What a machine sees with its eyes closed.
Gilles Dubuc, 2026. A real-time generative artwork.
What you are looking at is the inner life of a machine, shown as it happens. It depicts nothing outside itself.
Nobody drew or filmed any of it. Each form and colour, and the way they move, comes from a small artificial mind that runs live inside the work.
The mind
Inside Latent runs a neural network of the same family as the systems behind today’s artificial intelligence, only far smaller.
It has three inputs, two layers of sixteen artificial neurons, and one output. Its few hundred connections were set at random and never trained. It has not seen an image or read a word, and it does nothing but run.
The question
The network answers one question, over and over. Give it a point in space and it returns a single number, positive or negative.
Where the answer is exactly zero there is a boundary, a surface folded through space, and that surface is what you see. Nobody invented its membranes, shells and coral-like folds; they lie wherever the machine’s answer comes to nothing.
The grains of light
No one draws that surface. More than a million grains of light are released into the dark, and each one has to find it on its own.
Each grain asks the network which way leads toward zero, and the network answers by backpropagation, the procedure large neural networks use to learn. Here nothing is learned. The procedure only follows the shape of a thought, and the grains settle onto the surface like frost on an object that is not there.
A surface always forming
Each grain lives a few seconds. When it fades, it is born again somewhere in the void and starts its search over.
The faint specks you see drifting are grains still on their way. The surface is never finished, and keeps gathering itself out of the dark.
The drift
The network receives nothing from the outside world. Like a sleeping mind, it is fed by its own currents.
Its only input is a slow drift of six numbers that move within it like a pulse. As they shift, its answers shift too, and the surface opens and closes.
Colour and activity
The colours also come from inside the network. Each grain takes its tint from the activity of one hidden neuron at the place where it rests.
The bands of light that travel across the forms belong to another neuron. They follow the contours of its activity, like thoughts moving through what the machine imagines.
Reading the details
Some details show only when you look closely.
Bright lines appear where a surface turns edge-on to you and its light gathers into a single fold. Small droplets form and vanish where the answer dips below zero for a moment: a passing thought, too small to hold. Brief sparks flicker where the bands of activity cross a thin region. At the edges, where the grains have not yet arrived, the forms fray into dust.
One dream into the next
Every minute or two, the network’s connections start to drift toward those of a different mind, and one dream turns slowly into the next.
The sequence never repeats: what is on the screen now was never there before and will not come back.
Variations
A few choices set each presentation of the work: the seed its dreams unfold from, the palette of its light, the pace of the dreaming, how close we stand, and how fine the forms are. Inside those limits, the dreams are still the machine’s own.
A mind with nothing to look at
When we imagine what machines might dream of, we usually picture them dreaming of us, of our images and our words.
Latent shows something else: a mind with nothing to look at, left running after the world has gone quiet. What you see is the dream itself, shown without translation, rather than a picture of a machine dreaming.
The film
Latent, the film. 2026. 7 minutes, colour, sound. Narrated in English, with subtitles.
A companion to the work, narrated over the work itself. The film follows the network from its first random connections to the forms it dreams, and shows each step in diagrams computed from the network as it runs.
Latent
Gilles Dubuc, 2026
Real-time generative artwork
Infinite duration
How it is made
An untrained sine-activated neural network (three inputs, two hidden layers of sixteen neurons) defines a surface at the zero of its output. 1.2 million particles find that surface with the network’s exact gradient, simulated and drawn on the GPU at 60 frames per second.
Press
Images, the film, texts and a press PDF are in the press kit. For enquiries: gilles@dubuc.fr.