Motion

The sphere is a being; the two lobes are its intelligence, free to become whatever says what is happening. Everything below is the living mark the app runs, beside the candidates each form was chosen over.

Thinking and speaking, from scratch

The shipped thinking and speaking forms on the engine the app runs; the voice is simulated (the beads and ribbons that were live before 2 October were removed on 2026-10-16). Each shipped form was chosen from living prototypes by blind readings of unlabelled frames. The runners-up that lost (writing dots, leaping, mercury and swell for thinking; ripples and bubble for speaking; inflow for listening) were removed on 2026-10-16, and their scores are kept in the design system and the motion evidence.

In from rest and back home every eight secondsShow the reduced-motion stills

Thinking

Working on it

Three equal dots — shipped (2026-10-16)The owner's sketch: the being shrinks to a dot and each lobe curls up into one the same size either side of it, then a hop passes left, centre, right, once every 1.5 s with a pause between, and the three unfurl back into the exact logo. Seen alone it reads as typing (five of five blind model readers), accepted because the mark shows while the reply is prepared. Its still is the level ellipsis.

Speaking

Answering aloud

Call — shipped (2026-10-16)The being steps back and swells with each syllable, and the arms become two arcs in front of it, pushed outward by each burst: the loudspeaker sign. Six of six blind model readers said speaking, five with high confidence ("sound waves coming out"). It replaced the chorus of arcs, which three of five read as speaking.

The being and its intelligence

The sphere is the being: it keeps its place, its size and its light through everything. The two lobes are its intelligence, and they are free to become whatever says what is happening, so a person can tell what Sathi is doing without a caption. At rest they are the exact logo. Every change bends from wherever the lobes are, with their speed, into the next form; nothing waits on the logo between states and nothing finishes a lap.

Show the reduced-motion stills

restThe logo: the being, cradled by its intelligence. Exact, byte for byte.

listeningA cup: both lobes join behind the being into one deep C opening toward you, the being leaning into it. Each burst of your voice pushes the cup back like wind in a sail; the being nods once in each pause, and a held cup breathes. Only while the microphone is really open (2026-10-16). (Simulated voice here.)

thinkingThree equal dots: the being shrinks to a dot and each lobe curls up into one the same size either side of it, and a hop passes left, centre, right with a pause between, the way a reply is shown being written. It replaced the writing dots on 2026-10-16, which had replaced leaping and the rising thoughts. A thought claims nothing: no count, no progress.

loadingA half turn: the arms swing round the being, heads leading and tails trailing, and land on the exact logo, rest, and turn again (2026-10-16). Nothing is being thought about, so it is not the thinking form.

searchingA lens: one lobe rings the being and the other is its handle. It glides from spot to spot, dwells, and glides on, tilting a little into its travel (2026-10-16).

noticingOne kick, no keyframes: the arms reach out and the being lifts, a small hold, and the exact logo again by 2.7 s. When the app is opened or brought back, or a field is focused (2026-10-16).

sleepingThe being drawn in, the arms curled close, the whole mark smaller and dim; then only its light breathes, once every seven seconds (2026-10-16).

something went wrongA head-shake: the head swings side to side, decaying, then the logo. Its still is the logo and the words say it: a drooping sigh read as falling asleep (2026-10-16).

offlineA drift: the arms part along their axes and turn a little, everything dimmed, held while the connection is gone (2026-10-16).

speakingA call: the being steps back and swells with each syllable, and the lobes become two arcs in front of it, pushed outward by each burst: the loudspeaker sign (2026-10-16). (Simulated voice here.)

A request, from start to finish

One simulated sequence (a voice request, work, a search, a wait on your approval, a spoken answer, home), and a second one stopped while working, drawn by the engine the app runs. The events are scripted and the voices simulated.

1. at rest
2. hears a request (simulated voice)
3. works on it
4. looks through your things
5. waits on your approval (the words ask)
6. answers aloud (simulated voice)
7. home

at rest

Free forms — what the app runsThe sphere is the being; the lobes take whatever form says what is happening: a cup, three equal dots, a lens, a call.

The trace as stills

Every 1.2 s of the first trace, computed by stepping the engine at 60 Hz from the start: frames show what motion hides.

Free forms — what the app runs

0.4s

1.6s

2.8s

3.9s

5.1s

6.3s

7.5s

8.7s

9.8s

11.0s

12.2s

13.4s

14.6s

15.7s

Why free forms (measured 2026-10-03 against the grammars since removed)

Grammar	States apart (mean)	Closest two states	Thinking motion at 22 px	Ink flicker while thinking	Home after a cut
Free forms — chosen	0.46	0.17	0.4 px/s	0.02	0.32 s (at most 0.32 s)
Posture — comparison	0.21	0.00	5.9 px/s	0.03	0.63 s
Assembly — comparison	0.39	0.08	21.6 px/s	0.10	0.65 s
Beads — as shipped before	0.37	0.00 (no waiting pose)	20.7 px/s	0.03	0.68 s

States apart is one minus the overlap of two states' reduced-motion silhouettes, averaged over rest, listening, thinking, searching, waiting and speaking; 0 means two states look the same. Thinking motion is the mean speed of the lobes' samples and of the drops they let go of at a message-row size, flicker the variation of visible ink. Posture is the calmest but says the least: thinking and searching are one picture. Assembly and the beads say more but move far more where people read. Free forms say the most, move the least and come home fastest (home is the slowest of 80 cuts for the comparisons; for free forms, the median cut and, in brackets, the slowest). This is a geometric proxy; the thinking and speaking forms were also read from unlabelled frames by three model viewers (above), and a recognition test with people is still owed.