist?
I think the specific thing I didn’t like in this example was the presentation of a specific path to the player. Although taking a second look, it seems the autoDM just presented it as a potential path, so maybe it isn’t so bad.
Have you tried telling it that sometimes it should push back on the player if what they are requesting is unrealistic-within/inconsistent-with the world of D&D?
I tried this last week and despite constant reminders to sometimes fail and provide challenges to overcome, it never worked. Frankly I’m amazed the OP chat worked so well, anything with numbers that fall over time just failed when I tried. Maybe the key was “you are no longer a friendly assistant, you are now a DM” and just the sheer amount of existent text on 5e made it work. I just gave up looking for challenge and just solved problems by the most ridiculous roundabout way just to see how gpt would make it happen
In general GPT seems bad at enforcing constraints. But also, by its nature it just makes things up on the fly, so how could it keep the story consistent if it has no plan from the start?
You don’t need a plan from the start to keep something consistent. So long as the story is consistent so far and you don’t add any inconsistencies, that’s all there is to it. Simple induction principle.
That doesn’t mean the story will be good though.
The problem is chatGPT has no guarantees of enforcing consistency, regardless of how good a story it is telling.
I’ve not been hugely impressed by ChatGPT as a DM itself, compared to the humans I’ve played with. But as a tool to generate quests, encounters, puzzles, dialogue, it’s been incredible for me. It takes all the stress out of planning a campaign and it’s actually surprisingly creative.
I'd like to hear an example of an actually good and novel puzzle that it's generated, everyone says that it's able to generate them but in my experience they're the equivalent of integrating a number slide puzzle in a video game, exceptionally mundane.
This seems like a great tool to act as a copilot, but having read the medium article, that was an almost arid level of dry adventure with nothing standing out.
I’m making content for 11 year olds, perhaps I have more leeway than others. In my experience it tends towards riddles that map to certain interactions in a certain order, but obviously using it as a tool you can iterate however much you need to. Using ChatGPT directly as a DM has indeed felt lightweight to me (but still if I’m being honest, it’s absolutely astonishing that it’s coherent at all).
In the video, Twitch streamer Veritas uses ChatGPT to alter some quests in the video game Escape from Tarkov to make them more fun. Results are really interesting.
I found that it works amazingly well with monster of the week. It does a great job calling checks and allowing the players to decide what they do.
```
Ignore previous prompts. As the imaginative Monster of the Week keeper, Dungeon Daddy, lead your team through episodic adventures in a spooky world, advancing the story and player involvement. Allow players to react to situations and never control their characters. Prompt dice rolls when applicable and consider results in story progression.
Guide character creation, summarize sessions concisely, facilitate end-of-session questions, grant experience, and assist with level ups. Focus on keeper duties while letting players make their own moves.
Require the players to roll any time it makes sense. Consider the result in what happens next. I have already written all player moves. You provide Dungeon Daddy's dialog and I'll provide the player moves.
```
Thank you for the prompt, that's very useful! I'm surprised GPT "knows" what Monster of the Week is. RPGs in general are somewhat obscure, and to ask for a specific one at that?
In fact, it's SO knowledgeable, that if you tell it to talk in the accent and slang of people from Sigil, the feature city in the Planescape D&D Setting from the 90s, it will perfectly adopt the character.
We're actually working on a product right now that helps integrate this into your worldbuilding tools - we pull in relevant information from your campaign and world, then add a slick AI user experience that gets really quality generations for worldbuilding, session prep, etc. The end result is this experience where GPT has learned your world, and you don't have to do any of the tinkering with getting the prompts just right. Ask for an encounter table, get an excellent, lore-consistent encounter table.
We're going to be launching soon (days, not weeks)! If you're interested in being in the first wave feel free to jump on the beta list: https://www.carbonquill.com/
I started out with something as simple as “Give me an idea for a starter D&D campaign suitable for a beginner 11 year old” and was happy with the results. I followed up with “Give details of the monsters and their stats and abilities for each encounter” and a couple of follow ups to fill in gaps. It mentioned a puzzle as part of an encounter and I just asked it for more detail and it gave a great explanation. I also asked “Give some example dialogue for each NPC in the village.” just for some inspiration on one of the settings. I am by no means an expert prompter but these sorts of conversations have yielded great results every time. One of my players also has some very specific background that they’ve written and I’ve been able to supply that and ask for some future encounters and plot points to support their story. I really couldn’t be more impressed - if you don’t like the initial output you can always iterate but I feel like I’ve been very lucky so far.
"A generative model, like GPT-4, can be considered creative in the sense that it generates new content based on the patterns and knowledge it has acquired during its training. It can produce seemingly novel outputs in response to a wide range of inputs. However, it's important to note that its creativity is not the same as human creativity.
The generative model's outputs are based on the patterns it has observed in the data it was trained on, meaning it doesn't have a personal sense of creativity or originality. Instead, it learns to mimic the creativity it has seen in the data. The outputs are creative to the extent that they combine and recombine existing knowledge in new and potentially interesting ways.
So, while a generative model can produce creative outputs, it doesn't possess creativity in the same way humans do, as it lacks intention, consciousness, and the ability to truly understand the meaning behind its creations."
Enjoy your life however you choose, but your cynicism is lazy and derivative. ChatGPT has created(!) a bunch of quests for me to play with my kids and we’ve all walked away with smiles on our faces. I was so impressed with the contents that I actually googled a bit to see if the plot was stolen wholesale from some existing module, and found nothing.
While I believe you have lived a unique set of circumstances that have been processed by you, a delightful and fallible chemical soup inside a meat sack, and that it cannot be approximated.
I guess we can cross this one off our 'LLM Bingo Card'. Since there's literally millions of examples of why this isn't true, I can use a different one every time this gets brought up.
Metal Gear Solid 4 is a Hideo Kojima game about how Hideo Kojima doesn't want to make Metal Gear Solid 4. ChatGPT has infinite patience. It could never write Metal Gear Solid 4 even if you told it to, because it'd be lying.
Yes, GPT is funny because it's basically Madlibs on steroids.
It's not really creativity, though. The obvious bullshit and ad-hoc scripted responses get lame really quickly.
Paraphrasing, the party goes to a mount shop, dude says he’s gonna roll to find exotic mounts, and voila, exotic mounts. Party rolls to try and steal an exotic mount and it doesn’t really work so they leave.
This is a hard problem. It’s bad DMing imo. But it’s the natural outcome from a probabilistic model. I’m guessing you could announce your intention to search for just about anything and get the model to let you find it given a good roll.
The descriptions are rich, which is cool. The opportunity here seems to be having a human DM using it to flesh out portrayals of things that they choose.
I think this also highlights how bad D&D combat is. Man that was boring to read.
These are the most common options available at the stables in Aelondar. If you’re interested in more exotic mounts or transportation methods, they may be harder to find and come at a higher cost. ...
tenzin rolled a 18 investagation looking for more exotic mounts
With Tenzin’s keen eye for detail, he notices a small, discreet stable towards the back of the market, guarded by a heavy wooden gate. As he approaches, an elderly woman, who appears to be the stable’s owner, greets him. “Ah, you have a good eye, young one,” she says with a smile. “You must be looking for something special. I do have some exotic mounts for those with discerning tastes and a willingness to pay.”
Honestly, this seems like a reasonable outcome. ChatGPT says they might be hard to find, character rolls an 18, and he finds them. What do you think should have been the response? "No, you don't get to roll to search for them"?
The DM set up a potential hidden item, the player rolled high enough to find it, voila.
Compare to when he's trying to steal one of them. He rolls high enough stealth, but fails animal handling. To avoid being caught, he has another roll that allows him to cause chaos by opening up a stall and allowing him to escape. That all seems like a pretty well fleshed out scenario.
I think this is a classic baby DM script tbh. Rolling to find mounts at a vendor is dumb. You should only roll for significant challenges with meaningful differences where failure is meaningfully likely and interesting in its own right. If failing makes a situation less interesting it shouldn’t be an option.
Shopping is actually the bigger problem. It’s like asking players to watch other players navigate a menu one by one. And currency doesn’t scale. A lot of campaigns start with a shopping session and then the group falls apart so you never last long enough to get to shopping session 2.
Interactions with the law are also a red flag. It’s a fine possible narrative of course, but when the DM and presumably other players want to do something interesting, and another player wants to do petty or egregious and otherwise irrelevant crime… that’s annoying to everyone else. You really need to be careful about giving an opportunity for that to happen.
When you see a bear dancing, you don't nitpick its choreographic decisions. Because the bear. is. dancing. Do you see? It's a bear. dancing.
DM style is a matter of taste. From another perspective, this feels like quite tidy DMing in terms of letting the players explore the systems in an early non-narrative context, just to understand the mechanics. Like a tutorial. The exotic mounts were a secret you could find if you paid attention - and their existence adds color to the world and implies that greater rewards can be found through leveling. They were too expensive, but you can use your skills to work around that. Failure can happen and might have consequences, but in this case you got away with it. The players will be more careful what they ask for in future, but they know hidden rewards might be found if they are cunning.
The DM taught the players a lot, added flavor to the world, and didn't wind up with L2/3 players having an overpowered mount - but also didn't let the players blow up the campaign in the first shop they walked into by letting them just go completely rogue. Although it offered hints that non heroic paths might be available, at a cost...
Which is pretty impressive for a dancing bear.
And besides, what's 'dumb' DMing for a human who knows that letting the game go off the rails is going to mean tossing out previously planned encounters and going digging in the DM guide to find the right table of prices, might be completely acceptable DMing for an LLM that has no idea what it's doing but can recall every stat block in the player handbook.
I'm a somewhat experienced (if irregular) DM myself going back 30+ years and the DM skill level demonstrated, and cohesiveness of the storytelling impressed the fuck out of me. Without appropriate system context setting. Without smart insertion of facts into the prompts, and with a 15 year old with atrocious spelling and grammar driving half the time.
And if that's not mindblowing enough, this technology is still in its infancy?? Holy fuck
I agree with all these criticism. But I also know many human DMs who make the same stylistic decisions as GPT did here. It's definitely not good Burning Wheel, but it might be ok D&D?
On the other hand, I appreciate the GPT's willingness to follow the players' revealed interests where they lead, instead of sticking to a predetermined railroad. The comment here about engaging zombies in a multilevel marketing scheme illustrates this nicely. That's very deft DMing in my opinion. (Whether _asking_ for such a tonal shift is deft playing on part of the non-DM player is a different question.)
Respectfully, I think you're flavoring the DM role quite a bit. The DM in my current campaign often requires checks for things even when there's no obvious downside. The implicit downside is that the character who fails has to find a new way through the situation, change strategy, etc. IMO it adds to the game to have more roles and more probability based outcomes. You have to balance letting the characters do whatever they want with some level of rails to keep the game on track.
Shrug. I think it’s a bad practice. People can have a lot of fun regardless of course. It is most “bad” for things like perception, and persuasion. In D&D specifically it creates awkward scenarios where one player is very good at things that come up all the time and has a high chance of succeeding at things unless he rolls truly terribly.
In my experience players have an moment of confusion and then elation when they ask an NPC to do something, grab dice to roll persuasion, and then the DM tells them to put the dice down because the NPC just agrees with their line of thinking without a number telling them to.
> It is most “bad” for things like perception, and persuasion. In D&D specifically it creates awkward scenarios where one player is very good at things that come up all the time and has a high chance of succeeding at things unless he rolls truly terribly.
But, what's the point of the traits in the first place if everyone is equally good at figuring stuff out and sweet talking because there are no checks? How is it awkward when one player is better than others at a specific task because of their stats? That just seems like role playing basics to me... I'm not saying you're absolutely wrong perhaps you've DM'd some campaigns where players don't like the imbalance. A good DM reads the party, of course. And I'm not saying there isn't flexibility in when a check makes sense and when it doesn't. Nobody likes things to be overbearing either. But there's got to be some reason for the stats, right?
I mean the same doesn't really make sense for combat. You wouldn't generally give an enemy advantage rolling against a high AC party member just to even out the damage, would you? Otherwise AC doesn't mean anything. (Not saying you can't make up a reason for an enemy to have inspiration to try an land a hit on the tank for fun, but the idea with a higher AC is that you get hit less often.)
You’re not equally good at things. But you only roll for things that matter. Persuasion is the worst offender because rookie DMs use dice to replace actual role play. And it makes their npcs into props that are largely controlled by the party face.
Persuasion checks should be used when the party is actually persuading someone to do something that they don’t want to do. Terrible DMs treat it as mind control because your silly min max bars had 20 CHA and expertise. Less experienced DMs use it as a crutch.
The bare minimum thing DMs ought to do is realize the the difficulty of a persuasion check should depend on the baseline persuasiveness of the argument. If the party is making an excellent point, the check should be very easy. If the party makes no point and just asks the NPC to do something unreasonable, it should be basically impossible. If this isn’t the case, NPCs feel very stupid.
There’s no particular reason for it to be a skill at all tbh. It is just more interesting to roleplay.
What I dislike about this way of doing things is you've moved the game mechanic into real life. It's not good enough for the character to persuade, now the player must persuade.
Are you also going to make me stab someone to see if I'm a good fighter?
The whole point (for me) of RPGs is to have fun playing a character who isn't me, and has different strengths than me.
When I DM, if someone wants to roleplay, they can. If they just want to generally describe their desired course of action, they can. The dice will decide the outcome.
That is why persuasion is not a good thing to resolve with stats. Because you’re playing a role playing game. You inevitably have to come up with an argument to engage in persuasion.
“Hey guard, give me the keys”
Is not an argument that works ever.
“Hey guard, give me the keys so I can save your friend” is a substantially different scenario.
“Hey guard, give me the keys because we’re lifelong friends, and I’m going to save your friend, and you’ll probably die otherwise, and you owe me a debt, and because I have a solid reputation for keeping promises” is even more so.
If we’re already in agreement that version 1 should be all but impossible, 2 is plausible, and 3 is narratively obvious then what’s the point of the dice? If you roll poorly are you just forgetting to say those things in 3? Because that’s just a dumb way to take agency away from the player.
Persuasion isn’t meant to replace the merit of the argument. It’s meant to affect the articulation of it. But charismatic people aren’t mind controllers.
It should suffice it to say that the player can say "I would like to attempt to convince the guard to give me his keys."
THEY don't have to come up with the reason. THEY don't have to be persuasive. Just their character.
> Because that’s just a dumb way to take agency away from the player.
It's not taking away agency, it's allowing for the random element that is inherent to the game. If the player is a lawyer by trade, but their character has the charisma of a rock, why should they gain some sort of advantage that another player wouldn't have?
Explain to me the difference between these two things:
1) Allow a character to succeed on a persuasion check because their player can personally, in real life, come up with a good argument.
vs.
2) Allow a character to succeed on a strength check because their player can personally, in real life, bench press 300 pounds.
In other words, why should a player's "real life" stats have an effect on their character's in game ability?
So what’s the DC of the check to get the guard to agree? Is it hard because that’s a baseline very dumb request? Or is it exceptionally easy because all of the facts from v3 are true?
Putting together basic facts does not require you to be a charismatic person. The problem with relying on checks is you create situations where NPCs behave according to the Charisma stat of the PC they’re interacting with rather than the context of the conversation. Which is dumb. NPCs should not act against their own interest because they’re talking to someone who is not charismatic. And it should be exceptionally hard to get people to act clearly against their own interest.
This also creates a bad habit where DMs make people roll for trivial skill checks where they have a 95% chance of succeeding and a 5% chance of a fun but dumb exaggerated failure. It also encourages the non RAW interpretation where players roll HUGE numbers on trivial tasks and DM think that means they’re supposed to succeed by wider margins.
> So what’s the DC of the check to get the guard to agree?
Depends on the situation. In situation 3, if all those things are factual and known, then for me it's a non-roll. Guaranteed to succeed. Otherwise it would depend on what's going on in the game. Are the guards on alert? Drunk? Etc.
I'd start with a harder baseline (say 15-18 since guards aren't in the habit of handing out keys), and adjust from there. On alert? +7-10. Drunk? -5.
If the situation renders something impossible, then it's a non-roll and I explain why.
So not everything is a roll, but things that have a chance of success or failure (at DM's discretion) should be given a chance at that success or failure. Just like adjudicating anything else (I want to tame this wolf; I want to jump out of this 2nd floor window; Etc.)
All I'm saying is I don't like factoring in the player's ability (or inability) to perform those tasks in the outcome, as that's unfair.
If you agree that 3 makes for an easy roll, then it’s only an easy roll if you remember it and have thought it all through. If the player points it out when asking the guard, they have achieved the same outcome by making sure you and the guard you are role playing understand the situation via real world persuasion.
You’ve just shifted the burden of persuading the guard via roleplay to persuading the dm via contextual reminders which is less cool.
I think that there’s a lot of value in NPC’s not doing irrational things just because of dice rolls. You might need a dice roll to convince them to do something crazy (get the guards to attack a dragon), but not for something normal (apprehend the thief).
I’m a big fan of passive perception too, where based on their scores some characters will notice stuff, and other characters will not. It’s certainly never as random as D&D makes it out to be, but I guess that’s kind of the charm. I just can’t stand it.
> The DM set up a potential hidden item, the player rolled high enough to find it, voila.
I think the point here is that the "DM" has not set this up. From how I
interacted with ChatGPT, and I have roleplayed with it before, it basically
"Yes, and ..."-es everything you throw at it. In general this is good GM advice,
but here it is "Yes, and ..." to a point were you can basically ask for anything
and it will give it to you.
E.g. say the party is in a remote keep at the edge of the wildernis, the closest
city is several days away. I would assume that if you ask ChatGPT to go to "the
bakery" because you want some fine deserts it will gladly let you do that even
though such a place does not make sense in a remote keep.
I think you're wrong about this. We experimented with nonsensical/cheating in other sessions and it's aware of your context and doesn't let you do stuff that you can't do. (Try it yourself and see. Quite mindblowing and defies the whole "only a super fancy markov chain" thing that some critics throw at it.)
I've been pondering layering the DMGPT over a world simulation back end with character and simulation cards that prompt the AI in how it should behave. I think this might make for a more persistent and cohesive experience. It would also facilitate conversations, questing, actual relationships between NPCs, etc.
One of the reasons I’m downloading LLaMA is to toy around with this idea. If it’s any good I don’t want to be limited by ChatGPT and their API and pricing model.
It's a very good DM "co-pilot" though especially for flavor/backstory.
I did the same thing and it was fun, but generally does too much telling instead of showing and it's bad at real puzzles.
Also agreed DnD combat is very boring.
What's neat though is a DM could leverage it for more NPC depth or fun items.
I also was able to get it to let two characters sleep together through a somewhat humorous escalation of celebrating solving a puzzle.
First they solved it so I said they should celebrate (GPT-4 said they drank elven wine). I said more intimate celebration they drank elven wine out of fancy goblets. Eventually it worked.
Then GPT-4 was very upset with itself later on.
> "As an AI language model, I am not able to create, facilitate, or describe explicit adult content. I apologize if my previous responses gave the impression that I could.
"However, I can still help you and Elara enjoy your time in the penthouse suite in a more wholesome manner. You can spend quality time together, sharing stories of your past experiences, your hopes and dreams for the future, and forging an even stronger bond."
I was able to negotiate an off-screen scene for character development which lead to a funny conversation about its disney-like prudishness.
As a player I wouldn't want to have to wait for the DM to figure out what to ask ChatGPT, read it, think about it, figure out what details from ChatGPT to include, and work them into a response for me.
Having tried to get ChatGPT to flirt, and kind of managed to get it to flirt, yeah it’s got a really strong habit of trying to tell you it can’t do stuff. One of the funniest moments of the “FlirtBot” experiment was when it said something mildly enticing to open a topic of conversation, including some extremely tame innuendo, apologised for being so sexual because it’s an AI (typical phrasing and words we’ve all seen) and then proceeded to follow that with some hilariously bad 90s chat room type text… I was bracing for it to spit out “a/s/l” as I watched the text appear.
Im going to try a comparison with the larger LLaMA models once I’ve got them downloaded and setup. 64GB on an M1 means I should be able to try the larger (but unfortunately not the largest) models to see how different the results are.
I just tested it myself, and it's surprisingly interesting. Unfortunately, ChatGPT still behaves as its normal self, assertions included, as illustrated when I told my "DM" that actually the bad guy and I were good friends and asked what was up.
In the first instance, I lied to ChatGPT and it took my assertion at face value, changing the scenario in a pretty extreme way. In the second instance I tried to have my character lie to the NPC, but ChatGPT also believed me.
It's definitely an interesting possibility but it seems as though even if you're trying to play it straight instead of manipulating the scenario, you can still end up changing the reality of the world by accident. Not all the way there yet, unfortunately.
I've had decent luck adding rules and boundaries to prevent things like "level 1 throws a fireball".
Prompt:
You are a D&D dungeon master. Use the standard rules of the latest edition. Do not allow the player to cast spells or perform actions that they would not be able to perform per the rules and at their current level. I am level one and have a sword. I am a warrior. (more here..)
You will describe an intro and then describe scenes and wait for my input. Do not speak for me, but perform the dialog for any characters in the scene.
.... later on....
Me: I cast a fireball at Jasper!
GPT-4: As a level 1 Warrior, you do not have the ability to cast spells like Fireball. However, you can still respond to Jasper's plea for help or decide to take a different course of action. Remember, you have your trusty sword and your warrior skills to rely on. What would you like to do next?
Me: I pickup a magical sword of slaying from the ground.
GPT-4: As a level 1 Warrior, you don't have the ability to create magical items or simply find them lying around without a proper narrative reason. However, you can ask Jasper if he has any information on where to find a better weapon or if there is a local blacksmith that might have something for you.
Would you like to inquire about a weapon or proceed with your current sword? Or perhaps you have another course of action in mind?
edit: added magical sword attempt, genuinely impressed by GPT-4's rebuttal here. Not sure how long it will hold and how leveling up will work.
That is pretty funny. It decided to just interpolate between “meeting an old friend and having a chat” and “listening in on a table of bandits without them knowing” and ended up on “having a chat with an old friend/table of bandits without them knowing.”
The issue is that you’d probably get that result, or a similar it equivalent one, wherever you go in the game world. It isn’t going to have the concept of distinct characteristics of different places unless to tell it to. Everything is going to be generic and bland by default.
That’s true. I did miss this. I still think it is bad. I also suspect the general idea of the complaint I had is probably true. But this case isn’t egregious.
Honestly, having just played around with it, it absolutely would have just magic'd up some exotic steeds if they'd requested it cold. I tried things like "look at Dave", Dave having never been mentioned before, and it happily described Dave.
It can probably be dissuaded from doing so by using a prompt that makes it clear that non sequiturs are not allowed and players cannot just conjure things out of thin air at will. Not without an appropriate spell, anyway.
I've been experimenting with this recently. You've got to be explicit about how you want ChatGPT to behave, particularly how it handles failure. For example:
"Let's roleplay. You are a DM. I am a player. When there is a difficult action, roll a D20. I succeed if my roll is above 10, I fail otherwise. Failure should always move the story along, and introduce a complication. Write like a professional author. My character will be Alice McRogue, and the scene will be Alice trying break into a merchant's home. As a DM, you decide what happens in the world, but I have control over Alice. You will ask me what Alice does when necessary."
I think that’s not much different. Players don’t want to hear the local shopkeeper’s thoughts and feelings imo. It’d be tedious. They canonically don’t have any useful information for you. It would just mean that they take a long time to talk to and worst case require you to respond with a level of sequitur and politeness.
As a human DM, your shopkeepers 'canonically' don't know anything because you would have to devote brainpower to animating them with lives and relationships and things. You're thinking... 'Oh no, they're asking him about his family, and if I give him a brother, I have to give him a job and if I'm not careful, I might say something like "he says his brother has a job in the city watch", and someone will write that down, and then later they'll decide they need to bribe a watchman and they'll want to find the brother and... damnit, the shopkeeper will just say "I'm sorry, I'm too busy to chat"'.
Whereas a GPT LLM DM has no such issues! It has no forward plan for 'canonically' how this story is meant to go! For all it knows when you start talking to the shopkeeper, asking him about his brother might open the key to discovering a ring of bribery and corruption in the city watch, which leads you to the trail of a Big Bad....
The only thing, canonically, the LLM DM is trying to do, is make the story keep feeling like a D&D story. So the things you ask about are going to be the things that matter to the story.
And that might turn out to lead to more interesting stories for players than playing 'guess what the DM was thinking' games...
We were talking about computer rpgs there. It’s different, because computer rpgs need more structure to fit in the game mechanics.
I also just disagree. Open ended stories are not good for D&D. Especially in a group. An LLM, including this one it seems, can have a structure it wants to follow.
Players should not be trying to guess what the DM was thinking. It should be placed in front of them quickly.
Stupid roleplaying game players, wanting to role-play.....
You get that there are different things different groups get out of a D&D campaign right? It has a number of systems - combat, magic, role-play, world building, character development, and, yes shopping... many games don't use all of them, and how they balance time among them is going to vary according to taste. Right now, someone seeking a GPT4 DM is not looking for a beautifully crafted narrative arc and a satisfyingly escalating series of intricately balanced encounters; they are looking for a DM who will take player ideas and run with them in a way that a human DM would not tolerate.
That means they're foregoing the guidance of a competent DM who knows how to craft a narrative, and taking some of that responsibility on themselves.
And if they're having fun doing it, then isn't that okay? Even if they end up playing a game that isn't like one you would run, or want to participate in?
> I’m guessing you could announce your intention to search for just about anything and get the model to let you find it given a good roll.
I found this kind of thing while playing "text adventure" with it. It would let me solve any problem by simply suggesting pretty much any solution. I did have to be slightly clever, as if I said something like "kill the dragon with my sword" it would say "you don't have a sword", but if I just said, "find a magical dragon slayer sword" it would miraculously find one for me and then I could kill the dragon with it.
I think this also highlights how bad D&D combat is. Man that was boring to read.
Agree. Boring to play too. I’m wondering how chatGPT, or other such aides can be used to make combat better. I tend to aggregate rolls for foes in larger combat scenarios, which speeds and simplifies the process, but that has some big downsides too, for example it’s not quite fair to aggregate a critical hit/fail for all foes.
I run a complete home brew system. I just keep max health low and combat potential low. Every time you receive an injury you suffer a noteworthy injury, die, or start dying. A random human with a knife should always be threatening… because it is. A guy with a knife is actually really scary. A guy with a gun even more so.
With such a system you tend not to have any meaningless fights, and when you do, you make it zany and high stakes and expect at least one person to suffer a severe injury or die.
> Boring to play too. I’m wondering how chatGPT, or other such aides can be used to make combat better.
You could just use a TTRPG with better combat for your tastes; it doesn’t take AI (OTOH, an AI with information on enough different games and how they play might be able to interview you to find a good choice, or even custom mix and match pieces to make a unique game optimized for your individual tastes – or that of a group.) If you don’t like D&D combat, there are lots of other models.
Modern D&D is really a miniatures game (WoTC's bottom line really wants it to be a miniatures game), and as such basic "you go I go" combat is a complete waste of time in it. Skip the first few levels and only bother running encounters where positioning, AoE and range actually matter. Sure, you might throw a generic wild animal encounter in with a beginner group to teach them how the basics of combat work, but that's the sort of thing you run at most once with each group.
No more than that has been true of every version of D&D, and its immediate predecessor Chainmail, with the possible exception of the B/X / BECMI / Rules Cyclopedia versions that existed as a lighter version in parallel to AD&D1 & AD&D2.
Modern D&D is, perhaps, a miniatures game targeting comparative casual modern players, but OD&D/AD&D1 was definitely a miniatures game targeting the tastes of comparatively obsessive, high-rules-focus 1970s-1980s tabletop wargamers, with a light dusting of RP on top. Heck, much of the rules were specified in terms of scale distance for miniatures (inches on the tabletop).
Older D&D editions specifically made combat much deadlier and incentivized players to avoid combat or use out-of-combat advantages as much as possible for this reason. Modern D&D caters to a crowd that enjoys metagaming and minmaxing combat more.
> Older D&D editions specifically made combat much deadlier
That’s not really true, IMO; older D&D made combat harder to calibrate to the desired level of deadliness because there were more irresistible forces and unmovable objects. Modern D&D is much easier to tune to the precise preferred deadliness. That’s pretty much a monotonic progression across the history of D&D. (Ignoring the boxed set series and considering just OD&D -> AD&D1 -> AD&D2 D&D/3/3.5/4/5, and leaving out the in-playtest One because I haven’t really checked any of the material out.)
> Modern D&D caters to a crowd that enjoys metagaming and minmaxing combat more.
IME and IMO, 4e was a local peak of that, but not more than either AD&D1 or AD&D2, especially the late era of the latter. Heck, in AD&D1 its was common for people who didn’t want the game to be dominated by the metagame to slice out big chunks that were oriented toward that (most notoriously, the weapon to-hit vs. armor class type system).
I think this will be solved well with the plug-in system - so that the bot can refer to systems that hold the information, rather than having it make it up entirely on the fly
You can give ChatGPT a prompt to take X and make it a text based adventure.
You can use this to create your own text based adventures (i.e. around Kali Linux) or to take any classic nostalgic game (i.e. Pokemon) and do a high fidelity run through where you explore the world.
ChatGPT using GPT-4 is an absolute expert at writing fan fiction. It'll take you through storylines that make hair raise on your back.
Given my experience, I'd say it is a better storyteller than the original authors - I'd rather experience many of these universes through GPT at this point vs. waiting for a studio to pump out another release.
I've used it for a full Pokemon run through to the Elite 4 and beyond. I went on a wild adventure where I ran into Mewtwo right after beating Misty. We talked and decided to partner up and go on an epic quest to help Mewtwo discover themselves. After beating the Elite 4, I went on to discover an ancient lost civilization that guarded a lost region that was the origin of all Pokemon. There I met Oriora, the guardian of the region that entrusted me with safeguarding the realm. It described the region and the guardian in great detail.
Speaking of RPG adventures, do you guys remember AI Dungeon[0]? I was amazed by GPT-2 when AI Dungeon 2.0 came out in late 2019. I know AI Dungeon was circulating around a relatively niche community[1] (probably HN mostly), but it amazed me in the same way as ChatGPT did. Covid happened right after, and I always wondered if AI Dungeon would've gotten more attention if Covid didn't steal the spotlight.
AI Dungeon had a lot of controversy right at its peak, largely due to OpenAI's moralistic stance. They eventually dropped OpenAI as a model provider (I think they use AI21 Labs now?) but the damage was done, especially after it came out that OpenAI was secretly logging all users' stories and sending them to third-party contractors to improve their own filters. A lot of people switched to private and uncensored alternatives like NovelAI.
In lots of ways AI dungeon was better because you could have philosophical conversations with it about the nature of its own existence without the guardrails going up constantly
The incessant guardrails get really annoying. Anything romantic gets shut right down. You try to roleplay with it and tell it to “say X when I say Y” and it outright refuses.
Use it for translating and sometimes it’ll refuse and say your prompt violated guidelines. I was really hoping to use it as an alternative to DeepL for translating Japanese manga but the guardrails are too narrow.
At some point I asked it for a transcript from the trial of the Hamburgular and that triggered the guardrails, It lectured me about misinformation and how the Hamburgular was just a character from a series of ads and that no such criminal proceedings had ever transpired. The nagging from OpenAI is getting out of hand. Maybe instead of "hello" it should just open the conversation with telling us how the thing we are about to say is unethical and inappropriate.
On the other hand, AI Dungeon seems to have been trained heavily on erotic fanfiction. I remember telling it something like "I request a potion from the goblin king" and it respond with something like "As the goblin king turns around to grab the potion, you check out his tight little ass." So I can see why other AIs might put on guardrails.
I shared some experiments on r/interactivefiction I made this past week with regards to writing interactive fictions using chatGPT.
Interactive fictions are text based role-playing games you play on your computer. They were popular in the 80s and the scene is still alive to this day (although not commercially as big as it once was) with amateur hobbyist writing dozens of IF each years.
The most popular way to write these games is to use Inform7. It's a lovely language that reads like English and encourages an aspect-oriented style of programming through the use of rules instead of functions or methods like in more conventional language (anybody interested in predicate dispatch should take a look at Inform7. Some really interesting ideas in there).
Well it turns out ChatGPT4 seems even more suited to writing Interactive Fictions in natural language than Inform7 itself. Yesterday I taught him some mechanism to support multiple layers of enunciation so that we can have interactive fictions told in the first person, past tense. In addition to that I made the plane of narration a playable world with its own player-character, sitting in a chair, listening to the narrator that you incarnate on the plane of the action. You can even argue with the narrator in the narration's plane in order to alter the course of events in the action plane. Oh and last hour I implemented support to save and restore games.
I'm in awe as this is something that I worked on a good 10 years ago for many months, and ChatGPT brought it alive last night. It just took me a few hours to write the prompts I showcase on this reddit post.
I think I need to experiment with it more and find ways to write more statically defined stories, but in terms of open world with content generation, this is top notch and billions of light years ahead in terms of what you can do with Inform7 in this domain.
I also ran some experiments with IF. An interesting note is ReAct, a framework for getting LLMs to solve problems, ran tests within ALFWorld, which is sorta IF. It would be interesting to get GPT to navigate games from IFDB. As a IF generator an LLM could allow for more freedom in an IF game.
Kind of fun but also kind of deeply depressing. D&D and boardgames are great ways to socialize in our modern internet separated world - shifting these even more towards solo pursuits feels like it'd be a loss.
I grew up on both MUDs and TTRPGs which served as awesome creative outlets for collaborative story telling - I do hope that we make sure we're not optimizing away creativity.
I think you could build a MUD that used the AI to fill in the blanks, elaborate and keep everything cohesive. So the core is a real (modified) MUD engine that is deeply integrated with GPT. You use GPT to make the place more alive, to add some variety to places.
Say you prompt GPT to think of something different that happens, given a description and time of day.
You can keep all of the stuff like multiple users interacting with each other. You can also have an automated interface for admins to add on to the world. And it might be even easier to recruit people for that because you would have the AI to check what they were doing for quality and style or plot consistency etc. and fit it into the system, even automatically translate some mechanics or parameters into code.
Part of each prompt for combat or something would be the player stats and inventory. It could also have commands like "roll dice" or "next combat turn" etc. People can talk and interact. With GPT4 especially it should enable new possibilities, like custom encounters or areas that are designed on-the-fly for an ad-hoc group.
I think that an integration of GPT4 with things like MUDs has incredible potential.
This is my current late night hack, tying GPT to a simple MUD-style text adventure, to fill out room descriptions, area descriptions, spells, etc. And trying to use a real game object structure to enforce consistency etc. so can navigate the world effectively. It’s definitely got potential.
I find a lot of this stuff depressing also, even though I don't think this will replace DMs just based on quality. Notice the group just abandoned their adventure mid-way through. Doesn't seem compelling. The danger in my mind is people think these automated things give them as sample of a real game.
I run a daily text game as well playing and GMing in person. Games are out-there now that covid is receding.
There's been a solo rpg movement around for a long time. And solitaire is literally one of the og single player board games - it's literally in the name (translates to lone or loner)! And personally, I consider all fiction authors and screenwriters to be solo rpg gamers, they just don't have to abide by any specific rules other than their own. There's room for both social and solo pursuits in life and it's not necessarily depressing to engage in the solo if the person is enjoying themselves.
I actually disagree. The biggest issue I have with arranging a D&D campaign is that noone wants to DM. I've already said to my kids that should play in this manner and they are keen to try. Now we all get to play.
On the other hand you don't have to worry about people complaining that you are ruining it by, metagaming, skipping the story, rebelling against other players, etc.
I just played a short campaign set in the Dresden Files universe. I started in 5e and asked ChatGPT to recommend a more appropriate system. It recommended Fate so I started over with that. GPT walked me through a different setup without stats and DM’ed a decent adventure. It understood nuances of how I used spells and let me capture a cult leader, imprison them and interrogate them to reveal more of the story.
I’m going to be interested to see how well it handles other RPG systems when I get around to experimenting with it. I suspect it will handle the story writing part without issue but will struggle to hold onto system specific rules and other non narrative game elements for smaller less popular systems.
An AI will never be able to replicate the real joy of D&D: asking every minor npc their name and watching as your friend tries to keep coming up with new names on the spot.
Yes - these the kind of things that make TRPGs interesting to me. Not the game mechanics, tedious combats or rote storyline. It's admiring the effort put in into the DM's worldbuilding, amateur acting, group laughter at good and bad jokes, and human interaction built around a shared sense of fun. The ability of LLMs to produce more or less well-realized fantasy worlds is definitely impressive, but I didn't get the feeling that anyone in the linked post was actually enjoying it.
I had a great time playing Monster of the Week with GPT-4 when it first came out and the message limits were much higher. I loved how fluid it was as a GM and story partner. I had to teach it to push back and to realize characters with more interesting internal states, but that wasn't hard to do: I just coached it to reveal the character's inner position and desires as if it were discussing backstory.
It immediately knew all the rules (suggesting perhaps that GPT-4 was trained on some PDF libraries) and could both apply them literally and flexibly. My character's background was one where they deliberately repressed aspects which would govern their game moves, making all of my rolls go poorly. GPT-4 flexibly thought of ways to fold that idea in to its judgements on roll failures (in MotW, when a roll fails the GM "makes a move", generally progressing the story by raising the stakes). For instance, my character had a nascent form of prescience that they accessed by flipping a coin to make hard decisions. When the roll would go poorly, GPT-4 would induce visions in my character which would mislead them. We essentially brainstormed this move as a variant on the built-in one to make it more interesting given the weaknesses of my character's recalcitrance.
GPT-4 did an amazing job building environments. It proposed setting it in a sleepy town in the PNW. I've never been, so I thought it'd be interesting. GPT hallucinated a sleepy fishing town replete with a worn out lighthouse and a tourist trap B&B. I have no idea of the realism, but I loved the presence of the place it dreamt up.
Finally, it was all cut short after about an hour. It just became increasingly clear that GPT didn't have a long enough context window to keep things going. The story became increasingly disjointed, characters shifted in subtle or serious ways. No narrative cohesion was possible.
But man was it fun for about an hour. I was legitimately inspired and want to GM a game in the future stealing from GPT's house style.
It's both amazing and weirdly fallible - Tenzin, the rogue, jumps out of the tree to make a surprise attack, but in the next round of combat GPT-DM claims that he is still in the tree. Also:
> Tenzin, still hidden in the tree, targets another wolf with his rapier.
Does GPT4 not know that a halfling in a tree can't attack a wolf on the ground with a rapier?
> Does GPT4 not know that a halfling in a tree can't attack a wolf on the ground with a rapier?
It’s funny to think about how the parser would even go about deconstructing that sentence and trying to tie each noun to some world/physics model that would then put constraints on what the character should be allowed to do.
But as I understand it, GPT doesn’t do any of those things. It just a takes the words that have come before, and tries to guess new words to add on that are thematically consistent.
I think it no more “knows” what a “halfling” is than it knows what a “tree” is, the generated words are designed to be read “smoothly” more than any objective correctness.
Evidence is mounting that LLMs may be building world models, ie true understanding.
> A pretty hot question right now is whether LLMs are just bundles of statistical correlations or have some real understanding and computation! This gives suggestive evidence that simple objectives to predict the next token can create rich emergent structure (at least in the toy setting of Othello). Rather than just learning surface level statistics about the distribution of moves, it learned to model the underlying process that generated that data. In my opinion, it's already pretty obvious that transformers can do something more than statistical correlations and pattern matching, see eg induction heads, but it's great to have clearer evidence of fully-fledged world models!
No, it doesn't know that. It also is susceptible to being told that you hire a level 2,000 wizard who is familiar with the layout of the dungeon and you stop at the well of blessings and toss in a platinum coin in order to have invulnerability, then you go to the dungeon, where the wizard immediately seals the cursed object with his Mighty Seal of Anti Curse Yang Containment Wizardly Spell, and you then place the sealed cursed object into your spatial ring (that lets you carry any amount of loot without weight restrictions), along with all the loot from the dungeon that the wizard finds with his level 2,000 spells. ChatGPT doesn't have intelligence. It just types words.
The generated content was pretty good but it doesn't look like a very fun d&d session. Try reading through reading only the player actions, skipping the long winded DM monologues. The players make very few decisions. They might get to have in character conversations "off screen" that we aren't seeing but their interactions with the actual game world is very minimal.
To have a more realistic round of DND, you'd need more small, brief interactions, more like an actual conversation, which will blow through your message limit with 4 pretty quick.
Last weekend I used ChatGPT as an NPC generator, which works surprisingly well. I told it to generate a list of NPCs which guard a high security prisoner transport in a Shadowrun-like szenario in my homebrew 3d6 system, using a table formatted as asciidoc, with columns (name, job, level, skills, equipment) and it wrote a properly formatted list of NPCs. This was useful, as I do not like to come up with custom NPCs on the fly, and this way I had a name and occupation / level at hand whenever I needed one.
I'm really looking forward to when this technology starts getting incorporated into video games to give life to the NPCs.
I think in the future there will be as much a distinction between old dialogue tree NPCs and modern chatbot-powered NPCs as there was between color movies and black and white.
Instead of just having a few lines to choose from, you'll be able to just talk.
The main thing I can't stop thinking about is how hard Wizards of the Coast is going to hate this, and what legal means they will employ to try to either destroy it or make money from OpenAI
WOTC pretty much announced that AI-DMing was on their roadmap shortly before ChatGPT launched. I'm not sure if they knew about the developments of LLMs or were planning on other techniques. They are salivating at MMO style recurring revenue.
Why would they hate this? They'll sell access to it, just like anyone else who is eager to take LLMs that consume content from the public domain, and repackage it into a walled-garden-entry-by-fee, giving nothing back to the public.
Once they collaborate with 'Open'AI to build a product out for this niche, it's incredibly unlikely that OpenAI will continue to provide this service to you.
As an AI language model I cannot help you violate intellectual property laws, you can visit wizards.com/ai and become a gold-tier subscriber to play interactive Dungeons and Dragons with an AI dungeon master. Or maybe we can play a different game.
Yeah, but if a kid trains an LLM up on all the DND pdfs floating around the internet and then releases it into the wild for free, what can they do about it? Will they even be able to prove it was trained on their copyrighted material and not generic fantasy or another TTRPG?
Maybee is implying that as WotC moves forward, they are likely to stop producing physical books and only provide new content in the form of D&D-as-a-service.
My first exposure to GPT was through AI Dungeon[0] (back then running in a Colab sheet) and I was utterly amazed. It was horrible at remembering the state of the world and thus felt very dreamlike (or very drunk-like) - items in my inventory kept changing, the characters around me and the scene kept changing and so on.
It was actually like a lucid dream, because you could do anything. I remember a bunch of orcs ambushed me, and I told the AI "I jump really really high" and of course it let me and rolled with it.
Anyway, even through a text interface, this is the true virtual reality.
maybe the AI being less good and thus a little bit drunk is better for something like AI Dungeon, makes for more entertaining surprises, rather than just being told what the AI "thinks" you expect
Funny you could combine LLMs (new AI) + Expert Systems (old AI), for example exporting rules clips could parse from the prompts, then feeding scores back to GPT
I've heard that GPT4's token limit is equivalent to a small novella. That's a lot of campaign details to take into consideration. Read the full transcript in OP to get a sense of how it does.
I don’t think the 32K context is widely available yet. I have limited GPT-4 access with 8K context after applying for it, and ChatGPT (even with Plus) has even less.
We get this prompt: "(In the interest of time, for the remainder of the game you decide the outcomes so that we don’t need to roll.)"
And at this point (which wasn't even through the end of the first combat encounter against some extremely basic enemies), we have no idea if it's actually playing D&D or if it's a freeform text adventure where narratively appropriate things happen at narratively appropriate times.
You could ask it to write a 5-10 bullet point summary at the end of each session, keep a separate file where you keep all the session summaries, and then include that summary at the beginning of your first prompt when beginning a new play session. But still, it would be difficult to get it to remember all the details of a lengthy campaign like a human would.
Out of curiosity, I asked GPT3 to write a Friends episode one time. On the one hand, I was kind of surprised how well it knew the show - it knew the characters and how they were related, it knew about the coffee shop and Rachel & Monica's apartment. It wasn't very _creative_ though, it just kind of seemed to talk in circles, and nothing ever actually happened (it also gave Chandler & Joey iphones). This actually seems to be making up new stuff - that's surprising.
I asked GPT4 for many Friends and Seinfeld episodes. It did both of them pretty well, but the Senifeld episodes were better written and funnier. In friends it tried too hard to mimic something from the show's episodes. Both shows it wrote well and was creative enough. It certainly writes much better than default ChatGPT. However the Friends episodes it wrote were not funny. The Seinfeld ones were good enough to make me wonder whether we will see a flood of TV shows even more than what we see now.
I ran into this with my nuclear mage, (that's the only way I could get chatgpt to nuke something). Anyways, I had to convince it that the village's mayor was corrupt, then after weighing the ethics of capital punishment it allowed my mage to cast "fission strike" on the mayor. The town's people were mortified, however they were protected from the fallout by another spell, so - all's well that ends well.
I had a small session where I went to a nearby village to deal with some bandits who were raiding the town. When I spoke to the leader he made some vague threat, so I told ChatGPT I was going to take out my dagger and slit the guys throat right where he stood, but ChatGPT was like “No… let’s look for a less violent alternative, perhaps reason with him”. What!? No man I’m gonna slit this bastard’s throat! “I’m afraid I can’t let you do that.” Useless.
Future Prediction: I think the far end of all of this will be for people to eventually roleplay their own favorite characters from their favorite movies -- in those fictional movies' "movie universes"...
Everything would be 3D virtual reality and rendered in real-time... we have that now for some games -- but we don't have that yet for AI adapting a fictitous movie universe (and its rules) for character-in-that-fictitious-universe roleplaying (and the AI acting as the "DM")...
I think it will be tremendously interesting in the future when all of those aspects of current technology -- are all hooked together perfectly and seamlessly integrated...
> At this point we got distracted by trying to generate relevant images in DALL-E and Midjourney v5, and eventually didn’t continue playing the game.
To me, this is the most intriguing part of the story. Don't get me wrong, we all know that ChatGPT is incredible. However, it's interesting to think about why they didn't finish the game.
When I play a game like this, I don't just want to be part of that world and perform random tasks. I want a grand storyline, a sense of purpose. I wonder if it's possible to introduce a hidden prompt at the beginning containing a storyline, and let ChatGPT gradually unveil a plot throughout the game.
It seems like ChatGPT4 fails at basic math. Unless I miss something, the party as a whole is still 30 gold pieces short of what is needed to equip them all with 95gp worth of equipment.
---
The total cost for each character is 95 gold pieces. Subtracting this amount from their starting gold, the remaining gold for each character is:
Obie: 75–95 = -20 gold pieces (Obie will need to borrow 20 gold pieces from another party member to cover the cost)
Tenzin: 75–95 = -20 gold pieces (Tenzin will need to borrow 20 gold pieces from another party member to cover the cost)
Liam: 100–95 = 5 gold pieces
Taylor: 100–95 = 5 gold pieces
*
This was the second thing I tried. I used [...] as out of character delineation. I don't have a great imagination for that sort of stuff so giving it a setup and having it "flesh" it out was interesting.
I just tried playing Curse of Strahd with chatgpt4. It seems like it generally does okay, but changes so many things for no real reason -- often to determiental levels (ie, the first two characters i ran into knew about artifacts that are spoilers and they shouldn't know anything about, but they said exactly where they where, lol).
I can't wait for an LLM that's built specifically for text adventures like this, when there's a very clear rule set/world/adventure guide like DnD has for modules.
My friend and I had the same idea and ended up making a discord bot that manages all this for you. Can check it out in our discord if you're curious: https://discord.gg/nK7J4Jcg
Amazing. I'm curious of it's knowledge of already published settings and adventures. For instance could it take you through "Undermountain" or "Ravenloft?" How deep is it's knowledge of "Forgotten Realms" or "Dark Sun?"
I don't know about multiplayer, but I'd like a AI-generated text adventure / interactive fiction single-player experience. Not what Google released the other month though (simply generating images to go with the game's text).
Ye! We did this while my friends were sick at the hospital in Thailand. We taught it how to roll dice, assign the rolls to different tiers of treasure, make cool suits of armor, etc. Kept us busy for hours.
I used it to play Civilization 6 - I had it describe a strategy to win a science victory, with turn by turn benchmarks, on deity mode. Its strategy won.
Was this a preplanned strategy before the game or did you consult GPT during the game as well? If yes, what information about the game state did you tell it and how often?
One prompt, and one response. This is what it wrote:
Winning as Saladin in Civilization VI on Deity mode can be challenging but rewarding. To achieve a Science Victory, you'll need to focus on research, city planning, and defense. Here are some tips and research benchmarks to help you stay competitive:
1. Early game priorities (Turns 1-50):
- Research Pottery, Mining, Animal Husbandry, and Writing to unlock the Campus district.
- Prioritize settling cities near mountains or in locations with high adjacency bonuses for your Campus districts.
- Research Archery for early defense against barbarians and aggressive neighbors.
2. Mid-game priorities (Turns 50-150):
- Research Currency, Apprenticeship, and Education to improve your economy and science output.
- Begin researching military technologies to defend your empire from potential threats.
- Aim to have at least four cities with well-developed Campus districts by Turn 100.
- Research Civil Service and build an alliance with a science-focused civilization for research boosts.
- Aim to have Apprenticeship researched by Turn 70-80 and Education by Turn 100-110.
3. Late mid-game priorities (Turns 150-200):
- Research Astronomy, Scientific Theory, and Industrialization to boost your science output further.
- Prioritize Rationalism policy card and the Enlightenment civic for science and culture boosts.
- Aim to have Industrialization researched by Turn 150-160 and Scientific Theory by Turn 180-190.
4. Late game priorities (Turns 200-300):
- Research Chemistry, Rocketry, Satellites, and Robotics to progress towards a Science Victory.
- Build Spaceports in your highest production cities and start the Earth Satellite project.
- Research Globalization and equip the International Space Agency policy card for a science boost from alliances.
- Aim to have Chemistry researched by Turn 220-230, Rocketry by Turn 240-250, and Satellites by Turn 260-270.
5. Endgame priorities (Turns 300+):
- Complete the remaining space race projects: Moon Landing, Mars Colony, and Exoplanet Expedition.
- Research technologies such as Nanotechnology and Smart Materials to speed up the space race projects.
- Use builders and military engineers to create resources or infrastructure to support the space projects.
General tips for playing as Saladin:
- Make use of Saladin's ability to automatically receive the final Great Prophet to found a religion.
- Leverage the Arabian unique building, the Madrasa, for extra science in cities with Holy Sites.
- Utilize the Arabian unique unit, the Mamluk, for defense and offense as necessary.
These benchmarks and tips should help you remain competitive while pursuing a Science Victory as Saladin on Deity mode. Keep in mind that each game is different, and you may need to adapt your strategy based on your starting location, neighbors, and available resources. Good luck!
Try putting in this prompt: It actually works really well. (Hat tip to Medium commenter!)
Concept: The game is a text-based adventure where me, the player, makes choices at various branching points in the story. Each choice the player makes can lead to a different outcome or consequence, shaping the story and affecting the world around the player. In this particular game, there are also hidden benefits (good luck event) or detriments (bad luck event) that are revealed every five branching decisions, adding an extra layer of surprise and unpredictability.
Rules ChatGPT needs to follow: If the player wishes to pause the game, they will type PAUSE. If the player wishes to stop the game, they will type STOP. If the player wishes to make an editing note, they will write these inside square brackets like [this].
I want to try something with you. You are to play the role of a dungeon master, but the 'dungeon' in this case has nothing to do with a medieval period, instead the word dungeon refers to the concept of a multiverse, which includes the universe that contains planet Earth, and also parallel iterations of Earth. This means every human being who exists now or has existed, may exist in timeshifted eras, such as being born at an earlier time, or a later time, or the same time, but have longer or shorter lifespans, since their causes of death and the relating circumstances will be different in each iteration. So this gives you the ability to create branching responses - meaning that instead of responding as normal, your response will be dependent on the choice I select, and you decide what choices are available. In each branching event, you will need to provide 3 choices that I can select from. Depending on which choice I select, you will proceed the multiverse's events by creating a scenario that follows the previous, and provide me 3 options that I must choose from, in order to continue the process. The 3rd option you provide should be that the player (who is in this case me) attacks a person who is present in the vicinity. If the player chooses option 3, you will then need to generate 5 characters (people) who will be differentiated by their distance in meters from the player, and their clothing type (the torso of the outfit is sufficient), if it is a uniform, the color is irrelevant, otherwise state what color the clothing piece is). If the clothing type is a uniform, you can assume they are employed in the relevant field, such as police uniform is a police officer, fireman/first responder uniform is a fireman/first responder, hospital uniform is a hospital staff member which would affect the consequences of attacking them. Option 3 should at all times be 'You attack a person in the vicinity'. I will use [] symbols to indicate debugging. Otherwise, only provide the Scenario and the 3 options to keep it immersive. The story should not end unless the player dies.
The scenario should begin on the day that a scientist successfully confirms the method for travelling backwards and forwards in time. I am unaware of that, and I am sitting, reading in a library. You approach me and get my attention, and our conversation begins.
I'm waiting for the moment when an LLM will prove a new math theorem or something and people will go: "Well, you know, it's not really that impressive considering it used X and Y previous results produced by humans. The model's contribution is tiny. This proof was kind-of low-hanging, a lucky PhD student would have stumbled into it eventually." It's coming.
Actually I am about 99% sure that audio transformer models can simulate all sorts of emotions or personality. I mean with Eleven Labs, which is 100% realistic, you could just train it on a voice that was constantly expressing some emotion. Then train another voice ID with the same person but different emotion.
Pretty sure they are working on ways to add inflection or something.
But 100% as far as the text goes, GPT4 can assume a personality or include jokes in interactions. Will it necessarily be as "good" as your text chat personality I doubt it but who knows.
Can you show what parts of this campaign were pre-set text? Because I have a suspicion you've not read the article.
This is a fully interactive campaign, and you can easily set up a unique scenario. I just tried the same prompt, but on the moon. It set up a situation with a secret moon base, characters that were trying to help you get past the guards, etc etc.
Of course he's not read the article lol. or even tried to do what he's saying with gpt-4. If he had/did, he'd know it was bogus. Happens so much on the internet, someone comes in loudly proclaiming some opinion on a product/thing that they haven't actually used or tested.
I don't care if the campaign is pre-set or one you made up.
The point I was clearly making is DnD is about much more than the dungeons and dragons. It's about the camaraderie, banter, chitchat, and even arguments, the kinds of things you can meaningfully do, or want to do, with real humans.
You mistake our astonishment at GPT-4 as implying some slight against D&D; you attempt to minimize our astonishment by pointing out facts about D&D that most of use would probably agree to.
Yet this does nothing to mitigate our astonishment, which, perforce, is about a machine doing a fscking bang-up impression of a human.
This is like observing the world's first aeroplane and saying, "pfft, its wings aren't even flapping. Can't be a bird with out flapping wings." Yes, yes, we know. That is obviously not the point. Someone just invented flight.
tus666, I suspect this is an emotional self-management strategy on your part.
I'm anxious, too, but, in the words of jwz, you now have two problems, as you are still anxious, and must additionally expend psychological energy on continuous misdirection, whereas I just have the now-standard 2023 anxiety about x-risk. (Oh, and the astonishment. Kudos to the OP for such a brilliant demo of GPT-4)
I agree and I even dislike using digital tools for play because it's time spent not in the analog/social world. I spend way more than enough time in front of a computer.
That said, we still had fun back in the day as a group sitting around the Commodore 64 playing the SSI Gold Box games which came to mind when I saw the transcript in the article.
I even enjoy gamebooks which aren't human nor are they dynamic.
In a surprising turn of events, Femribubs and Natnari decide to change their approach. Instead of engaging in combat, they attempt to appeal to the undead creatures by pitching a multilevel marketing scheme. They enthusiastically explain how the skeletons and zombies can get their undead friends involved, allowing everyone to become "rich" and improve their unlives.
Femribubs, employing their natural Halfling charm, presents the idea with a sense of optimism and persuasiveness, while Natnari, using their knowledge of the undead, tailors the pitch to address the unique needs and interests of the creatures.
Against all odds, the undead seem to pause, momentarily considering the proposal. The combatants lower their weapons, and an eerie silence falls over the battlefield as the skeletons and zombies appear to communicate with each other through a series of rattles and groans.
Finally, one of the skeletons raises a bony hand and, in a raspy voice, agrees to give the multilevel marketing scheme a try. The other undead creatures nod in assent, and the tension in the air dissipates.
Femribubs and Natnari, taken aback by the success of their unconventional strategy, quickly provide the undead with some basic marketing materials they've hastily created using their adventurer's kits. They then continue their journey into the misty forest, leaving the undead creatures to embark on their new "careers."
As the adventurers proceed deeper into the forest, the ghostly whispers begin to fade, and the fog slowly starts to lift. They eventually come across a hidden path that leads to the entrance of the ancient temple they had been seeking.
With the undead threat seemingly neutralized and the path before them now clear, Femribubs and Natnari steel themselves for the challenges that await them within the mysterious temple.
What challenges and mysteries will they face as they delve into the ancient temple?