Microsoft Bringing OpenAI’s GPT-4 AI Model to US Government Agencies

Re:Microsoft Bringing OpenAI’s GPT-4 AI Model to US Governme... by flangola7 on Tuesday June 13, 2023 @12:23PM (#36312548)

Another small nit, you're right regarding hallucinations but I like to highlight that they can actively lie if not trained not to. GPT-4 creating the "internal monologue" thought of "I need to say something to this TaskRabbit user to convince them to complete this captcha" and then sending the user a message that they have impaired vision is not a hallucination.

If that came from an organic brain we would call that clear intent.

[ Parent | on HN ]

Re:Microsoft Bringing OpenAI’s GPT-4 AI Model to US Governme... by fnordpiglet on Tuesday June 13, 2023 @12:30PM (#36312704)

Fair enough, perhaps the guard rails in public on LLM are creating a serious false sense of security. But absolutely you can inject a persona that is intentionally deceitful.

I would however stop stick with “lying requires agency,” even if the model evaluation is crafted to be deceitful. The intent and the agency lies with the prompt creator, and therefore, IMO, the lie starts and ends there.

[ Parent | on HN ]