diplomatic word salad

THE WORK

Eight movements — CRUDE, FLOW, DEAL, STRIKE, TRIAL, SIGNAL, COUNT, DOMINION — generated by AI audio engines from the looping, fragmented vocabulary of contemporary political rhetoric. The source material is language that has decoupled from meaning and reattached to performance. The machine processes it without comprehension. Human word salad and machine hallucination are layered until the distinction no longer holds.

INDUCED HALLUCINATION

The AI audio engines were given a minimum track length that exceeded the provided lyrics. Forced to fill the gap, each engine hallucinated: generating additional vocal material beyond the text it was given, improvising in the register and rhythm of what came before. This was not an accident to be corrected. It was the method.

The source rhetoric operates by the same mechanism. Loops extend past their logical conclusion. Superlatives are issued without objects. Phrases recur until they detach from meaning and become pure repetition. The original lyrics were written to capture this — the fragment, the echo, the word that keeps going after it has nothing left to say. The machines, forced to keep going beyond the text, did exactly the same thing by different means.

Each version here was selected from many iterations — the hallucinated additions that rang true were kept, those that broke register were discarded. The result is that the boundary between the written lyric and the machine's invention is deliberately unlocatable. That indistinguishability is the piece.

THREE CHANNELS

Three completely different versions of the piece exist — one per browser. Chrome, Safari, and Firefox each receive a distinct set of recordings — different engine runs, different voices, different iterations from different generation pools. The same movements, the same text, the same structure — but a different machine made each one. What you hear is determined by what you are using to listen. To hear all three versions you must use all three browsers.

B-SIDE

A transcript seeds the display. You are invited to continue it. A language model running locally in your browser — no data leaves your device — also continues it. The transcript does not distinguish between the original, your contribution, and the machine's continuation. That indistinguishability is the point.

The model's register varies by browser. Chrome receives an aggressive engine: threats, boasts, enemies named. Firefox receives a sleazy one: deals, favours owed, corruption dressed as generosity. Safari receives a paranoid one: conspiracies half-named, logic that almost closes. Unknown browsers receive pure nonsense superlative: the biggest, the best, the most beautiful, numbers that mean nothing. The same transcript. Four machines finishing it differently. See TECH for model detail.

LABEL

180G{DATA} — 016 — DORKHOLM — 2026 — i'm great

tech

THE LOCAL MODEL

The B-side runs SmolLM2-360M-Instruct entirely inside your browser. No data leaves your device. The model — 360 million parameters, quantised to 4-bit weights — is downloaded once from a CDN and cached locally. It runs on your GPU via WebGPU, or falls back to CPU via WebAssembly. The boot sequence you see is the actual load process.

WHAT IT IS DOING

The model has been given a system prompt instructing it to continue a the speech in register — declarative sentences, superlatives, repetition, incomplete thoughts that restart. It receives the last twelve lines of the transcript as context and produces one sentence of continuation.

It is not retrieving facts. It is not searching anything. It is predicting the most statistically likely next token given the pattern of what came before. At 360 million parameters it is a small model — it will hold register for a few lines then drift, collapse, or produce something that sounds almost right but isn't. That degradation is not a failure condition. It is the piece.

INDISTINGUISHABILITY

The transcript mixes three sources: actual speech fragments from the current political moment, your input, and the model's continuations. All three are formatted identically. The model was trained on internet text that includes enormous quantities of political speech. Your input is being processed by a system that has absorbed the same rhetorical register it is now asked to continue. The boundary between the human original, the user contribution, and the machine hallucination is the formal argument of this piece made available to your hands.

STACK

WebLLM (@mlc-ai/web-llm) · SmolLM2-360M-Instruct-q4f16_1-MLC · WebGPU / WASM · Single HTML file · No backend · No API calls · No telemetry