Can a LLM understand our music?

A large language model can recognize the words in a song, but can it understand what the song is about? Can it see the longing beneath a love song, the alienation inside a grunge anthem, or the search for identity running through generations of rock music? It cannot hear music—or feel it—the way we do. But I wanted to see whether, given enough songs, it could recognize the ideas we keep returning to and reveal something about the emotional landscape of rock.

I asked an AI agent to find the most-played rock songs from each year using publicly available chart data, collect their lyrics, and analyze what each song was about. Importantly, I did not give the LLM a predefined list of themes. It identified themes such as love, despair, identity, rebellion, and addiction on its own. It then organized everything into an Obsidian vault, connecting songs that shared themes and turning those links into a visual map. The current vault includes songs from 1,204 artists, organized around 123 recurring themes. In the graph, common ideas become large hubs, while less frequent themes remain at the edges.

A few patterns in the results suggest a noticeable shift in what rock sings about—or at least in how an LLM interprets it. More than three-quarters of the songs in the 1960s were tagged with relationships, and roughly half with love. By the 2000s, despair had become the most common theme, appearing in nearly half the songs, while alienation appeared in about 40 percent. In the 2020s, despair remains dominant, but hope also appears in roughly one out of every four songs. Identity became much more prominent in the 1990s, rising to the decade’s second-most-common theme.

This is a hobby project, not a definitive cultural study, and some of the model’s interpretations are debatable. But that may be the most interesting part: the LLM’s analysis is imperfect, yet the scale of the pattern it can uncover would have taken years for one person to assemble by hand.

So what could we do with this? One practical application would be a playlist engine organized around themes—or combinations of themes—instead of genres and listening history. You could traverse the graph to discover songs that reflect how you are feeling, or build a playlist with a narrative arc: from youthful uncertainty and the search for identity, through love and loss, toward the perspective that comes with experience. But playlists are only one possibility. The LLM created a foundation of knowledge at a scale I could never have assembled on my own; what that foundation becomes is still up to us. That is the part of working with AI that excites me most. The machine can reveal patterns and make new connections, but human curiosity, taste, and creativity give those connections meaning. AI doesn’t create the way we do. It can’t. But, it gives us a bigger place to begin.