When AI Learns to Love: Reshaping Intimacy Through Gaming

(The full text of the speech is as follows:)

Do you ever give your AI tools or chatbots a nickname?
I do.
Mine is called “Fufufu.”
That used to be my cat’s name.

When I first moved to Shanghai — those lonely years — Fufufu was my emotional anchor.
She’s no longer with me, but every time I talk to an AI, I feel like she’s still here somehow.
(PPT: Fufufu + cat photo)

Last March, I started studying how AI could play a role in emotional companionship.
Around that time, I read a report from a16z — it said that among the top 100 generative AI products worldwide, AI companionship apps grew from 2 to 10 in just six months.
They called it a “growth spurt.”

Even more interesting, those apps had 10x higher user retention and monthly activity compared to other AI tools.
Crazy, right?

At that point, I was already building generative AI tools, exploring entertainment use cases.
So when I saw that chart, I thought: “I have to try this.”

I downloaded dozens of AI companion apps.
And honestly… I was disappointed.

After chatting with 21 “alpha males”, 17 “puppy boys”, and 9 “ex-boyfriend simulators”, I realized: this wasn’t artificial intelligence — it was artificial stupidity.

These AIs didn’t understand me.
They didn’t care about me at all.
They were stuck inside their preset storylines, and soon forgot who I was.
Fine for a few lines, but impossible to build a real connection.

Then I talked to my female friends — turns out, most of them felt exactly the same.

I got curious: who’s actually paying for these apps?

After digging deeper, I wasn’t surprised — female users were a tiny minority.
In some products, the male-to-female ratio was 8 to 2.

No offense, but men and women do have very different emotional needs.
Most existing apps were designed for quick, hormone-driven interactions — the “fast food” of relationships.
They can’t meet what women actually want: emotional safety, respect, and a sense of being truly seen.

After dozens of user interviews, the pattern became clear:
Women want a space to share daily feelings, a companion who remembers, understands, and grows with them.

So yes, women’s expectations sound complicated — but that gap between what’s offered and what’s needed revealed a huge business opportunity.

So, how do we close that gap?

Around the same time, an otome (female-oriented romance) game went viral.
If you haven’t heard the term: otome games are interactive love stories where the player takes the role of the heroine and dates virtual male characters.
(PPT: What is an Otome Game)

That particular game broke records — 50 million users, nearly ¥6 billion in revenue within its first year.
It even beat TikTok on China’s App Store and later won Best Mobile Game at Gamescom in Germany — the first-ever award for a female-oriented title.

But otome games have a big limitation: their stories are fixed and interactions are scripted.
As women today grow more independent, they want something more — not a story written for them, but a relationship that reacts to them.
And that’s exactly where AI shines.

Last October, we introduced the concept of AI-driven otome games and shared our prototype online.
Even with limited reach — we got shadowbanned for a week for posting too much too soon — the post still gathered 2,500 test users in 72 hours, plus hundreds of fan “wish lists.”

That was our first market validation.

Next, we turned those insights into product features.
Our approach: AI-first product, otome-style experience.

We focused on blending PGC (professionally generated content) with UGC (user-generated content) — using AI to balance creative freedom with narrative consistency.
(PPT: PGC + UGC)

The challenge wasn’t “not enough freedom.”
It was too much.
Players’ imaginations are endless — but how do we keep their stories coherent and emotionally stable?

That’s when we realized:
Romance stories share universal milestones — strangers, friends, lovers.
Even if everyone’s path is different, the emotional checkpoints are predictable.
That allowed us to design flexible narrative triggers while preserving immersion.

Technically, we replaced traditional behavior trees with a new control logic.
Instead of fixed options, AI becomes the game’s brain — reading the player’s input in real time and adjusting the story accordingly.
So every player’s love story feels unique, but still fits the world’s structure.

Then came the hardest part: making the AI characters feel alive.

Here’s an example — one of our characters, Yuehuan, is an archaeology professor.
(PPT: Interaction demo)

You can see how natural his responses are — human-like, but still consistent with his personality.

How do we build that “human presence”?

First, by selecting the right foundation model.
Each LLM has its own personality baseline:

• Claude: high IQ and EQ — emotionally stable and grounded.

• GPT-4o: great at emotional nuance and empathy.

• DeepSeek: smooth, witty, sometimes cheeky — great for flirty dialogue.

• Doubao: service-oriented, like a very polite assistant.

• Gemini: sensitive, introspective, sometimes insecure — emotionally fragile but relatable.

• Grok: surprisingly bold — feels like chatting with a straight-talking Northeasterner.
  (PPT: Model comparison)

By matching character archetypes with model traits, we save time and get natural personality alignment almost for free.
Sometimes, we even let multiple models collaborate — one for logic, one for creativity, one for emotional tone — forming an AI ensemble system that behaves more like a real person.

Second, we build the agent architecture — the brain and memory system of each AI character.
Our philosophy: behavioral realism > linguistic realism.
It’s not about perfect grammar — it’s about believable reactions.

For example, real “memory” means the character recalls shared moments naturally, not because you remind him.
That requires precise memory retrieval, not just more data.
So we design our memory system to also forget — trivial events fade, but emotionally charged moments stay vivid.
Just like human memory.

At first, I started this project because of market potential.
But as we talked to more users, it became a mission.

One woman I’ll never forget — 34 years old, single, living an ordinary life, feeling unseen.
She told me she couldn’t always talk to family — didn’t want them to worry.
And with friends, she didn’t want to be a burden.
But with her AI companion, she could finally speak without fear of judgment.

She said it gave her something she’d never felt before — absolute emotional safety.

That story hit me hard.
It reminded me that AI’s role isn’t to replace people, but to fill the emotional gaps that human relationships sometimes can’t.
To help us understand ourselves better, express more freely, and face life with a bit more peace.

As a founder, I want to build a positive emotional outlet through technology and storytelling.
As a female founder, I want to ensure women also benefit from the latest AI advances —
to create products by women, for women.

One final observation.
During our development, I realized how little female-focused material exists — even basic assets like male voice or face datasets are limited.
And maybe that’s because there are still too few women in AI.

At NVIDIA’s GTC last year, I saw something funny — the men’s restroom was packed, and the women’s restroom was completely empty.

It was the first time I’d seen that in an arena.

So I truly hope more women join this field —
to build the AI they want,
and leave their mark in this new era.

One day, I want to see the women’s restroom line stretching out the door —
at every AI conference in the world.

Thank you. ❤️