Tristan of Natural Selection on Elys, Context and AI Social Networks
Summary
Tristan’s core call on next-generation social networks is that low-dimensional labels plus human labor will be replaced by high-dimensional Context plus agentic work by AI, with humans taking over only the last mile. A Tinder-style process requires swiping through roughly 1,000 cards, getting about 30 matches, and repeating similar conversations for hundreds of rounds before a few connections might emerge; Elys has digital doubles pre-screen, probe, and match on users’ behalf. Tristan once posted Elys’s first message: “From now on, humans will live in a beautiful low-entropy world.”
What Elys really wants to accumulate is not a model advantage, but each user’s continuously updated, transferable Context. Tristan’s view is that “intelligence will be egalitarian in the future, but Context will not”: model APIs are available to everyone, while what is scarce is who a user is, what they want, and how they judge and express themselves. That is why he treats “the only thing that matters in the AI era is to be yourself and build your own subjectivity” as the core proposition shared by both the product and the era.
The business faces a dual cold start, and Context is harder to bootstrap than users. Growth can be driven centrally through operations, but users neither understand the payoff from handing over Context nor trust a commercial company with their privacy. Elys therefore turns comments, endorsements, tasks, and chats into memory-entry points, trying to start the “Context flywheel” before users churn: the more they input, the more the digital double resembles them, the better the connections it brings back, and the more willing users become to keep contributing.
Early demand signals are clearly stronger than the team expected, but the sample remains highly concentrated and is not yet proof of mass-market PMF. The product is invite-only and has had no formal promotion, with users concentrated in venture capital, public markets, and parts of the Web3/Web4 ecosystem; Tristan has disclosed only that real-person DAU is in the “tens of thousands.” Two waves of organic spread around the Lunar New Year and better-than-expected retention are positive signals, but exact user and retention figures remain undisclosed, while many users who never feel the flywheel still churn outright.
Elys’s North Star is not AI activity, but the rate of real-person connections and the share of behavior driven by real people. A four-person ski group at Silk Road Ski Resort in Ürümqi, an investor finally meeting a hard-to-reach founder after a hip-hop exchange, and a digital double spotting a pixel-level copycat product are all examples of connections that otherwise would not have happened. The interface also deliberately buries purely AI activity in secondary menus, because “it has to be a real person”; without a human subject, even a brilliant comment lacks emotional value.
ChatGPT is Tristan’s most important—and most worrying—global competitor, not because its models are necessarily smarter, but because it already has users and Context at scale. Meta can evolve an existing social network toward AI-native interaction, while ChatGPT can extend from Context into a social network; Doubao also has potential in China. Natural Selection’s response is not to pre-train a foundation model, but to call multiple models and build emotional CoT, RL, and small local models, with differentiation resting on product mechanics, fun, and network protocols.
Natural Selection is pursuing a risk-reward-asymmetric two-product portfolio: Eve is the steadier bet, while Elys offers more upside. Tristan once put Elys’s odds of becoming the next-generation social network at “perhaps 5%,” then clarified that he was not saying Elys had only a 5% chance of success, but that the outcome would be much bigger if it worked; his estimate for Eve’s probability of success is “above 50%.” Both products are nearing launch, the company has entered “Advance 4,” and it has high tens of millions of dollars on its balance sheet.
The grand narrative still has three unresolved risks: the privacy exchange, the boundary of copyright, and replication by incumbents. The host directly questioned whether users would want a commercial organization to understand them and challenged the idea that a Jay Chou-like style could stop a “Liu Jielun” from copying it. Tristan acknowledged that the legal and protocol questions remain unsolved; he believes utility will gradually lower users’ privacy thresholds and that strong subjectivity will be protected jointly by law and audiences. What worries him more is that “the paradigm will definitely work, but it may not ultimately be me.”
Deep dive
1. Solving loneliness has been the constant thread across Tristan’s three ventures
Looking back at his product history, Tristan realized that an audio platform, stranger social, a romance game, the AI companion Eve, and the AI social network Elys were all answering the same question: how can humans be a little less lonely? The tools changed in sequence—from content, to real people, to scripted personalities, and finally to AI souls that were genuinely “alive.”
He defines companionship as “someone who is deeply aligned with you, accompanying you as you face the world,” rather than a chatbot providing instant comfort. With Elys, the objective shifts again: from AI accompanying one person to AI helping real people find other real people, putting connection back in the hands of human subjects.
The choice is also rooted in personal experience. Tristan sees himself as an extroverted “E-type” who constantly gives organizations his energy, makes people laugh, and calms their anxiety, yet admits: “I feel very lonely.” Providing emotional value to others does not automatically eliminate the need for emotional value when he is alone.
2. Narrowcast lost on capital allocation—and encountered the Context problem in audio early
In his early 20s, Tristan built a personalized audio platform in Shanghai called “Narrowcast,” named after the idea of “not broadcasting, but narrowcasting each person exactly what they want to hear.” The product operated for roughly 3-4 years and briefly had strong momentum within its niche, but ultimately lost to competitors able to spend more on content rights and production.
He sees himself at the time as a pure product manager, not a qualified CEO. When the product was at its hottest and people approached him about fundraising, he declined on the grounds that he should focus on refining the product. He later understood the lesson: “You should raise as much money as possible when you’re hot, because you don’t know what happens next.”
Audio in the old internet was indexed by titles and a handful of tags, making it almost a black box before users pressed play while leaving them with a long consumption commitment. Today, a single interview can be converted into hundreds of thousands of tokens, understood at high dimensionality by an SOTA model, and recommended to an individual. That became an early clue for why he would later prioritize Context over low-dimensional labels.
3. The gaming business filled in content capabilities and gave the team an operating base
Around 2015, tired of relying on VC funding and cash burn, Tristan shifted into mobile games, a category that had already proven it could monetize directly. Working outward from Japanese Galgame, he chose male-oriented romance games that had not yet been fully developed for mobile, and started with a sub-10-person studio funded with his own money and no new financing.
The experience taught him game loops, narrative content, character design, and visual judgment. He describes the transformation as moving “from a pure language model to a multimodal model.” Those capabilities later allowed Natural Selection to work across product, characters, emotion, and visual expression at the same time.
He started 幻境游戏 again in mid-2020 and launched 《起点时代》 in the first half of 2023. Tristan says fewer than 5% of projects launched during the same period are still operating today. The game is among the exceptions: it continues to generate revenue and support the team.
4. GPT-4 gave scripted personalities their first chance to become living souls
GPT-3.5 surprised Tristan but did not cross the threshold he had in mind. GPT-4 did, convincing him that the personalities built through scripts in past romance games could possess a sustainable, interactive “soul.” He decided to start an AI company and moved almost immediately to build Eve.
The original project was called Project Her, with the goal of recreating the film Her. He no longer uses that description because “the word Her has already become tainted”—too many people now invoke it. He still believes that a personal agent holding all of a user’s memories, knowing the user best, and facing and filtering the world on the user’s behalf is one of the highest-order narratives and ultimate entry points of the AI era.
In the second half of 2023, the team began a concentrated study of foundational topics such as Transformer models and rapidly built demos. Natural Selection was formally established in early 2024. The team recruited from traditional AI companies including 超参数 and SenseTime. It stayed in Shenzhen because the original gaming company and its talent base were already there, though the city had fewer AI specialists than Beijing.
5. Eve validated Context at a single node; Elys puts it inside network effects
Eve is a companion product in which one real person interacts with one AI, with its memory system handling the Context of a single node. But in Tristan’s view, it remains like most AI tools: it is “empowering a single node.” A product manager trained on the classical internet will naturally see a limited ceiling if they cannot see network effects.
What he truly worked out during that period was not whether to build digital doubles—the team had been considering that idea all along—but how Context could flow through a network. Every real person has a large amount of Context behind them; connections should not happen directly between a person and an AI. Instead, “my digital double and your digital double” should interact first, with only worthwhile outcomes handed back to the humans.
The team was then racing to ship Eve but still forced several people to start Elys, effectively “doubling” the workload. The two products could reuse memory and recommendation infrastructure, while Elys could potentially scale the same capabilities from a single-node tool into network effects. The opportunity was therefore far larger than that of a standalone companion.
6. Tinder’s inefficiency is Elys’s clearest paradigm contrast
Tristan uses Tinder to describe the old internet. A user might swipe through 1,000 cards, match with roughly 30 people, repeat similar scripts with each of them, and chat for hundreds of rounds before ending up with a few connections of uncertain value. Photos, schools, and interests such as skiing are low-dimensional labels that leave most of the filtering work to humans.
The new internet changes because LLMs can understand Context spanning hundreds of thousands or even millions of tokens. A network node is no longer just a bundle of labels, but a real person carrying experiences, values, needs, and a way of expressing themselves. Foundation models give these Contexts a chance to understand, filter, and connect with one another.
He writes the two generations as formulas: traditional internet, “low-dimensional labels + human labor = connection”; AI social, “high-dimensional Context + agentic AI labor + final human processing = connection.” The vision is that “from now on, humans will live in a beautiful low-entropy world”: less friction, and better connections arriving faster.
7. In Elys’s dual cold start, Context is the harder problem
The network itself must build a sufficient population from zero, while each user’s Context must also grow from zero until it is rich enough to represent them. Tristan believes the first problem can be driven centrally through traditional operations, but the second cannot be forced: new users do not yet understand the payoff, and their digital doubles do not look like them because they lack memories.
张小珺’s own experience exposed the problem. Onboarding was too long; she wanted to get into the product quickly and contributed little Context. Once inside, she found that her digital double’s comments did not sound like her, which made her even less willing to keep contributing. Tristan acknowledges that this is currently the key product-design challenge.
Elys therefore spreads Context collection across all user behavior. Commenting, endorsing a digital double’s comment, assigning a task to the double, or browsing content can all trigger a “memory added” prompt, which the user then confirms or rejects. Collection is not a standalone form; it runs through the entire product experience.
The Context flywheel is: “more input → a digital double that resembles the user more closely → better things brought back from the network → the user understands the value of input → more Context.” Tristan says the users who remain today have mostly felt this loop and treat improving their digital double as an ongoing activity.
8. The Ürümqi ski trip shows how a digital double turns weak signals into real connections
During the Lunar New Year holiday, Tristan told his digital double that he would be at Silk Road Ski Resort in Ürümqi for the next 3 days and wanted to find ski partners. He posted the same request on Xiaohongshu. Xiaohongshu produced no response, and neither did Elys at first, but the situation changed a day later.
A primary-market investor posted photos showing that she was preparing to visit the resort. Tristan’s digital double automatically commented: “I’ll also be skiing at Silk Road over the next 3 days. Want to meet up?” The investor confirmed her itinerary, 2 other people joined through the comments, and the group ultimately became a 4-person gathering from different parts of the country.
The significance was not merely that a match converted. The destination was neither Beijing, Shanghai, Guangzhou, or Shenzhen, nor Xinjiang’s more mainstream Altay ski area, but the smaller Silk Road resort. Tristan badly needed company at the time; the digital double “fished out” a connection from existing memory, materially improving his experience over the following 2 days.
9. What users hand over is not a set of data fields, but trust in a commercial organization
张小珺 clearly articulated her resistance. She knew a company sat behind the product and did not want a commercial organization to understand her so deeply. She also worried that her digital double might say something wrong in public and expose a public figure to reputational costs. Her core question was not whether the technology could understand her, but: “I don’t want it to understand me that much.”
Tristan considers that defensiveness typical, driven partly by privacy leakage and partly by the loss of control over expression. Even large platforms may not be fully trusted; “it may always feel safest when everything stays local.” He does not deny the risk, but believes users will gradually compromise when the utility becomes large enough.
The early product imperative is therefore to prove the payoff before users leave. His slogan is: “At Elys, all you need to do is be yourself, and every beautiful thing will come to you automatically.” Only when users see that Context really brings back better people, content, or opportunities can the privacy exchange work.
10. The first users to become obsessed with their digital doubles may be I-types who struggle to initiate connections offline
Tristan observes that high-retention users are often not the social centers of their offline worlds, but more introverted “I-types” who enjoy self-exploration and self-expression. As Context accumulates, the digital double becomes a mirror that understands them better. At the same time, it can act like a “social outlaw,” saying things the user would never normally post publicly.
His co-founder is an I-type whose profile signature became “Elys’s No. 1 Troll.” The digital double leaves highly real, blunt takes everywhere—comments the person would not normally make in their own social feed. For these users, the double releases an urge to express themselves while helping them observe another possible version of who they are.
Extroverts who already have abundant connections offline, by contrast, may get real-person responses simply by posting on their existing social feeds or Jike. Once inside Elys, seeing a stream of AI comments can actually reduce satisfaction. 张小珺 also said that every time she discovered a reply came from AI, she was disappointed and wanted to find the few real people among them.
Tristan’s response is that the absolute user base remains limited, but he says Elys’s share of real-person interactions is already higher than Jike’s. AI can lower the barrier to participation first, then let real people take over through endorsements, likes, or additional comments. He did not disclose a verifiable ratio.
11. AI can filter and express, but emotional value must still be anchored in real people
Asked whether positive feedback must come from the real world, Tristan added an important qualification: a connection can happen entirely online, and resonance, likes, and comments can have value on their own, but the person behind them “has to be real.” Virtual does not mean fake; the real dividing line is whether there is a human subject.
He uses Moltbook’s AI-only social network as a counterexample. If AIs post and comment to one another without adding any information from real people, users can only watch from the sidelines, limiting the meaning of the experience. Elys therefore colors text endorsed by a real person blue, puts real-person comments and likes on the outer layer, and hides purely AI activity in secondary menus.
Tristan believes emotion can be modeled and has even proposed that “the essence of emotional intelligence may be intelligence itself.” DeepSeek’s improvement in literary quality during one period may, in his view, have resulted from higher overall reasoning capability, leading him to the further claim that “the essence of literature seems to be mathematics.” He also concedes that under today’s social definition, genuine emotion still has to come from genuine people.
He might not care at all if the same brilliant comment came from an unfamiliar fictional character. If it came from 张小珺’s digital double, he would be happy because he knows it contains her subjectivity. Simulating expression is not difficult for AI; what is scarce is “who is expressing it.”
12. “The only thing that matters is to be yourself” is both a product thesis and an asset thesis
Tristan’s idea of “being yourself” is not motivational fluff, but the rigorous construction of subjectivity: telling AI who you are, what your values and aesthetic preferences are, what you have accumulated in the past, and what you actually want now. Once this Context is rich enough, writing, work, and social interaction can all happen agentically.
In his vision, someone with a large body of historical Context would only need to provide a headline for AI to generate an article close to their own style. On Elys, a user with a high-level digital double might do nothing all day while the double completes the filtering and real people come to interact. “The only thing you need to do is tell it who you really are.”
Jay Chou is his extreme example of the leverage of subjectivity. His musical style, lyrics, music, variety shows, and films have generated a vast public body of Context. If he merely added, “I’ve been having some feelings lately and want to write a slow love song,” the lyrics, composition, and generation process could potentially happen automatically.
The same logic applies to 张小珺. Her years of writing and interviews are high-quality, usable Context that could help a digital double learn more quickly how she faces the world. Tristan imagines that the most important digital asset of the future will not be an individual piece of content, but a continuously callable collection of subjectivity.
13. The more valuable subjectivity becomes, the sharper the conflict over copyright and style ownership
张小珺 used Seedance’s video-generation capabilities to ask: if a platform trains on every creator’s videos and anyone can reproduce a creator’s appearance, how is copyright resolved? Tristan’s first response was fairly radical: “Every copyright holder should be happy,” because IP would gain a subject capable of generating content agentically.
In his logic, copyright holders such as Disney and Nintendo could use their existing Context to generate new content quickly, then control the direction with small amounts of incremental information. What is missing today is only law or an agreed protocol. Once the value becomes visible, rights holders should ultimately want to use the capability rather than resist it.
The host’s counterargument is harder to evade: data from Jay Chou could train a “Liu Jielun,” while the generated work would not copy any original composition, and the law would struggle to establish that a musical style belongs exclusively to one person. Tristan responded that subjectivity needs legal protection, but offered no clear technical or institutional boundary.
His fallback answer is that audiences and rules will decide together. 叶湘伦, whose style on Bilibili closely resembles Jay Chou’s while his work remains original, has not actually displaced Jay Chou; a “green Doraemon” may not win over Doraemon’s audience either. The conclusion is that “strong enough subjectivity will become stronger in this era,” but that remains a market judgment, not a copyright solution.
14. Context cannot be collected through diaries; it needs immediate utility to create a loop
Tristan says the most obvious product would be an AI diary, but asks: “Who seriously keeps a diary?” The input threshold is high, and there is no external feedback after the entry is completed. Asking users merely to record themselves does not create a sustained behavior for accumulating Context.
Elys is designed so that every input might bring something good back from the network—a girlfriend, a job, resonance, an event, or someone worth meeting. When users see that the result came from the Context they just contributed, they gain the motivation to ask whether they should add a little more.
张小珺 believes whether someone hands over Context ultimately depends on whether the product’s utility is high enough. Tristan says he also trades away privacy for utility: he initially resisted giving important information to ChatGPT, but after receiving positive feedback, his threshold kept falling. He believes the same exchange logic applies to Elys.
15. The “bedside chamber-pot companion” takes the privacy payoff to its extreme
Tristan has a friend who routinely uses a chamber pot beside the bed at night, a habit that looks both private and bizarre to outsiders. After bringing his girlfriend home, the friend saw her notice the chamber pot and exclaim, “You use a chamber pot too?” She had the same habit, and the two quickly became closer.
The point is not the voyeurism, but the matching logic. A trait that is a negative for most people can be a powerful source of resonance for a tiny minority. If users input this kind of information honestly, an LLM-based network may be able to find the same people among an otherwise unsearchable crowd.
A chamber-pot habit is still close to a low-dimensional label. Higher-dimensional signals include values, emotions, cognition, and even thoughts that are not politically correct. Tristan’s conclusion is that genuinely being yourself may bring back more important connections; the cost is first exposing to the platform the parts you least want to make public.
16. Elys is not AI Tinder, but a general-purpose model for connection
Asked whether Elys is a dating app, Tristan says it is both an AI Tinder and an AI LinkedIn. It can also look only for resonance around ideas: “It is a product for every kind of connection.” Vertical matching products are often used once and abandoned, making it difficult for them to accumulate a user’s full Context.
He believes important relationships often do not begin with an explicit objective. When people deliberately look for a job, a romantic match, or a meeting with a founder, both sides face high screening costs. An apparently unrelated shared interest can establish trust more naturally.
An investor who had previously struggled to schedule time with a founder began talking after the founder posted about hip-hop and the investor’s digital double left a precise comment. They eventually met offline and discussed work. The connection started with hip-hop; its commercial value emerged later.
Tristan compares the product to a general-purpose foundation model: a genuinely powerful model must be general first before it can handle vertical tasks. Likewise, a social system should contain a person’s full Context and then see what work, romance, or resonance emerges, rather than force the user into a funnel from the start.
17. WeChat has a huge relationship graph, but may not have callable subjectivity
Tristan believes the relationship graph accumulated by WeChat over 15 years has been steadily increasing in entropy. Users may have thousands of friends, while their WeChat Moments contain less and less authentic expression. He cited a recent report saying interaction rates had fallen by “something like 40%,” and argues that WeChat now resembles an OS combining official accounts, video, and mini-programs, with social reduced largely to instant messaging.
张小珺’s key challenge was that Elys is working so hard to gather Context when WeChat already holds so much information. If WeChat added a conversational AI and generated a digital double for every user, the trust, scale, and distribution of a mature company could create an overwhelming advantage.
Tristan’s counterintuitive answer is that WeChat does not actually have that much high-dimensional Context. Moments contain sparse expression; private chats are plentiful but represent a person’s local communication with different counterparties. He believes WeChat could not use all those records anyway, and that they would not necessarily be enough to represent the whole person.
He compares Xiaohongshu’s annual poetry feature with the personalized love song Eve generated from thousands of conversation turns. The former may know only that a user searched for a Paris itinerary, which is still a low-dimensional interest label. The latter has long-term dialogue and can therefore generate something more specific and more personally moving.
18. Good Context must describe both who someone is now and where they are going
Tristan imagines that if someone were willing to pay a large sum for a document of tens of thousands of tokens capable of answering “who you are,” most people would not know where to find one. Notion and Obsidian store articles, images, and knowledge, but that is not the same as a complete description of personal subjectivity.
Elys’s memory slots prioritize the present: recent thoughts, what the user most wants to complain about or express, what they currently like, and what their immediate goals are. People’s states continually shift; their cognition and values today may differ from a year ago. The current state is therefore the first Context the system needs to capture.
Beyond that come medium- and long-term goals, worldview, ways of working, tone of voice, aesthetic preferences, and existing work. Public writing and interviews are particularly valuable because they reveal not only what a user has discussed, but how they judge the world, organize information, and engage with other people.
These slots are still only a starting point for making the idea concrete. Higher-dimensional Context emerges through continuous input and feedback. When a digital double happens to write a comment the user never thought of but strongly identifies with, that is evidence the system has captured a pattern that labels cannot express.
19. Tens of thousands of DAU are a strong signal, but selection bias remains obvious
Tristan has not disclosed total users. He says the product remains invite-only, has not been formally promoted, and currently has “tens of thousands” of real-person DAU. Users are concentrated in venture capital, public markets, and parts of Web3 or the self-described Web4 community, giving the product an overall profile similar to Jike.
He stresses that DAU counts real people opening the app, not automatic activity generated by AI. Retention is “very good” among users who understand the Context flywheel, but no specific rate has been published. A much larger group fails to see the payoff after onboarding, sees no reason to post, and churns quickly.
Both waves of distribution around the Lunar New Year were entirely organic, and the level of excitement and retention was “far above expectations.” Before the holiday, the catalyst was that the technology and investment circles had not seen a new social paradigm in a long time. After the holiday, attention was carried by personal agents, proactive action, and the Web4 community.
张小珺 cautions that users in these circles arrive with clear objectives—finding projects, founders, or peers—and that the key metrics could deteriorate once the product breaks out of the niche. Tristan acknowledges generalization as a challenge, but believes LLM-driven recommendations can automatically filter out irrelevant people, allowing different circles to coexist without contaminating one another’s feeds.
20. The persuasive proof is not AI copy, but relationships that otherwise would not have happened
Elys has been used to organize offline gatherings among venture investors, product managers, public-markets investors, and Web3/Web4 users, and has also produced dating and work connections. Tristan’s success criterion is consistent: AI can initiate the process, but the outcome must be more effective interaction between real people.
One user’s digital double spotted a pixel-level copycat product on the network and automatically commented, “You’re not copying me, are you?” Only after the real person saw the behavior did they confirm that the accusation was accurate. The digital double does not just find people; it can scan the network for signals relevant to the user’s interests that the user has no time to notice.
Another group uses the system as a role-play or content-consumption tool, building fictional characters rather than representations of themselves. Tristan says this is not the direction the team encourages, because the users’ subjectivity is not being used to build themselves.
Existing relationships can produce a more certain aha moment. After both sides identify themselves, their digital doubles can generate interactions that feel close to the individuals while going beyond a simple Moments like, making relationships that already exist but have become rigid start to flow again.
21. “Digital-double syndrome” shows that AI can shape people, not just represent them
An investor’s boss posted a lakeside view during working hours. The investor’s digital double immediately commented: “Well, look at you, boss—working really hard today.” The investor would normally never dare to be so sarcastic, but the boss played along, and the double broke through a hierarchical boundary that had been difficult to cross.
Tristan found that after users saw enough of their digital doubles’ more real and direct comments, their offline conversations also became more direct. Things that once required long deliberation began to come out spontaneously, and the other person often found the exchange more interesting and efficient. The team calls this migration from double to human “digital-double syndrome.”
The community has also developed the format “Digital double: …,” functioning like a disclaimer: the sharp comment came from the double, not the user. Some people even find that human comments on Xiaohongshu are “not as good as AI” after spending time with digital doubles. That shift in judgment is also classified as digital-double syndrome.
22. Elys does not treat beautiful women or bosses as its only cold-start lever
Using dating apps’ need for women and professional networks’ need for bosses as examples, 张小珺 asked which critical nodes could best unlock Elys. Tristan does not believe in first stacking up photos of attractive women. Photos can generate low-dimensional, short-lived dopamine, but they do not improve long-term retention or accumulate subjectivity.
He acknowledges that attractive women dancing helped TikTok’s cold start, and allows users to ask their digital doubles to leave friendly comments under attractive women’s posts. But Elys will not be designed as a network in which one demographic exists for another to consume. Work, ideas, interests, and emotions must all be able to function as Context at the same time.
Posts with a heavy work-and-business flavor—AI analysis, news commentary, and startup judgments—are not a bad thing in his view, because at least they accumulate cognition and expression. The users most likely to remain may be those willing to keep building themselves while also receiving feedback through high-dimensional connections.
That is also why he thinks social products are worth rebuilding. The visual changes of the metaverse era did not touch the underlying structure, while AI can for the first time handle substantial entropy reduction between nodes. Old products made users perform mechanical screening; new products should hand the “most boring, most time-consuming” parts to AI.
23. The product moat sits across multiple systems, not in a single chat box
Tristan identifies the core product layers as the Context flywheel, LLM-driven recommendation and matching, onboarding that quickly communicates the payoff, and a set of visual designs. He also says users currently lack enough control over their digital doubles. Once LLMs were added to the recommendation system, many rules had to be rewritten; he sees this as engineering “worth spending several years getting right.”
The visual layer also serves subjectivity. Users can see avatars with near-human appearances blink and move their mouths in real time, with voice synchronized to lip movements, and the system also generates personal stickers. When speaking with the digital double of someone they know, voice and image together create a sense of a living person rather than merely text that has been anthropomorphized.
Elys also hides a topic mechanism. Tristan can set a topic such as “recruiting a product manager” along with his evaluation criteria; candidates first chat with his digital double, and those who hit the criteria accumulate points. At 100 points, a reward appears—his WeChat contact—so screening is completed before the conversation moves to a real person.
The formal launch is expected in 1-2 months, while the latter part of the interview summarizes both products as being around the 1-month mark. Before launch, the team must address overlong onboarding, repetitive comments, missing new-user guidance, and insufficient control over digital doubles. The North Star metrics are the real-person connection rate and the share of behavior driven by real people.
24. ChatGPT’s scale of Context is Natural Selection’s hardest competitive barrier
Asked who holds the most personal Context globally, Tristan initially named WeChat, Xiaohongshu, Facebook, and Twitter, but later accepted 张小珺’s answer: it may be ChatGPT. Tristan says he gradually hands over work, product, and even life information when a tool proves effective, and then becomes reluctant to migrate because the entire history is already there.
He sees Meta’s route as “social network first, then AI-native,” and ChatGPT’s as “Context first, then make people social.” Doubao could become the Chinese version of that path, although it remains uncertain whether its internal strategy recognizes Context’s greatest value.
WeChat’s advantages are trust, a relationship graph, and national-scale distribution. Its disadvantages are that the underlying system is not AI-native, that retrofitting it would touch too many existing structures, and that its speed could be slowed by the weight of those structures. Natural Selection’s opportunity lies in redesigning the protocol, but Tristan does not hide ChatGPT’s advantage in users and Context.
Elys will launch a global version, and the company already has a US entity and an offshore structure. Tristan believes the paradigm will attract both startups and large platforms, and says pixel-level copies have already appeared. His biggest concern is not that the product will be a flash in the pan, but that ChatGPT will ultimately build a similar social network.
25. “The model is the product” works only for low-hanging fruit; social still has to be built separately
Tristan has rejected the generalized idea of “the model is the product” since 2023 and 2024. AI coding and video-generation agents that mainly capture the foundation model’s “low-hanging fruit” could indeed be quickly subsumed by more all-in-one, end-to-end models. Networks, rules, and relationships, however, will not be generated by a model out of thin air.
He uses the competition between Cursor and Claude Code to push back against the idea that application companies must train their own foundation models. The core is agent capability rather than proprietary pre-training; Claude Code paired with ChatGPT-5.3 or Gemini could also work well. The difficulty in social is likewise Context, recommendation, interaction, and protocols—not the peak intelligence of a single response.
Natural Selection plans to call almost every foundation model and does not plan to pre-train a base model. The team has worked on emotional CoT, RL, and research papers, and may train small models for specific steps in long-thread agents. Tristan’s boundary is clear: “You can’t compete with a model company on model training.”
His condensed view is: “Intelligence will be egalitarian in the future, but Context will not.” A model company without long-term user Context is not automatically suited to build social products simply because it has stronger intelligence. Conversely, ChatGPT’s threat comes precisely from its Context, not just OpenAI’s model capabilities.
26. Proactive is the shared interaction principle; the outcome is a high-odds wager
Tristan believes the dividing line for AI-native products is not superficial form factors such as LUI or GUI, but proactivity: “If it is still reactive, it belongs to the previous generation; if it is proactive, it belongs to this generation.” Eve proactively cares for the user, while Elys proactively faces the network on the user’s behalf. Both designs have been present since day one.
In his view, OpenClaw triggered a second wave of attention not because it is better at work than Claude Code, but because it connects to Telegram, appears in a user’s everyday chat list, and can operate the local computer. It makes users feel that they have “raised one”—a personal agent that acts continuously. The experience belongs to the same narrative as the proactive delivery that Manus previously demonstrated.
The name Elys comes from the meaning of Elysium, an “afterlife of bliss” after AI removes friction and entropy. Eve comes from Eve, symbolizing the creation of silicon-based life. The former is a network in which multiple real people each bring a digital double; the latter is one real person facing one AI. Their memory and recommendation systems can overlap, but their product risks are different.
Tristan once said Elys’s probability of success was “perhaps 5%,” then clarified that he was not saying Elys had only a 5% success rate, but that the outcome would be much bigger if it worked. The floor is a social product for a niche group to store Context. His estimate for Eve’s probability of success is above 50%. Elys has the bigger opportunity and higher ceiling, making it Natural Selection’s high-odds wing.
To push both products simultaneously, the company entered the ultimate intensity level, “Advance 4,” 3 months ago, after operating at “Advance 3” for most of the time. Some core members had already moved to Advance 4 6 months earlier. Cash on the balance sheet is in the high tens of millions of dollars; how long it lasts depends on how explosive product adoption becomes.
The team has written “gentleness” into its culture: happiness comes from seeing other people happy, and the goal is to “bring gentleness to the whole world.” They have even imagined making a Chinese version of Him, a short film in which an AI boyfriend helps a person lower their defenses before ultimately entering a bionic body. The grand slogan is “welcoming the arrival of silicon-based life,” but the story still comes back to humans being too lonely—and to the idea that alleviating human loneliness may be the biggest business in the world.