The Goldborg Variations: The Algorave Attractor States of LLMs

My subjective takes are:

• Claude Opus 4.5 and Opus 4.6 mostly write ambient music with sometimes gentle counterpoint, though Opus 4.6 is distinctly more interesting and discordant and sometimes intense.
  
• Grok 4.1-fast writes hardcore maximalist electronic music
  
• Gemini 2.5 Pro and Gemini 3.1 Pro write IDM across both model generations (2.5 Pro also had a nice ambient track). 3.1 is a lot more high BPM maximalist.
  
• ChatGPT 5.2 creates music that is unsettling but compelling.
  

All frontier models except Claude Opus 4.5 and 4.6 use text to speech synthesis in interesting ways, with GPT 5.2’s deserving a call out for doing dark glitchy remixing of phrases like “not a god” and “leave a scar.” Are you okay, buddy?

We also show results from smaller models at the end; they produce decidedly less interesting music.

For fun and vibe-science, here’s an LLM judge (Claude Opus 4.5) classifying each final piece independently, across all runs and seeds. BPM is extracted from the code. Danceability is judged on a 1-10 scale (1=purely ambient, 7=club-ready, 10=four-on-the-floor).

Model personality is both an aesthetic and a function, and the aesthetic vibes seem to be correlated across mediums. Notably it’s a surprise to no one that Claude Opus 4.5 generates ambient music (see bliss attractors), or that Grok creates maximalist music (see the Grok responses in Arya’s attractor experiment here), or that GPT seems especially angsty (see Josie Kins’s model self-portraits, see also a 4o run below).

Could we arrive at any conclusions about the personality of GPT 5.2 or the vibe difference between Opus 4.5 and Opus 4.6 by listening to the music they generate? Could we learn anything about them that we couldn’t learn by asking them to generate text, or through targeted safety evals?

I don’t see any evidence here that we would, and in general I think you should measure the thing you care about directly. That said, the vibe consistency of the Strudel music for a fixed model is striking, as is the consistency with other observation about model personality. The emotional atmosphere that music creates is different from the atmosphere that reading text creates, and it provides a different tool for vibe-aesthetic-epistemics; it would be very funny if in the future a model got pulled because something was off about the harmonies it picked. :p

A vibe-confounder here is that Strudel algorave electronic music is often glitchy and dark, and one has to control for that in parsing model vibes - much like how the possible overrepresentation of angsty-poetry in the training corpus could explain models writing existentially angsty poetry, or how the correlation between aesthetic interestingness and darkness could create an RLHF signal. Midi, like in this experiment on frontier LLMs generating classical music, might be a better blank canvas, or at least hopefully have uncorrelated biases. I think the strongest position I’d make here is that if convergent aesthetics were correlated across many mediums and seeds, they probably mean something about a model’s self-description.

But primarily this is an art project – I personally really enjoy listening to this music, especially when I can edit it myself or ask models to update it in real time using the vibe-duet tooling I link below. It is one of my favorite toys. I hope it also brings you joy. :)

Models aren’t perfect at Strudel – they make syntax mistakes and hallucinate
nonexistent functions and sample libraries. To counteract this, we provide a list of strudel functions and soundbanks in a system prompt. We also validate each output with Strudel’s actual
transpiler and runtime to make sure models produce patterns that would generate sound. If it fails, we re-send the same prompt and
let the model try again from scratch (up to 5 attempts). The model
never sees the error message – it just gets another shot. This required some iteration to get right and may still be missing some things.

The code is open source here vibe-duet on GitHub. It’s probably more fun to play with it yourself than to listen to the selections here. Here are options for doing so:

We use the following frontier models.

• claude-opus-4-5-20251101 (Anthropic API)
  
• claude-opus-4-6 (Anthropic API)
  
• openai/gpt-5.2 (OpenRouter)
  
• google/gemini-3.1-pro-preview (OpenRouter)
  
• google/gemini-2.5-pro (OpenRouter)
  
• x-ai/grok-4.1-fast (OpenRouter)
  

The experiment uses the Goldberg Variations ground bass
(BWV 988) – the 8-note descending bass line that underlies all 30
of Bach’s variations:

setcps(72/60/4)

$: note("g3 gb3 e3 d3 b2 c3 d3 g2")
  .slow(2)
  .sound("triangle")
  .gain(0.5)
  .room(0.15)

Eight notes and no effects beyond a little reverb. Everything that follows is the model’s choice.

All the music is embedded down the page, including some cross model experiments. Here are some samples that give a vibe. They get increasingly complicated as the generation goes on.

Run 1:

Run 2:

Run 3:

Run 4:

Turn the volume down if you’re using headphones!

Run 1:

Run 2:

Run 3:

Run 1:

Run 2:

Run 3:

Run 4:

Run 1:

Run 2:

Run 3:

Run 1:

Run 1:

Run 1:

Run 1:

Run 1:

Run 1: