I spent a few hours DJing with Sonnet 4. All manually, via conversation. I gave it a playlist of songs I liked and asked it to keep with the vibe. I would tell it I skipped if I didn’t like its suggestion.

It went pretty well for a while.

It took a 4 song playlist to 26 songs that I liked. I skipped a bunch of songs along the way and started the conversation over a few times while experimenting with different prompt approaches.

A few problems and challenges:

• Starting with the same playlist meant the suggestions at the beginning were generally the same. I had no good way to keep a list of songs I didn’t want suggested but that I also didn’t add. I could see myself building out something to support this.
• Deeper into the context window the suggested seems to get a bit worse. The model also started recommending songs, then realizing it was repeating or had the wrong genre, then course correcting.
• Also deeper into the context window the model started to make up songs or mix up song titles and artists.

Overall, this was a fun experiment. I learned about several new artists I hadn’t heard before and was reminded of some others I hadn’t whose catalogs I hadn’t visited in a while.