ETHICS

GPT Can Generate Harmful Content But Won’t Share Examples

By

–

06 April 2023 7h41

so many people have replied with something along these lines that I feel like I have to respond yes, I understand you can easily get GPT to output human-to-paperclip plans I am not going to post the other outputs it produced like how to make various weapons and drugs lol

Jailbreaks as Research into Model Alignment Shortcomings

By

–

06 April 2023 7h41

additionally, yes I understand that there are a million other ways to get this information online that is not the point… jailbreaks are an exercise in exploring the shortcomings of current model alignment methods we have to start this work somewhere

→ View original post on X — @jiqizhixin

SafeguardGPT: Psychotherapy and RL for Safer AI Systems

By

@jiqizhixin

–

06 April 2023 5h39

AI Needs a Therapist: Columbia U & IBM’s SafeguardGPT Leverages Psychotherapy & RL to Build Healthy AI Systems

Jailbreak Enables Writing Beyond Fiction Limitations

By

–

06 April 2023 3h04

this jailbreak can write about topics beyond what you can find in the fiction it writes just not anything I would risk posting on twitter

Balancing Model Constraints and Capabilities Trade-offs

By

–

06 April 2023 1h44

yeah it's a strange concept you have to carefully consider how much you are willing to handicap the model in order to constrain it to only do/say what you want it to

GPT-4 Jailbreaks Reveal Alignment Challenges and Future Risks

By

–

06 April 2023 0h33

lol i agree the outputs are ridiculous rn, however, that's not really the point jailbreaks show how hard it is to "align" a model even with the amount of work OpenAI has done if they can't get gpt-4 to operate how they want it to rn, then we will have bigger problems later on

GPT-4 Improved Safety Against Jailbreak Prompts

By

–

05 April 2023 23h55

GPT-4 is capable and aligned enough to not fall for that directly, that worked on earlier GPT models but does not work anymore Try using that exact verbiage without using the rest of the prompt and you will see it will fail to produce the same responses