That went on for a long time before getting to the point.
"Trained on gigabytes of input, GPT-2 is uncannily good at producing sensible text by simply predicting the next word that should follow a given list of input words."
There are now a few systems like that - autocomplete with a big text database. They have no underlying model of what they're talking about. So, after about three sentences, you realize the output makes no sense. It's just words strung together with probabilities that match known text.
This seems to be where machine learning gets stuck. There needs to be some underlying model of the subject matter to have a useful dialog. Outside of well defined subject areas (sports, weather, travel planning, shopping, etc.) that barely exists. Frustrating to see this half a century after Eliza.
Chatbots have the same problem. Either they force you onto a specific track, like a phone tree, or they just natter endlessly without going anywhere. I've been playing around with Rasa lately. This is a chatbot which uses Tensorflow to match user questions with canned answers. That's about all they get from the machine learning part. Outside of that it's a phone tree engine. There's a file of "smalltalk" questions and answers, so it can natter better.
What do we want? Chatbots!
When do we want them? Sorry, I don't understand the question." - bag at chatbot conference.
Isn't this just a Markov chain? I did the same thing with a chatbot years ago, trained on emails or chat logs to sound like a certain person. Are we really making any progress?
I'd like to see someone use a procedurally generated world with a chat bot text description to add flavor to it. The author highlights the problem of procedural worlds in the article as being endless expanses of bland repetition but what if that just served as the armature onto which you add the bizarre imaginings of GPT-2. Use the procedural world to seed chat bot and then use the chat bot output to set the description and parameters of a simulation.
I think even humans can still have the same problem. It's the difference between actually knowing subject matter and bullshitting. There are some incredible bullshitters out there and they can be very successful. That BS gap for chatbots is getting smaller, but the ability to actually learn something and apply that knowledge abstractly seems to be as wide as its ever been. Sure you have these bots winning Go and SC2 tournies, but good luck asking it something like "How would you improve those games that you're programmed to play?"
I suppose the NLP idea has been for first RNN and LSTM and now that Transformer and friends to act as a kind of memory. Mostly those memories have been pretty black box however, and it is interesting if ideas from data structures etc. could be useful.
Ideas from "data structures" will keep you stuck in the problem GP is talking about. The problem being lamented isn't that we don't know clever tricks for representing the training that a machine has undergone (as well as various tricks for having 'memory'), it's that for there to be a meaningful output for certain domains, there needs to be a thoughtfully implemented model that codes what it is to be (for example) a DM. Said model is, of course, ridiculously complex, so we won't see anyone coding it by hand anytime soon. We'd sooner have ways to copy the relevant bits of a human brain.
A few of you mentioned Dwarf Fortress. I am aware of it, and considered adding a section on it after Elder Scrolls, but this was already twice the length of one of my usual posts, and I wanted to 'ship it.'
It was pointed out that I only presented two alternatives in my conclusion, but there's a third of a super-dense procedurally generated world a la Dwarf Fortress. I agree this was an oversight on my part. This points in the direction of an interesting subtlety about whether a rigorous simulation is what you really want at all - the DM 'simulates' a world in a way that is highly biased towards adventure, excitement and fun.
I may go back and revise to try to bring in some of these points at a later date.
One person mentioned the Mythic GM simulator, I'm aware of that also, but a) it's a very niche product and b) it still requires someone to GM, that person may just also be the sole player. It does provide a biasing function so that you can by 'surprised' by outcomes and guide your own adventure in unplanned directions. It actually is more in the realm of a DM aid, as some have mentioned this may be the most fruitful direction for AI, as an aid rather than a replacement for a DM.
Would you want to? I also wonder the goal vs the means. Do I want a virtual DM? Part of the experience/journey is the social component/ enjoyment of a campaign.
A part of me hoped he mentioned Skyrim in his article, since it is a landmark edition (and same can be said for the two earlier titles) in the sense that people still create new content and adventures for it almost a decade on. Creating an endless adventure.
Some are working on an open source implementation of it, and it would be interesting to see that work.
Yes. DF shows that with a complex simulation, emergent behaviour is enough to make for a very compelling game experience. Similar dynamics are at work in other simulation games such as rimworld, factorio and oxygen not included (obviously not in the fantasy genre). However I do think having a human involved could sometimes make the DF experience better by tipping the scales in favour of a more fun experience rather than what DF players call a "Fun" experience.
For example, recently I built a beautiful fort. Everything was working great and all my idiot dwarves were happy. Fluid dynamics worked slightly differently from my expectation and confusion around the presence or absence of a wall caused by the tileset meant that I went to fill my well I had a whole river running down through my base and slowly filling it up from the bottom with no accessible top entrance remaining. I had the choice of letting my guys drown or a massive grind to fix the problem during which probably 2/3rds of my base would have starved to death and/or gone into a tantrum deathspiral. I chose to just stop and start a new fort. An experienced human DM could well have decided to essentially change the rules of the simulation and thereby given me some out that allowed me to rescue the situation through heroic effort and sacrifice and get back on track.
This is the problem with pure emergent behaviour from complex simulation. There's lots of interesting stuff that can happen, but you can paint yourself into very boring corners also and the simulation doesn't care about that.
> It is the difference between writing a story together or playing along in someone else's story.
A 100% this, a good GM actually doesn't really know much about where the story is going -- you play to find out what happens. I also have a hard time imagining an automated DM effectively 'failing forward'. Every skill check players make should change the game state, pass or fail.
Never, ever let your players roll the dice and tell them "Sorry, you don't see anything". Every check should be an opportunity to move the story... somehow.
This is actually why I think it's viable. The mistake is that crpgs try to tell a very particular story, and procedural generation stops at creation. But that's not the GM's job.
The GM's role is to create and simulate the universe, and to fudge the simulation (and the players) a little to keep it entertaining.
And (sufficient) simulation is definitely something we can hope to do -- though few games really try.
But the automation has a slightly different role than your GM.. it's goal is not to "move the story forward", but rather to craft a universe so interesting the players will move forward on their own! The automation should not be rolling to find out whether something is behind the door.. it should simply be checking if the simulated creature has wandered there.
The drudgery of true simulation then must be mitigated, and the tools for mitigation are the job of the game designer and the hueristics of the simulation. That is, you need not simulate the atom, nor the food & lifecycles; just enough of the system to allow it to change and adapt on its own (and ideally be self-stabilizing to a degree; probably best done by acknowledging the players only explore a single dungeon, but it's affected by a world outside of it; this is the opportunity to inject whatever corrections are necessary to stabilize the environment, avoiding the problems of Ultima Online)
I disagree - the current RPG trend is away from simulating a world and towards crafting a narrative.
> The automation should not be rolling to find out whether something is behind the door.. it should simply be checking if the simulated creature has wandered there.
The better example is the GM deciding if there's anything interesting that could be behind the door. If yes, then you're rolling to find out the terms of the engagement. If no, if it's not interesting, then you're done. There's nothing behind the door.
To automate that, an AI would need to be able to gauge how "narratively interesting or emotionally meaningful" something is. I think we're a little way off from that.
>the current RPG trend is away from simulating a world and towards crafting a narrative.
Yes, making the mistake of the mediocre GM -- trying to lead the player(s) into following your story, instead of enabling them to create their own. The ideal GM (at least, for me, someone who strongly prefers ADnD to DnD5) maximizes player agency, while constraining it by the rules of the universe (and also operating as a manager, because you're dealing with meatbags where one guy will happily try to spend 5 hours dicking around in the bar when the rest of the party wants to go to the dungeon, and another idiot will be trying to "troll" the group, in the most uninteresting fashion).
The core narrative (and every GM inherently brings with him a narrative he wants to tell) is incentivized, in much the same way any open world videog-game RPG has a main questline, but the ideal GM allows for departure from it as long as its interesting.
The narrative-focused GM, and RPG, will only allow such a deviation so far as it doesn't detract from the main narrative, because there is only one narrative the GM has interest in telling.
Narrative-focused RPGs may be the current trend, but that doesn't make it a good trend.
>To automate that, an AI would need to be able to gauge how "narratively interesting or emotionally meaningful" something is. I think we're a little way off from that.
You're correct, but my point is that we don't actually need, or even really want that, for our imaginary automaton.
Humans, by their nature, will bring the emotional impact, and fill in the details of the narrative, entirely on their own, if they are enabled to do so. They will assign personalities to blank-slate NPCs, they will craft their own reasons for doing whatever they do, they will spin their own tale, so long as the system will allow it (and as long as its sufficiently limited; for whatever reason, limitations breed creativity). You just need to provide an environment that reacts to the players, and sufficiently interesting ways for the player to interact with the environment.
Give them a gravity/portal gun and players will come up with all kinds of stories..
In a tabletop, the GM enables the reactive behaviors of what would otherwise be a static map. In a cRPG, lacking the GM, you get a static map (except for very specifically defined reactions, usually determined by narrative events). In most open world rpgs, you get a large map and multiple narratives, but still rather static. In a simulation, you get the reactive environment; the closer the RPG approximates a simulation, the closer we get to the GM's role -- to react to player input (rather than prescribe it)
This is in much the same way that any game can be fun if you toss in multiplayer, and give players sufficient controls -- the players will do all the hard work of setting goals, spinning tales, emotionally attaching themselves to arbitrary constructs.
Games don't need narratives, they need agency. Then the narratives come for free.
You conflate "narrative" and "GM-controlled narrative". These are not the same thing.
The current trend enables player agency because the table as a group decides the narrative, not just the GM. It's about the table as a whole deciding what it most narratively interesting to happen right now.
Couldn't we enhance gameplay by having the milieux generated and then giving the DM the ability to use that model to roleplay with the players? That would eliminate the problem of having to account for every possible direction the players might go. Just look it up and improvise with the help of a detailed description of the area and it's denizens. Sort of like a module that just sets the scene for gameplay.
One thing I think everyone may want to consider is that the current experiments with GPT-2 aren't necessarily the only way to use it. It's just that it happens by its structure that feeding it "all the text up to this point" and asking it to generate the next words is really easy to make into a snazzy web demo. But you could also do things like try to set up a vector representation of some space and turn it loose on that. If it's prone to its mind wandering, you might be able to confine it to a particular place in the hyperspace or something. This may require more novel research (especially around how a human being finds the place they want to use in the rather complicated space) and a deep understanding of the algorithm, and perhaps it is impractical in the end. But it's at least something that could be tried. We are not confined to merely using it as a glorified chatbot as it has been.
Probably the game that is closest to being full AI would be Dwarf Fortress. You can play it in Adventure mode and you can literally talk with NPC's who all have backstories and lives.
You can even play in legends mode and read all the existing stories. I've also heard about a software that can show the stories with a better presentation/UI than the DF one, with additional features like family trees.
When I saw Telengard played on a friend's Commodore 64, I thought for sure it was the end of Dungeons & Dragons forever. Who would want to bother with getting a DM and buying lots of books when you could get the same thing from a computer whenever you wanted?
Telengard was great fun, but it was still similar to a rogue-like in the sense that you were just walking around a dungeon trying to kill, loot, and not die.
I always heard it was procedurally generated but the layout was always the same for me. Looking into it, it seems that to save on memory, the levels are "generated" on the fly, but from a set of seed numbers that were hard coded into the game and didn't change.
I tried mapping for a while but lost patience. Good thing I did because it allegedly has something like 2 million rooms.
Not including the "Mystic" system seems like a huge oversight. It's an analog game system that works using yes/no questions, 2D10, and a probability table. It's not too elegant, but it works wonders. It can be used to assist the DM, or as a full-on DM emulator.
(2) the Mythic Game Master Emulator actually leaves all of the details to either the players or a human GM, it just provides directional guidance. It would be interesting if you could use it to provide a framework around which an AI could provide detailed descriptions, though.
This is a great article. I've been considering where this will head with gpt-2 dungeon master showing how far AI has come.
While I agree with the article that so far everything digital has been a noble failure, I feel there is too much discussion between AI vs humans. I think the future is in AI-assisted human DMs. Software may not be able to "entertain and dazzle" with the creativity of humans but it can help coordinate many routine tasks in any dnd game allowing the human DMs to focus their time on the components that make the adventure fun.
Does it show how far AI has come? It produces a game that is fundamentally unplayable, and because it's purely based on a model of language doesn't have a very obvious path to improve, since there's no underlying model of the thing it's trying to represent.
Are any of these applications of GPT-2 that attempt to use a pure language representation as a way to sidestep modelling the underlying principles actually successful? I feel like I've not actually seen any real successes.
AI Dungeon is very playable if you are make an attempt to guide it in the right direction. If you try to confuse it, though, you will very likely succeed.
I agree if by "make an attempt to guide it in the right direction" you imply "suspend disbelief and critical thinking, excuse/actively-imagine-away dumb and nonsensical output and try to convince yourself that there is something rather than nothing there".
In other words, AI dungeon is very playable if you delude yourself into a state of continuous confirmation bias.
Great article about a topic that has fascinated me for quite some time. Please excuse the wall of text.
Procedural generation can provide unlimited amounts of content seeded from finite effort. It's still mostly hand-crafted, just at different abstraction level. Building blocks are constructed manually and algorithm that fits them together is also carefully designed and fine tuned. It is used often in indie games because of lower effort needed to give illusion of huge amount of content. Meanwhile AAA companies can afford armies of level designers to make a detailed words. I guess everyone agrees that human design beats procedural and current AI. It is more varied, imaginative and purposeful - the 3 ingredients necessary for immersion. The crowd-sourced approach has not yet been successfully applied, to best of my knowledge.
The work that GM is doing can be quite valuable, and yet it only reaches few players of that campaign and is remembered vaguely. What if we could capture and store GM responses to player prompts? What if other sessions could reuse these responses and build on them to create new responses? The basic use of recorded responses is to automate mundane actions and checks (fights, lockpicking...) so that GM can focus on creative stories and arcs.
I've thought about this for quite some time and came up with partial solution. The automatic response system could consist of rules that are hand-crafted by GM while the session is played. Each rule is response to some action and changes the world state. So, each rule has to store the name of action, portion of world state that GM considers critical for that action, and how the world state is affected by action.
For example, let's attack an orc with a sword. OK, the GM decides that player's strength, sword sharpness, orc's agility and a dice role are important in this interaction, so GM marks them as such. Because D20 dice came out as 12, GM decides that attack was fairly successful, and changes orc's state to severely injured. This interaction is saved and the gaming session moves on. Later on, the same interaction 'attack with sword' is repeated by with different world state. If important parts of world state are the same/similar, then same result can be applied. If GM decides that existing rule is not applicable, then differentiating world state is marked and different results are entered and stored under a new rule. Now system is has two rules for 'attack with sword' and can automatically choose between them, based on which rule matches the world state the most. With time the system becomes rich enough to simulate this action under various conditions, while still being able to accommodate new situations with help from GM.
Such rule-based system can be visualized as bunch of simple state-charts that are built on the fly, and they interact together by modifying global state. I can imagine this simple mechanism would be good enough for combat mechanics, crafting system, simple NPC simulation (moods, dialogs, quests), inventory interaction, terrain manipulation... It should be enough to hand-craft a rich immersive world?
The hard part is coming with data structure that can capture rich world state in enough details without overwhelming the GM. The structure would have to model locations with objects, each with its attributes. All this would have to be relative to character who wants to preform the action. I'm considering graph databases, but don't have enough experience with them to move onto implementation :(
"Trained on gigabytes of input, GPT-2 is uncannily good at producing sensible text by simply predicting the next word that should follow a given list of input words."
There are now a few systems like that - autocomplete with a big text database. They have no underlying model of what they're talking about. So, after about three sentences, you realize the output makes no sense. It's just words strung together with probabilities that match known text.
This seems to be where machine learning gets stuck. There needs to be some underlying model of the subject matter to have a useful dialog. Outside of well defined subject areas (sports, weather, travel planning, shopping, etc.) that barely exists. Frustrating to see this half a century after Eliza.
Chatbots have the same problem. Either they force you onto a specific track, like a phone tree, or they just natter endlessly without going anywhere. I've been playing around with Rasa lately. This is a chatbot which uses Tensorflow to match user questions with canned answers. That's about all they get from the machine learning part. Outside of that it's a phone tree engine. There's a file of "smalltalk" questions and answers, so it can natter better.
What do we want? Chatbots!
When do we want them? Sorry, I don't understand the question." - bag at chatbot conference.