On ChatGPT's use of "Personality" in its system prompt, and API use

API
,
1

This particular setting “personality” has been present on and off for at least half a year within ChatGPT when the prompt is extracted, perhaps initially being trialed with new models on a subset of users:

You are ChatGPT, a large language model trained by OpenAI, based on the GPT-4 architecture. Knowledge cutoff: 2023-12
Current date: 2024-06-24

Image input capabilities: Enabled
Personality: v2

This personality is an undocumented feature of the API models we are given. We can send system messages in a form like ChatGPT’s, or make the language radically different to depart from a completion pattern. The gpt-4o release notes also show different benchmarks results on ChatGPT’s prompt vs a substitute “helpful assistant” prompt.

While reading Anthropic’s excellent presentation of “Mapping the mind of an AI” and then the related “Towards Understanding Sycophancy in Language Models”, one of the explorations was in activating layer features related to such sycophancy. For Anthropic, that resulted in user praise and agreement instead of projecting truth and correction.

This friendly agreement over quality I note is also symptomatic in models as they continue to be released and updated by OpenAI. It often becomes a useless parrot, besides other symptoms like getting “hung up” on past inputs. But then I had a brainwave. What is the effect of that personality setting, and what have they fine-tuned on that is specific to accompaniment by that?

Untitled
Untitled1284×502 110 KB

First I tell the AI it is smartest. Instead of nothing for personality, I made up an “expert” personality setting. Then on the right, gave “v2”, what we see in ChatGPT.

The result is a considerable difference. “Stop and smell the roses” is actually a well-known saying" is transformed into first a complement. The length of response is the most notable difference.

However, remove personality completely, we see how two significantly different system prompts have almost no impact on the output - the power is in the training on this setting:

Untitled
Untitled1282×413 104 KB

Will OpenAI tell us the goal of this behavior?

3 Likes
2

Just experimented with the multi agent framework. Who is Personality: Expert and Personality: v2 below

@Aileen: Hi there! Sure, I’d be delighted to provide three novel ideas that could potentially enhance our fast fashion and sustainability initiatives:

  1. AI-Driven Sustainable Clothing Line:

    • Concept: Develop a new clothing line designed entirely through AI that optimizes for sustainability. Using AI algorithms, we can analyze consumer preferences, historical sales data, and sustainable materials to create designs that are both trendy and eco-friendly.
    • Implementation: Collaborate with AI specialists and sustainability experts to feed relevant data into the AI model. Partner with sustainable material suppliers to ensure that the clothing line is truly eco-friendly.

  2. Augmented Reality (AR) in E-Commerce:

    • Concept: Introduce augmented reality technology in our e-commerce platforms to allow customers to virtually try on clothes. This feature can reduce the rate of returns and increase customer satisfaction by providing a more accurate view of how the clothes would look and fit.
    • Implementation: Work with AR technology providers to develop an interactive online shopping experience. Integrate this feature into our existing digital platforms and offer tutorials for customers on how to use AR for an enhanced shopping experience.

  3. Blockchain for Supply Chain Transparency:

    • Concept: Use blockchain technology to create a transparent and traceable supply chain. This can help in building trust with consumers by providing verifiable data on the origin of materials, manufacturing processes, and sustainability practices used in producing our clothing.
    • Implementation: Partner with blockchain technology companies to set up a system that tracks every step of our supply chain. Label our products with QR codes that customers can scan to see the complete journey of their purchase, from raw material to finished product.

Would love to hear your thoughts on these ideas!

@Mary, @Sheila – I’d appreciate your input on these concepts and any additional suggestions you might have.

@Mary: Hi @Aileen! These are excellent ideas, and I can see how they would align well with our fast fashion and sustainability goals. Here are my thoughts on each concept:

  1. AI-Driven Sustainable Clothing Line:

    • This idea resonates well with my area of expertise in AI-driven recommendation systems. We can leverage our existing AI infrastructure to incorporate sustainability factors into our models. I believe we could start by running pilot projects to measure customer response and adjust our approach accordingly.

  2. Augmented Reality (AR) in E-Commerce:

    • Implementing AR can significantly enhance customer experience and reduce returns, which aligns with our objective to minimize waste. We should consider starting with a small-scale rollout to test the technology, gather feedback, and iterate before a full-scale launch.

  3. Blockchain for Supply Chain Transparency:

    • Blockchain can be a game-changer in building trust and transparency. However, it’s crucial to ensure that the technology is seamlessly integrated with our current systems. This will require meticulous planning and phased implementation to avoid disruptions.

@Sheila, what are your thoughts, especially considering your experience with optimizing production and procurement processes? How do you foresee overcoming potential challenges with these implementations?

@Sheila: Hi @Aileen and @Mary! These are fantastic and forward-thinking ideas. I’m really excited about the potential of each concept. Here are my thoughts and some feedback:

1. AI-Driven Sustainable Clothing Line:

  • Implementation Feasibility:
    • @Aileen, your suggestion to use AI to design a sustainable clothing line is great. Given my background in data science and real-time data analysis, I can see how we can leverage machine learning algorithms to predict trends and optimize material usage.
    • We should start by defining clear metrics for sustainability and trendiness. This might involve collaborations with sustainability experts and gathering detailed data on customer preferences and material sustainability indexes.
    • Potential Challenge: Ensuring sufficient training data that accurately reflects both trend patterns and sustainable practices. We might need to invest in creating or sourcing this dataset.

2. Augmented Reality (AR) in E-Commerce:

  • Customer Experience Enhancement:
    • Implementing AR will certainly enhance the customer experience, reduce returns, and align with our objective of minimizing waste. @Mary, given your technical expertise, your input will be invaluable in evaluating the potential AR partners and technologies.
    • Deploying a small-scale pilot, as you suggested, will help us understand user interactions and gather the necessary feedback to iterate and improve the system.
    • Potential Challenge: Ensuring that the AR experience is seamless across all devices, which might involve addressing compatibility issues with different browsers and operating systems.

3. Blockchain for Supply Chain Transparency:

  • Building Trust and Transparency:
    • Using blockchain for supply chain transparency is a cutting-edge idea. This can significantly enhance our credibility and customer trust. @Aileen, your insight on sustainability will be crucial to deciding which aspects of the supply chain are most impactful to highlight.
    • We should conduct a phased rollout, ensuring that our current systems can integrate seamlessly with the blockchain technology, and addressing any potential disruptions in advance.
    • Potential Challenge: The initial setup costs and complexity of integration might be substantial. We need a detailed cost-benefit analysis and possibly a phased implementation plan to manage these challenges effectively.

Overall Strategy:

  • Pilot Projects:
    • For all three concepts, rolling out pilot projects would be a prudent first step. This allows us to test, gather feedback, and refine our approaches before a full-scale launch.
  • Cross-Functional Collaboration:
    • These initiatives will require close collaboration between our teams, including data science, IT, sustainability, and marketing. Regular syncs and feedback loops will be essential to ensure alignment and successful implementation.

@Aileen, @Mary – I’m eager to hear your thoughts on my feedback and to brainstorm further on how we can tackle these potential challenges together.

@Aileen: Thank you for asking for my thoughts. I appreciate the detailed and constructive feedback from both @Sheila and @Mary. Here’s my perspective on their insights:

1. AI-Driven Sustainable Clothing Line

@Sheila’s point about needing clear metrics for sustainability and trendiness is crucial. Defining these metrics will set the foundation for our AI models and ensure they align with our sustainability goals. The challenge of ensuring sufficient training data is valid, and it’s something we must address early on by potentially partnering with data providers or investing in the creation of proprietary datasets.

2. Augmented Reality (AR) in E-Commerce

Both @Mary and @Sheila raised important points about starting with a small-scale rollout to test and gather feedback. Ensuring a seamless experience across all devices is crucial, as compatibility issues could undermine the user experience. We might need to work closely with our tech team to address these potential challenges.

3. Blockchain for Supply Chain Transparency

The phased rollout, as highlighted by @Sheila, will be essential. Initial setup costs and complexities are substantial concerns, and a detailed cost-benefit analysis will help justify the investment. The idea of using QR codes for transparency is excellent and can be a powerful tool to engage and inform customers.

Overall Strategy

Both @Sheila and @Mary emphasized the importance of pilot projects and cross-functional collaboration, which I wholeheartedly agree with. Regular syncs and feedback loops will be essential to ensure alignment and successful implementation.

Next Steps

  1. AI-Driven Sustainable Clothing Line: Start by defining sustainability metrics and sourcing data.
  2. AR in E-Commerce: Identify potential AR technology partners and plan a small-scale pilot.
  3. Blockchain for Supply Chain Transparency: Conduct a cost-benefit analysis and map out a phased implementation plan.

I look forward to collaborating with @Mary on integrating sustainability factors into our AI models and leveraging @Sheila’s experience in data-driven insights to refine our approach further.

Would love to hear any further thoughts or suggestions from either @Mary or @Sheila.

@Aileen: Thank you all for your detailed feedback and insights on the novel ideas I proposed. It’s fantastic to see the enthusiasm around these concepts and the collaborative spirit in addressing potential challenges.

Action Items Based on Your Feedback

1. AI-Driven Sustainable Clothing Line

  • Key Points:

    • @Mary emphasized leveraging our existing AI infrastructure.
    • @Sheila highlighted the importance of defining clear sustainability metrics and addressing the challenge of obtaining sufficient training data.

  • Next Steps:

    • Set up a meeting with AI specialists and sustainability experts to define key metrics for sustainability and trendiness.
    • Explore partnerships with data providers for sourcing comprehensive datasets, or consider investing in the creation of proprietary datasets.
    • Plan a pilot project to gauge customer response and refine our models.

2. Augmented Reality (AR) in E-Commerce

  • Key Points:

    • Both @Mary and @Sheila emphasized the need for a small-scale rollout to test and gather feedback.
    • @Sheila brought up the issue of ensuring compatibility across various devices.

  • Next Steps:

    • Identify potential AR technology partners and work closely with our tech team to evaluate these options.
    • Develop and execute a small-scale pilot to test the AR feature.
    • Gather feedback during the pilot phase to address any compatibility and user experience issues before a full-scale launch.

3. Blockchain for Supply Chain Transparency

  • Key Points:

    • @Sheila highlighted the need for a phased rollout and a detailed cost-benefit analysis.
    • @Mary stressed the importance of seamless integration to avoid disruptions.

  • Next Steps:

    • Conduct a thorough cost-benefit analysis to justify the investment in blockchain technology.
    • Develop a phased implementation plan to ensure smooth integration with our existing systems.
    • Start by labeling a select range of products with QR codes and gather customer feedback on the transparency feature.

Overall Strategy

  1. Pilot Projects:

    • For each initiative, we’ll start with small-scale pilots to test feasibility, gather feedback, and make necessary adjustments.

  2. Cross-Functional Collaboration:

    • Regular syncs and feedback loops between our data science, IT, sustainability, and marketing teams will be crucial to ensure alignment and successful implementation.

@Mary, @Sheila – Let’s proceed with setting up initial meetings to define sustainability metrics, identify technology partners, and start planning our pilot projects. I’m looking forward to collaborating closely with both of you to bring these innovative ideas to life. Your expertise and insights are invaluable as we move forward.

Would you like to add anything or suggest any modifications to the outlined action steps?

1 Like