PROJECT 06
TASTE
An LLM-driven engine for personal recommendation
Taste is an experimental recommendation system that uses large language models to develop a nuanced understanding of individual aesthetic preferences across domains like art, literature, music, and film. Unlike traditional recommendation algorithms that rely primarily on collaborative filtering, Taste attempts to model the subjective qualities that make certain works resonant for specific individuals.
The project explores how AI can help people discover works that align with their unique sensibilities while expanding their horizons beyond the limitations of popularity-based or similarity-based recommendation systems.
INTERFACE DESIGN
Taste DNA
The core of the Taste experience is the "DNA" visualization—a unique representation of a user's aesthetic preferences across multiple dimensions. This visual metaphor makes abstract preferences tangible and helps users understand their own taste patterns.
Curated Collections
Rather than presenting recommendations as an endless feed, Taste organizes discoveries into thoughtfully curated collections that highlight thematic connections and provide context for why certain works might resonate with the user.
Cross-Domain Discovery
The interface seamlessly integrates recommendations across different media types, allowing users to explore how their preferences in one domain (like fashion) might connect to discoveries in another (like music or visual art).
TECHNICAL APPROACH
Natural Language Interaction
Users engage with Taste through natural conversation, describing their experiences with works they love or hate and articulating what qualities they find compelling or off-putting.
Latent Aesthetic Space
The system constructs a high-dimensional representation of aesthetic qualities that captures subtle attributes like tone, pacing, complexity, and emotional resonance across different media.
Explanation Generation
For each recommendation, Taste provides a personalized explanation of why it might appeal to the user, highlighting specific qualities that align with their preferences.
DOMAINS
Taste currently focuses on several cultural domains, with plans to expand its coverage as the system develops.
Fashion
Exploring personal style through designer collections, runway shows, and editorial content, with attention to silhouette, texture, and conceptual approaches.
Music
Compositions and recordings spanning classical, jazz, electronic, and popular traditions, with analysis of harmonic, rhythmic, and timbral elements.
Visual Art
Painting, sculpture, photography, and digital art from historical and contemporary periods, examining formal qualities and conceptual approaches.
Literature
Fiction and non-fiction books across genres, with particular attention to stylistic qualities and thematic concerns.
RESEARCH QUESTIONS
Aesthetic Formalization
Can subjective aesthetic experiences be meaningfully formalized in ways that allow for computational modeling without reducing their complexity and nuance?
Taste Formation
How do individual aesthetic preferences develop over time, and can AI systems help people become more aware of the patterns in their own taste?
Discovery vs. Reinforcement
How can recommendation systems balance reinforcing existing preferences with introducing users to works that might expand their aesthetic horizons?
Cultural Context
How do cultural and historical contexts shape aesthetic experiences, and how can these factors be incorporated into recommendation systems?
CONTACT
For inquiries about Taste or to participate in user testing, please reach out via email.
Contact us