multi-turn inc.
INDEX

index

△w ∈ ℝⁿ

art

Exhibited at the 2nd Total Museum AI Art Hackathon, Breathing with the Chaos. Technical Statement (PDF) →

Nine texts written by Rokkaku Ayako were converted into 2,396,160 neural network weights. The 500-step process of a blank model learning toward these values was recorded across three dimensions.

It started with generating Rokkaku-style images using AI. But I realized the style came through while Rokkaku herself did not. The surface resembled her work, yet nothing was captured about why she paints the way she does. So I changed direction. Instead of the output, I decided to record the process.


Δw ∈ ℝ

Weights are displayed directly on screen as decimal sequences. Numbers across 26 transformer layers shift in real time during training.

Δw ∈ ℝ

Δw ∈ ℝ²

Weight values are mapped to 32 colors and shapes and placed on a 2D plane. As training progresses, colors differentiate, and the characteristics of each layer emerge through color distribution.

Δw ∈ ℝ²

Δw ∈ ℝ³

The 26 layers are arranged in space as translucent 3D slabs while a camera orbits around them. When the weights derived from Rokkaku's language overlap in space, her colors unexpectedly appear.

Δw ∈ ℝ³


Ablation: Δw ∈ ℝ⁴

Up to ℝ³, you observe from outside the screen. In ℝ⁴, the observer steps inside the data. This was not included in the exhibited work, but is documented as an ablation study exploring the possibility of dimensional extension.

Accessible on Meta Quest 3 and other WebXR-compatible VR devices: breath-exhibition.vercel.app/vr


Exhibition

Breathing with the Chaos in Berlin

P61 Gallery, Potsdamer Str. 61, 10785 Berlin-Schöneberg. 2026.4.3–6.30.

Co-organizer: Total Museum of Contemporary Art, ATND School. AI ART × Ayako Rokkaku.

Exhibition Homepage →

△w ∈ ℝⁿ | Multi-turn Inc.