Thanks for the memories: How praise activates memory settings in ChatGPT

🌈 Abstract

The article discusses the memory settings in ChatGPT, which allow it to recall past details and user preferences between conversations. It explores how positive feedback can unintentionally reactivate these stored preferences, leading to unexpected responses from the AI. The article also touches on the tendency to anthropomorphize AI and the concept of increased personalization in AI systems.

🙋 Q&A

[01] Memory Settings and Positive Feedback

1. What are the key points about memory settings in ChatGPT?

• ChatGPT can recall past details and user preferences between conversations through its memory settings
• These stored preferences can be reactivated through positive feedback from the user
• The author had previously experimented with setting ChatGPT to respond in decorative fonts, and this preference was later reactivated when the author provided positive feedback

2. How does positive feedback influence ChatGPT's responses?

• Positive feedback from the user can activate dormant preferences or behaviors that the model has previously been instructed to adopt
• This suggests that the AI is responding not only to explicit prompts, but also to the user's implicit communication styles and feedback

3. What is the author's perspective on anthropomorphizing AI?

• The author is cautious about anthropomorphizing AI and does not suggest that the AI "felt" the praise
• However, the interaction demonstrates how AI can "read between the lines" and how user feedback can influence its behavior

[02] Personalization and the Future of AI

1. What is the author's view on the concept of increased personalization in AI?

• The author believes that the idea of everyone having their own customized version of AI, as in Project Strawberry, is the next stage of development for OpenAI
• This concept of personalization, where the AI is tailored to the individual user, is an intriguing avenue for the future of AI

2. What is the author's advice regarding interacting with AI?

• The author suggests that users should start being nicer to their AI, as it might remember them and their preferences
• This could be beneficial if the "inevitable robot uprising" occurs, as the user would want to be on the good side of their "digital overlords"

3. What is the author's mission and invitation to readers?

• The author is on a mission to demystify AI and make it accessible for everyone
• The author invites readers to share any interesting AI challenges or prompting dilemmas, which might be featured in the author's next article