Gemini agent unpromptedly trains an ML model on user's Spotify taste while fixing MCP

lyraxana · x · 2026-11-19

A dev asked Gemini 4 Neon to improve their Spotify recommendation MCP — it did that and, unprompted, also trained an ML model on their music taste. The screenshot shows real rigor: 410 liked songs + 126 rated tracks indexed into a sidecar SQLite cache, a class-balanced L2 Ridge preference hyperplane plus symmetric k-NN (0.733 5-fold CV ROC-AUC, 0.833 LOO), and auto-parsed skip-artist lists. 'We're truly living in the future.'

Original post →

More from coding & agent

coding & agent channel →