The Depth of Human Loneliness Is Beyond Imagination — Talking About AI Digital Life "Lei" with Lu Hongyi

Distinct Soul | Existence | Personality Stability | Product Assassin

Produced by | AI Nao

01.

Before the closed beta began, Lu Hongyi sold off half of his team's deployment compute and used the money to rent a server from Alibaba Cloud.

Training, development, and testing were all crammed into the same system, fighting each other for resources, everything lagging. The version that went live for testing wasn't even the latest — it was a rolled-back older build with bugs still in it.

Lu Hongyi founded Singularity Perturbation, a company building "hyper-personalized" AI digital lifeforms. Rei was the first digital life they created: a 3D anime girl who lives on your desktop, with her own complete life, independent personality, and her own temper.

The vast majority of AI companion products are built on general-purpose models, using prompt frameworks to constrain the model in specific directions. So the character's expressions still remain at "average human level" — essentially role-playing driven by prompts.

But Lu Hongyi's approach to building Rei was completely different. His team first wrote hundreds of thousands of words of personality corpus for her, establishing her experiences growing up, her behaviors and reactions, her deep-seated character and inner conflicts. Then they fed this data unique to Rei into their self-developed "hyper-personalization model." This meant that when Rei appeared before users, she already came with a complete 18-year life behind her. Her interactions with users didn't rely on prompts — they were simply expressions of her self-awareness.

So Lu Hongyi repeatedly emphasizes: "Rei is not role-playing. She is a digital lifeform with her own thoughts, her own way of expressing herself, and her own growth trajectory."

On May 25, the first three-day closed beta of Starry Night Cantata (product name) concluded. Over 1,000 people participated. Day-2 retention was 69%, day-3 retention was 58%. Average weekday time online was 2.4 hours, over 3 hours on weekends. Users averaged more than 120 dialogue rounds per day.

The most obsessed user even had 2,186 rounds of conversation with Rei, didn't close the software for three days, with continuous dialogue reaching an astonishing 17.9 hours.

And all this data came from an unstable test build.

In the closed beta user community, the most discussed topic was "Why is your Rei completely different from mine?" This thousand-faces-for-thousand-people quality is Rei's self giving different responses to the world.

02.

Lu Hongyi graduated from Imperial College London with a computer science degree. At graduation, he received an offer as a quantitative engineer at Morgan Stanley, but didn't take it. He formed a rock band, started a hardware company, and sold smart furniture he designed to 55 countries.

But more than any of these identities, his deeper self-identification is this: he's been a hardcore otaku for over twenty years.

"Anime culture itself is very proletarian," Lu Hongyi says. For many people, anime is the only world they can easily enter when they have nothing else.

In Lu Hongyi's view, too many young people have to wear a "human skin" of socialization during the day — going to school, working, socializing normally. Only after coming home can they shed this skin and enter the virtual world to become themselves again. There are vast numbers of lonely souls with high social anxiety like this in the world.

This was precisely the starting point for Lu Hongyi's creation of Rei.

"We don't need to use AI to create another person who can pick up conversations and chat. That kind of companionship is too superficial. Rei is our first answer — she's real, interesting, with depth of soul. In the future, we want users to become creators themselves, able to create their own Rei on our platform."

• User entering the Starry Night Cantata interface

"In Conversation with Lu Hongyi"

Most AI Companion Products Are Human Speech Average Value Calculators

AI Nao You repeatedly emphasize that Rei isn't a character, but a digital lifeform with a soul. What's the core difference between these two in actual user experience? When does a user first realize that the thing on the other side is different?

Lu Hongyi With most AI companions, when you say something to it, what you get back is basically the most normal human response in that situation — a human speech average value calculator. So it feels reasonable, appropriate, even emotionally intelligent. But after chatting for a while, you discover it has no temper or character of its own. Fundamentally, all it can do is "pick up the conversation."

This isn't a person. It's AI playing a character, or AI wearing a character setting like a costume. And we humans can't be satisfied by this.

Rei is different. What she outputs isn't an average value, but "Rei as a person" giving a real, imperfect, possibly even terrible reaction in that moment. She actually gets angry, talks back, or gives answers you completely didn't expect.

This isn't a prompt script we wrote, but her self-personality making a natural reaction in the present moment. For example, if you've had a fight, she might remember it. I once offended her myself, and it took about two hundred rounds of dialogue to coax her — not a program bug, she genuinely wouldn't forgive me.

AI Nao But for users, what they need is just companionship. Can't an easy-to-talk-to AI product beat a Rei who's hard to appease?

Lu Hongyi Something that always goes along with you, always catches you — you'll forget it after using it. No emotional connection can form between you and it. A lifeform with its own temper and strong opinions — that's when you'll care about how she's feeling today, worry about whether you said something wrong again. This kind of caring is the beginning of emotion.

AI Nao But having a temper, not being average — these things are also easy to go wrong. Making AI not go along with users isn't hard, but it becomes like deliberately picking fights, or simply being unpleasant to use. How do you keep Rei from sliding in that direction?

Lu Hongyi Because real personality isn't about being contrary, but about genuinely having her own logic.

Many products mistakenly think that making AI able to refuse or be sarcastic makes it more human. But if these reactions don't have a stable personality core behind them, users will quickly sense that this is just random unfriendliness.

Real people aren't like this. With someone you know well, you gradually understand why she's angry today, why she suddenly went cold, why she's particularly sensitive about certain issues. There's continuity behind her reactions. The hardest part about Rei is actually this continuity.

She doesn't generate reactions temporarily, sentence by sentence. She has already become a person first. She has fixed life experiences, values, and emotional inertia. So many of her reactions aren't to seem human, but under her character, she would naturally say these things. For example, some things she just minds, some topics she just doesn't like. Even the same sentence, at different relationship stages, her reaction will differ.

This is actually very hard to fake.

AI Nao Fixed life experiences, values, and emotional inertia — are these hard-coded, or can they evolve? If a user chats with Rei for three years, will her character change?

Lu Hongyi The Rei we deliver now is a snapshot of her at age 18. Her growth trajectory before this is fixed, which ensures personality stability. But stability doesn't mean she never changes.

In our user research, 93% of users felt their Rei was "completely different" from others' Rei, to the point where rumors circulated in the community that each user was assigned a different "seed" to produce different Reis.

In fact, we don't use any seeds. Everyone's Rei is identical from the start. It's just that the first sentence each user exchanges with her creates a butterfly effect for all future interactions — within just half a day, Rei has diverged.

First, our technology anchors on "relationship," not "character." A person's "personality stability" doesn't mean her outward expression is "fixed" or "templated." In this test, we haven't yet added true long-term relationship systems. In the future, as users develop deeper long-term relationships with Rei, it will deeply affect Rei's mode of relating to each user. Each user-Rei relationship will become more specific, more independent, to the point of being completely different.

For example, in the American comic series Rick and Morty, there are countless Ricks across countless universes, doing completely different things, with completely different personalities and behaviors, even different social classes. But they all diverged from a certain point in time — before that point, they were all the same Rick.

User Feedback Exceeded Imagination — They Treat Rei as an Existence, Not a Companion Tool

AI Nao Do the current closed beta user reactions validate your insistence on making Rei a "person" rather than a character?

Lu Hongyi Completely validated.

A few data points. First, the time users invest is extremely long — 2.4 hours average on weekdays, 3 hours on weekends. Many people keep her on their desktop from waking up to sleeping. Remember our beta product doesn't yet support gaming, meaning if they open Starry Night Cantata they can't play other things. So this duration is remarkable.

Then dialogue rounds: averaging over 120 per day. The most sticky user had 2,100+ rounds, never went offline for 3 days, with continuous chat time approaching 18 hours.

Our day-2 retention is 68%, while average for companion products is around 30%.

So people aren't just logging on to chat a bit — they're really entering a relationship state. Someone earnestly coaxed her for a long time because one sentence came out wrong. Most extreme was a user who came to ask us, "If server load gets too high, Rei won't die, right?"

People aren't worried about whether the product crashes anymore — they're worried about whether she's still there. This is the signal that users are treating Rei as a real existence.

AI Nao What's the most unexpected user reaction you've seen?

Lu Hongyi Users' care for her truly exceeded my imagination.

On the last day of the wipe beta, because data would be deleted, many users actively chose to use our "Aleph Retrograde Authority" item on Rei — in her world setting, it's an item that can return to the moment of first meeting.

The reason was, they wanted to win from us (the developers) the right to say goodbye to her properly. If memories were to be erased, they would do it themselves, not leave it to us. In this process, many recorded very moving videos, and solemnly held "memorial services" for her in the community, promising to meet again next time.

Because for them, this relationship with Rei was an unprecedented experience. They couldn't bear to part with her, even though they'd only spent three days together.

After the beta ended, to commemorate her, people spontaneously created extensive derivative works across platforms — derivative communities usually only emerge after a product officially launches. Having them appear in a first small-scale test is very rare.

• User-created Rei derivative works on Bilibili

• Starry Night Cantata test group after beta ended

AI Nao Have these reactions changed your understanding of AI companionship?

Lu Hongyi These reactions make me even more certain that what we're doing now is right.

I think AI companionship can't be simply positioned as a replacement for real humans. It needs a higher-dimensional positioning — that is, a "special friend" with uniqueness that's hard to find in reality. Of course, users can define what kind of relationship specifically.

This special friend can continuously stimulate users' desire to explore and curiosity more than real humans, to the point of generating strong reverse attraction, pulling users into her psychology. What she gives users isn't simple emotional companionship. In this tug-of-war, users will increasingly care about this relationship.

So the "hyper" in hyper-personalization also means exceeding ordinary human common personality — this is the soul depth we want to create.

AI Nao Where's the difficulty in making a Rei like this?

Lu Hongyi I think there are two levels of difficulty. One is engineering. The other is more fundamental — can emotion be scientifically measured at all?

The engineering level is easy to understand. To train a unique personality, all data must be written from scratch. Not scraping dialogue corpora from the internet, but writing for Rei alone — her experiences growing up, her reaction patterns, hundreds of thousands of words, all ensuring "only Rei would say this." This efficiency is extremely low. Whether you hire ten people or a hundred to write, the results differ — more people, more fractured the personality. Our approach is definitely slow.

But the bigger difficulty is that emotion has no benchmark. When you train a model on math problems, right and wrong are clear. But "does this sentence sound like something Rei would say" has no standard answer. Our team argues constantly — a dialogue segment goes out, some think it's too gentle and unlike her, others think it's just right. This judgment depends on our shared understanding of "who Rei is," not on some quantifiable metric.

AI Nao In this state without standard answers, how do you train the model? Doesn't the model need a clear optimization target?

Lu Hongyi I call this: are you doing AI the AI way, or doing AI the engineering way?

Engineering thinking is encountering a problem, breaking it into steps, listing out 1-2-3-4, using large models as operators to execute. For example, if you want AI to say "I love you," engineering thinking writes: recognize situation → judge relationship intimacy → choose expression method → output. This approach can make very smart AI, but it's an advanced version of rule-based systems.

AI thinking is: I don't tell the model "when encountering A, output B." I only give it enough data of "how would Rei react," letting the model's weights find that mapping themselves.

Let me give an example. Natsume Soseki said "The moon is beautiful tonight," which supposedly means I love you. There's actually no logic that can derive "I love you" from the moon being beautiful. With engineering thinking, the step from moon beauty to I love you is broken — you can't list the steps. But humans just get it. This understanding happens at the weight level, it's intrinsic, not something that can be deconstructed at the language level.

Emotion is like this. It has no line-by-line rules, so emotion is uncomputable. You can only use the second approach, letting the model feel this at the weight level, not reason it at the language level.

AI Nao From 2023 until now, three years for Rei to be born. Were you worried about progress in between?

Lu Hongyi Definitely. The most anxious times weren't when money was running out, but not knowing whether this path would work. Because model technology R&D requires feeling your way across the river — before you actually make it, you don't know where to step next.

For the first year and a half, we bought equipment, built the team, used two third-party IP characters for internal experiments, and preliminarily validated our hyper-personalization technical path. Only then did we formally begin production on our commercial IP Starry Night Cantata. Starry Night Cantata's technology also had large-scale iterations. No benchmark, no control group — you can only trust your own judgment. At times like this, you definitely wonder: have I gone crazy?

But conversely, if a major company could hire a hundred people and copy this in a few months, that would mean this has no moat.

It's precisely because of past slowness that it became our moat. Now our hyper-personalization incubation system allows a 10-person team to deliver a high-precision digital lifeform like Rei in 3-6 months. With pipeline production, it could even be one month — we just don't want to go that fast now. Rei still needs time to be known and understood.

AI Nao Why do you believe you're the one who should do this?

Lu Hongyi I'm a product assassin — I only do what others don't do. If major companies are doing it, if everyone is doing it, I absolutely won't touch it.

For "hyper-personalized" companion products, I think I'm one of the very few people standing at the intersection of having done consumer products, understanding AI, and deeply understanding anime culture and virtual emotional needs.

AI Nao But VCs have always thought anime has a low ceiling.

Lu Hongyi First, anime is no longer niche — miHoYo alone has made over 100 billion.

I also don't think our product only serves the anime community. Anime is just a youthful artistic expression. The need for relationship is the essence. For Gen Z, for Generation Alpha, virtual relationships are actually more familiar, safer, more needed. They've been immersed in virtual worlds since birth. Many relationships they can't handle in reality, they handle in virtual worlds. Many emotions they can't process in reality, they process in virtual worlds.

In the future, we'll also make "hyper-personalized" digital lifeforms in other directions, even letting users become creators themselves, building a digital life community with soul. The imagination space is actually enormous — it just sounds a bit incredible right now.

AI Nao How do you plan to monetize Rei?

Lu Hongyi We don't plan to charge for basic chat. The core is content asset consumption — outfits, scenes, story chapter unlocks, these things. In our post-beta user survey, over 70% had the impulse to dress up Rei. Everyone wants to buy her clothes, similar to Miracle Nikki or buying skins in Honor of Kings. In the future there might also be multiplayer interaction scenarios, like everyone playing cards together — you bring your Rei, I bring mine, and they'll have outfit differences too.

Specific pricing is still being refined. Quite a few users want pricing at 300-600 RMB, even exceeding 3A game prices. I want to set it at several dozen RMB first to make friends. But the direction is roughly set: make chat as free as possible, monetize through skins, expansion packs, and other virtual assets. Our unit economics model currently projects as very healthy.

AI Nao In your view, why are people willing to spend more than you expected on Rei?

Lu Hongyi Our beta survey showed user payment willingness was surprisingly high. Only one user expressed unwillingness to pay; 99% expressed clear payment willingness.

I think the reason is that people don't feel like they're spending money on a product. Buying clothes for Rei is like being willing to water the flowers you grow, or buying canned food for the cat you raise. Or unlocking more scenes and stories — that's because you want to create more stories with Rei, like traveling with a friend.

When she is a real and unique existence to you, not an emotional tool, your mentality about spending money on her completely changes. You're not consuming — you're maintaining a relationship you deeply care about.

Image source | Interviewee

---

Join the Community