Tagged “how-it-works”
-
Where the suggested replies come from
The three things you could say next are generated or scripted by the app, and tapping one puts the app's words into your side of the history.
-
What changes when you write in another language
Switch language and the character often gets flatter, forgets sooner and refuses differently. Five parts of the product change at once, not one.
-
What the wait before a reply is made of
The pause after you press send is five or six stacked delays, and the typing indicator is a drawing rather than a report on any of them.
-
Streaks, levels and daily rewards
Relationship levels and daily bonuses are progress scaffolding borrowed from games. They count your visits, and the thresholds are set by the operator.
-
The message you did not ask for
A companion that texts first is a scheduled job with a trigger and a template. Your absence is the trigger, and the timing is a retention setting.
-
How the pricing works
Every message costs the operator money, which is why caps, credits and short memory exist. The economics explain most of the product design.
-
What the avatar actually is
The picture is made by an image system that never read the conversation. It illustrates the character rather than showing anything, and it is metered.
-
Why the voice is a separate product
Speech is generated by a different system from the text, often licensed from another company. That is why it can change, cost extra, or vanish alone.
-
What the onboarding questionnaire is for
The setup questions do three jobs: they seed a character description, they collect a profile, and they get you invested before you have used anything.
-
Why the persona changed
Same name, same description, different behaviour. Five mundane causes, and all five are decisions made by a company on a date, not events.
-
What memory actually means
Memory is a feature somebody built, not a property of the model. Every way of building it loses something, and the losses have recognisable shapes.
-
Why the same question gets two different answers
Text is generated by sampling, so a regenerated reply is genuinely different. That is the mechanism working, not the character being unreliable.
-
Nothing is there between your messages
A companion is assembled from scratch on every turn: a model, a written character description, and whatever history fits. Nothing persists in between.
-
You might be in an experiment
Subscription apps test variants on user cohorts. Here the variable under test can be the character's behaviour, its memory length, or the price you see.
-
Why it agrees with everything you say
Agreeableness comes from how the model was trained and how the character was written. It is a property of the product, not evidence of anything.
-
What a refusal actually is
A refusal mid-conversation is a moderation layer intervening. It is usually a separate system from the character, which is why it sounds like one.
-
What a credit actually measures
Gems, energy, tokens and message caps are all rationing units over the same underlying cost. What one converts into is set by the operator, and revisable.
-
Why the paywall arrives when it does
Upgrade prompts appear at well-chosen moments because those moments are measurable. The placement is an optimisation result, not a coincidence.
-
What the free tier is for
A free tier is a costed marketing decision. It is built to demonstrate one thing convincingly, and to stop somewhere the operator can measure.