Vibe-coded a silly app today: it plays the twelve-bar blues, then I play 12 bars on my guitar, then the AI listens to my phrase and plays its own solo, then we repeat. I'm trying to replicate the experience of jamming in a group where the players take cues from each other.
Claude tried a few different models on Fireworks and eventually settled on Minimax 3 for the right tradeoff on speed and price. The real challenge here is that you want the other player to take some cues from your phrase, but you can't send the model your entire 12 bars because it wouldn't respond in time.
So I eventually settled on it listening to my first 6 bars to write its first 6, then it gets the full 12+6 bars for its final 6. Seems to be working so far although it's a bit derivative.
It's kind of funny to see its thought process; I never thought of my guitar playing in such clinical terms:
"Echo their low E drone figure (45/59/64 stacked) as a sparse call, then climb into their crunch chord-tone clusters on the A7 quick-change — single-note answer"