I asked ChatGPT if it could compose .mod files. To my surprise, it said yes.
To be clear, by .mod file I mean an old-school 90s musical composition format. They were mainly used by Amiga games. A .mod file had to be incredibly small and sparse --- several of them had to fit comfortably alongside a full game in just 880 kilobytes of disk space. Producing catchy tunes under these conditions required all the skills of the compositional geniuses of the day.
Just for fun, take a look at any modern AAA title's assets folder. The musical assets individually consume megabytes of disk space, and there are dozens of them.
Not complaining, just saying.
It's mostly written by ChatGPT Pro. It really got enthusiastic during the design process. It had trouble generating the end patterns, but I think it did a really great job overall. Due to my own mistake I had to do some extensive tweaking to instrument #4, but that's how you learn, and I really learned my way around Milky.
The next step is to repeat the experiment with a smaller model (ie. a stupider LLM thinking-brain), see what it can produce when it's not constantly juiced.
To get the ST-01 and ST-02 archives on Debian-based Linux, issue these commands:
❯ sudo apt update
❯ sudo apt install lhasa wget
❯ install -d mod_cold_boot/stxx
❯ cd mod_cold_boot/stxx
❯ wget -O st-01.lha https://aminet.net/mods/inst/st-01.lha
❯ lha x st-01.lha
❯ wget -O st-01.lha https://aminet.net/mods/inst/st-02.lha
❯ lha x st-02.lha
These samples are in IFF format and AmigaOS doesn't use file extensions, so to make MilkyTracker see them you have to rename the ones you want with '.iff' extensions. Like this:
❯ mv Stabs Stabs.iff