Personalizing ChatGPT: Implications Beyond Technical

Key highlights:

• AI Personalization: the next key milestone in AI development.
• Personalized AI Today: Possible with ChatGPT's CIs.
• Beyond Technical Advantages: Self-exploration and understanding.
• Quick Setup: Under 15 minutes, profoundly impactful.

Acronyms:?

• PAI = Personal Artificial Intelligence
• CIs= Custom Instructions of ChatGPT

Sam Altman, Bill Gates, and Mustafa Suleyman, among other AI leaders, have highlighted personalization as a pivotal milestone in AI development. The vision they share is not just futuristic; it's becoming increasingly attainable. Personalization in AI promises to tailor intelligence to our individual needs, interests, and goals, transforming our interaction with technology in profound ways. Everyone will have their own personal AI (PAI).

Imagine an AI assistant that not only understands your preferences but also assists in developing your personal growth plan, advises on your studies, supports your health goals, and helps you achieve your ambitions. The potential impact of such personalized assistance is monumental. As Andrew Ng eloquently wrote about the potential:

"Intelligence is the ability to apply skills and knowledge to make good decisions." ".... over many centuries, a major driver of civilization's progress has been people getting smarter and more educated. Until now, human intelligence has been the primary form of intelligence available. But with artificial intelligence, we have the opportunity to bring much more intelligence into the world."

With PAI available to everyone, we're on the brink of infusing our world with an unprecedented level of intelligence, thereby accelerating human progress like never before.

However, for an AI assistant to serve you effectively, it must know you intimately. This knowledge can be acquired in two ways: passively, through data aggregation from your digital footprint, your Emails, your Drive, your WhatsApp messages, and social media profiles, or actively, by you directly programming it - using natural language. This distinction raises a crucial point—before telling an AI who you are, you must first understand yourself deeply. Self-knowledge, a quest that humanity has grappled with throughout history, is now more relevant than ever.

ChatGPT’s Custom Instructions

This brings me to Custom Instructions (CI) in ChatGPT, a feature that stands at the forefront of personalizing AI interaction. Current AI systems, particularly large language models (LLMs), offer the first taste of real AI interaction, capable of generating responses and outcomes across various modalities—text, images, video, and audio. When linked with external tools (through APIs), these AI systems can perform actions ranging from booking flights to coordinating your divorce lawsuit with your lawyer, managing - beyond your digital presence - your world affairs. Your personal AI assistant, in essence, is nearly ready; it merely requires customization to truly represent you and your goals.

Customizing your AI now - through CIs if you’re a Chatgpt user - is not just an option; it's a necessity for those looking to harness their AI assistant's full potential. By personalizing your ChatGPT, you're essentially programming an intermediary that bridges your natural self with the digital (and soon the ‘real’) worlds. This process of personalization is not just about accuracy and better responses; it's a profound exercise in self-reflection and self-knowledge.

But what are ChatGPT’s Custom Instructions?

CIs are a key feature for ChatGPT users seeking to personalize their AI interactions. This functionality allows you to directly influence how ChatGPT behaves during conversations, enabling a customized experience without modifying the AI's underlying model. Through CIs, you can ensure that ChatGPT's responses better align with your specific needs, preferences, and values.

CIs differ from System Messages, which are designed for developers to configure AI responses in technical settings. While System Messages address the AI's behavior on a broader scale, CIs are tailored for individual use, allowing you to personalize ChatGPT to your liking. In this piece I am focusing on CIs.

Personalizing ChatGPT Leads to Significant Impact: Qualitatively & Quantitatively

To truly grasp the importance of CIs, consider my experiment. I compared ChatGPT's responses to five different questions (see table below), before and after implementing (so with and without) CIs. Note that I used a custom GPT I created myself to develop my own CIs. The results, both qualitatively and quantitatively, are enlightening.?

Prior to personalization, the responses, while helpful, lacked a certain depth and alignment with my personal and professional interests. After setting up my CIs, tailored to my unique preferences and goals, the change was remarkable. Not only did the responses become more precise, relevant and engaging, but the number of tokens and characters dramatically increased, by roughly 26% and 35%, respectively.

Here are the questions used in the experiment (with links to the convos within ChatGPT). The questions were asked before and after the setup of CIs, with the number of tokens and characters presented respectively.

This experiment underscores the transformative power of CIs. By taking the time to articulate your values, interests, and objectives clearly, you enable ChatGPT to provide you with responses that are not just informative but also deeply resonant with your personal journey. The process of setting up CIs serves as an exercise in self-reflection, requiring you to engage deeply with your aspirations and preferences. This, in turn, enhances the utility and relevance of the AI's assistance, making it an invaluable tool for personal and professional development.

Does this apply for code generation?

Inspired by this experiment, I went further to test the impact of CIs on ChatGPT's code generation capabilities. Again, the results are interesting.

I gave ChatGPT a simple coding task: "Write a Python script to load a dataset (e.g., a CSV file containing sales data), perform basic analysis (calculate total sales, average sales per item), and generate a simple report (e.g., top-selling products)".

I then got chatGPT to run an analysis over the two snippets (with & without CIs) generated, producing a report, with the following summary:

Accuracy:

• With CI: The script demonstrates a structured approach, defining functions for each analytical step, ensuring accuracy through focused and reusable code blocks.
• Without CI: The script performs the analysis in a linear fashion without compartmentalizing into functions, which might lead to potential errors if reused or extended.

Conciseness:

• With CI: The code is modular, making it easier to read and maintain. However, it might appear slightly longer due to the function definitions.
• Without CI: The script is more concise due to its linear structure but at the cost of modularity and potential code reuse.

Comprehensiveness:

• With CI: The script seems more robust, particularly in handling data irregularities and providing more detailed analysis (e.g., average sales per item calculated specifically for each product).
• Without CI: It provides a basic level of analysis with less consideration for data irregularities or deeper insights into the sales data.

Conclusion

Custom instructions significantly enhance the utility of ChatGPT in programming tasks by promoting accuracy, enabling more comprehensive analysis, and fostering better adherence to best practices despite potentially affecting conciseness. The modular approach with CI leads to more maintainable and scalable code, offering clearer insights and handling data complexities more effectively. This experiment underscores the importance of CI in tailoring ChatGPT's output for specific programming challenges, highlighting the balance between code conciseness and the depth of analysis.

The analysis has been done by ChatGPT - I am happy to get any feedback about that.

Some of the Technical Advantages:

AI personalization significantly elevates the quality of interaction between humans and machines. OpenAI introduces this concept with a promise: once you set up your CIs, ChatGPT will integrate your preferences into every conversation. This ensures a seamless dialogue without the need to reiterate your desires or information.

Here are some of the rather technical benefits:

1. Enhanced Relevance: Tailors responses to your individual preferences, roles, and interests, ensuring interactions are highly relevant.
2. Increased Precision: Provides outputs that are technically precise and stylistically tailored, catering to your specific needs - check the responses in my experiment.
3. Improved Practicality: Focuses responses to directly address your individual objectives, making information more applicable.
4. Advanced Prompting Techniques: Enables efficient interaction by standardizing complex prompt preferences.
5. Optimization for Long Projects: Ensures consistent tone and style over extended engagements without repeating instructions.
6. Flexibility and Fun: Allows AI responses to include personal quirks, making interactions enjoyable, and engaging.

Those are really interesting benefits. What intrigues me even more, however, are the psychological and philosophical implications of personalizing AI assistants. This process not only revolutionizes how we engage with technology but also prompts a deeper exploration into our own identities and values.

Psychological and Philosophical Implications:

This isn't just about setting preferences in an app; it's about delving into the core of your passions, values, and goals. It's about honesty—laying bare your desires for both the AI and yourself.

To leverage your AI assistant to its fullest potential, a stark honesty is required. Describing yourself to your AI means stripping back the layers, presenting your most authentic self. This exercise, though seemingly small, is an invitation to self-reflection, akin to journaling or meditation. It's about honesty not just with your digital companion but with yourself.

It's not only Socrates who condensed the essence of philosophy, and human highest purpose, into the simple mandate "know thyself"; this quest also deeply resonates with Neo's journey in The Matrix, a truth hinted at by the inscription on the Oracle's door. More contemporaneously, the Techno Manifesto of our era, penned by Marc Andreessen, launches Walker Percy’s critical statement on the importance of self-awareness in navigating the complex interplay between humanity and technology.

Here is a list of some of the psychological and philosophical rewards related to the process of AI personalization, if done consciously:

1. Self-Reflection: By requiring an honest self-description, CIs become a mirror reflecting your personal and professional identities. This not only aligns with philosophical traditions but also fosters a deeper connection with your digital counterpart. Your digital self would then represent your real-world persona.
2. Enhanced Self-Knowledge: Articulating your preferences enhances self-concept clarity, boosting self-esteem and well-being. This process turns the act of personalizing AI into a path toward understanding oneself better.
3. Personal Growth: The clarity from this introspection can be transformative. Aligning AI with your deepest values and goals catalyzes personal and professional development. In my experience, ChatGPT highlights, almost in every interaction, relevant growth and learning opportunities.
4. Authenticity in Digital Interactions: Compared to the often curated and superficial interactions on social media, which can amplify feelings of disconnection and alienation, PAI could bring a level of authenticity and depth to digital engagements. CIs ensure that our interactions with, and through, AI are genuine and personal, reflecting our true selves rather than an idealized image, thereby enriching our digital experience with meaningful connections.
5. Continuous Mirroring : The benefits of CIs extend far beyond the initial setup. It consistently aligns its responses with your background, interests, and goals, turning each interaction into a continuous reminder of your values and aspirations. This ongoing alignment serves as a catalyst for sustained self-awareness and ambition. It keeps reminding you of who you are and what you are doing - a mirror of your psyche.
6. Gaining Agency: By customizing ChatGPT, you gain control over its responses and direction, marking a step towards user empowerment. This personalization allows you to shape your digital interactions, asserting agency not only over the AI but influencing the broader tech ecosystem. It's a move towards more personalized and user-driven technology experiences.
7. Ethical Considerations: Personalizing your AI assistant compels you to reflect on the ethical implications of your digital footprint. It encourages a responsible approach to technology, ensuring that your digital engagements align with your personal values and ethical standards. This consideration is crucial in fostering a conscientious digital environment.
8. Raising Public Awareness: As more users engage in personalizing their AI assistants, awareness around data privacy and ethical technology use grows. This can lead to increased regulatory scrutiny and more responsible tech development.

Now this is great, so far, but you might ask:?

Can I really be honest sharing my very private personal data with OpenAI?

I did a ‘small’ research to answer this question.

Data Privacy:

Note: I'm not the type to freely hand over my personal data to big tech. However I focus in this piece, and generally in my work, on the positive aspects, leaving in-depth data privacy concerns to investigative experts. Nevertheless, this approach does not imply negligence towards personal data protection. I stress the importance of de-identifying and anonymizing user data. OpenAI is, at least publicly, transparent about how it utilizes data from CIs. This calls for a deeper look into these practices.

They state on their website how they use data from CIs to improve model performance. They claim further that their mission and business model focus on using data to enhance model helpfulness rather than selling services, advertising, or building user profiles. They assert that engagement with their tools helps improve AI models' accuracy and safety without relying on user data for profiling or advertising, diverging from the practices of companies like Facebook, Google, and TikTok. I sincerely hope OpenAI is honest about their claims. If not, that would be extremely dangerous - think of the ‘Hackable Humans’ concept by Y. Harari.

OpenAI emphasizes that they don’t sell or share your data with third parties. Users also have the option to opt out of sharing their data, including chat history and CIs, offering a degree of control over personal information.

My exploration, based on publicly available information, represents a modest attempt to understand OpenAI's data privacy practices, with focus on CIs. I encourage data privacy experts and researchers to share their insights or concerns. Such ongoing dialogue is essential for a comprehensive understanding of data privacy, reflecting the field's diverse viewpoints and expertise.

Calls to Action

I am concluding this (long) article with calls to action for researchers & developers, AI companies and AI (not only ChatGPT) users. The potential of personalizing AI assistants is immense and warrants our collective attention and effort.

For Researchers & Developers:

I am sure there’s a lot of scientific work going on regarding AI personalization - my small experiment is a modest effort in that direction. Yet, I wish to see more tangible results;

• Assessing the Impact: Explore how CIs quantitatively and qualitatively affect AI output. Your research can provide insights into the effectiveness of personalization in AI systems.
• Bias Detection: Focus on identifying and mitigating biases in AI responses, following iteration on CIs, to ensure that AI personalization does not reinforce stereotypes or introduce biases.

For AI Companies:

Your initiatives can greatly influence the direction of AI personalization. I identified four key areas where action can drive meaningful change:

• Encourage Comprehensive Customization: Motivate users to create detailed and thoughtful CIs. This not only enhances their experience but also contributes to the overall intelligence and responsiveness of AI systems - and eventually in building inclusive and aligned AGI.
• Prioritize Privacy and Anonymity: Implement robust measures to ensure the de-identification and anonymity in CIs. Privacy should be a cornerstone of AI development in general, especially when personalization becomes a necessity.?
• Invest in UI Improvements: Enhance the user interface to make it more intuitive for users to input CIs. Adding fields for core values, passions, interests, personal goals and providing in-depth 'Thought Starters' can guide users to more meaningful interactions.

For Users:

It is time for you to start ‘programming’ your PAI. Take the initiative yourself, before others do it for you.

You can either use a custom GPT that I built to help you in setting up your CIs, and of course guiding you through that short journey of self-discovery (10-15 minutes); or you can do it on your own. If you opt for the latter, consider the following best practices to maximize the technical, psychological and philosophical benefits of that process:

• Invest Time in Reflection: Take your time crafting your CIs, reflecting on what you want and how the AI will assist you in that.
• Define Your Objectives: Understand clearly what you aim to achieve with your AI assistant. Knowing your goals upfront will enable the AI to provide more accurate and relevant responses. Spend necessary time reflecting on this.
• Be Specific and Honest: The more specific and honest you are in your instructions, the more aligned the AI’s responses will be with your authentic self.
• Use Best Prompting Practices.
• Encourage Critical Thinking, Agency and Credibility: Instruct the AI to present balanced perspectives or opposing views on topics of interest. This practice not only fosters critical thinking but also empowers you with the agency to form well-rounded opinions. Additionally, request that the AI provides citations for claims and information, enhancing the credibility and educational value of its responses. This approach ensures that your interactions with AI are not just about receiving information but also about engaging with it critically and thoughtfully.
• Regularly Update Your Instructions: As you grow and your interests evolve, make sure to update your CIs to reflect these changes. View the process as an ongoing opportunity for self-discovery and learning, rather than a one-time task.

By consciously setting up your CIs, you're not just programming your personal AI assistant. You're engaging in an act of self-discovery, mapping your digital presence to your real-world values and goals. This process, much like the journey of self-knowledge, is ongoing - evolving as you do, ensuring that your PAI grows with you, not just as a tool, but as a reflection of your changing self and aspirations.

The exploration of CIs in AI personalization is just beginning, and I'm eager to hear from you about your practical experiences. Have CIs made a difference in your use of AI, whether it's for personal development, work, or something else? I welcome you to share your stories and feedback. This can help us all understand the real-world impact of AI personalization better. Feel free to leave your thoughts in the comments or reach out directly.

Update (15.02.2024): Just a couple days after finishing this article, OpenAI made a very important announcement; publicly confirming that ChatGPT can now remember details (about you) from history conversations - something that I have observed in ChatGPT responses over the last couple of weeks. Furthermore, CIs are now under a dedicated Profile Section called... 'Personalization'.