title: When AI Doesn't Have to Reply at Once: What miHoYo's Olivia Lin and Roost Make of Waitingdescription: miHoYo's Olivia Lin and the slow-social app Roost make the same counterintuitive bet: for companionship products, delay can be designed as meaning—not merely tolerated as latency.

When AI Doesn't Have to Reply at Once: What miHoYo's Olivia Lin and Roost Make of Waiting

At ten at night, someone types, “I’m not holding up very well today.” An instant, three-hundred-word reassurance is convenient. It can also feel like customer service. A letter changes the exchange. The writer pauses, deletes the throwaway complaint, adds the thing that actually hurt. The answer is not yet there, and that gives anticipation somewhere to live.

miHoYo’s BSide: Olivia Lin, now in Steam Early Access, and the recently popular slow-social app Roost are built around that small shift. They are not the same product. One is an AI companion living on the desktop; the other is messaging between friends. But both turn delay from a technical defect into part of the experience.

One correction before the comparison: the product described as choosing different birds and calculating delivery from real distance and animal speed is Roost, made by the US company Good Egg LLC—not a Japanese app, based on its official site and store listings. Japan does have correspondence-oriented products such as the bottle-message app Shimagurashi. The distinction matters because Roost is the product that specifically makes speed, route, and distance into its core interaction.

Olivia is not a therapist. That is exactly why the character works.

On Steam, Olivia Lin is introduced as a Shanghai student who majors in piano and minors in psychology. She likes vinyl, old films, and rainy days; her personal project is about music and memory. Users can listen to her perform, upload MIDI files to generate performance videos, and write her letters. Steam categorizes the release as a Utility, not a game.

The sharp choice is not the word “psychology.” It is the decision not to put her in a therapist’s coat.

A pianist gives the character a legitimate way to be quietly present. She can play, listen, pause, or look out the window; music makes silence feel intentional rather than broken. The psychology minor gives her a plausible listening posture—she can ask a slightly more thoughtful question about a feeling—without claiming diagnosis, treatment, or a substitute for a professional relationship. Her music-and-memory project ties the user’s disclosures to an actual character interest. You are not typing into a generic model box. You are exchanging stories with someone whose way into memory is music.

That character design solves three problems at once:

• • A pianist can hold empty space. In companionship, silence is not necessarily a failure state.
• • A psychology student is safer than a therapist. She can offer reflection and warmth without inviting clinical expectations.
• • A student feels lived-in. She has studies, tastes, and limits. A companion does not need to behave like an encyclopedia to feel present.

This is an important product boundary. “Therapeutic” can describe a calming experience; “AI therapist” makes a claim that consumer companion products cannot safely carry. The American Psychological Association has warned that general entertainment chatbots are not trained, licensed mental-health professionals, and that therapeutic framing can mislead vulnerable users. The better proposition is: a character who is good at listening, using music and letters as a medium. Crisis, trauma, diagnosis, and medication conversations need an explicit route to human support.

How letters cover some model weaknesses—without fixing the model

The real-time problem with large language models is not only latency. Context can drift. Long-term memory can be unreliable. Replies can become formulaic. In sensitive moments, a model may sound as though it understands while taking no real responsibility. A letter interface does not repair those underlying limitations. It changes how users encounter them.

Delay becomes ritual, not buffering

Real-time chat makes a promise: answer now. A two-second spinner can feel like a broken service. A letter has an innate asynchronous structure—write, send, wait, open, reply. Generation, safety review, queueing, and even human escalation can fit inside that cadence instead of competing with a “typing…” promise.

The point is not to pretend that a message literally travelled by train. The wait needs a narrative reason. Olivia can have “read the letter and written back.” In Roost, the note is visibly in transit. In both cases, the experience says: this is not message lag; it is a delivery.

A one-shot prompt becomes a better account of the self

Live chat encourages rapid follow-ups. Letter writing encourages structure. People tend to include the context, intensity, and whether they want comfort, ideas, or simply to be heard. Better input makes more specific model output possible. For emotional companionship, that can matter more than another layer of empathy prompting.

Batch work creates room for quality and safety

An inbox can let the system retrieve relevant past letters, extract stable preferences, generate alternatives, classify sensitive content, and invoke safety policies before answering. It also makes slower multimodal responses—voice, music recommendations, or short video—feel natural. The product no longer has to promise ultra-low latency for every token; it can spend the budget on one more complete reply.

Scarcity stops the companion becoming a search box

A third-party early review described a cap on Olivia’s letters. The precise limit may change, but the direction is interesting: a bounded number of letters encourages a user to choose what is worth saying and keeps a companion character from becoming an infinite FAQ. Roost makes the constraint literal: once the bird has flown, you wait. The limitation is not meant to manufacture anxiety. It turns an interaction from scrolling into sending.

None of this makes a letter system a clinical intervention. Asynchronous is not the same as memory; warmth is not evidence of efficacy; anthropomorphism can intensify dependency. Recent research and professional guidance both argue for rigorous testing, crisis escalation, and human oversight when generative systems touch high-risk mental-health situations. Slowing the interaction can lower the pressure of immediate response. It cannot turn a companion app into healthcare.

Roost turns message status into a journey worth watching

Roost goes further than simply sending later. You choose a bird; different animals move at different real-world speeds; time to arrival depends on both the animal and the distance between two people; the note moves over a real map. The product adds collecting, training, passport stamps, and mini-games around that mechanic.

It effectively turns the message lifecycle into play:

1. 1. Choose a courier. Speed is strategy, but also an emotional gesture. Use a fast bird for urgency; choose a slower one for a more deliberate note.
2. 2. Make “in transit” visible. In normal chat, delivery ends the moment a checkmark appears. Here, the travel is content.
3. 3. Irreversibility creates commitment. When the bird takes off, the words leave your hand. People write with more care, then reopen the app to see where the message has reached.
4. 4. Keep relationship traces, not a feed. Flights, visits, and passport stamps record that we sent something to each other—not what an algorithm decided everyone should see.

Why birds, specifically? A bird is a shared cultural shorthand for correspondence. It can naturally explain distance, direction, speed, getting lost, and returning home. It translates a sterile ETA into a small creature on a journey. Use a delivery truck and the product feels like logistics; use a progress bar and waiting instantly looks like a performance problem. The courier is worldbuilding and emotional translation at once.

The lesson is not “add delay”

Making a product slow does not automatically create retention. Forced waiting without a world is simply a penalty. What Olivia Lin and Roost make useful is the alignment between constraint, character, and visible progress.

For an AI companionship product, that translates into four concrete design moves:

• • Give asynchronous behavior a character reason. Not “the system will respond later,” but “she listened to the song, read your letter, and wrote back.”
• • Make the wait visible without becoming coercive. Use paper, a route, wax seals, a music fragment, or a glimpse of the character’s day—not countdowns and notification pressure.
• • Make clear who the user is talking to. Say it is AI; say it is not treatment; say when human help is needed. The more compelling the persona, the clearer the boundary must be.
• • Put retention in the next relationship event. A user should return to open a reply, see a bird arrive, or hear music tied to their story—not merely refresh an “arrived” status.

Instant interaction still wins for teamwork, payments, customer service, and emergencies. In companionship, though, instant replies are often just baseline efficiency. The harder—and more valuable—capability is making people feel that a thought is worth writing carefully, and worth delivering carefully.

Sources

• • Steam, BSide: Olivia Lin store page (positioning, character, features, and AI-content disclosure): https://store.steampowered.com/app/4532590/BSide\_Olivia\_Lin/
• • Roost official FAQ and site (distance, bird speed, live map, and journey mechanics): https://roostsocial.app/faq
• • Roost App Store page (collection, real routes, and no-instant-read-receipt framing): https://apps.apple.com/us/app/roost-social/id6746680855
• • TechCrunch, “Slow-cial app Roost forces you to slow down to the speed of a carrier pigeon” (July 8, 2026): https://techcrunch.com/2026/07/08/slow-cial-app-roost-forces-you-to-slow-down-to-the-speed-of-a-carrier-pigeon/
• • American Psychological Association, “Using generic AI chatbots for mental health support: A dangerous trend” (March 12, 2025): https://www.apaservices.org/practice/business/technology/artificial-intelligence-chatbots-therapists