I was bored waiting for a medical appointment, so I opened up Thaura mobile and had a little vibe coding session, and added a better beat generator. Later, at home, getting it to run hit this venv situation again. I told Thaura to stop it, and we debugged code together, sort of.

(If you want to skip to the music part, click here.)

The venv situation is that there’s something about how Thaura Work executes programs that just doesn’t work with Python venv. It thinks the venv is broken, and tries to fix it, and eventually thinks Thaura Desktop is the Python interpreter.

So, for now, I’m trying to make it stop doing that.

Resuming the prior Work session, I told it to read the new code files I added.

I ran it, and got some syntax errors. I showed it the errors, and it fixed the code.

To debug things, I told it to put print statements in the code. It did that, and then tried to fix the venv.

It tried to fix the venv. I hit stop, and told it:

And, so I did. I copy-pasted the output

It worked on the code, and then told me to run a command. I did, and pasted the output. We repeated this a few times.

Then, the code worked!

Here’s all the code and generated midi files in a zip file:

The Music is Better, but not good

I don’t know anything about playing music, writing music, beats, patterns, playing drums, and can’t play an instrument.

So I’m unqualified to debug the algorithm, or even use the algorithm.

I wanted better beats, so I asked Thaura to tell me about beat generating software. From the list I found something about HATEFish and Euclidean rhythms.

The LLM, however, read the Euclidean Rhythm paper, and generated the code to produce beats.

Then I had it write some code to generate a script to generate some patterns, and convert those patterns to .mid files.

(Thank you to all the music bloggers and programmers who had your work snarfed up by the AI training spiders.)

The end result is in the zip file, above, and it’s better than the previous midi file.

I can’t tell if the code is correct and corresponds to the paper.

I can’t judge the output to know if it’s “good” because I don’t know drumming.

You can produce output that would fool a fool, if you know how to work the LLM. I suspect that’s what Elon Musk’s “DOGE” cost cutting crew was doing.

Here’s a sample of the beatmaker’s work:

I think I spent around 2.5 hours on the work presented here.