Nayuta: The Transformer

A multi-channel software installation that performs a complete BERT-base forward pass over a 58.8-million-token corpus of institutional art-world texts, serialized into individually observable floating-point operations and rendered across synchronized screens and audio channels.

At the installation pace (1,092 ops/second, derived from a structural constraint of the model rather than chosen for effect) the machine takes on the order of 159,000 years to complete. No one will see it finish.

The work turns a measurement apparatus into the artwork: rather than visualizing analytical results, it exhibits the raw labor of computation, where every visual, auditory, and temporal event is a direct, unembellished translation of one underlying operation. One floating-point operation equals one observable event. No smoothing, no interpolation, no decorative invention.

Architecture

bert.c (engine)
  ZMQ PUSH/PUB  ── FMA op batches ──┐
  OSC           ── structural events ┤
                                     ├─> graphics_consumer (C++/Dear ImGui)  visual screens
                                     └─> SuperCollider                       audio engine

The engine streams every operation. Consumers render it. Client machines run the visual consumer; the deployment spans a main host plus desktop and Raspberry Pi clients, each driving one screen.

Layout

• bert_inference/ — computation core. The engine bert.c is a submodule at engine/ (repo nayuta_bert); the surrounding python pipeline builds the engine's corpus and weight inputs.
• graphics_consumer/ — C++ / Dear ImGui visual consumer; one screen per role (helix, op-stream, scanner, clock, entropy, mandala, whisper, nebula, rain).
• supercollider/ — audio engine (whisper word-cloud + drone), OSC-driven.
• whisper_tree/ — offline pipeline turning the corpus into the whisper graph consumed by the audio engine at runtime.
• cross/ — Raspberry Pi cross-compilation (Docker + QEMU) for the Pi clients.
• scripts/ — deployment packaging (package.sh).
• docs/ — design documents and decision records (the spec lives here).

Build

git submodule update --init --recursive

# engine (installation mode: ZeroMQ + liblo)
cmake -B bert_inference/engine/build bert_inference/engine
cmake --build bert_inference/engine/build --target bert_install

# visual consumer
cmake -B graphics_consumer/build graphics_consumer
cmake --build graphics_consumer/build

Cross-compile Pi client binaries with ./cross/build.sh aarch64, then bundle deployment tarballs with ./scripts/package.sh. See cross/README.md and bert_inference/README.md for details.

Related repositories

• nayuta_bert — the pure-C engine (submodule).
• nayuta_web — the browser version.
• nayuta_crossword — the crossword companion.

License

Code is MIT. Non-code material (text, design, documentation) is CC BY-NC.