21
votes
AI Dungeon 2: a text adventure game that uses OpenAI's GPT-2 model to respond to any actions that you enter
Link information
This data is scraped automatically and may be incorrect.
- Title
- You can do nearly anything you want in this incredible AI-powered game
- Published
- Dec 6 2019
- Word count
- 477 words
Looks like it's down until he can figure out why it's so expensive to run: https://twitter.com/nickwalton00
Holy shit. He's saying it's costing $10,000 per day.
https://twitter.com/nickwalton00/status/1203370250030350338
To put that in perspective I just looked at wikipedia's financial statements and it looks like they pay "only" $6417 per day in hosting.
Apparently AI is really computationally expensive.
Wow, that's insane. I don't think it's even computational expenses, it sounds like it's just the cost of transferring data around. In the replies to that tweet he says it costs 30-40 cents per user in data transfer fees when it "downloads" the model to Colab.
Aww. I understand if it has to go subscription based, but it was kind of wild to just be able to log into a website and have the AI spin up adventure prompts for you.
Anyone know enough about how this works to tell whether you could just download it and run it locally? Isn’t the expensive part of AI usually the training?
You can run it local, but it's going to be a dog without mega GPU clusters to generate responses on the fly.
EDIT: http://familiarcycle.net/aidungeon-on-windows.html
The latest update has a P2P version running, kinda sort of? Hopefully that works as a stopgap measure.
Seems more of a gimmick than a game to me, it's like talking to Cleverbot.
Yes, although perhaps more impressive if you think "collaborative role-playing partner". It might not be coherent but it understands "yes, and...".
It specializes exclusively in "yes, and." I've tried writing stories together with GPT-2, and based on that and based on the adventures I've read from AI Dungeon 2, it seems like it really, really struggles with contradicting anything. Like someone deep in the twitter threads pointed out, jedi mind tricks are a nearly unstoppable superpower. Seems like GPT-2 accepts anything that's part of its prompt as unshakeable truth. The words it tries to predict next will always be based on the reality created by the prompt, so you can warp its reality by prompting it with a different one.
Maybe they need to add training data that shows a player failing at something, so it will generate failure messages occasionally?
But here's another way to look at it: although it appears you are the player, maybe you are actually the game master, and you're controlling a NPC? You get to decide whether the NPC succeeds or fails in any situation. (You could even roll dice if you want.) Then type in the result, and the game environment automatically tries to adapt.
But this "game" is still just an interesting experiment. I think there needs to be some kind of coherent world model (with persistent characters, places, and items) before it gets more game-like. Someone needs to figure out how to combine an evolving but coherent world model with a babble-bot.
I think it's fine to have a game that you can't lose when it's not skill-based in nature. In this case, it's an adventure; the whole point of the game is to enjoy the journey.
In any case, it's not like the AI always lets you have your way. Check out this amusing transcript for an example where it doesn't seem to let the player do exactly what they want all the time.
Oh that's great!
So a later version could have success/failure conditions for a more realistic experience or would that take away from the experience of "solve this however you would like?"
Probably be hard for it to judge the odds and consequences of success or failure. That'd be interesting, though, and I really have no idea what kind of stuff they have layered on top of just a pre-trained GPT-2 model. I don't necessarily think this way is wrong or bad though. It just means that the player needs to be mindful of "godmodding" because the AI is a complete pushover. Or they can shove the AI all over the place and make a world of their own if that's what they prefer.
From a subreddit that's popped up for this game, a little tip to help the AI overcome its "yes, and..." tendencies if that's what you want:
Yeah the examples were pretty janky. But I think it probably works better if you use discrete actions vs vague ones, eg "bandage leg" vs "heal"
This could probably be a big deal, even if it has a similar sense of "lolrandum" to the robot generated chapter of Harry Potter that made the rounds some years ago. The sheer flexibility offered is impressive and tempered down version would allow for all sorts of "I seduce the guard" solutions that would be impractical to put in.
So I have been playing this for a little while and decided to actually do that to Orgrim, the leader of the orcs who were invading the town. Needless to say it soon became porn.
Honestly, I wonder why they bothered with the fantasy setting at all when they could have just skipped the middleman and turned it into a cybersex simulator. They would have much more luck monetizing that service. :P
Mobile Versions are out!
iOS: https://apps.apple.com/us/app/ai-dungeon/id1491268416
Android: https://play.google.com/store/apps/details?id=com.aidungeon
Reviews seem kind of mixed. I guess it works sometimes?
It seems like they are getting hammered. I can't go more than a couple prompts in before getting the Jason error.
Just another client then? Getting it actually running on a phone would be impressive.
It would also take an hour between prompts.
That's the first game, no? Not AI Dungeon 2.
Links on the website up top lead to those apps, so I think they're the second version.
From the article: