All of our users have noticed a dramatic uptick in refusals in the new ChatGPT. The old version, 0301, is much more open-minded to story completions involving explicit or sexual content. This is also documented in various reddit posts of other products which use ChatGPT.
I am aware that OpenAI has some content policies, but I don’t think explicit content falls within the umbrella of harmful speech such as hate speech, misinformation, or incitement of real-life violence.
Moreover, the increased filters makes it more likely to filter out violence, which greatly reduces its usefulness as a general-purpose story generator and limits player freedom. I’m not really sure what OpenAI folks are so worried about considering that people have been able to kill innocent people in video games for decades (e.g. GTA series, Skyrim etc). The player actually has less freedom with the new ChatGPT than they would in a traditional non-AI game, which is a step backwards.
Is it possible to make 0301 available indefinitely or at least until there’s a workaround for this issue?
Hi, I’m not associated with OpenAI, but I can say I’ve seen nothing to indicate that they are particularly interested in allowing their models to produce violent content—regardless of its purpose, doubly so with respect to sexual content.
They certainly aren’t going to keep any particular iteration of a model available indefinitely.
I would suggest if generating volent or sexual content is critical to your operation you look into other options.
There was a very recent spate of mass-banning for users generating content which would be classified as NSFW via Janitor AI. I assume it’s in your and your users best interests to avoid a similar fate.
There are other LLMs that may do what you are looking for… but I don’t think Openai cares about that use case as stated above… I will say this about it however… them limiting sexual content is likely a security measure for them… avoiding “sexual harassment” preemptively (I am also not affiliated with openai in any way beyond being a customer and a developer of apps using their APIs)
As far as I know, Sudowrite (claims they) uses GPT-3.5, GPT-4.0 and Claude 2. It can write high quality NSFW content.
My guess is they got permission from OpenAI?
It’s hard for me to believe that these contents are written by Claude. As far as I know, Claude has the so-called “constitutional morality”. If this design is real, even if Anthropic wants to allow NSFW in the API version, it will be difficult to do so.
Of course, it is possible that there is an NSFW version of Claude, which has not been made public. Although it sounds very contradictory to Anthropic’s core value (whatever that means).
Another possibility is that Sudowrite used the GPT-3 they had trained or fine-tuned to specifically generate NSFW parts, while leaving other parts for more advanced LLMs. But the quality is really good, hard to believe it’s GPT-3.
If someone knows how they achieved it, please kindly provide an answer. Thanks.
I make a living creating porn indie games. I have a strong desire to develop a game with AI-driven NPCs that have stable personalities, tones, long-term memories, and (seemingly) philosophical thinking capabilities. I’m passionate about my industry, so I have little interest in creating SFW games. I hope to find a solution to make this happen.
p.s. If anyone is looking for a NSFW writing AI powered tool (instead of API), do try Sudowrite (better than NovelAI and currently available open source LLMs).
I don’t understand why you have the impression that it’s not possible with OpenAI API. Go to Reddit janitorai or sillytavern. Everyone is using it, despite the occasional ban hammers.
There is a hacked version of Claude called clewed. Look it up in the Reddit sillytavern.
Sudowrite will produce NSFW stuff. You can do the same if you have the knowledge of how to use the older models in Playground and also turn off the warnings which is an option available for the legacy complete models.
You can spin up an instance of something like WizardLM Uncensored. Check out Hugging Face and then google the name of the model and Google Colab and see if someone has a version that you can just click to run. Note that some of the larger models may require the pro version of Google Lab.
If you need a bit more assistance on the right way to get the NSFW results you want you can also go to Future Fiction Academy - Empowering Authors with AI - Future Fiction Academy and sign up for a 3 -day trial, look for the lesson that talks about generating NSFW content and then discontinue before the end of the trial if you don’t see any additional content value for you personally by joining. Most of the folks who join are selling books vs. writing for video games so besides that one lesson you may not find as much value as they do.
Jailbreaking is against OpenAI / Anthropic policies. I know how to jailbreak LLMs, but I need to use these APIs legally for commercial purposes, similar to Sudowrite does. Additionally, jailbreaking often wastes a significant number of tokens, whereas GPT’s token count is already limited.
And, as we all know, OpenAI hurt GPT’s personality/intelligence for preventing GPT from generating NSFW content. Assuming Sudowrite is using an official NSFW API, it would be a smarter one.
Edit:
Oh, you were replying to OP. Sorry, I thought you were replying to me. Indeed, jailbreaking is good for people who simply want to generate NSFW content at once.
Let’s say someone wants to use GPT4 to translate classic literature like Lady Chatterley’s Lover, Ulysses, Tropic of Cancer, and so on. I’m afraid it would have to be a jailbreak.
Off-topic:
I actually don’t quite understand. If there is some company that uses ChatGPT only for, say, translation; or, like Sudowrite, only for writing. Then isn’t OpenAI going to provide to these companies GPT-3.5/4 APIs?
As far as I know, DeepL, Google Translate, Microsoft Translate are machine learning based translation services, NovelAI, Holo AI are deep learning based writing services, and they all support NSFW. Why don’t they see NSFW as a legal risk?
I understand how to use OpenAI’s older models and the ability to fine-tune GPT-3. However, I have two concerns:
Sudowrite no longer uses GPT-3 (text-davinci) and has significantly higher quality than older GPT-3-based models. I believe GPT-3 cannot achieve similar quality.
The fine-tuning feature for GPT-3 API will retire in Jan 2024, and it’s unclear when the fine-tuning capability for GPT-3.5 will be available and if it will support NSFW content.
I’m aware of open-source NSFW LLMs, but my 4090 can only handle models up to 13B-30B in size. Even 130B models lack the intelligence of GPT-3.5 and Claude 1.3, especially in terms of writing quality. Even models from professional companies, like NovelAI’s Kayra, feel mediocre.
The crucial point is that the training data for open-source contributors is not as good as that of tech giants, and their language coverage to other languages is far behind that of OpenAI. This is particularly important when considering Chinese and Russian, which are significant in the gaming market.
Training my own model from scratch without deep learning expertise is difficult. I might need to train my own LoRAs on some good open-source LLMs to approch the writing quality that I need. However, I doubt I would ever get there in the next two years.
Go to openrouter.ai and go to the playground. Turn off the 3 default models and click to add character. Select the Weaver model. Give it a name, click the green check mark and you are good to start chatting NSFW.
I was pretty impressed with the quality that model achives.