Writing AIs

Since it appears that AI Dungeon has gone off the deep end with excessive censorship and predatory creators, we should probably look for alternatives.

I’m not sure if it would be better to have just one thread for all of them, or individual threads for each AI, but I hope this thread is a good place to start.

One AI that I found, but haven’t tested yet, and is currently available for free, is Kobold AI.


One alternative that I’ve seen all over the aidungeon subreddit is NovelAi. It seems like it’ll be a good alternative, if not direct upgrade to aidungeon in more ways than censorship. They’re currently preparing for the open beta, which will be paid and probably observed, seeing as that’s the whole point of a beta, but after that when it fully releases it’ll probably be perfect.


Wait what, when did this happen? I don’t use it super often, so this is just news to me, is why I’m asking.

Edit: Nevermind, I googled it. lol


I’v been using regularly for fetish content and got only one issue when the ai fucked up. Is this really that big of an issue?

A robust and totally accurate text filter is impossible as we understand it now (see: the Scunthorpe problem.) It’s even worse when you try and apply it to a language model, as you don’t end up filtering based on strings or substrings so much as tokens (which are even less precise). Filtering tokens can and does have adverse effects. At worst, it can result in incoherence. At best, it gives the model weird quirks (see: agAPE.) So yes, when you combine the innate problems imposed by filtering with some of the other news about Latitude, I would say it’s a big enough problem and have already jumped ship.

To be honest, I gave up on AI Dungeon way before any of this stuff. I just found it way too aggravating to use.


That’s likely a consequence of coming in and expecting a game. Latitude seems to be hellbent on gamifying as much of the experience as possible but it’s just not viable lmao. it has been and always will be more of a “writing assistant” than anything else (though the quality of assistance has declined in some respects.)

However, Novel and Holo AI seem to better understand the nature of their services and are marketing appropriately. They both seem to have different focuses in the respect of finetuning and it will (most likely) be easier to get AIs that aren’t grossly overfitted to follow whatever threads you desire.


I’ve been thinking of finetuning a model (just GPT-2 774M, the GPT-Neo stuff would require some more processing power and I’d need to get a TPU from google cloud) on weight gain related stories, where should I look for training data?

FYI Eleuther has just recently released a 6B parameter model known as GPT-J. In my own experimentation thus far, it seems to be a vast improvement over Neo even without finetuning of any sort. It’s too beefy for most of us to execute it locally, but Holo allows you to use it right now if you have a premium membership (though there seem to be some technical difficulties as would be expected.) There was also the KoboldAI + Colab option, but it’s currently being throttled due to traffic.