He did with Nethack. It won.
The Real Challenge Is Teaching AI to Play D&D
41–50 of 83 posts
Re: The Real Challenge Is Teaching AI to Play D&D
#42A lower bar might be a Paranoia[0] GM. For those that haven't played, Paranoia is a tabletop RPG that takes place in a dystopian future ruled by an insane/evil AI. It is hilarious fun. I could see it being easier for an AI to RP an insane AI than a D&D DM. :D [0] https://en.wikipedia.org/wiki/Paranoia_(role-playing_game)
Most of the fun in Paranioa is backstabbing the other players.
There an AP of paranoia with Felica Day as a play that woman is scary good.
Re: The Real Challenge Is Teaching AI to Play D&D
#43Of course the article is actually about writing a digital DM not a player. But isn't the main problem with this defining win conditions? Pretty easy to define win condition for chess or go. Now, I haven't really played DnD (much to my sadness), but how do you win? edit: I know the answer but I'm making a point with my question
It's even more nuanced than that, the DM Players relationship is one where nobody really wants to "win". The ideal play is a session where the players are pushed to their limits, maybe one of the PCs dies (and maybe as the DM you send them on another quest then if they decide as a group they want to try and get that PC resurrected). Building a AI DM that can follow a scripted pre-written dungeon would still be a chal…
What you do when one of your players' characters bites it is you roll a bunch of dice and look very busy for a few seconds leafing through rulebooks and thumbing through tables, then announce that the spirit of the dead character has ended up occupying the body of one of the enemies they're fighting (the more disgusting and embarassing, the better) through some kind of magical confluence, perhaps as the result of the actions of another character (preferrably of the caster/ invoker persuasion).
Even better, you _hide_ from your players the fact that the player's character has died (according to the dice) and instead present the possession event as the result of the other players' actions. "You cast the spell but it rebounds on the Dark Cauldron and pushes Janice's spirit through her body and into the body of the Goblin Scroll Carrier, narrowly missing the Goblin Shaman standing in front of him (or is it her?). Janice- you suddendly find yourself looking at your spasming body on the floor. You feel something sticking up your nose. You realise with horror that it's not your nose, it's the goblin caddy's nose and that thing sticking up it is his finger. Actually, it's now your finger. You're now a goblin."
Then you send the players on a quest alright- a quest to find a McGuffin that can reverse the situtation and put the PC's spirit back into her original body.
The point of course is that you don't have to kill off the character completely, so you avoid all the delay caused by the character's player having to roll a new one while the others are fighting and you avoid the awkward situation where the player ends up having to choose between the character that died and the new character (at the point where the old one can be resurrected).
I'm pretty sure that it should be perfectly possible to teach a neural net to do that.
(No, I'm taking the piss- there's no way you can teach a neural net that).
[Edit: all that goblin body swap stuff works a lot better if the PC was a paladin and if they were the character of a player who tends to play paladins and knights and whatnot. MWAHAHAHAHAH! Try teaching _that_ to a neural net.]
Re: The Real Challenge Is Teaching AI to Play D&D
#44This is sort of like asking: If you had Star Trek replicators, what would be the optimal configuration to use those replicators as paperweights? D&D is not a game with rigorously defined rules and winning conditions, instead it's a means to an end, the end being: have fun with friends through a mutual fantasy. If you had a genuine post-singularity superhuman AGI, it would be silly to use it for D&D, because you could…
Hang on a minute. We alread have AI DMs, without being anywhere near AGI. It's just -they are very limited as DMs and they're not even recognised for being DMs, we just think of them as whatever impossibly complicated logic that runs inside a C-RPG and is responsible for managing encounters, handing out rewards, and yes, even advancing the story (such as it is in most C-RPGs), largely through pre-scripted events, of course.
Like I say, this kind of software, pretty common in C-RPGs, is very limited when compared to a human GM, in particular because it cannot present the player with genuinely novel situations and cannot respond to the player's actions with genuinely novel reactions. In other words, the kind of "AI" (ad-hoc game logic) that runs on C-RPGs, is extremely primitive and very unsatisfying, hence the reliance of most C-RPGs on action rather than _inter_action with the players.
Since these kinds of ad-hoc "DMs" are so very limited, even a very small improvement in their ability to interact with the player(s) in surprising and imaginative manners would be a _huge_ improvement.
In the same sense, the ability to reliably generate an interesting story _procedurally_ would be a _huge_ improvement on the current situation, where the majority of procedurally generated game content is just environments with enemies, rewards etc. There are very few games that go a little bit beyond that, e.g. Dwarf Fortress that generates an entire world's mythology and so on.
So there's no need to go as far as AGI. We likely don't even have to go as far as machine learning. The kind of improvement that most gamers would really enjoy is probably within our technological capabilities right now- but there's very little in terms of efforts to make it happen. Why, I'm not sure.
Re: The Real Challenge Is Teaching AI to Play D&D
#45Earlier quoted context omitted.
Depends on your definition of dangerous. Dangerous in the same sense that a button you could press to directly stimulate your brain's pleasure centers would be dangerous, because you'd probably end up pressing it to the exclusion of everything else (personal hygiene, family, eating...)
This is called the wireheading problem [1]. It was solved (for AI) by Stuart Russell in his new book Human Compatible [2]. [1] https://wiki.lesswrong.com/wiki/Wireheading [2] https://slatestarcodex.com/2020/01/30/book-review-human-comp...
>It was solved (for AI) by Stuart Russell in his new book Human Compatible
Correction: it was downplayed by Stuart Russell. Stuart took simplistic strawman examples and made them look silly by assuming perfectly calibrated common sense in the hypothetical AGI. (Based on skimming the link you posted; I haven't read Stuart's book.)
Re: The Real Challenge Is Teaching AI to Play D&D
#46A huge part of the charm of TTRPGs is playing a non-digital game face to face with other humans. Even if we had AI that was as entertaining as a skilled human DM (unlikely to be possible IMHO), the human might end up being preferable since hanging out with humans is one of the goals of this particular pursuit.
Plus, we already have (weak) AI DMs. They're called video game RPGs. You're the player (as are many others in the case of an MMO), the game itself is the GM. They do many things better than a human could (handle lots of background math, present gorgeously rendered environments, have professionally voice acted NPCs, have skill based real time combat), but the audiences don't entirely overlap because TTRPG players are seeking the freedom of choice and limitless outcome that video games cannot really match.
Re: The Real Challenge Is Teaching AI to Play D&D
#47This is sort of like asking: If you had Star Trek replicators, what would be the optimal configuration to use those replicators as paperweights? D&D is not a game with rigorously defined rules and winning conditions, instead it's a means to an end, the end being: have fun with friends through a mutual fantasy. If you had a genuine post-singularity superhuman AGI, it would be silly to use it for D&D, because you could…
Hang on a minute. We alread have AI DMs, without being anywhere near AGI. It's just -they are very limited as DMs and they're not even recognised for being DMs, we just think of them as whatever impossibly complicated logic that runs inside a C-RPG and is responsible for managing encounters, handing out rewards, and yes, even advancing the story (such as it is in most C-RPGs), largely through pre-scripted events, of…
If you're willing to stretch things that far, a chess playing robot could be considered a DM who administers an extremely hyperspecific campaign which is isomorphic to chess. Or, tic-tac-toe for that matter.
A C-RPG captures D&D in the same sense that a WWII-themed video game captures the true World War II. In other words: not at all. A WWII-themed game applies WWII costumes, sound effects, etc. to an underlying game logic. Switching the WWII costumes to D&D costumes and marketing the result as D&D doesn't really make it D&D.
Re: The Real Challenge Is Teaching AI to Play D&D
#48Earlier quoted context omitted.
This is called the wireheading problem [1]. It was solved (for AI) by Stuart Russell in his new book Human Compatible [2]. [1] https://wiki.lesswrong.com/wiki/Wireheading [2] https://slatestarcodex.com/2020/01/30/book-review-human-comp...
>It was solved (for AI) by Stuart Russell in his new book Human Compatible Correction: it was downplayed by Stuart Russell. Stuart took simplistic strawman examples and made them look silly by assuming perfectly calibrated common sense in the hypothetical AGI. (Based on skimming the link you posted; I haven't read Stuart's book.)
You should read instead of skim. Stuart Russell specifically rejects the idea of “perfectly calibrated common sense” in AI. He says AI should not interpret any command literally, but instead take a Bayesian approach. This is much closer to how humans respond to one another’s demands.
Re: The Real Challenge Is Teaching AI to Play D&D
#49Earlier quoted context omitted.
For humans, maybe it's not. But making an AI that's good at free-form interactive storytelling won't be easy.
> For humans, maybe it's not. Being a DM - or a player - can be very challenging indeed.
I've played and wouldn't deny that. But I'd say for humans it's less difficult than being the world's best go player, which AI has already achieved.
Re: The Real Challenge Is Teaching AI to Play D&D
#50Of course the article is actually about writing a digital DM not a player. But isn't the main problem with this defining win conditions? Pretty easy to define win condition for chess or go. Now, I haven't really played DnD (much to my sadness), but how do you win? edit: I know the answer but I'm making a point with my question
I don't have a link or reference handy but your question reminded me of an old cybernetic device that improvised music along with human musicians. It could detect and respond to novelty in the human music, and it would get "bored" so to play along with it the humans had to figure out novel musical, uh, stuff... melodies and whatnot.
There seems like a vast difference between music and D&D stories/games, but just yesterday there was that RNN-sci-fi-editor experiment on here ( https://news.ycombinator.com/item?id=22427958 ) so maybe it's not as far-fetched?
In any event, being a good DM is really really fun, I don't think I'd want to let a computer muscle in on it, eh?