solaris — a standalone 4D artwork.
You type a thought. A small language model (MiniLM-L6, 22.7M parameters, q8) reads it — once, in a single forward pass, entirely on your own GPU. The sculpture is the geometry of what the model holds together: at each of its 7 layers, every token's hidden state is compared with every other (cosine similarity), and the association graph is embedded into space so that tokens separated by bottlenecks land far apart. The body is the smooth surface fused from those points, its regions colored by association neighborhood. Layers stack into a fourth axis you can fly through, rotationally aligned so the same word traces one continuous path. Early layers hold surface associations; late layers compress toward abstraction. That re-organizing is the exhibit.
This is not attention — it is association geometry, and the piece says so. It is also not a chatbot: the model never writes a word back.
Then the thought materializes, the way the ocean in Lem's novel imitates what it has seen. The imagination has a memory of about 15,000 forms — statues, heads, towers, animals, machines, stones, each one compressed to a small latent code by a 3D autoencoder trained for this piece. The model's reading of your thought is matched against a description of every form, and its neighbourhood becomes the sculpture: one form per layer, the deepest layer holding the closest match and the shallow layers looser associations of it. A 1.7M-parameter decoder grows each one back into a signed-distance field.
The layers are not a slideshow. Together they make one four-dimensional body: between neighbouring layers the forms melt into each other through a continuous fourth axis, w, and what you see is a three-dimensional slice of that body. Scroll to move the slice along w. Tilt it (shift + drag), and the slice cuts obliquely through time — one side of the form is already becoming the next while the other still remembers the last, the way a 3D object turning through a plane looks to a flat observer.
Same thought, same weights, same sculpture — forever. Nothing you type ever leaves this machine; the weights ship with the page and after the first load it works offline.
Forms: Objaverse models under CC-BY, CC-BY-SA and CC0 (credits; each form is also credited in the readout). Descriptions: Cap3D (Luo et al. 2023), ODC-By.
scroll: travel along w · shift + drag or shift + scroll: tilt the slice into w ([ ] keys, 0 resets) · drag: orbit · hover: reveal a word · click: pin it · enter: a new thought (first ~128 tokens count)