Show HN: I trained an AI model on 120M+ songs from iTunes (maroofy.com)

Article summary

A user-trained AI model analyzes raw music audio from over 120 million songs to generate recommendations. The model produces embedding vectors as output, which are used to find similar-sounding songs. The discussion revolves around the model's performance, potential improvements, and comparisons to existing music recommendation services. The model's training data and scalability are also questioned.

Main themes

• AI music recommendation
• Model performance and improvement
• Music streaming services comparison
• Scalability and copyright issues
• User feedback and interaction
• Music discovery and eclecticism

What commenters say

• The AI model is effective in finding similar-sounding songs, but may not always capture the nuances of human music preferences.
• The model's recommendations are often more eclectic and less influenced by marketing than those of popular music streaming services.
• Some users find the model's recommendations to be inconsistent in terms of tempo, genre, and instrument timbre.
• The model's performance could be improved by incorporating additional metadata, such as lyrics and cultural context.
• The use of a custom AI audio model and vector database raises questions about scalability and potential legal issues with music copyright holders.
• Allowing users to upvote or downvote recommendations could help improve the model's performance over time.
• The model's ability to find obscure and unique songs is seen as a major advantage by some users, who feel that popular streaming services often prioritize heavily marketed music.