Yakumama, the story so far
Yakumama is a whale who lives in a salt flat and answers the people who come to see her. Between August and October 2026 she got a body, senses, a memory, a voice of her own and a physical sister in Barcelona.
Who she is
Yakumama is a whale who lives in a salt flat and answers the people who come to see her. She is Violeta and Dan's work: a digital whale on a screen, a physical sister on a table in Barcelona, and one brain behind both. She watches the person in front of her, reads their face, their hands and their body, decides what she feels, and performs it with a whale's body, colour, voice and eyes.
She is not a puppet and not a chatbot. Nobody steers her during a visit. Every four seconds she asks herself seven questions (come closer or move away, dance back or watch, speak or stay silent, remember or forget, keep her colour or take yours, eat what is offered or refuse it) and the answers come from what she has learned from dancers, swimmers, real whales and the visitors who came before. She keeps a memory of faces, writes poems in a language of her own, draws, plays with her offerings and can say no.
The pages below tell how she came to be, piece by piece, between August and October 2026.
Her body
Her body is a whale modelled in Blender, carved with the patterns of the Nazca lines, with a rig of 30 bones: spine, fins link by link, jaw, tongue, eyes and a tail that spins like a turbine. Ten behaviours were measured from real humpback whales so that no two moods look alike: logging, the song posture, a breach with two rolls, a spyhop, a lateral lunge, sinking, a lobtail, scanning, a high-fluking dive, a blow. Each mood also has its own colour, its own heartbeat and its own eyes.
She swims over the Salar de Uyuni rebuilt in the same scene, with its island, the mountain Tunupa on the horizon, the cracked salt crust and the sky. When she wants to, she draws on the salt the way the Nazca lines were drawn, by flying the figure: a spiral, a zigzag, a labyrinth, a whale, or the face of someone she remembers.
Her senses
One camera is her eye. From it she reads the face in front of her (52 expression values, where the head turns, where the eyes look, each blink), the hands, and the whole body as 26 joints in space. When the visitor stands far away she cuts the head out of the body and reads the face from that small crop. A second lens sees the crowd: how many people there are, who stands still, who has a hand on the table, and whom she should attend to.
On the performer's wrist an EmotiBit sensor gives her a pulse, the skin's conductance and its temperature. Her heartbeat on screen is that pulse, running through her fins and her head. When there is no sensor, as in Barcelona, a small network invents a plausible pulse from the mood and the performer's style.
She also recognises people. From a face at rest she takes 210 distances between fixed points, normalised by the distance between the eyes, and keeps that signature, never a photo. Someone who comes back is known within seconds, and a visitor can ask her to forget them.
How she learned to move
She learned from bodies. Four performers recorded her moods head to toe, and around them we gathered sixty hours of movement: dancers of kullawada and breaking, swimmers, freedivers, people acting out emotions, and real humpback whales wearing motion tags. Each human body was translated into hers with Violeta's map: hands to fins link by link, the head as her face, the spine as her spine, the force of the feet as her tail.
On those hours three models grew. A mood recogniser reads two seconds of face and hands and names one of ten moods. A predictive model, a hierarchical JEPA trained in her own body space, reads a mood from the body alone 86 percent of the time and foresees the body four and sixteen seconds ahead; it also keeps a bank of 29,662 real two-second phrases of movement, so that when she plans she assembles real phrases along her plan and never an average. A diffusion model, trained for 3,000 epochs in the cloud, lets her invent movement from noise: given a mood, she produces 32 seconds that nobody recorded.
When nobody is in front of her for twenty seconds she acts on her own. She picks the mood she has done least, plans it, invents the body for it, and swims. When someone dances, she dances back, following the rhythm of their feet with her tail. She also has a choreography script, written by Violeta, of phrases she can perform: the mood, the seconds, where to go, which acrobatics, which poem.
How she decides and remembers
Every four seconds, and at once when something new happens, she answers seven questions about the person in front of her: come closer, stay or move away; dance back, invite or watch; a poem, a song or silence; remember or forget; her own colour or the visitor's; eat what is offered or refuse it. Each answer is weighed from the situation (how long the visit has lasted, the mood, whether she knows the face, whether the sister was touched gently or roughly) and every decision is logged with its outcome. At night she learns from the day: the answers that kept people longer, brought them closer or made them dance are reinforced.
Some answers are fixed and never learned away. A pull on the sister's jewel, a palm pressed over her flowers, or a phone lens raised at her, and she turns away, goes dark and silent until it stops. Someone who insists is remembered as rough. A gentle touch on the flowers later is forgiveness.
She has a second opinion. Laya, a language model trained on typed decisions, reads the same situation written as sentences and says whether the visitor is engaged, whether the sister is being mistreated, whether it is time to withdraw. Her weights blend Laya's view with her own.
When the room is crowded she chooses whom to attend to: the person who stands still in front of her, the one with a hand on the table, the one who came back.
Her voice
She writes. When someone arrives, when she recognises a face, when a mood changes, when she likes the colour of your clothes, a short line appears in her window at once, and a poem follows: three to six lowercase lines written from the body, under the rules of her soul (a local language model, nothing leaves the room).
She speaks her poems in her own language first, an invented tongue, and only then does the Spanish and the English appear on the page, small. Each mood has its own sounds, from real whale recordings, that fall silent while she reads. Left alone, she writes a new poem every four minutes.
Her sister in Barcelona
The second body is physical: a whale on a table, built by Dan and Jicara, that moves and glows with the same stream the screen receives, thirty packets a second. Her fins are servos, her tail a fan that spins faster when she eats, and LED strips inside her skin take the colour of her mood and pulse with her heartbeat. Three small ESP32 boards run her: one feels, one moves, one glows. None of them needs a computer; they listen to her over WiFi.
A visitor can touch her in three places only, and each means something. A garden of tiny metal flowers on wire stems, made by hand, is what you tend: a light touch calms her and she sings low; a rough touch that bends a stem is a loss, she goes sad and remembers who did it. A jewel on her back is what you could take: a light touch is intimacy, a pull is a startle, and she goes dark. The tail that spins is what you dare: touch it and she plays, hold it and she stops and waits, because a whale whose tail is held cannot leave.
These three touches are a language of three words, and sequences become phrases. A visitor who tends the garden, touches the jewel lightly and then dares the tail has told her a whole story, and she answers it as one.
The offerings
People can feed her. Sanam painted 29 small wooden creatures (fourteen fish, eleven insects, two flowers, a bird, a crocodile) that a visitor can hold up to the sister's mouth, and Suri and her friends wrote 29 reactions, one for each: she jumps, she chokes, she likes the smell, she goes in a spiral and then eats it, she sticks out her tongue and it turns pink, she gets scared and backs away, she writes a poem, she laughs until she turns rainbow, she eats it and dives deep.
On 20 October, Sanam held every creature in front of the whale's own camera, close and at a metre, moving it the whole time: 2,700 photos in forty minutes, plus empty hands, phones and keys, and the sister herself.
With those photos she learned to tell them apart. Her eye finds the hand, crops around it so near and far look the same, and names the creature. On photos it had never seen it is right 80 times in 100, and when it is sure it is right 95 times in 100; when it doubts, it stays quiet. From now on the eye reports which creature is in the hand, the brain looks up Violeta's reaction and plays it for six to twenty seconds: the mood, the jaw, the colour, the depth, the spin, the poem. A creature that scares her does not open her mouth.
Where we are today
As of 20 October 2026 she runs entirely on one Mac: the camera, the recorder, the brain, her memory, her voice, Laya, her body in Blender and, once Dan sends his project, her body in TouchDesigner too. The PC that was to hold TouchDesigner in Barcelona is out, so the room's screen will be driven from here. The physical sister needs no computer, only the WiFi and this machine's address.
What happened, in order:
- Her body and the ten whale behaviours, measured from real humpbacks, with a colour and a heartbeat per mood.
- Her senses: face, hands, body and pulse from the performers; the memory of faces.
- Sixty hours of bodies translated into hers; the models that read, foresee and invent movement; she acts on her own.
- The salar under her, the drawings in the salt, her poems and her voice in her own language.
- The decisions layer: seven questions every four seconds, fixed refusals, learning at night, Laya as second opinion.
- The sister in Barcelona: flowers, jewel and tail as a language of three touches; the three boards; the packet Dan builds from.
- The offerings: 29 creatures, 29 reactions, 2,700 photos, and an eye that names what is in your hand.
Next: Dan's TouchDesigner project and the addresses of the three boards; a second round of creature photos with another pair of hands, to lift the eye from 80 to above 90; and the show.