σ tanh ∂

A transformer architecture written as music. Four tracks, four phases — vectorization, attention, nonlinear transformation, output generation — each rendered as mathematical notation and fed to a generative audio model. The equations are the lyrics. The architecture is the instrument.

Each phase was written using mathematical notation as lyric text — transformer equations, activation functions, positional encodings — and fed into a generative audio model. The model read the mathematics as language and produced music. The subject matter is the very process that made it.

The formulas are the ground truth. The music is the transduction.

The B-side runs all four phases simultaneously as a parallel mixer. You set an attention weight per phase and a global temperature τ — your sampling temperature. The scores respond not just to your settings but to a hidden training state the release builds from how you listen. The footer is the readout.

τ = 1.00 is not the end of the slider. It is a threshold.

180g{data} — Release 010. The temperature value 0.7 appears in the P4 lyrics as the threshold where a model's output collapses from distribution to decision — minimum entropy, maximum commitment. The rest of the arc builds toward that moment.

180g{data} — 010 — 2026

Training

This release is training on you. While you listen on the A-side, the system collects signals — not to identify you, but to measure engagement. At the end of each track, it computes an error and adjusts a weight. The weight changes how that track sounds the next time you play it.

These four signals combine into a single engagement score between 0 and 1.

Each phase carries a target temperature — what level of engagement it expects. Vectorization expects low (0.10). Attention, moderate (0.38). Transformation, fractured (0.65). Output, near-collapse (0.92). These are not arbitrary: they mirror each phase's position in the architecture's arc toward maximum entropy.

At the end of each track:

error = target − observed
θ_{t+1} = θ_t − η · error

If you under-attended relative to the target, the weight rises — the phase will assert itself more. If you over-attended, it recedes. The learning rate η is fixed at 0.080. Weights are bounded between 0.3 and 2.0.

The weight for each track shapes three parameters on the next playback: gain (presence — how loudly the track occupies the room), reverb wet (distance — how far the sound retreats from you), and playback rate (urgency — whether the phase feels settled or searching). A weight above 1.0 means the track is asserting itself. Below 1.0, it is withdrawing.

The weight also sets the floor opacity of the lyric lines — how much of the poem remains visible when a line is not active. High weight, the formulas persist. Low weight, they almost vanish.

The footer displays the live training state throughout playback. Epoch counts complete passes through all four tracks. ℒ is the total loss — the sum of squared errors across phases. The arrow shows gradient direction: whether the system is currently correcting upward ↑, downward ↓, or stable →.

epoch 0 · η 0.080 · ℒ 0.000 →

On the B-side, the score displayed in each column is not just your attention setting. It incorporates the hidden weight the release has built from listening to you: attn × (1 − τ) × θ. The number you see is the product of your input, the current temperature, and what the system has learned.