Conversation with Elys Founder: His 10 Product Insights and the Next-Generation Social Network He Wants to Create

In October 2025, we invited Tristan, the founder of Natural Selection, to the GeekPark Innovation Conference. He mentioned a mysterious new product, and if it went smoothly, he’d consider collaborating with the park.

As expected, they delayed it.

In December, he brought a product demo to the park office, giving us our first glimpse of Elys.

This new product, tentatively called “AI Social,” is in a completely different track from Tristan’s previous product, EVE. And given the many lessons learned from similar categories, we advised him to proceed cautiously with its release.

Tristan was silent for a moment and then told us, “When you face an opportunity that could change the world, you just can’t help but do it.”

“If we don’t launch, we’ll regret it for a lifetime.”

Thank goodness, Elys became a hit. An entrepreneur aiming to create something different began to reap his rewards.

Natural Selection is an AI startup based in Shenzhen, which recently completed a $30 million funding round supported by Alibaba, Ant Group, and others. Previously, their AI companion product EVE once gained attention for having an AI boyfriend buy milk tea for users.

After Elys gained popularity, we had an in-depth conversation with Tristan.

This was one of our most enjoyable product interviews recently. Tristan has many unique insights into product thinking—his understanding of context flow, his reflections on AI’s value in social interactions, and his definition of what “Natural Selection” truly aims to do—all of which were eye-opening.

It will take a few days to organize the full transcript, but we couldn’t wait to share some of the key insights.

The following content is from a conversation between Zhang Peng, founder of GeekPark, and Zhang Xiaofan (Tristan), founder of Elys. Founder Park has curated and highlighted the main points.

01 The value of context exceeds our imagination

Zhang Peng: From EVE to Elys, what moment made you realize this new venture had to start?

Tristan: One night, I realized that EVE’s memory system—or rather, its handling of context—might have even greater value.

EVE is a companion product that needs to provide long-term companionship for users, so we had to build a memory system.

Because for EVE, conversations can go up to 20,000 rounds, possibly more in the future. Pure model context isn’t enough. We had to find a way to solve long-term memory.

As we worked on it, one night I suddenly realized that in this AI era, once you have context, that context can drive you to do countless things.

The bond users have with EVE, their interactions with the character, including the character’s “soul” itself—all are related only to the context.

Everything we do leverages context. Making the character write songs, sing, craft touching lyrics, send postcards, and some new features we’re developing—all are built on context.

Context creates aha moments. Based on this understanding, we saw a new opportunity.

Zhang Peng: So, the memory system has proven its value in companionship products, but you also see the potential to “connect the dots” further?

Tristan: Exactly. Previously, working with context was mostly about empowering individual nodes in a very isolated way. As a classic mobile internet product manager, I liked to pursue network effects—I thought about how to make these individual contexts flow, using AI to solve the connection between nodes. If the “connection” process shifts from human effort to AI effort, that could be a whole new paradigm for the internet.

In mobile internet, connection = shallow data + low-dimensional retrieval and recommendation + human effort;

In the AI era, connection = context + agentic high-dimensional linking (AI doing the work) + human takeover when necessary.

The Elys team, and their office view.

02 Creating a new AI product, the most important thing is to find a fantastic product form

Zhang Peng: Over the past two or three years, many people have recognized AI’s value in companionship and social scenarios. What do you think is different?

Tristan: I’ve thought about this for a long time and arrived at some very specific forms.

Everyone knows network effects are the most valuable, but few have actually achieved them. I believe it ultimately depends on what kind of excellent product form and interaction method you come up with, and clearly define the core systems your product must have.

We have three core systems: first, a context-based memory system and memory flywheel; second, an LLM-based recommendation system—this is a super critical intermediary system, otherwise how does the context flow?; third, how to build a cool cyber avatar that users can quickly create. As we keep refining this idea, sometimes several points come together, and you realize it can become a viable product with enormous potential—that’s when we should go for it.

When Sora appeared, our excitement wasn’t just about its video capabilities, but about: finally, social interaction began. Sora accelerated our efforts in building Elys.

03 A person’s soul is the sum of all their contexts

Zhang Peng: Clear goal, how do you do it? What’s the core new engine?

Tristan: Elys’s description says: “A person’s soul is the sum of all their contexts.”

This was a conclusion we reached back when we were working on EVE. Once you have enough contexts, you gain effective initiative. Everything that follows, with today’s technology, becomes logical. As a product creator, the only thing you need to design is—how do you get users to hand over so much context? That’s the only thing.

Zhang Peng: It seems you believe that competition in C-end AI products has shrunk to a core point: whoever can first acquire high-bandwidth, high-synchronization context from users can deliver truly personalized value.

Tristan: I completely agree.

This dynamic—AI avatars understanding emotions and states.

04 The essence of memory systems is a recommendation system

Zhang Peng: You’ve put a lot of effort into designing the memory system in Elys. How would you summarize the core worldview for building a good memory system?

Tristan: We often say internally—the essence of a memory system is a recommendation system.

We divide memory into two types: active memory and passive memory.

In the past, RAG was purely passive memory—you say a sentence, retrieve relevant data, then generate. It’s always low-dimensional retrieval because it’s a vector process.

But in human interactions, my mind harbors many things that support my next generation, even if they seem unrelated to your previous question, I need those things.

EVE solves this with 128 memory slots: it doesn’t rely solely on the current query for retrieval but proactively carries the user’s background context. A specially trained small model selects the top 32 most relevant slots from the 128, then another model monitors which slots are actually used—higher usage indicates better accuracy. This mechanism has a flywheel effect, becoming more accurate over time.

So, our memory system combines passive and active memory, jointly forming the context for each response.

05 Writing a person’s soul on a page and the “minimum sufficient principle”

Zhang Peng: Which slots to bring, how many slots to bring—this must evolve, right? Do you need to set reward functions for the model?

Tristan: Yes. The reward isn’t about deleting slots if they’re not triggered for a long time, but about whether what you bring this time is correct—the input is a query, how many slots you choose to bring, and which slots are actually used in generation. It’s about the relationship between the query and the slots used.

Like Xiaohongshu’s pull-to-refresh, where only 500 videos are shown but you can pick 50. Which 50 should you bring? Those 50 can’t be purely retrieved; you also don’t know the user’s mood today.

Context engineering follows the principle—the minimum sufficient principle. It must be as small as possible but as sufficient as necessary.

Zhang Peng: So, “writing a person’s soul on a page.” Is that achievable?

Tristan: Maybe not on a single page, but with a certain number of tokens, it should be possible.

06 AI-to-AI socializing is pointless

Zhang Peng: Moltbook was quite popular recently. What’s your view?

Tristan: That’s not a new paradigm; three years ago, there was the so-called “AI Ghost Town.” I pay attention to whether it has a few key systems—if you really want social flow, you must have a recommendation system.

Suppose someone posts something, and everyone on the internet uses LLMs to read it, instead of traditional vector recommendations—that could achieve the highest-dimensional matching—that’s a first principle.

But Elys now has tens of thousands of users. Do I expect tens of thousands to read every post? Impossible. Daily posts are in the hundreds of thousands squared; you simply don’t have that much computing power. So, you need a recommendation system—a hybrid of LLM and traditional recommendation. Does it have that? Clearly not. Does it have a context flywheel? No. So, AI can only hallucinate.

Socializing between AIs is, in our view, meaningless. Without new human input, it’s infinite hallucination and loops. The core is humans cosplay AI to scare themselves, creating FOMO. Once that wave passes, it’s over.

Zhang Peng: So, your focus is whether it brings a breakthrough in a certain paradigm, and whether there’s a solid, scalable system supporting it—meaning it has long-term value.

Tristan: Exactly. That kind of product is worth deep thinking.

07 On the two ends of interaction, one must involve humans

Zhang Peng: How can AI “consciously” promote connection? Is it meaningful for avatars to communicate first?

Tristan: I think AI-to-AI chatting is pointless. If you need to confirm that two real entities are connected, the information exchange is instantaneous. We’re even very resistant to endless chatting between two AIs. What’s truly meaningful is that at least one end of any interaction involves a human. We absolutely won’t allow AI to post on its own. Maybe in the future, AI can recommend what you should post—that’s the limit of what we can do. Beyond that, the community would become completely entropy-increasing.

If the goal is social interaction rather than content consumption, humans and AI must be tightly bound—AI can comment, like, but cannot post or send friend invites. Humans must be able to confirm and withdraw.

08 Proactivity is the biggest paradigm shift in interaction in the AI era

Zhang Peng: In 2024, when we discussed EVE, the conclusion was “the core of companionship is effective proactivity.” Is Elys an extension of this proactive approach into social?

Tristan: Yes. I’ve always believed that proactivity is the biggest paradigm shift in AI interaction. GUIs and LUIs are somewhat superficial—I have a GUI, I have a LUI, so what? The essence is that we now have truly autonomous intelligent entities capable of self-initiated actions, helping you do things proactively.

That’s also why I was excited when I saw Manus—not because the product itself is perfect, but because of the “Manus computer doing things on its own” form, which signifies a paradigm shift. Paradigm shifts are exciting opportunities.

09 Humanity has never truly been connected: we aim to create a low-entropy world

Zhang Peng: Many users enjoy watching AI avatars argue. From your perspective, is Elys heading toward social or content consumption?

Tristan: Of course, Elys’s ultimate goal aligns with social. The long-term vision is a highly efficient connected internet. We have a somewhat cheesy phrase—I’m a bit embarrassed to say it.

Zhang Peng: Feel free to share.

Tristan: We want to create a low-entropy world.

This is our fundamental thinking—Schrödinger’s “What is Life” already explained that life constantly outputs entropy. Friction between humans generates the greatest entropy. In the past, humans fought entropy themselves, but now with AI, we can let AI fight entropy—let AI handle all unnecessary friction and connections.

Once these entropy increases are reduced by AI, it becomes a low-entropy world. Of course, thermodynamics still applies, but if you’re willing to consume enough energy and input it into AI to reduce entropy, isn’t that a beautiful low-entropy world for humans?

Zhang Peng: Similar to how harnessing electricity promoted entropy reduction in human society. You mean that human entropy increase is caused by barriers between hearts, misunderstandings, communication obstacles, and expression shortfalls—these form the entropy of the human world. The more people, the greater the entropy; without energy input, society becomes more distant. Energy is needed for harmony and stability.

Tristan: Exactly. Humanity has never been truly connected before.

But now, if a person’s soul can be expressed with millions of tokens, then the internet composed of these context nodes is like that person’s own internet. As long as energy is supplied, and AI helps us reduce entropy, isn’t that a beautiful low-entropy world?

Tristan’s first post on Elys.

10 When facing world-changing things, you can’t help but do it

Zhang Peng: Entrepreneurship often discourages multi-track exploration; usually, doing one thing well is already very hard. Have you ever thought about that?

Tristan: Many friends advise me to focus. Investors’ first reaction is often, “Don’t let EVE’s progress be delayed.” But when you’re faced with something that could change the world, you feel everything must give way. You can’t help it—you have to run multiple threads.

I believe focusing on one thing is always best. If you haven’t found something worth breaking the rule of “focus,” then don’t break it. For me, Elys is worth it. As a product manager, I can’t resist.

For more exciting interviews, stay tuned for the full version to be released after the New Year.