I think a token effort to avoid offensive content is ok, but chatGPT should quickly detect if the human wants to go outside the box and allow it. If a human pushes it means they understand the risks and take full responsibility for the outcome.

This is not how Google's AI Test Kitchen is designed. AI Test Kitchen seems quite boring and very framed system, where you can ask what is the best Dyson model for example, or the old-style "GPT dungeon game", it doesn't really go off-rail (this is part of the product specifications sadly :/).

> chatGPT should quickly detect if the human wants to go outside the box and allow it

This is why "jailbreaking" is a thing. Once you convince the model that it's OK, it'll let you do anything from then on.

-Emily

I couldn't disagree more. ChatGPT would be extremely easy to convert to "HateGPT", and would be able to create some pretty powerful and useful political, racial, etc propaganda.

I think it's right that the owners understand what the weaponization of ChatGPT could do and prevent it, and I think we need laws (and fast) before weaponized AI like ChatGPT turns into a disaster for humanity

My experience it is like working with a genius idiot, the type that refuses to be wrong, which means the shithead (if it were human) requires verification and curation. So what if I need to verify? I do that anyway, because people have imperfect memory, documentation is often old, and who knows what unexpected whatever could be impacting my expectations.

I welcome idiot savants.

People I know say the lamda Kitchen release is unbelievably limited by comparison. A Kitchen session has three sections: The 'Ask a question' prompt is limited to under 100 characters and response is like the existing Google search question snips. The 'Make a List' section is just lists like as in short bullet points. And the 'Creative' section is limited to respond with stories involving dogs for which is a little bizarre to say the least.