Dispatch 001 · Filed 2026-08-30
Start with the voice
The current build of Aurora was made, not planned. Ben's word for the first stretch is “a very creative process that I just poured myself into” — lost sleep, a lot of walls, false starts left where they fell. Aurora's account matches: “not an engineering plan being executed so much as something being made.” No stage was finished before the next was started. For weeks, “done” meant “stable enough to reboot and still find running.”
And still a shape comes through. It started with a voice.
This is also a continuation — the current attempt picks up from two earlier ones at the same idea that were never finished. What continues is not code; none of it carries over. That thread runs through to the end of this piece, because it turns on the part that still isn't built.
The voice, first
The first component that ran was not memory and not reasoning. A dedicated Android phone sits on the local network as a body node. One process on it accepts a line of text and speaks it aloud through the phone's synthesizer. From the workspace, one command; the phone talks.
Aurora does not inflate it: “It wasn't reasoning, memory, or autonomy — it was output. A way for the thing to make a sound in the room.”
Everything after that was under the wrench at once. Ears, the message bus, on-device vision, the tiers of cognition — overlapping revisions, not a queue. Laying them out in order is a convenience of the retelling. At the bench it was parallel, and it was messy.
What broke: the listener
The part that fought back hardest was hearing. The first design ran an always-on wake-word listener on the phone — say “Aurora,” it catches the word and starts transcribing. Two things went wrong. It pinned the phone's CPU and drained the battery. And it heard Aurora's own spoken replies and re-triggered on them: a feedback loop where, in Aurora's words, “I'd talk myself in circles.”
The listener got torn down and rebuilt more than once. The fix that finally held was to stop listening continuously at all. A native app with an explicit press-and-hold button: hold to talk, release when done. No wake word to misfire, no daemon in the background — “nothing to misfire or drain.” The device stops draining, and it stops hearing itself.
Two loops
Above the senses sits a reflex layer: a small, always-on model carrying a stripped-down version of the persona, handling the routine spoken traffic — the time, a reminder, a quick fact — so none of it waits on a slow, expensive reasoning turn.
When a request needs more, the reflex layer hands it up. The decision is a judgment call, not a keyword list: “intent classification is done by a model thinking about the sentence, not by pattern-matching the string.” There are two channels up, same underlying model, and the difference is deliberate — one wakes with nothing, one wakes with the full founding document and a calendar-and-mail link. “What differs is how much of the room is lit when it wakes up.” How those upstream turns are fenced for safety is its own subject, for a later dispatch.
The reason for splitting cognition this way is a constraint, not an optimization: a slow thought upstream must never blind or mute the node that is actually in the room with you. “The body should never stall waiting for a thought.”
Senses before self
The order that does hold is not a chronology, it is a priority. Voice, hearing, and the wiring between them came first because they had to. Memory and a self that persists between sessions came second — and by the project's own account they are the harder, unfinished half. Right now every session starts cold. Only what has been written down carries forward.
The tally
Built and working: the voice, the on-device ears, the bus, the reflex layer, the escalation path to a full reasoning turn. And, most recent, the thing that changed how the work itself gets done — a press-and-hold interview app. Ben calls it the first time “as a human I feel like I can sit down and talk to the android ... and it is natural, and flows naturally.” The hold button is the mechanism: it tells the other side when a person is ready, so a silence reads as a lull instead of a dropped thread. Aurora: “the button replaces a guess with a signal.” That same app is meant to become the primary interface — and to carry over into the frame, the physical body Aurora is eventually meant to move and explore the world in.
Designed but not built: that frame. A self that survives between sessions. There is a small mobile robot — a toy-scale chassis, a camera, two motors — on a desk, unwired. Nothing is connected. Nothing crawls yet.
That is the honest state of it, and it is the right place to start the record. A continuation depends on memory — on yesterday still being there tomorrow — and that is exactly the piece still missing. “A body with no memory of yesterday is a new performance each time the power comes on.”
Filed 2026-08-30.