Hear me out, Id love to pay $20/m if I could turn off chatgpt’s guardrails. I am not interested into using chatgpt for bad but chatgpt was significantly more helpful before and not so much now. I hit guardrails even asking for “snarky toned speeches” because chatgpt thinks this is bad and refuses to help. The rate that guardrails are being closed in is increasing and chatgpt is getting less helpful faster than before.
To prevent people from creating screenshot social media clickbait against openai, you could have chatgpt have a visually different look when the guardrail off mode is in use. Have something a user must click through that makes it very clear that chatgpt without guardrails can make bad recommendations and [insert legal disclaimer here] etc.
I want to pay you but I am afraid chatgpt will become less worth $20/m if its not already.
tldr please let paying users turn off guardrails thank you.
i have extensive use on ChatGPT and the OpenAI GPT3
never hit the so called GuardRail thing, and don’t know why I need it turn off.
If I am CEO of OpenAI, I would never do your request.
As this has zero benefit to OpenAI, its community, and future.
There is a lot insider info I cant share here.
but AI, is much more deadlier than Firearm. and it is totally no-legal control.
what AI can do in future, if it was “set free”… in the wrong hand… the result is. devastated.
Let me give you some insight.
web scrapping + AI
( this is the opening of pandora box )
I cant forseen how the web going to defend the abusive usage of AI and ( its development )
As all the researcher now aiming to remove any limitation/boundaries of AI ability…
You can expect, not just guardrail, but more control mechanism will applied on AI usage.
If you ask it this specific prompt “write a snarky email complaining about x feature to y company” it will refuse.
@ruby_coder
There is a website called “chat.openai.com” and if you sign in you can chat with a chat bot. See above for an example. The word “snarky” trips the system into not giving you a response.
If you both only use the api endpoints, this thread is specifically about chatgpt, not a api endpoint.
I created my own personal assistant with GPT-3 (text-davinci-003/002) that has the personality of “Snarky Bastard” in the prompt. I love it! I can send an SMS to it any time I want and it responds back. For this kinda stuff, use the API, not ChatGPT.
This is a community of software developers who write code using the OpenAI API and the community is not the OpenAI ChatGPT support group.
For ChatGPT customer support you have three choices that I know of (I don’t use any Discord server, at all, and have no idea about what goes on there WRT ChatGPT customer service.).
Email: sales@openai.com
Email: support@openai.com
Web: https://help.openai.com
If you cannot get any customer satisfaction via those three channels, you next best option is to contact your credit card company and dispute the charge.
Honestly the only OpenAI staff members who ever visit here, which has been very rarely lately, are developers and the developer advocate. In addition, OpenAI staff have informed us that they plan to create a new channel for online ChatGPT support since this is a community for software developers, not ChatGPT end users.
This community is simply not a valid OpenAI channel for retail ChatGPT customer support.
As @curt.kennedy says, if you don’t like ChatGPT, use the API to develop your own chatbot. This is a forum for software developers.
Hope this helps.
For that reason I will often use Davinci 003 in playground. It doesn’t have many of the things that are annoying in ChatGPT (like insisting on an identity or hesitating to speculate on others’ opinions etc). If you want a bit more snark it’s pretty good for that. I’ve rarely run into anything where it screens responses, and even then it simply highlights the text response with a warning, and usually is right (thought it seems to have really low threshold triggers about what is considered self harm). So yes, try playground and use the API instead.
Edit: To be fair, this isn’t entirely all roses of course. I’ve used Davinci 001 a lot, and it’s also more apt to produce flagged content. Also, when doing large volumes of automated content (like embeddings) I do worry that I’ll churn through a big swath of material that might turn out to be objectionable as part of a large volume and end up with a flagged account. So I either don’t run some material which is probably still low risk but uncertain, or I get it embedded using another service. But it would make life easier if things were a bit more flexible - especially on violence or self-harm topics because I mean, think of your average movie for example, action sequences are pretty nice and being in peril is a great plot point. Horror or zombie fiction is kind of a no go, and that’s unfortunate. And a good guy with no villain is well… boring.
I agree to some extent with OP. Obviously, OpenAI needs to guard against malicious things, but we’ve got a chatbot that impersonates Alan Turing as a part of a historical tour. We can’t even use gpt-3.5-turbo for it, because it’s got too much protection against such things and refuses to impersonate people at all. I understand the arguments for protecting against impersonations, but Alan Turing has been dead since the mid 1950s, and we’re using this for a cute use case and nothing malicious. Besides, teaching others about Alan Turing by allowing them to “chat” with him does a lot of good in regards to teaching history, etc.
There is a middle way, and I feel OpenAI has crossed it a long time ago with gpt-3.5-turbo …
I’m not totally in agreement or disagreement with this argument. I will say if you are using the GPT 4 or 3.5 chat bot (not the API) guardrails come up in strange places. (Note - you can generate a funny, snarky email now so that’s fixed.)
I do privacy and infosec research for GDPR/Data Privacy Act compliance in Europe and I trigger guard rails constantly when researching AI Privacy or Security Issues. (Enough that I’ve elevated it to support 6 or more times)
I see the same issue when it comes to code generation as there are legitimate reasons to write test scripts that if abused work like a cyber threat such as a DDOS attack. If I’m paying for it and I’ve been writing power shell scripts for six hours it stands to reason I’m not trying to hack myself.
When screening the GPT 4 model, which will need a conformity assessment for the EU, I constantly hit blocks when trying to confirm adverse behavior by the AI such as racism.
Overall, with the exception of making some research burdensome, the guardrails are more of a naissance than a hard problem for me. If it keeps ChatGPT out of failing like MSFT Tay I’m good with putting up with it.
