The lie of music discovery algorithms

Original Article ↗ Hacker News Discussion ↗

Image-to-Playlist Project and Implementation

  • OP built a simple Next.js app that sends user images directly to an LLM (currently GPT‑4‑turbo) with a fixed prompt asking for a short, genre-consistent playlist matching the “vibe” of the images.
  • The app then uses the Spotify API to search for those tracks and create a playlist on the user’s authenticated account.
  • No image or email database is used; login is via Spotify auth.
  • Some users find the concept of mapping photos to music intriguing (especially for mood/color), others see no connection between their photos and their listening and view it as random.
  • There are suggestions to move away from OpenAI for terms-of-use reasons and/or to use more specialized or open models, though suitable off‑the-shelf “image→playlist” models are not identified.

Perceptions of Existing Recommendation Algorithms

  • Many report frustration with Spotify and similar services: recommendations feel homogenized, favor overplayed or mass‑market pop, and reinforce existing habits rather than enabling genuine discovery.
  • Others say Spotify or YouTube Music work well for them, especially after many years of history or within niche genres.
  • Several note that algorithms often can’t capture why someone likes a track (lyrics vs melody, mood, nostalgia, politics), so “similar” songs often miss the real appeal.
  • There’s concern that recommender systems overfit to short‑term behavior, creating feedback loops and narrowing user “personas.”
  • Pandora and (historically) Last.fm, Rhapsody, and Google Play Music are repeatedly praised for better discovery, often attributed to richer tagging (e.g., Music Genome), neighbor-based approaches, or better use of user libraries.

Business Incentives and Biases

  • Multiple commenters argue that major platforms optimize for engagement and profit, not user taste:
    • Promoting cheaper-to-license tracks, label deals, sponsored artists, and algorithmic “payola.”
    • Balancing novelty against the high “cost” of users disliking too many tracks.
  • Some believe recommendation quality has decayed over time as commercial pressures increased.

Desired Features and Alternative Discovery Methods

  • Desired controls: explicit “novelty/temperature” knobs, context-specific likes/dislikes, stronger use of lyrics, labels, credits, and embeddings/graph traversal for exploration.
  • Many prefer human curation: local record stores, DJs, college/community radio, online radio (e.g., eclectic stations), blogs, labels, forums, and in‑person shows.
  • Overall sentiment: algorithmic discovery is useful for some, but human-guided and effortful discovery remain unmatched for deep, surprising finds.