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

[Clint_L]

Okay, so these are tentative thoughts, because who knows where this thing is going. A year from now, everything I write today will probably look ridiculously besides the point and unimaginative. Yet, I put myself out there (in other words, if you disagree with me and think I am wrong, that's okay, I probably am, and I don't mind being told so).

As far as use in RPGs goes, at the minimum I hope we see a revolution in how one player computer games, like Skyrim etc., are designed when it comes to NPC interaction. Those games are already entertaining and immersive enough to sell gazillions of copies, and that is with storytelling that is, fundamentally, not far removed from "pick a path to adventure" story books (i.e. choose from a few prescribed options and follow the branching path). I put months of time in the Baldur's Gate games, and had memorable interactions with NPCs who typically had 2-3 dialogue options in any given situation. An integrated chatbot will allow NPC interactions that feel much more like a conversation with an actual person. The main problem I foresee will be establishing limits that don't constantly break the fourth wall (i.e. when the player tries to make the NPC do sexy talk or whatever).

A level up from that, the AI will be able to integrate dynamic dialogue with a more flexible plot. In other words, maybe you can actually persuade the NPC to do something that isn't already scripted into the game. Note that ChatGPT was easily able to handle this on the examples posted earlier, so again the question becomes one of limits. How do you do this without breaking the game, or getting the game to do things that the publisher doesn't want it to do, and the the publisher REALLY doesn't want to be held accountable for.

Like, what happens if you can use the AI to turn Baldur's Gate into a StormFront fascist fantasyland, a wackadoodle porn parody, or a rip-off of someone else's IP? Does the publisher become liable?

The next level up would be something like a chatbot equipped version of a D&D adventure, with an option to have the chatbot run the game as the DM. I think we are a long way away from a real general purpose intelligence that could be as good as Matt Mercer, but I don't think that is what is needed for success in this field. Again, games with almost no AI, like Baldur's Gate, are immersive experiences. So I don't see any fundamental reason why a chatbot-run D&D session could not be immersive for the players - humans are very good at using our own imagination to fill in gaps.

Again, though...the risks to the publisher could be extraordinary.

 

[UngainlyTitan]

Okay, so these are tentative thoughts, because who knows where this thing is going. A year from now, everything I write today will probably look ridiculously besides the point and unimaginative. Yet, I put myself out there (in other words, if you disagree with me and think I am wrong, that's okay, I probably am, and I don't mind being told so).

As far as use in RPGs goes, at the minimum I hope we see a revolution in how one player computer games, like Skyrim etc., are designed when it comes to NPC interaction. Those games are already entertaining and immersive enough to sell gazillions of copies, and that is with storytelling that is, fundamentally, not far removed from "pick a path to adventure" story books (i.e. choose from a few prescribed options and follow the branching path). I put months of time in the Baldur's Gate games, and had memorable interactions with NPCs who typically had 2-3 dialogue options in any given situation. An integrated chatbot will allow NPC interactions that feel much more like a conversation with an actual person. The main problem I foresee will be establishing limits that don't constantly break the fourth wall (i.e. when the player tries to make the NPC do sexy talk or whatever).

This is a cost issue, I know that stuff like Chat GBT need serious amount of time on supercomputer clusters to chunk through the training process. I have no idea as to the processing power then need once trained but it will be a barrier to entry for some time but will eventually be overcome.

A level up from that, the AI will be able to integrate dynamic dialogue with a more flexible plot. In other words, maybe you can actually persuade the NPC to do something that isn't already scripted into the game. Note that ChatGPT was easily able to handle this on the examples posted earlier, so again the question becomes one of limits. How do you do this without breaking the game, or getting the game to do things that the publisher doesn't want it to do, and the the publisher REALLY doesn't want to be held accountable for.

Some stuff can be done, if the game has procedurally generated maps and can populate them on the fly to a quest template(s) then the NPC can generate isolated quests on the fly for ever but interacting with the main plot is a hard problem and I suspect outside the capabilities of this type of AI.

Like, what happens if you can use the AI to turn Baldur's Gate into a StormFront fascist fantasyland, a wackadoodle porn parody, or a rip-off of someone else's IP? Does the publisher become liable?

If the AI Art is not copyrightable then I would not see the publisher not responsible as long as the acted to stop it on appropriate notification.

The next level up would be something like a chatbot equipped version of a D&D adventure, with an option to have the chatbot run the game as the DM. I think we are a long way away from a real general purpose intelligence that could be as good as Matt Mercer, but I don't think that is what is needed for success in this field. Again, games with almost no AI, like Baldur's Gate, are immersive experiences. So I don't see any fundamental reason why a chatbot-run D&D session could not be immersive for the players - humans are very good at using our own imagination to fill in gaps.

Again, though...the risks to the publisher could be extraordinary.

I think moderately conversational NPCs are doable, as are infinite linear procedurally generated quest, sort of at Diablo or slightly better level. Even at current tech level, but they might not be cost effective.
I think that a genuinely creative experience is probably beyond them. What might be possible is to communicate to designers what the players are doing and have to world evolve around the consensus play. Sort of like the old Living Greyhawk idea.

 

[fnordland]

The price of using AI is dropping fast, but when you can go through many conversations over months of gameplay time, API usage of $0.02 can quickly add up. Perhaps a monthly all in fee for the game would absorb the compute costs.

OpenAI Price Drop: [Another] Megathread

Today's 90% drop on ChatGPT API – aka gpt-3.5-turbo – is another before & after moment in AI

cognitiverevolution.substack.com

Project development life cycle tends to run in years for the big games companies, change is hard, mistakes get made, they take even longer to fix. Project development is a balance between luck, art and science.

 

[Fanaelialae]

Cost is certainly a potential barrier, but probably not an insurmountable one for a well to do studio. AI dungeon exists, and it's basically the AI part divested from the video game.

I think the far bigger question is whether they could get the NPC responses tuned to a degree where it doesn't break down into absurdity and detract from the game.

Imagine that you're conversing with a proud elven innkeeper and you ask, "Are you a dwarf?"

The AI responds "No, I'm an elf."

The player responds, "You seem like a dwarf disguised as an elf."

The AI, responds, "You've seen through my disguise. Yes, I am a proud dwarf born in Rockholm."

And your interactions with that NPC just start degenerating as the AI hallucinates. The reality is that because AIs lack any capacity for comprehension, they can be "tricked" by the right prompts into saying just about anything. With the current models, you can try to limit the paths that a user might take to get it to say something, but because you can't predict every potential prompt a user can come up with, it'll never be bullet proof.

Will someone do it anyway? It wouldn't surprise me. A game like that will get a lot of hype, and that might drive initial sales. However, I think once people play with it a while, they'll notice cracks in the veneer, and it probably won't be something that everyone else rushes to replicate. I think, at least for the next decade, a game like that is likely to be more akin to the VR of the 90s than the VR of today. An interesting novelty, but with a lot of, much needed, room for growth.

 

[Clint_L]

I agree that the inability of the AI to actually think is going to make it prone to abuse and incoherence when pushed. So I think the issue becomes one of guardrails. For example, what if the AI comes with a preprogrammed list of features for that NPC. One of which is that it is an elf. So you can't convince it otherwise.

Basically, I am wondering if you can can curtail the AI enough to still allow for some of the flexibility of ChatGPT while preventing things from going completely sideways.

The problem here is that the chatbox does a very good job of emulating human creativity in many ways - way better than most of us thought possible just a few months ago - but not of emulating human judgement.

 

[Fanaelialae]

I agree that the inability of the AI to actually think is going to make it prone to abuse and incoherence when pushed. So I think the issue becomes one of guardrails. For example, what if the AI comes with a preprogrammed list of features for that NPC. One of which is that it is an elf. So you can't convince it otherwise.

Basically, I am wondering if you can can curtail the AI enough to still allow for some of the flexibility of ChatGPT while preventing things from going completely sideways.

The problem here is that the chatbox does a very good job of emulating human creativity in many ways - way better than most of us thought possible just a few months ago - but not of emulating human judgement.

No, because the AI simply doesn't understand what an elf or a dwarf is. You can certainly align it towards responding as an elf would (assuming you train it on "elven" responses). However, if you want it to be able to have an ongoing conversation, then there's no way to prevent prompts from eventually deviating it from that course. At least not without kneecapping its capacity to carry on a conversation. Sure, you could stop it from talking about dwarves entirely, but then it wouldn't be able to converse if you asked the innkeeper how he feels about dwarves.

I admit that I'm certainly not a professional in the field of AI. But I have worked professionally as a software engineer for over a decade, and I have done a bit of research into how these AIs work. It's really interesting (even though I'll totally admit I don't understand all the particulars) and impressive.

However, once you know how it works, you realize there are certain technical limitations to the current model that no amount of training or tuning is likely to ever overcome. It's kind of like how someone like Usain Bolt is an incredibly impressive athlete, but he's never going to sprout wings and fly, no matter how hard he trains. That's not to say that a better model isn't waiting to be discovered, but rather that it's likely to be at a level of complexity that's still decades off (true AI, or at least something that's reasonably close to it). A program that doesn't merely fake comprehension, but can at least approximate genuine comprehension, such as by sanity checking its own responses. Which, to the extent of my understanding, does not exist in the current model, and is likely to be quite the technical challenge to implement.

 

[Hussar]

I think the goal here though isn't so much to replace a DM, but, to give DM's unbelievable tools to automate a lot of the "work" of being a DM. A DM+an AI is an incredibly powerful tool.

 

[fnordland]

This is a twitch stream that is AI driven with questions from the chat, answers are given by AI versions of famous people.

Imagine you are the DM and you want to add in players based on famous people. How about if you want to play with a group of people based on generic gamers, I want a tactical, casual, indie storyteller, etc.

I think there is potential in the future to create your own game with the cast of tactical role or hit tv show.
Imagine the holodeck of Star Trek, you are there and playing a role.

 

[Prime_Evil]

I'm mostly using AI to automate some of the drudgery. For example, I am generating a subsector for a Cepheus Engine game and am using ChatGPT to help me randomly design each inhabited system. Consider the following prompt:

Act as a planet designer and describe a planet suitable for a science fiction tabletop roleplaying game. Begin by detailing the star this planet orbits, including its spectral class, population, mass, luminosity, surface temperature, and radius. If the star has a companion, provide details of it as well, but also the periapsis and apoapsis of its orbit. If the star is a variable star, provide some details. Determine how many other planets there are in the star system. Mention any gas giants, with details of their location and size. Specify where the habitable zone lies in the system. Also specify the location of the system’s frost line or snow line. Next, provide details of the mainworld in the system. Specify orbital parameters such as mean orbital period, orbital eccentricity, orbital semi-major axis, perihelion, and aphelion. Ensure the planet is in the habitable zone for the system. If the planet has any natural satellites, provide plausible details of them. Mention the sidereal rotation period to measure the length of the local day in standard hours. Consider the effects of tidal braking. Specify the orbital period in both Standard Terran Years and local days. Mention the planet’s axial tilt and the effects on seasonal variation. Ensure all details are scientifically plausible.

The output is something like this:

Star System: Kestra-71
Star: Kestra-71A
Spectral Class: G2V (main-sequence star)
Population: I (metal-rich)
Mass: 1.05 solar masses (M☉)
Luminosity: 1.1 solar luminosities (L☉)
Surface Temperature: 5,800 K
Radius: 1.03 solar radii (R☉)

Companion Star: Kestra-71B
Spectral Class: M4V (red dwarf)
Mass: 0.2 solar masses (M☉)
Luminosity: 0.005 solar luminosities (L☉)
Surface Temperature: 3,200 K
Radius: 0.25 solar radii (R☉)
Periapsis: 35 AU
Apoapsis: 70 AU

Variable Star: No

Total Planets in Star System: 7
Gas Giants: 2 (Kestra-71d and Kestra-71f)
Location: 3 AU (Kestra-71d, Size: 1.2 Jupiter masses (Mj)), 9 AU (Kestra-71f, Size: 0.6 Jupiter masses (Mj))

Habitable Zone: 0.9 AU - 1.3 AU
Frost Line/Snow Line: 4.5 AU

Mainworld: Kestra-71c
Mean Orbital Period: 1.05 Standard Terran Years (383.25 local days)
Orbital Eccentricity: 0.02
Orbital Semi-Major Axis: 1.0 AU
Perihelion: 0.98 AU
Aphelion: 1.02 AU

Natural Satellites: 2 (Aira and Rheni)

• Aira: (Radius: 1,350 km, Mass: 0.006 Earth masses, Orbital Period: 35 local days)
• Rheni: (Radius: 800 km, Mass: 0.001 Earth masses, Orbital Period: 80 local days)

Sidereal Rotation Period (local day): 23.2 standard hours

Tidal Braking: Mild, the two moons' gravity has slowed Kestra-71c's rotation over time but not enough to cause significant tidal locking.

Axial Tilt: 23.5 degrees, causing Earth-like seasonal variations.

Kestra-71 is a binary star system with a main-sequence G2V star (Kestra-71A) and a red dwarf companion (Kestra-71B). The mainworld, Kestra-71c, is a terrestrial planet located in the habitable zone of Kestra-71A at 1.0 AU. It has a slightly eccentric orbit, with a perihelion of 0.98 AU and an aphelion of 1.02 AU. The planet's orbital period is 1.05 Standard Terran Years or 383.25 local days.

I then follow this up with a series of additional prompts to develop further information about the system. For example, the next prompt is:

Describe the geophysical properties of the planet from the previous request. Ensure the output describes a planet suitable for a science fiction roleplaying game. Begin by specifying the planet’s mean radius, equatorial radius, polar radius, and oblateness. Calculate the planet’s surface area. Specify the mean density both in grams per cubic centimetre and compared to Earth’s mean density. Next, calculate the surface gravity both in Gs and in m/s squared. Also calculate the planet’s escape velocity. Estimate the distance to the horizon on the surface for a human of average height. Provide the planet’s geosynchronous orbit distance and velocity. Give all measurements in both metric and imperial units. Ensure all details are scientifically plausible within the framework of current planetary science.

This gives me the following output:

Planet: Kestra-71c

Mean Radius: 6,500 km (4,038 miles)
Equatorial Radius: 6,510 km (4,043 miles)
Polar Radius: 6,480 km (4,025 miles)
Oblateness: 0.0046

Surface Area: 530 million square kilometers (205 million square miles)
Mean Density: 5.5 g/cm³ (0.99 times Earth's mean density)
Surface Gravity: 0.95 G (9.34 m/s²)
Escape Velocity: 10.8 km/s (6.7 miles/s)
Distance to Horizon (for an average human): 4.7 km (2.9 miles)

Geosynchronous Orbit Distance: 42,200 km (26,220 miles)
Geosynchronous Orbit Velocity: 3.05 km/s (1.90 miles/s)

Kestra-71c is a terrestrial planet with a mean radius of 6,500 km (4,038 miles), an equatorial radius of 6,510 km (4,043 miles), and a polar radius of 6,480 km (4,025 miles). Its oblateness is 0.0046, indicating a slight equatorial bulge. The planet's surface area is approximately 530 million square kilometers (205 million square miles).

The mean density of Kestra-71c is 5.5 g/cm³, or 0.99 times Earth's mean density. The surface gravity of the planet is 0.95 G (9.34 m/s²). Kestra-71c's escape velocity is 10.8 km/s (6.7 miles/s).

For a human of average height, the distance to the horizon on Kestra-71c's surface is approximately 4.7 km (2.9 miles). The planet's geosynchronous orbit distance is 42,200 km (26,220 miles), with a geosynchronous orbit velocity of 3.05 km/s (1.90 miles/s).

Kestra-71c has two natural satellites, Aira and Rheni. The larger moon, Aira, has a radius of 1,350 km and orbits the planet every 35 local days, while the smaller moon, Rheni, has a radius of 800 km and orbits every 80 local days. The gravitational interaction of the moons has caused mild tidal braking, slowing the planet's rotation over time, but not enough to significantly impact the planet's day-night cycle.

The sidereal rotation period of Kestra-71c, or the length of a local day, is 23.2 standard hours. The axial tilt of the planet is 23.5 degrees, which leads to Earth-like seasonal variations as the planet orbits its star.

And so on...I'm slowly developing a set of prompts that can randomly generate entire star systems and spit out detailed write-ups. This is no different to randomly generating systems and doing the math by hand but it's a lot faster and provides a wealth of info if you want a simulationist game. Heck, I can even ask ChatGPT to do stuff like the following:

Provide a detailed analysis of the planet's habitability and suitability for human colonisation based upon the information generated so far. Give all measurements in both metric and imperial units. Ensure all details are scientifically plausible.

And this gives me:

Kestra-71c Habitability Analysis:

Atmosphere:
Kestra-71c's atmosphere is composed primarily of nitrogen (76%) and oxygen (22.5%), with trace amounts of argon, carbon dioxide, and water vapor. The atmospheric pressure at sea level is approximately 1.1 atm (1113 hPa), making it suitable for human habitation without the need for pressure suits or additional oxygen supply.

Temperature:
The minimum, mean, and maximum surface temperatures of the planet are -20°C (-4°F), 15°C (59°F), and 35°C (95°F) respectively. These temperatures fall within the range of human tolerance, and while some regions may experience more extreme conditions, overall the climate should be suitable for human colonization.

Hydrosphere:
With 70% of the planet's surface covered by liquid water, Kestra-71c offers ample water resources for human colonization. The average salinity of the planet's seawater is 35 g/kg (3.5%), which is similar to Earth's oceans and can be easily desalinated for consumption and agriculture.

Magnetosphere:
Kestra-71c's magnetosphere provides adequate protection from solar radiation and cosmic rays, ensuring a safe environment for human colonization. The surface equivalent dose rate of ionizing radiation is low enough not to pose significant health risks to the inhabitants.

Gravity:
The surface gravity of Kestra-71c is approximately 1.1 Gs (10.8 m/s²), which is slightly higher than Earth's gravity. While this might cause some initial discomfort and require some adaptation, it should not pose a major obstacle for human colonization.

Natural Resources:
Kestra-71c offers a variety of valuable natural resources, including rare minerals, silica-rich sand, unique photosynthetic pigments, bioluminescent fungi, and bacteria, as well as geothermal energy. These resources could support a human colony and provide opportunities for trade and economic development.

Native Biosphere and Ecology:
The native biosphere and ecology of Kestra-71c, while alien and diverse, do not appear to pose significant threats to human colonization. The described species are primarily non-aggressive, and the biomes can support human habitation. Careful ecological management and conservation efforts would be necessary to preserve the unique environments and prevent negative impacts on the native species.

Conclusion:
Based on the information provided, Kestra-71c is a scientifically plausible and suitable candidate for human colonization. Its atmosphere, climate, hydrosphere, magnetosphere, and gravity are all within the tolerable range for human habitation. The planet's natural resources and diverse biosphere offer opportunities for economic development and scientific exploration. As with any colonization effort, care must be taken to minimize the impact on the native ecosystems and maintain the delicate balance of this unique world.

There's nothing here I couldn't do myself with a computer and a few science textbooks, but it greatly accelerates the process. I wouldn't use everything in the game, but it's great if you want a hardish SF feel. I don't like a lot of the text it spits out - it's full of passive voice and subordinate clauses. But as a starting point, it's pretty decent.

 

[Andvari]

A Skyrim mod was made that communicates with ChatGPT to enable freeform "voice chat" with NPCs in the game. An interesting tech demo of sorts for what we might see in future games.