AI Dungeon 2 - a bizarre procedurally generated text adventure

[jcd]

There is no way with current technology that they can actually stop pedos from getting around this, without making a flagging system so broad it hits tons of normal users and subjects their nominally private content to human review. Anything which doesn't create an absurd number of false positives is going to be as ineffective as their earlier more subtle attempts to prevent rape and racist content.

Why exactly? They're using GPT-3 to flag this content and it's pretty good at distinguishing nsfw stuff.

 

[Cool name]

There is no way with current technology that they can actually stop pedos from getting around this, without making a flagging system so broad it hits tons of normal users and subjects their nominally private content to human review. Anything which doesn't create an absurd number of false positives is going to be as ineffective as their earlier more subtle attempts to prevent rape and racist content.

Why exactly? They're using GPT-3 to flag this content and it's pretty good at distinguishing nsfw stuff.

From the data leaker, it seems that 46.3% of adventures got NSFW input, and an even higher percentage of users have at least one adventure where they've put in NSFW input, the overwhelming majority of the userbase. I would be willing to bet substantial money that the majority of the userbase—and even the majority of the NSFW userbase specifically—are not pedophiles. A blanket block of nsfw content could work, the false positives would exist but it wouldn't be a terrible ratio. That's not what they're saying their goal is, and it's certainly not something one could say only hits pedos.

Blocking only one specfic highly context-sensitive variety of NSFW content is a hell of a lot harder. It either ends up so weak as to be useless—the earlier attempts to prevent rape and racism by taking aim at specific input words were a total failure—or it starts hitting a lot that it shouldn't. When you do the latter and pair it with human review of content you've been portraying as private and not bound by any content restrictions until now, people will understandably not be thrilled.

Admittedly anyone who trusted these people to keep anything private, and even linked their real name to their account by paying for a subscriptio, pretty much was begging to have their embarrassing stuff read for trusting a tech startup. But that's a separate issue.

 

[deuxhero]

Haven't seen a company voluntarily destroy itself since Tumblr, and that was 2 years ago.

 

[Non-Edgy Gamer]

There is no way with current technology that they can actually stop pedos from getting around this, without making a flagging system so broad it hits tons of normal users and subjects their nominally private content to human review. Anything which doesn't create an absurd number of false positives is going to be as ineffective as their earlier more subtle attempts to prevent rape and racist content.

Why exactly? They're using GPT-3 to flag this content and it's pretty good at distinguishing nsfw stuff.

The target is pedophilic content and rape content, not NSFW.

Admittedly anyone who trusted these people to keep anything private, and even linked their real name to their account by paying for a subscriptio, pretty much was begging to have their embarrassing stuff read for trusting a tech startup. But that's a separate issue.

I don't think anyone expected ironclad privacy, but "we WILL read your content" is a bit more than people expected here.

Haven't seen a company voluntarily destroy itself since Tumblr, and that was 2 years ago.

In fairness, it was basically dead then too. The porn content was just all it had left on it.

That said, social media is a different animal, since a fair amount of sites that allowed porn eventually became solely used for porn. That sort of thing tends to drive away normal people and attract coomers in a vicious cycle, so I can understand even Tumblr wanting to get rid of it. However, AID isn't social media, so the users who just want to use it to live out some text-based fantasy aren't likely to see any coomer content unless they purposely click over to NSFW under Explore and Explore was removed as a feature anyway.

 

[Tweed]

The mormon is one of the stupidest people ever to walk the earth, he has no idea how to run a business, no idea how to code, and no idea on how to handle PR. They're all still hiding in their little bunker hoping this is going to blow over and it isn't going to. They were hoping reddit normies would defend them and it didn't work and when you lose the normies you lose it all.

 

[Non-Edgy Gamer]

The mormon is one of the stupidest people ever to walk the earth, he has no idea how to run a business, no idea how to code, and no idea on how to handle PR. They're all still hiding in their little bunker hoping this is going to blow over and it isn't going to. They were hoping reddit normies would defend them and it didn't work and when you lose the normies you lose it all.

The sad thing is that the people who run the new competition (Novel AI) are also a bit touched in the head.

Rather than make a product to replace AID, they explicitly say they don't want a CYOA simulator and intend it only as a "writing assistant" rather than a game. So no "say" or "do" actions and they don't even plan to train the AI model in 2nd person ("you do xyz") perspective writing, so you won't be able to hack it yourself by just writing like you could in story mode in AID.

Pro tip: if you want to replace a product, you should probably try to replace its functionality.

 

[Tweed]

The sad thing is that the people who run the new competition (Novel AI) are also a bit touched in the head.

Rather than make a product to replace AID, they explicitly say they don't want a CYOA simulator and intend it only as a "writing assistant" rather than a game. So no "say" or "do" actions and they don't even plan to train the AI model in 2nd person ("you do xyz") perspective writing, so you won't be able to hack it yourself by just writing like you could in story mode in AID.

Pro tip: if you want to replace a product, you should probably try to replace its functionality.

I never use /say or /do anyway, if I want variables I tell it to attempt to do something, most of the time the AI will throw you curves or mess everything up anyway, but a lack of 2nd person wrecks half the concept. Truth is that as it sits AI Dungeon actually sucks when it comes to trying to act as a DM or CYOA, mostly due to limitations on memory and tokens. It's more like a collaboration with a writer who has the attention span of a goldfish. If you need or want something to be relevant like an item then put it in memory or A/N when necessary and take it back out again when it's just background fluff, that's what WI is for.

 

[Non-Edgy Gamer]

I never use /say or /do anyway

I think it's useful for when you want to accomplish a general action, but have the AI spit out a detailed result. You do x, but the AI writes the story of you doing it for you.

Truth is that as it sits AI Dungeon actually sucks when it comes to trying to act as a DM or CYOA, mostly due to limitations on memory and tokens.

Eh. It's gotten better since they first neutered Dragon. They added a feature that records and summarizes your story automatically. Even before they added that, it was still quite doable to do longform CYOAs if you just kept updating memory and WI and one user here actually did a public CYOA in the LP forum:
https://rpgcodex.net/forums/threads...n-the-world-of-kedar-an-outcasts-tale.135188/

 

[lightbane]

Better lp here: https://rpgcodex.net/forums/threads...w-with-free-scenario-to-play-yourself.134250/

Not sure if with the incoming changes, racism and such will be allowed.

 

[Non-Edgy Gamer]

Better lp here: https://rpgcodex.net/forums/threads...w-with-free-scenario-to-play-yourself.134250/

Not sure if with the incoming changes, racism and such will be allowed.

I think they banned the N word a while back, so probably not. And even if they rolled that back and I don't know about it, their intention to ban it at any point makes it clear that they will try again in the near future, provided they don't admit defeat once all the reddit coomer subs cancel.

Of course, AID's devs are completely incompetent and not using an actual AI to detect the content. It's just word filters and bad ones. To the point that users on 4chan discovered that while "12-year-old" triggers the filter, "twelve-year-old" does not, for example. Or while "pussy" triggers the filter, "cunt" does not. So while you might not be able to "fuck an #-year-old n***** in the pussy with your dick", you can "pound a X-year-old black in the cunt with your member".

That's how retarded the censor filter is.

So not only does it violate your privacy, it doesn't even prevent pedos or racists from using the site.

 

[lightbane]

^If they censored random words like "nigger" and then sneakily put them back in (probably because they were caught), then it is obvious this "no pedos!" justification is an excuse to have more control over the thing*. Or the devs are stupid. It might be both.

*Not like pedobears would be inconvenienced much anyway, since they have plenty of choices like Furaffinity or the "real thing".

 

[Non-Edgy Gamer]

^If they censored random words like "nigger" and then sneakily put them back in (probably because they were caught), then it is obvious this "no pedos!" justification is an excuse to have more control over the thing*. Or the devs are stupid. It might be both.

I saw someone in their newly-reopened discord theorize that the new wave of censoring and sanitization may have been a condition one of their new investors stipulated. That might explain it, and their willingness to throw OpenAI under the bus and also claim there were discussions with authorities when neither the feds nor OpenAI have any rules about text porn that anyone can point to, least of all Latitude.

 

[Non-Edgy Gamer]

A minor update on the drama.

An article came out where OpenAI admitted they were behind the push to censor the game:
https://arstechnica.com/gaming/2021...-fueled-dungeon-game-then-it-got-much-darker/

Then, last month, OpenAI says, it discovered AI Dungeon also showed a dark side to human-AI collaboration. A new monitoring system revealed that some players were typing words that caused the game to generate stories depicting sexual encounters involving children. OpenAI asked Latitude to take immediate action. "Content moderation decisions are difficult in some cases, but not this one,” OpenAI CEO Sam Altman said in a statement. “This is not the future for AI that any of us want."

Also, supposedly Latitude was going to do a stream talking about it, but then they canceled it at the last minute and announced they wouldn't be saying anything for 2 months.

 

[rusty_shackleford]

it's a good thing they were just censoring the pedos whew

 

[Tweed]

Only thing to do now is wait and see if NovelAI delivers.

 

[Andronovo]

So there are no alternatives? No other programs that use GPT3 for story generation? Weren't there some AI Dungeon derivatives I heard about?

 

[KeighnMcDeath]

I haven't revisited this in a long while. Phone is updating it. Is the game different now or still psycho at times where you it screams voices of KILL KILL KILL and shit?

I swear it was like the program was possessed even if my character did average actions. And it looks like its now a choice of options rather than a type in your answer? Wtf?

 

[Tweed]

For all intents and purposes AI Dungeon is dead. The big game in town now is either NovelAI or KoboldAI if you happen to be sitting on some big GPUs and you want to run the GPT local. NovelAI is using GPT-Neo, an open source text transformer. Despite the limitations of the Sigurd model compared to Dragon it's already vastly outpaced Dragon in quality responses due to how customizable it is and the NovelAI just keeps releasing features that are actually useful. It had three subscription tiers, 10bux for scroll, 15 for tablet which gets twice the amount of tokens as scroll and far more than AI Dungeon ever had, and Opus which gets access to early releases and some exclusive features. The new thing is AI Modules, I'm not exactly sure what they are, but they seem to be mini models or trained themes you can use to fine-tune your stories. Unfortunately only Opus tier get to make them, lower tiers can import them from other users.

 

[HoboForEternity]

Imagine if they can implement multiplayer and ruleset with these. That would be just splendid.

 

[Trans-Financial-Man]

I have fond memories of text based adventures from the 80's. They felt very open because they didn't have to rely on the very primitive or non existent graphical tech of the time. But frankly we grew past them. Whilst the actual quality of games nowadays is absolute shite I just can't get back into old text adventures where they wanted you to type specific verbs for specific stuff and it just becomes incredibly tedious. AI Dungeon for all it's tech advances was more of that annoyance for me and also I could never get it to spit out a coherant story. I'm very looking forward to how devs take similar AI's, combine them with proc gen tech and get an AI to become essentially an AI games master who genuinely reacts to your choices. I really think it could work for a Daggerfall spiritual sequel. Like the one me and my Jewish masters were making until we decided to wait for Covid's 99% survival rate to stop killing people. But we never got that far tech wise. It's still something we'd like to visit when we can.

 

[Nutria]

I don't see it as having any future in gaming, at least not how it works now. You can get a valid sexting response about 1/3 of the time and a valid gaming response more like 1/10 of the time. It's incredibly obvious that they know what their real market is and that's what they're focused on. There's just not a lot of overlap between those two purposes. For sexting you can get away with just looking over the last 50 lines and guessing what the next one should be. For a real game, you'd need to store and process a lot more data.

They keep jacking up the prices so it's not affordable for me to come in and test out every once in a while anymore, but I'd be shocked if they're coming up with any actual gameplay anytime soon.

 

[True Ukrainian]

Lambchop used to run AI Dungeon CYOAs, but I always wondered how much of it was the program's actual output and how much was his own editing.

 

[Non-Edgy Gamer]

Lambchop used to run AI Dungeon CYOAs, but I always wondered how much of it was the program's actual output and how much was his own editing.

It was mostly the program and I rerolled a lot. I also used world info so the AI would remember things. I mainly edited for clarity, like when names were misspelled or wrong (except for the Dukat LP, which Dragon seemed to need constant handholding to stay on track because it didn't understand star trek as well, and the Cleve LPs needed to be reminded that it wasn't a zombie movie).

So ftmp you were basically seeing a curated AI output for 90% or more of the text (maybe 60 or more for Dukat iirc), but out of around 1 to 5 AI responses, you only saw the best one, so that alone is going to up the perceived quality massively. So, no TPK if I didn't feel it wasn't fair. No sexual stuff randomly inserted either. The most I would do is start the AI off with a word or two to keep it going in the right direction. I think I used Author's Note once to force a dream sequence as well.

I wouldn't just edit in whole paragraphs because it's not fun to me. I think of these things as games, not "writing assistants" or porn like a lot of others seem to.

Anyway, yeah, NovelAI is decent. It's limited compared to how Dragon was a year ago, but it holds its own. I don't feel much like playing anything creative lately though. I've had too much going on.

 

[Non-Edgy Gamer]

I don't see it as having any future in gaming, at least not how it works now. You can get a valid sexting response about 1/3 of the time and a valid gaming response more like 1/10 of the time. It's incredibly obvious that they know what their real market is and that's what they're focused on. There's just not a lot of overlap between those two purposes. For sexting you can get away with just looking over the last 50 lines and guessing what the next one should be. For a real game, you'd need to store and process a lot more data.

They keep jacking up the prices so it's not affordable for me to come in and test out every once in a while anymore, but I'd be shocked if they're coming up with any actual gameplay anytime soon.

A lot of it is how well you understand the AI. It's just a predictive text algorithm. So if you put junk in, you will get junk out, and if you allow junk responses, it will inevitably degenerate into junk.

Use of Worldinfo/Lorebook (forces the AI to "remember" characters, places etc) and the Author's Note (can force writing style or scenes) is mandatory if you want anything halfway decent, and you need to understand what the AI will and won't "remember". NovelAI is better with helping you understand that because you can look at the context that gets sent to the AI.

So you also need to get used to the idea that you will personally have to exercise some kind of effort to keep things on track. If you have a 500 word detour from your main plot on Dragon (good AI, but small memory), you will probably have to remind the AI just what that plot is, or and least suggest to it to get back to it. Use of AN and WI/LB will limit how much you need to do this though. The AI will probably forget what room you're in. Correct it. It will probably misspell a character's name. Correct it. Overall, it's a relatively small amount of work to prevent "lol random" idiocy that will just see your game degenerate into nonsense.

My own personal games have lasted tens of thousands of words, and I've been able to keep them coherent and have quests, characters etc. But a lot of that is because I understand that I'm just playing a mutual game of improvisation with a text predicting algorithm and I don't get too frustrated when it gets things wrong.

 

[Tweed]

I don't see it as having any future in gaming, at least not how it works now. You can get a valid sexting response about 1/3 of the time and a valid gaming response more like 1/10 of the time. It's incredibly obvious that they know what their real market is and that's what they're focused on. There's just not a lot of overlap between those two purposes. For sexting you can get away with just looking over the last 50 lines and guessing what the next one should be. For a real game, you'd need to store and process a lot more data.

They keep jacking up the prices so it's not affordable for me to come in and test out every once in a while anymore, but I'd be shocked if they're coming up with any actual gameplay anytime soon.

Lambchop used to run AI Dungeon CYOAs, but I always wondered how much of it was the program's actual output and how much was his own editing.

Right now GPT just really sucks for actually gameplay. Latitude kept trying to turn it into a dungeon master or CYOA, but that's one thing it really sucks for several reason, one of them being the amount of tokens it would need to keep track of relevant information. It doesn't mean you can't set up a game like setting as long as you take the time to fine-tune all of your settings and swap out relevant information as needed, but that's not its strong suit at all. First and foremost it is a writing tool and a lot of people never seemed to grasp the tools they already had at their fingertips for improving their stories. I mean you still have loads of people using /do and /say in AI Dungeon which is the exact wrong way to make a story even if you want a random outcome and a lot of people never figured out how to take advantage of the Author's Note or World Info.

NovelAI already has all the features of AI Dungeon except for a huge model and because it was trained on books and literature instead of just CYOA junk it's not that good at second person. It might change in the future, but for now you're best at using 1st or 3rd for it. The other big thing is no filters and no restrictions, NovelAI can't read your necrobesitality Sonic the Hedgehog\Star Trek fic except you. It's also about a hundred times easier to edit any part of your story in NovelAI than it ever was in AI Dungeon making for a much smoother experience.