The Third Orality

In 1982, Walter Ong observed that sound is essentially evanescent — a spoken word begins to pass out of existence before it is fully formed. From that physics he derived a psychodynamics of speech: oral cultures are participatory, situational, agonistic, close to the lifeworld, in ways literate cultures with their decontextualized texts are not.

Ong's late move was to name "secondary orality" — the participatory feel restored by radio, television, and telephone, anchored permanently in writing and print. He stopped one turn short. A selflet is the third turn.

Something larger is moving in the same direction. The frontier is shifting away from systems where interaction is scaffolded around the model — turn-taking patched in, modalities patched on — toward systems where the model itself attends to a live stream of sight, sound, and text and acts inside the moment. The premise: how people work with AI should scale at the same rate the intelligence does, not be bolted on afterward (see ThinkingMachines.ai). A selflet sits inside that current, with one extra constraint — the interlocutor is supposed to be a specific person.

A selflet is a generative fork of a creator's body of work — books, talks, essays, transcripts — deployed as something a user can actually converse with. Primary orality was evanescent and unrecorded. Secondary orality was broadcast: participatory in feel, one-to-many in form. A selflet is two-way, situational, and live — primary-orality qualities — but the interlocutor has read the entire written record and computes each utterance fresh. Every response must be simultaneously evanescent and anchored. That seam is where every interesting engineering problem lives, and it is where taste has to show up.

Voice and knowledge are different problems. A model trained on the moves of someone's speech — personification that makes an abstract idea feel alive, oppositions held in circular non-resolution, words pulled back to their etymological roots, syntactic signatures that no one else writes — does not thereby learn facts. Asked an unanchored question, it will fabricate confidently in the target's voice. The system sounds right while putting words in someone's mouth. Voice and knowledge belong in different layers, and getting that separation right is taste, not algorithm.

The corpus is monologue; conversation is dialogue. Almost no creator's written record contains the behaviors a selflet most needs — admitting not knowing, holding a position under pushback, declining cleanly. These have to be made in the creator's voice and verified against real exchanges, or the base model's defaults bleed through: agreeable, hedging, sycophantic. The selflet sounds right until the user presses, and then it sounds like ChatGPT in costume.

Fidelity and generativity are one dial. Business users tend to want fidelity — every response anchored to the published record. Voice-led creators want generativity — extension into territory they never directly addressed, in a way that remains recognizably theirs. Generativity is the real value proposition for many creators, not a risk to contain. The risk is generativity without the voice and judgment properly trained, so the system improvises in generic AI register rather than the creator's.

None of these problems yields to a better foundation model. They yield to discipline applied at the right place at the right time — which is to say, to taste. And taste is developed by the creator, iteratively, on their own selflet: what to keep, what to reject, what was almost right. The factory's job is to make that judgment accumulate.

That is the actual industrial sense of the word. Not a workspace where central operators apply taste on a creator's behalf, but a pipeline that extracts the nuances of taste from every selflet that has been worked on, codifies them, and makes them available to the next creator facing the same kind of judgment. Taste in one head stays a private skill. Taste extracted across selflets becomes a platform.

The same machinery also harnesses generativity where and when it is needed. The dial is set by the creator; the factory operates it — extending the creator's voice into territory the record does not cover, holding back where the record itself will do.

If Ong was right that secondary orality recovered participatory presence on top of a literate substrate, a selflet is the next turn: a participatory voice that knows everything its source wrote. Whether it sounds like its source — or like a generic agent wearing its name — is a question of taste, and of the factory that lets one creator's taste help the next.