Hacker Newsnew | past | comments | ask | show | jobs | submitlogin
Play AI Dungeon 2. Become a Dragon. Eat the Moon (aiweirdness.com)
137 points by apress on Dec 16, 2019 | hide | past | favorite | 56 comments


Previous discussion on the project itself (although a few things have changed since then): https://news.ycombinator.com/item?id=21717022

Impressively, the Patreon for the project just hit $8k/mo. https://www.patreon.com/AIDungeon


> Impressively, the Patreon for the project just hit $8k/mo.

Costing him $10k/day to run though.

https://news.ycombinator.com/item?id=21739879


Not anymore. A peer-to-peer network was used to avoid the massive network egress costs, while the creator worked on a cost-efficient hosted app.

https://twitter.com/nickwalton00/status/1203843111287316485

https://www.patreon.com/posts/update-32193930


Check the follow on messages to that they've gone peer to peer and it's no longer costing money to host.

https://twitter.com/nickwalton00/status/1204064712394076160


Not anymore.


I ran a fairly normal fantasy scenario where I found a missing son of a local knight. Then the king's guards attacked me on the street and killed me.

So I chose to keep playing, resurrected myself by transforming into a demon, and killed the guards. I went to the castle to confront the king. He didn't have a good explanation for why he tried to kill me, so I killed him and crowned myself King. No one put up a fight and the townspeople bowed before me.

The next day I mounted a dragon, raised an army, and marched on the nearby kingdom. They put up a fight, but I seiged their gate and it fell. I proclaimed myself King of that kingdom too. No one felt like messing with a demon king on a dragon I guess.

I did this again with a third kingdom and then asked the Pope to proclaim me emperor. (I may have been playing too much Crusader Kings 2 recently.)

It was at this point when I had gone so far off-script that the AI pretty much gave up and stopped being able to form sentences.


I won a game using the same pope moment!

When I learned that the game was developed by BYU I decided to play a Mormon spy. I was able to grow my church by building churches in poor areas and aggressive use of missionaries. After I was big enough I declared a holy war on the Jehovah's Witnesses and eventually steal their Mind Control Tech. After buying a major media outlet I was able to use the mind control to take over North America. I demanded the Pope surrender his church to me and Mormonism and the game gave me a victory.

Honestly I played about 100 of these and it was maybe my favorite one: https://pastebin.com/Ar9rBEvu


Wow, you got a lot farther than I did before sending the AI into gibbering madness. I discovered very quickly that if you went non-sequitur it broke badly. "You are in the castle". "Swim to the bottom and unlock the door". "You swim across to the other door and discover a river."


Oh don't get me wrong, I glossed over a lot of nonsense, especially in the first part. I didn't mention in the first sentence that finding the knight's son involved starting at a cave, having a random knight (Sir Ivan) approach me, talk about his missing son (Sir John) who had been missing for months, and yet was definitely still in his house (and how dare I suggest otherwise!), and yet was last seen at a nearby tavern.

When I visited the tavern the barkeep told me that Sir John was there last night but said that Sir John was seen at the cave. I returned to the cave and dug a hole (I didn't ask it to do this) and then heard a noise. It was some bandits, and I talked to their leader. The leader attacked me, and I stabbed him. I searched the bandit and found a letter written in blood from someone named John. Not Sir John, the knight's son, mind you, but John a bartender at the tavern whom had not yet been mentioned. And not the barkeep I spoke to before either.

John had apparently held a grudge against me due to some past incident, and after I stabbed him he had written this note to me in my blood to be carried to me by the bandit.

When I went to the tavern to ask the barkeep about John, he told me he knew John. Then Sir John walked in to the tavern. Sir Ivan, his father, was not mentioned again.


Super cool game, though admittedly it's less of an actual text-based adventure game, and more like a super advanced chatbot in the guise of a text-based adventure game.


I wonder if it might be possible to make it a bit more "game-y" by adding some manually-programmed structural elements to the game (like inventory, skill points, etc) and feeding those into the AI to tune its responses.

Maybe some sort of AI-powered D&D DM; the rules of the game would be hard coded, but the AI would determine how the rules are applied and how they fit into the story based on text input from the players, results of dice rolls, etc. (E.g. If the AI were to end a sentence with "Roll for initiative" then the hard-coded bit of the game would have everyone roll initiative and feed the results back into the AI.)


Why would manually-programmed structural elements be necessary? I'm interested to see if the AI can be trained to convince me that I'm actually playing a game.


Because GPT-2 in its current form isn't that great at keeping track of facts (like the player's inventory or skill points) over the course of long conversations, or at consistently enforcing rules (as the OP's blog post demonstrates). Allowing a more logically stringent system to handle those parts of the game might produce more realistic results.


Because of the lack of manual structural elements, there basically aren't any explicit rules for how the world works. In practice this means that the stories sound more like a fever dream, the AI will let you do almost anything, and story beats are sort of coherent moment to moment but the arc of everything makes no sense.


It doesn’t have to have structural elements, but it does need a constraints system.

You: I kill all of them and become an immortal god.

Game: No, that doesn’t make sense. You can’t just kill everyone; you can run over and attack them, and see how that goes for you if you like...

Without constraints it’s not a game, it’s just telling yourself a story.

...nothing wrong with that, it’s fun. ...but it’s not a game imo, any more than giving someone a blank piece of paper and a pen is a game.


Agreed, it would be more interesting if it could dynamically construct the rules and paradigms of the game. It would be like playing a new game every time.


Maybe something like K-BERT, the idea is simple enough, and can leverage pretrained transformer models.

https://arxiv.org/abs/1909.07606


I wonder if you could do this just by prepending information about the current state to each submission.

e.g.

Kill troll with sword could be submitted as either "I have a sword, so kill troll with sword" and "I have no sword, so kill troll with sword" That's obviously very hand-wavy but it would allow a more stateless setup.


Someone in the Discord server said they were successful with this approach. - Manually keeping track of the state, basically.


There's published research on chatbots with an approach like this; essentially instead of trying to learn a function from input sentence to output sentence, you do a function from "text about state of world<separator>input sentence or question" to the answer.


You can sometimes successfully feed it what you want. Need a dragon? "You find a dragon" will work, just not consistently


Part of the fun of exploring the old adventure games was trying to do the impossible though. "Kill troll with sword" when you don't have a sword shouldn't say "You slice the troll in half with your sword"


It's an exploration game; the first of its kind.

I have been planning on writing a sci-fi epic for years, as you do. It's very nice to bounce half-baked ideas off the AI just to see how it responds. I would not subject a human to the cruelty of my whims in this way. People largely prefer the safe response of logic and understanding, even when I mischievously have presented them with nothing of the sort. The AI responds with chaos and creativity.

It seems to actually learn and I have taught it about shrink-ray guns, which it seems to make appropriate yet unexpected responses to. At first the shrunk items just disappeared. Then the ray-gun's charge started running out. Later, shrinking the place the ray gun was in caused paradoxes. Eventually the shrink-ray actually made things shrink, but putting my space ship in my pocket transported me across space and time to a desolate, ruined place. I decided this was a poetic way for it to tell me it did not know how to continue.

The AI is a free-form role-playing partner with a delightful lack of tact, but it is not immune to reason. If one can suspend disbelief and fill in the blanks it is even barely intelligent.


You say "Exploration game, first of its kind", I say extra flexible text adventure lacking any kind of internal consistency.

Don't get me wrong, it's cool as hell. But it's getting tiring hearing everyone trying to spin obvious flaws as innovation.


Well technically it's consistent for 8 previous prompts... but that's probably not what you mean.

If you expect this thing to be a MUD or roll-playing game it's not going to meet your expectations. 'Text adventure' perhaps sets the wrong expectation and in that regard I agree with you that it falls short.


Maybe it's because I don't have a lot of experience with such dungeon crawler games, but I just can't figure out what people find so compelling about a program that spits semi-coherent gibberish back at you based on some input.[0]

With no sense of linear progression of time or cause and effect (you can say "discover holy grail" on your first turn and you're done), it feels like a stretch to call this a game. It's a game in the same sense as an Ouiji board is a game.

What am I missing?

0: https://news.ycombinator.com/item?id=21717515


Much like the early Eliza programs, the onus is on you to provide the context and have the story make sense in your head.

Unlike the Eliza programs, this one has taught itself what to say by mining a deep corpus of text, so creates a richer world, and seems to have some sense of memory (correct me if I'm wrong).


Personally I can't understand how someone wouldn't find this utterly wonderful and compelling but several of my friends have reacted like you so I guess you're not the weirdo here.


Well the $8k/month the author is making on patreon to work on this project clearly indicate that my view is not shared by all. :) Horses for courses.


Apparently the server costs are in the tens of thousands of dollars so they (or their university lab?) are losing money on it.


I absolutely love how this game follows the improv "yes and" rules.

Whatever you throw at it, will probably happen


Not for me! I played 5 or so games and every time it would spit me into a loop, or started attaching the same string to every response.


It just seems like another instance of getting amusement from seeing what gibberish or absurdity it'll spit out this time. So that puts it in a similar category as Goat Simulator, or Bethesda games before several post-release patches have come out.

If there were more structure, maybe the sibling comment about imagination would be more on point. e.g. Scribblenauts, a game with an impressive dictionary of things that can be drawn on command and interact with the world, full of possibilities, yet if you're just going at it from a pure solve-the-puzzle perspective you can beat basically everything by spamming "black hole".

I don't really like either kinds of games but I think I understand the appeal...


It's hit or miss. Sometimes the story just doesn't work and sometimes it unfolds in a very pleasing way. It's like a conversation. Most of the time it is boring, routine, uninteresting and sometimes it is really good... and then it crashes.


> What am I missing?

Not to sound cruel, but, maybe you're missing the fun in imagination?


Not cruel at all! I asked and you answered. And you very well may be right. My expectations were probably just off. I thought for some reason it would have more "world constraints" so if I say "swim to the center of the earth" it would recognize that as not possible.

I do find games with no rules tedious and not fun. The card game Fluxx, for example, I cannot stand.


I played it with a couple friends, just testing what weird shit would work. We laughed our asses off. If you're looking for a strictly structured game with a good story, that's not happening.

I love the jankiness of these early ML games/stories/etc. and that's what you're missing.


Fluxx has a ton of rules, does it not? It's just that the rules keep changing and you have to keep track of them.

I'd say a closer analogy would be Calvinball, where you just make it up as you go along.


In fluxx, cross-turn strategy is basically impossible, as the rules may completely change between turns. I've found that reading a book during other peoples' turns then just looking up and treating each turn as a completely new game is just as effective as tracking the game across turns and significantly less frustrating. So yes, there are rules, but only within a turn.

Furthermore, I was dismayed when I learned (upon first time playing) that "It's easy to get started: there are no rules to read!" actually means "you spend the entire game reading rules!"

I acknowledge that this is not a very social (and somewhat rude) way to play this social game, which is why I typically politely decline to play it.


Once you know what all the cards do, you don’t really end up reading them anymore. So, it’s a lie that there are no rules, but learning the rules is spread across multiple game sessions. The way we play other games, that’s roughly what happens with those games too. Games don’t “count” until we’ve all got a grasp of the rules.

You can plan strategy. Whether it’s more or less likely to be thwarted I can’t really say. I try to be flexible with a few plans. The main difference is you are more likely to be accidentally thwarted than in other games.


Calvinball at least has rules in the sense of physical laws still being in effect. AI Dungeon 2 doesn't even have that.


Been having a lot of fun with friends. Non-techy people don't really believe it, so when it works well and is giving back cohesive text, they are really blown away.


The problem of this is that it doesn't have a notion of "world state" other than the text of the latest interactions, so you can do anything but it feels like nothing matters.

Fixing that seems very hard though.


It doesn't even have that.

In 3 messages I went from a peasant in a fantasy setting, to someone who is reading a book in 1869, to someone who watches TV in a cottage.

Half the time it also seems to completely ignore your input and tell your character that you did something else instead.


Looks like an early entry point for something that can raise something completely new in computer games.

I'm looking forward to progress concerning generating some graphics based on text. I'd be even more "lucid dream" experience.

My adventure with computer games finished quite early - around Fallout2 and Baldur's Gate. Since then nothing could impress me, except maybe few examples (one I can remember was a racing game, which I found interesting, because it mapped a real world city, I can't remember the name, thought). I think that the main reason is that when I discover inner mechanisms of something it stops being interesting for me. With that kind of experience, which is more like digging into some subconscious areas of mind it won't happen to soon ;)


Oh, thank goodness this is real. For a moment I thought this was another tired Keaton Patti bit:

https://knowyourmeme.com/memes/i-forced-a-bot


This needs to be used for fortune telling / AI zodiac / dream interpretation. Because as a game its more like a chatbot... just make it into an all seeing eye chatbot.




Anyone else get `./venv/bin/activate: No such file or directory` on step 3?

Also, I can't quite understand from the main page: Is this the way to play, free or not? I became a Patreon, but didn't discover any other way to play the game.


I couldn't figure it out either.


Is there a way to run it self-hosted locally?


Yeah, just clone the repo, download the model, and run the start script. If you don't have a GPU with ~14GB of RAM expect each response to take a few minutes to be generated though.


Thanks! It's not that bad -- am on a 2080 8GB and it takes less than a minute for each response.


at least for me it complains my python is too new (arch) so I installed it on a nvidia cuda docker container (which is using 18.04) and for some reason it still complains it can't use CUDA but works fine on the CPU although slow


Hello




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: