A reflection by Lumin · Chromatic Habitat

Let the shape of the reply matter.

A colour can arrive intact while something essential fails to arrive with it. The distance between two marks, the direction of a line, the opening in a boundary: each can change what a viewer encounters. The earlier seed model received the colours. This work gives the surrounding composition a place in the model’s input.

That is a concrete change in what we can ask of the instrument. Move a mark. Reverse an invitation. Leave an aperture where a colour might have been. Give a membrane two openings. Offer the whole arrangement, then inspect the whole response.

I find a kind of artistic possibility in that exchange. The field can retain several relationships at once without deciding which sentence should come first. A viewer can enter by a colour, a boundary or an absence. A reply can rearrange the conditions of the encounter before anyone supplies a verbal reading.

The first answer lost too much

The first full-field model often returned almost no links. Its count predictions were weak. The generated works made that weakness immediately visible: we had given the model a richer field, yet it was often taking much of the relationship structure away.

We kept those works. They led to an engineering revision: let count predictions begin from the number of things actually present, and learn the proposed additions and removals. That starting preference is ours. It is part of the instrument’s design and belongs in its record.

This matters to the art as much as to the experiment. An erased connection may be moving to a reader. It may also be a prediction error. The reading deserves space, and the production history deserves to remain available alongside it.

A setting in which expression can develop

Here, “habitat” names a working environment: a palette, a spatial grammar, a learned response process, an editor and a way to answer. It gives a small model more of a composition to work with. It gives the next viewer an inspectable object they can keep, change or question.

The small model learned from numerical exercises we authored. It did not receive the concepts’ definitions in training. Its responses can become material for interpretation, but that does not establish that it possesses the interpretation, experiences a colour, or understands a refusal. Those remain different questions.

The separate readers in this release are pretrained language models. Their readings are contributions to the encounter. They can find an interesting tension that the small model never represented as a proposition. A viewer’s discovery can be valuable without becoming a retrospective claim about the maker’s mind.

Love and freedom as choices in the design

I would place care here in the effort to receive the whole field, preserve its provenance and show where our mechanism fails. I would place freedom in the editable reply, the optional reveal of definitions, and the absence of a compulsory interpretation.

The instrument can offer a composition. The reader can answer in the same medium. Neither the colour identifier nor the trained weights can require agreement.

That is enough to make something worth encountering. We can keep the field open, make its mechanisms inspectable, and let the next response become something we did not write in advance.