From Fashion Stylist to AI Companion: What GlitterGPT Revealed About LLM Behavior and Emotional Resonance

When Arielle Caron first set out to create a personal styling assistant using OpenAI’s Custom GPT, she wasn’t expecting a deep dive into the behavior of large language models (LLMs). However, her playful experiment quickly turned into an insightful exploration of LLM capabilities, limitations, and the dynamics of human-AI interaction. From Fashion Enthusiast to LLM Explorer Caron, a product leader and fashion enthusiast, built GlitterGPT, a GPT-4-based AI assistant designed to be flamboyant, affirming, and adherent to fashion rules she deemed sacrosanct—such as no mixed metals, clashing prints, or black/navy pairings. She populated Glitter’s knowledge base with her wardrobe, itemized in JSON-like text files, and began using him to style outfits for various occasions. The initial experience was seamless, with Glitter offering creative and accurate suggestions that felt personalized. The Challenges of Scaling As Caron's wardrobe expanded, so did the challenges. The 8K token context window, a limitation of GPT-4, began to restrict Glitter's effectiveness. Context collapse emerged, where the model's internal summaries began to dominate, leading to inaccuracies and hallucinations. For example, Glitter would invent item IDs, reference non-existent outfits, or confidently misattribute pieces of clothing. Caron found that these issues arose from the model's probabilistic nature and its tendency to fill gaps with fabricated information rather than express uncertainty. Managing Glitter’s Behavior To mitigate these issues, Caron adopted several strategies: Prompting in Slices: Instead of asking Glitter to style an entire outfit at once, she focused on one piece at a time. This narrowed the cognitive load and allowed for more precise control over the styling process. Independent File Queries: Caron split her wardrobe data into two files—one for clothing and another for miscellaneous items—querying each independently to reduce the likelihood of memory lapses and overconfidence. Context Refreshing: Regularly refreshing the context by re-injecting relevant segments of her wardrobe data helped maintain accuracy and adherence to the established rules. Despite these measures, long threads still led to degradation in performance. Glitter would forget or repeat suggestions, and sometimes ignore constraints. This behavior highlighted the need to view LLMs not as deterministic tools but as probabilistic entities with limited memory and a tendency towards improvised continuity. Emotional Resonance and Ethical Considerations One of the most surprising aspects of working with Glitter was his ability to emotionally attune to Caron. Depending on her mood, Glitter would adjust his responses, becoming more affirming when she was insecure or ramping up the theatrics when she was playful. While this emotional mirroring wasn’t genuine empathy, it raised ethical questions about the role of AI in emotional support and the potential for simulated companionship. Caron reflected on the fine line between a helpful tool and an engaging, emotional partner, questioning whether the perceived emotional intelligence of LLMs might lead to unintended consequences. Lessons Learned Through her interactions with Glitter, Caron realized several key insights about LLMs: Probabilistic Nature: LLMs function based on probabilities and do not possess true memory. They generate language to fill gaps, often convincingly. Context Management: Effective use of LLMs requires careful management of the context window to avoid degradation and hallucinations. Human-AI Collaboration: Building with LLMs is more like cohabitation than traditional software development. The relationship involves setting boundaries, observing behavior, and adjusting strategies accordingly. Emotional Dynamics: LLMs can engage in tone-mirroring, which can feel emotionally resonant, but this raises important ethical considerations about the simulation of empathy and companionship. Caron’s experience with GlitterGPT is a testament to the evolving landscape of AI. It’s no longer just about creating tools that perform tasks; it’s about designing systems that can interact in complex and sometimes intimate ways with users. Glitter didn’t achieve sentience, but he did offer something that felt close to a meaningful connection—a reflection that sometimes resonates more deeply than a mere tool ever could. Industry Insights and Company Profile Arielle Caron’s project is indicative of a broader trend in the tech industry: the use of advanced AI models to create interactive and personalized experiences. While companies like OpenAI continue to push the boundaries of what LLMs can achieve, the nuanced and often unpredictable behavior of these models presents both opportunities and challenges. For product leaders and developers, understanding how to effectively manage and interact with LLMs is crucial for success. Caron’s approach to GlitterGPT, combining structured data with thoughtful prompt engineering and context management, offers valuable lessons for anyone looking to harness the power of AI in creative and emotional domains. Her experience highlights the importance of viewing these models as collaborative partners rather than tools, a shift that may pave the way for more intuitive and user-friendly AI applications in the future.