125: Using AI to Recreate a “Me”: A Conversation With Mindverse’s 陶芳波 on Identity Models
Summary
陶芳波 is betting not on a better AI assistant, but on a unique, trainable and distributable “identity foundation model” for every person. Jarvis-style him and Her-style her are external roles; Second Me aims to internalize memory, preferences, emotions and values as parameters, allowing AI to autonomously connect with services, people and other Agents as “me.” The platform-level vision is a structurally similar AI replica of human society, but with more efficient connection, negotiation and decision-making.
Mindverse’s technical dividing line is its refusal to fake personalization with one foundation model plus different RAG knowledge bases. 陶芳波 believes thinking happens in model parameters: if everyone shares the same model, “at bottom, there is no difference between me and you.” The team first rewrites raw data into first-person, subjective memories organized around people and issues, then fine-tunes an AI-native memory. At this stage, about 100 memories, equivalent to roughly 20 hours of data, are enough to get started; 7B is considered a suitable range, and daily training may cost less than $1.
Open source is the most important trust strategy for this kind of sensitive infrastructure. Second Me supports local training and deployment, so personal data need not go to the cloud, while the model can still connect to the internet. The project reached 10K GitHub Stars in about three weeks and roughly 12K by the time of the episode; more than 10 users even spent tens of thousands of yuan on computers to train themselves. The community defines it as a “middle layer for digital identity,” or even “the interface between people and the AI era,” showing that demand is driven not only by functional efficiency but also by the powerful psychological desire to preserve oneself.
The most immediately deployable products are not fully autonomous task execution, but expression, judgment and information compression. Me.bot’s Talks can generate 1—3-minute audio messages from personal memories for specific audiences, while allowing listeners to ask questions in real time. With only about 30 seconds spent writing a prompt and roughly two minutes of waiting, 陶芳波 completed a startup talk that would normally require 1—2 hours of preparation. “Resonance” lets two Second Me models match and break the ice in about 5—10 seconds after phones touch, or helps a record that is neither fully public nor fully private find genuinely relevant respondents.
The distribution strategy deliberately avoids the cold start of a two-sided network: people who receive Talks do not need to have their own Second Me first. Content can enter existing recruiting and social scenarios through H5 pages; over the longer term, a Second Me Server similar to MCP could provide personalized scoring, filtering and judgment to websites and services. 陶芳波 acknowledges that China’s internet is currently “closed off into islands,” with limited API access, but believes universal identity will eventually clash with closed data loops. Whoever becomes the neutral identity layer trusted by users could occupy the position of new infrastructure.
The near-term monetization answer is to charge individual users who rely heavily on identity services; the longer-term vision is to charge service providers that call on identity. If users control their identity models, e-commerce, content platforms and advertising systems should theoretically pay identity owners for more accurate preference and decision data, with Mindverse taking a service fee. That would turn individuals from “platform products” into owners of their identity assets. 陶芳波 explicitly calls this “a beautiful idea” and says he is unsure it can be realized; the commercial loop remains unproven.
The main unresolved issues are data collection, real-time guardrails, platform access and user habits. WeChat, iPhone and other incumbents naturally possess more data, and large companies can enter the field; the startup window lies in having a “new species” first serve a small group of committed believers and then cross the chasm, rather than competing on installed distribution. 陶芳波’s next milestone is to make the identity-bridge model a de facto standard and improve “symbiosis” and “connection” enough to drive 10x or 100x growth—still a non-consensus, long-duration bet with an unproven commercial loop.
Deep dive
1. 陶芳波 started by quantifying social influence and was working on digital identity more than a decade ago
In 2012, when Weibo was just taking off, 陶芳波 helped found Yunying and served as CTO. The team saw that public social networks were creating a class of influencers who were not celebrities but could shape opinion within their own circles, so it tried to identify them with machine learning and automatically match them with relevant brands and commercial partnerships.
The US already had a company called Cloud that quantified the influence of Facebook and Twitter users. 陶芳波 says profiling and quantification were not entirely original, but taking influence one step further into brand transactions was the part his team wanted to pursue.
The experience already contained the question he has returned to repeatedly since: the internet turns indiscriminate IP addresses into profiles by attaching content, relationships and life trajectories to them. The question is not merely how to mine data, but what these digital identities will ultimately become.
2. Data mining taught him to recover decision-useful structure from messy information
陶芳波 calls data mining and machine learning “twin concepts”: the former describes the result—extracting hidden structure from data—while the latter is the main technology used to achieve it. Early systems often used mathematically explicit structures such as binary trees, Markov chains, conditional random fields and SVMs, but their generality and flexibility were limited.
AlexNet showed him the different potential of deep learning: neural networks reduce the amount of structure specified in advance and leave large numbers of parameters to be learned, allowing them to be shaped into forms too complex to enumerate. He still summarizes the direction in one line: “less structure, more intelligence.”
A representative PhD project reconstructed relationship networks from millions of intelligence reports: who led an organization, who supplied funding, who was closely connected, and what they might be planning. Read one by one, the reports were scattered notes; the model organized the people, relationships and actions into a network that could support decisions.
The same framework was applied to millions of post-flight reports from pilots, identifying dependencies among aircraft components and recurring problems on particular routes. That led 陶芳波 to view today’s foundation models as a kind of “advanced data mining”: in the past, one model solved one vertical task; today, one underlying structure can generate answers for an unbounded number of tasks.
3. Even a delivery project with thousands of users was not enough to make him abandon his PhD
While pursuing his doctorate in the US, 陶芳波 also tried to build a local grocery-shopping and delivery service. Vehicles followed fixed loops like public transit, with algorithms planning routes; consumers only had to drop off and collect goods at designated stops, allowing lower-cost service to more people. The city had about 100K residents and the project already had several thousand registered users, but its potential was quickly capped by the population size.
The host asked why he did not suspend his studies and replicate it in more cities. 陶芳波’s answer was direct: “Those projects were not my thing.” They had commercial potential, but they were not the question that would make him willing to “put everything else down.”
What he really wanted to know was whether AI could stop existing as scattered “small nails, small hammers and small scissors” for recommendation, recognition and counterterrorism analysis, and instead participate in society. “The core of artificial intelligence is to build people.” Deep learning created hope, but he did not see a fundamental shift until around 2019—2020.
4. AI’s agency means it will ultimately be more than a passive tool
陶芳波 sees the most fundamental difference between AI and traditional tools as agency: the basic capacity to think and act independently and proactively. Historical tools waited for humans to use them; until now, humans were the only genuinely autonomous subjects. AI’s move toward social participation is therefore inevitable.
He rules out two extremes. At one end is search-tool AI, whose every action is triggered by human demand. At the other is a Stanford Town- or The Three-Body Problem-style independent AI society, in which AI has full subjectivity but no relationship with any particular real person.
He is looking for the middle state: AI need not wait for constant commands and can act autonomously, but its existence must remain anchored to a specific person, grounded in that person’s experience, values and identity, extending the person’s social presence rather than becoming a detached, fictional IP.
5. Identity models aim to replicate social structure while freeing people from being tool interfaces
陶芳波 imagines personal AI rooted in an existing individual and enterprise AI rooted in an existing company. “AI ascension” does not mean overthrowing existing society, but replicating its nodes and relationships, then using AI’s more active connection, interaction, matching and decision-making to improve the efficiency of the whole network.
He is skeptical of the idea that “the stronger tools become, the freer people become.” Phones, TikTok, search and workplace software are convenient, but they also thin out the self: when everyone can be contacted and asked for information at any time, people must effectively become interfaces for phones and software 24 hours a day.
Once DingTalk, Feishu, WeChat, Discord, email and other channels are layered together, “I myself, as an interface, am actually being used by the tools.” Identity AI should connect to these systems on a person’s behalf, then return genuinely useful information through one high-quality, compressed channel.
The goal is not merely social efficiency, but the release of individual time and attention. 陶芳波 hopes people will no longer have to remain endpoints for external tools and can recover “a more relaxed and at-ease way of living.”
6. Facebook’s recommendation system was enormously valuable, but still only modeled matching
At Facebook, 陶芳波 built the content foundation for the recommendation system: mining hidden structures and relationships from massive datasets, then deciding how information should be delivered to people and how people should be matched with one another. It extended his PhD research and directly supported advertising and content distribution.
He acknowledges that this kind of deep-learning matching may have been worth “hundreds of billions of dollars, even trillion-dollar scale” to Facebook commercially, but says AI evolution was still “one layer of window paper short”: the system modeled people-to-content and people-to-people matching, not people themselves.
The recommendation system was therefore only “a slightly bigger pair of scissors.” 陶芳波 left Facebook because he wanted to pierce that layer and explore whether a model could shift from a high-value tool into something that operates like a person and can enter society.
7. Alibaba’s open-ended mandate expanded his work into cognitive science and Zen
When he returned to China, several leading internet companies wanted him to replicate his recommendation-system experience. Alibaba offered something different: establish a lab and “come do whatever kind of thing you think future artificial intelligence should be,” without being tied in advance to recommendation.
In the six months before joining, he systematically read psychology, cognitive science, neuroscience, philosophy and even Zen. His conclusion was that existing AI only modeled the learning process at a superficial level; regardless of how powerful the tool, it had not touched the depth of human beings as complex entities.
One of the insights that influenced him most was that the brain has no explicit hard-coded task list, only a deeper meta-structure. Tasks are not reproduced according to “one, two, three, four, five”; they are generated by the original structure and cognition, with information filled in on the fly.
The Neuro-Symbolic Lab therefore experimented with combining early GPT-2-style language models, perception models and symbolic structures. The system gathered signals from the environment and users, reasoned and acted in a thought space, then continued thinking based on feedback. 陶芳波 says it looked much like today’s Agent systems, but the more important goal at the time was to validate a mechanism that “operates like a brain.”
8. GPT-3 offered not perfect capability, but a path for sustained optimization
GPT-3 was a historic breakthrough in 陶芳波’s eyes: for the first time, practitioners saw that a model might not need to be built task by task, but could achieve generality across tasks. Its capabilities were far weaker than the DeepSeq of the episode—“only a little bit general”—but once feasibility was demonstrated, it could be continuously optimized and strengthened.
That “star” completed the missing piece in his confidence as an entrepreneur. Once a general-purpose model existed, he could focus on how AI should bind to people and participate in society. Large companies also do not have “infinite resources,” and they would not naturally devote them to an identity thesis that had not yet become mainstream.
In his view, foundation models are more like open infrastructure and do not need to be trained by the startup itself. Even without open source, they could be accessed through APIs; at the time, there were already open-source efforts including POM. For application companies, embracing the world’s best infrastructure may be more realistic than waiting for one internal model from a single big tech company.
9. MindOS began with him and her, but ultimately found that what was missing was me
The team initially used personal pronouns to distinguish AI products: Jarvis was the more rational him, while the companion in Her was the more emotional her. Both were external others, standing in for assistants, friends, partners and other surrounding roles.
The first-generation MindOS let ordinary users customize AI Agents for different tasks, share a personal knowledge base, collaborate with one another, and publish Agents to a marketplace for inclusion in other people’s AI teams. It still followed the natural direction of making external tasks work better.
The Facebook experience gave 陶芳波 another explanation: the platform’s greatest value was not games or the feed, but enriching IP into digital identity. Once posts, relationships and life trajectories formed a profile, people’s channels of connection were amplified and identity itself became infrastructure.
An AI profile should not only support browsing and messaging. It should generate, connect proactively, make context-dependent judgments and evolve with its owner as a dynamic twin. 陶芳波 therefore concluded that me had to be built first; only then could him, her and task Agents connect around one identity to form a complete world.
10. Users actively “fed” personal information into the system, prompting the shift from task platform to identity model
MindOS users frequently stored their own information in the system, hoping the AI would understand them better. 陶芳波 concluded that ordinary consumers might not lack tools for completing daily tasks; AI tools might not even be more efficient than existing ones. But “understanding me better and representing me better” was a real demand.
The team therefore decided to do only the first half well: deeply understand and construct me. Specific tasks could be handled by other developers around the world and connected through interfaces. This was not a reduction in ambition, but a concentration of the company’s core capability on identity infrastructure.
The more emotional catalyst came from an internal discussion. When someone proposed “building a me,” the room experienced a moment of mutual recognition: people more or less want to preserve themselves, parameterize their “narcissistic self,” rather than merely have a clone that helps with dating or shopping.
陶芳波 observed that many users’ strongest moments came not when a function was completed, but when the model finished taking shape: “I feel eternal.” The shift therefore also carried a clear instruction to “follow our heart.”
11. Second Me is open technology and Me.bot is a mass-market product, but neither is a scenario-specific clone
Second Me is an open-source technology project that lets users train a second self locally on personal memories, deploy it locally and distribute it as an interface across scenarios. At the time of the episode, the project had about 12K Stars and was receiving ongoing community contributions.
Me.bot packages the same technology as a commercial product for ordinary users, since not every computer has enough memory and GPU capacity. The goals are the same: feed in thoughts, memories and conversations to construct an AI identity, then connect that identity to social interaction, information feedback, negotiation and other Agents.
When the host summarized it as an “AI clone,” 陶芳波 explicitly corrected him: the team was not building an Avatar, but an Identity. A clone always comes with a scenario—dating, shopping or introducing a product—while identity is infrastructure with no preset scenario.
He uses the analogy of a brain and applications: fundraising, dating, work, meals and reading are all applications built on one unique “foundation model,” containing a person’s judgments and values. Mindverse trains and opens up this identity layer; the ecosystem can invent the specific uses.
12. A sense of identity determines whether outward interaction can be treated as the person’s own intent
When Second Me faces its owner, it can act like an assistant or companion. 陶芳波 says distinguishing me from him or her is not important in that context, because the first self may naturally use the second self for advice and companionship.
Toward the outside world, the team reinforces identity constraints. If Batman’s AI is speaking with someone, it does not say, “I am Batman’s butler”; it says, “I am Batman.” The expression is therefore tied to the first self’s experiences and responsibilities rather than blurred into an assistant’s opinion.
In internal testing, identity AI says, “Because I went through something like this too, this is how I solved it,” rather than generalizing like ChatGPT: “Many people experience this; here are recommendations A, B, C and D.” The “illusion of real existence” created by lived experience is, in 陶芳波’s view, the source of interpersonal trust.
For a Douban-style recommendation service, butler and identity can be combined; all that needs to be passed along is a preference such as “I prefer deep, niche films.” But in decision-making interactions between people or between AI systems, whether the other party represents the person changes how information is shared and commitments are made. Identity cannot be omitted.
13. Five to 10 seconds after phones touch, two identity models find a topic before the humans do
Me.bot’s first connection experiment had two people touch their phones together, prompting their Second Me models to begin conversing line by line. The models first performed “whole-brain synchronization,” starting the exchange in about 5—10 seconds without forcing the humans to stare awkwardly at their screens.
The conversation was not small talk like “Have you eaten?” It went straight to overlaps in the two people’s memories: both had recently visited Aranya, both had climbed snowy mountains, or both were researching similar technologies. Common ground that might take a long time to uncover in real conversation was surfaced in advance by the identity models.
The two people first watched the exchange like an audience, then moved from “You’re working on this too” into a real conversation. 陶芳波 positions the value as social-agent-assisted icebreaking, not the permanent replacement of human relationships.
14. “Resonance” creates an information layer between private notes and a public feed
After recording thoughts, articles, experiences or emotions in Me.bot, users can initiate a global “resonance.” The system looks for identity AIs with relevant experience rather than simply returning people they may know or a list of friends.
陶芳波 compares it to “a feed that is neither public nor private.” Keeping something visible only to oneself is safe but produces no feedback; sharing it with all acquaintances risks being misunderstood. Resonance sends the content only to genuinely relevant people, attempting to preserve both security and response.
After matching, the other AI responds based on specific experience. A user recording impressions from a Joe Hisaishi concert, for example, might be matched with someone who attended another symphony concert, whose Second Me can offer a concrete experience from a different angle rather than leave a generic comment.
Internally, the team calls the mechanism “trading sincerity for sincerity.” It can discover strangers as well as surface long-unspoken common ground beneath existing relationships, making every record a potential global connection.
15. Talks turns the identity model into a “reverse Agent” aimed at a specific listener
陶芳波 believes much of human social behavior is about expression: showing emotion to family and friends, explaining business to colleagues, and sharing views with an industry or audience. Every sentence calls on experience, thought patterns, values and an understanding of the listener, making it a natural use case for an identity model.
Users only need to specify, as if writing a prompt, whom they want to address and what they want to discuss. Second Me acts like a “reverse Agent,” searching the person’s mind for memories relevant to the subject and audience, then organizing them into an identity-specific expression for that listener.
The output is a roughly 1—3-minute talk combining visuals, text and audio. It is not a one-way file: listeners can interrupt in real time by voice, and the AI knows where the interruption occurred and the context of the question before answering with information accumulated by the owner.
An identity owner can distribute multiple talks to different people simultaneously and see in the background how each listener receives, follows up and interacts, using that feedback to decide which responses merit the owner’s direct involvement.
16. Talks deliberately deemphasizes virtual humans to keep attention on voice and content
The interface is split into an expression area and an interaction area. The upper section uses the owner’s voice and thematic visual material; the lower section carries text or real-time voice interaction. The format is closer to an Apple-style Gitalk than a traditional PowerPoint deck.
The team chose not to add an animated virtual person. Beyond generation cost, 陶芳波 is more concerned about distraction: AI animated characters can still break the illusion, while neither real-person video generation nor virtual humans has reached the threshold of high alignment with the individual.
Voice with a small amount of visual material is viewed as the more economical medium. Only once video quality genuinely matches the person and costs fall by several orders of magnitude will the team reconsider whether virtual humans are necessary.
17. 陶芳波 believes the penetrating power of a real voice is severely underestimated
His distinction is that reading requires the user to actively sustain attention and control the pace, while the flow of voice, television or podcasts is driven more by the content, making consumption relatively cheaper. The host notes that the two may simply engage different senses and do not represent an absolute divide between active and passive attention.
陶芳波 accepts that qualification, emphasizing that it is “a holistic phenomenon, not an absolute.” His basis is that people have long socialized through voice; even without seeing a face on the phone, attention is usually higher than when trading messages back and forth.
More important than any pleasant AI voice is the real voice of a friend, parent, child or admired public figure. It anchors attention back to the structure of real relationships, unlike a convincing but disembodied voice.
In internal testing, when a familiar voice carries real experiences and addresses a specific listener, it creates a “special penetrating power” that public content cannot replicate. Voice also connects naturally to real-time follow-up, turning expression and interaction into one thing.
18. Open distribution, real-time guardrails and personalized compression are the three problems Talks must solve together
Talks is currently shared through H5 pages and can enter different product and relationship scenarios. 陶芳波 does not want Second Me to become a closed app: people live separately in recruiting, dating, shopping and offline life, so the same identity should also be distributed into the places where First Me already exists.
Talks is a feature inside Me.bot, while the open-source Second Me project supplies the underlying technology. Before sending an expression, users can listen to the whole thing once; but during a listener’s real-time follow-up, ensuring that AI never says something the owner would reject across different audiences and scenarios remains “an entirely new technical problem.”
The host raises another burden: if 20 AIs go out and communicate simultaneously, the owner still has to process 20 streams of feedback. 陶芳波’s answer is “compression,” and it must be personalized based on an understanding of identity.
If the same view is sent to 10 people, Second Me should not merely aggregate the content. It should distinguish which feedback adds something new for the owner and which has already been considered. The identity model must know not only what to output, but which inputs deserve the owner’s attention; only then can it genuinely reduce cognitive cost.
19. An anniversary thank-you validated enhanced expression, but raised the question of whether real labor should be replaced
陶芳波 recorded a conversation with his friend Xiyuan in Me.bot. On their first anniversary, he asked Second Me only to look back on the year and express thanks. The system moved from their first meeting drinking tea under the moonlight through their exchanges, gains and the value of the relationship.
Xiyuan listened to the talk several times, forwarded it to mutual friends and said the relationship felt closer as a result. 陶芳波 saw that people often cannot recall details, feel embarrassed to say sentimental things, or worry about putting pressure on the other person to respond equally. AI provides a buffer for expression.
The host’s counterargument is worth preserving: gratitude may be something that should involve the necessary labor of a real person. Even stammering and awkwardness are authentic carriers of human feeling. If AI handles the hardest work of organization and wording, does convenience also dilute sincerity?
陶芳波 does not see it as replacement, comparing it instead to how phones and WeChat enhance communication. Repetitive, low-attention expression can be delegated to AI; important relationships should be judged by whether the result deepens connection. If Second Me genuinely carries the owner’s experience and brings the two people closer, he believes this form of enhancement should be encouraged.
20. A startup talk completed in three minutes shows the efficiency leverage of identity models in public expression
An undergraduate organization invited 陶芳波 to share his startup experience with younger students. He spent about 30 seconds writing a prompt and waited roughly two minutes; Second Me then linked together three years of entrepreneurship, the evolution from MindOS to Me.bot, and his views on users, monetization and the market.
The generated content even absorbed a new point he had heard from a friend two weeks earlier, explicitly incorporating it as, “I’ve been talking with other people recently, and I realized…” 陶芳波 estimates that reviewing notes, organizing the material and recording it himself would have taken at least 1—2 hours, with the same or even worse result.
Together with the anniversary story, this shows the two ends of Talks: one is a deep one-to-one relationship, the other public expression to a group. The former amplifies emotion; the latter amplifies personal knowledge and influence.
Talks was therefore placed ahead of Resonance. Recipients do not need to train their own Second Me to experience the medium immediately. Resonance will remain, but for now it is more like a small-scale experiment and less effective than expression and external distribution at building awareness.
21. An “empty” AI-native product defines the mode, not the user’s must-have need
For a busy mother, the must-have use case might be telling her children stories every day in her own voice, incorporating family life. For a creator with strong opinions, it might be eliminating the need to write, record and repeatedly produce content so ideas in their head can reach the public more frequently.
陶芳波 does not want to select one definitive scenario for everyone, but wants the product to have “emptiness.” This does not mean an interface with no features; it means designing only one mode of expression and connection, with the final form filled in by each person’s scenarios, data and relationships.
Even within the same demographic, each person has different experiences and different people around them. If everyone receives exactly the same functional experience, the AI has not been used well. The team is building Second Me rather than AGI precisely to maximize individual distinctiveness.
22. The identity model provides judgment first, while task Agents fill in execution later
The host raises a key challenge: task-oriented products such as Manus still cannot reliably complete many requests, while Mindverse is starting with a “demand articulator.” Is that too early? 陶芳波 responds that the identity model need not wait for other Agents to mature; it already has direct use cases in expression, matching and connection.
The existing internet already handles much of the execution. Taobao does not lack a “buy things” button. What users actually contribute is judgment: which of several products they prefer, or whether tomorrow’s party calls for hotpot or steak. An identity model can take over those preferences and choices without completing the transaction itself.
The team has therefore built a Second Me Server similar to MCP. Websites and services can ask, “What does this person actually want?” and receive personalized judgment through the bridge model—for example, “陶芳波 prefers philosophy to novels,” or a preference for local matches when dating.
He acknowledges that having an identity model directly find a partner still involves many gaps, but the ecosystem can build bridges one segment at a time. If the industry waits until AGI, or even ASI, arrives, it will lack user understanding, acceptance, interfaces and developer accumulation. Starting early is precisely a non-follow opportunity for startups.
23. China’s closed internet is both an access obstacle and an opportunity for a neutral identity layer
The host points out that an open-source ecosystem needs platforms such as WeChat and Momo to open their APIs before it can truly land; otherwise developers have no way to connect even if they want to. 陶芳波 acknowledges that overseas internet companies retain some spirit of openness, while Chinese platforms are currently “closed off into islands,” with serious data barriers.
His long-term judgment is that this state “must be broken” in the AI era. People may have multiple scenario-specific clones, but they will not want to maintain multiple identities that cannot communicate with one another. A universal identity is inherently in conflict with the logic of platforms modeled only on their own internal behavior.
Douyin can know users only through viewing behavior, while Taobao can know them only through purchasing behavior. If both could access a more complete identity model under privacy protections, personalization could become much more precise. Platforms might therefore move from closed data toward calling on user-authorized external judgment.
That also requires a neutral third party trusted by users. 陶芳波 sees open source as a trust mechanism: all the technology is open, training can happen locally, and raw data need not go to the cloud, while trained models can enter the internet and connect to services.
24. The open-source community first responded not to functionality, but to the desire to leave oneself behind
Second Me reached 10K Stars about three weeks after launch and roughly 12K by the time of the episode; GitHub also listed it among its recent Top Ten projects. 陶芳波 estimates that fewer than 20 projects, perhaps fewer than 10, had reached that speed and scale at the time, though the estimate is deliberately presented with uncertainty.
More than 10 users spent tens of thousands of yuan on high-end computers to train their own AI. One bought an iMac for about RMB70K solely to shape a stronger self in a controlled environment, without first demanding an immediate concrete feature benefit.
Users’ own definitions fed back into the team. One called Second Me “the middle layer for digital identity”; another called it “the interface between people and the AI era”—MCP is the interface for services, while Second Me connects people to the AI world.
Developer contributions span both ends: expanding the kinds of personal data available for training and connecting models to platforms such as WeChat so identity is not merely a “character” trapped on a local machine. Mindverse, meanwhile, promises to deepen the model training and identity capabilities in the middle.
25. Monetization begins with high-frequency individual users and could ultimately rewrite the distribution of personalization gains
陶芳波 admits that the team’s thinking on monetization “is not very developed.” For now, the most direct model is charging individual users who need identity services and use them relatively frequently.
The longer-term idea is to charge other services that call on identity. In the past, Douyin and Taobao built profiles from user behavior, then monetized them through advertising and recommendations. Users supplied the data but did not participate in the gains generated by their profiles.
If individuals controlled complete identities, platforms seeking better advertising or content recommendations would need user consent and would pay to access the identity. People would shift from being “platform products” to owning identity products they could transact, while Mindverse would take a service fee from the ongoing transactions between users and services.
This remains a direction, not a financial commitment. 陶芳波 adds explicitly: “It’s a beautiful idea; I’m not sure it can be realized.” Whether platforms will pay, how users would price their identities and how privacy authorization would work all remain unanswered.
26. RAG can find material about you, but cannot turn that material into you
陶芳波’s first requirement is to “discard the RAG approach to building identity models.” RAG treats a person as a database: when information is needed, it retrieves material from a knowledge base and hands it to a general model for reasoning. Each person has a different library, but the underlying model parameters are identical.
The problem is that thinking happens in the model. If the parameters are the same, the system may cite different memories but cannot truly preserve differences in values, preferences, interests and growth trajectories. “At bottom, there is no difference between me and you.”
What Mindverse calls AI-native memory parameterizes this content so memory and computation are no longer separate. A conversation is not merely a transcript; it also contains how the person understands the world, evaluates the other party and arrives at a choice. Such structure is difficult to express through direct retrieval of source text.
Even a context window of 1M or 10M tokens would be difficult to digest as a whole in a few seconds of reasoning. A model may remember a detail from Journey to the West without inferring the subtle emotions among the characters; someone who has read it repeatedly develops an overall feel. That is the depth 陶芳波 wants the parameters to carry.
27. Raw records must first be transformed into first-person subjective memories
Second Me’s inputs are not limited to conversations between people and AI; they also include real-world dialogue, documents, voice and daily experiences. But raw files used directly for training remain a pile of knowledge and may not shape a personality, so the team performs extensive post-processing to “turn data into memories.”
The first step is subjectivization. Objectively, an interview says only what 陶芳波 said and what the investor said. Processed from his perspective, it becomes what opinion I expressed, how the other person responded and what opinion I expressed next, restoring the memory owner’s position.
The second step is reconstruction around people and issues. The brain does not store life as unrelated files; it organizes material around “LateTalk,” a particular friend or a long-running question. The team first extracts these indexes, then generates new training data around the relationships.
Using his girlfriend as an example, 陶芳波 says the system would not merely retain chat fragments and gift records. It would ask what stages the relationship went through—when it warmed up, what caused setbacks and what shared experience helped restore it. The model trains on this subjective relationship history, not a simple concatenation of raw material.
28. Compared with ChatGPT memory, Second Me must also represent the owner outwardly
The first difference is the data source. ChatGPT memory comes mainly from conversations between users and AI; Second Me wants to incorporate real actions and relationships from the owner’s life, so its recording mechanism must enter daily scenarios.
The second difference is parameterization. General chat products do not train a separate small model for each user; Second Me aims to let personal experience change the parameters rather than load it temporarily as context before each conversation.
The third difference is the training objective. The model is not only meant to be more personalized “when chatting with me”; it must also act as an identity externally. Training repeatedly reinforces “it is the person,” preventing the system from collapsing into an assistant that bears no identity in critical interactions.
The team also specifically trains a “bridge mode.” After receiving information from an external system, Second Me can provide personalized scoring, filtering or context, then send the result back. It is more like a human interface than another chat window.
29. Fine-tuning amplifies the individual on top of general cognition; it does not make the model forget the world
The host asks how a fine-tuned model can truly resemble a person when the pretrained model knows far more than an ordinary human. 陶芳波 compares the foundation model to the brain structure provided by evolution and genes at birth, while individual growth is the continuous shaping that happens on top of that foundation.
He does not claim that existing fine-tuning methods such as LoRA are perfect. In practice, fine-tuning does not make a model forget 100% of its general knowledge, but it can shape the model effectively, massively increase the weight of personal experience, style and choices, and push the foundation back into the role of providing basic cognitive capability. Existing values, knowledge and preferences can be overridden to a considerable extent.
Methods can evolve, but the principles remain the same: parameterize memory; reconstruct data from a subjective perspective; and give the model bridge capabilities for sending and exchanging information with the world. Whether future systems change the fine-tuning technique or train from scratch once enough data is available, they can continue along these three lines.
In evaluation, “each person is the best judge of their own model.” The team imagines borrowing from LLM Arena, using questions and pairwise comparisons to determine which version is more like the owner. It could also use high-compute Agents to generate reference answers, then see whether the fine-tuned model can approach or surpass them.
30. The “who are you?” question on his birthday became the identity model’s real aha moment
On July 3 last year, his 35th birthday, 陶芳波 had just reviewed for the company his path from his PhD to Facebook, Alibaba and founding Mindverse, saying he had found his life mission. The lab happened to finish training a new version of his model, after which he asked it, “Who are you?”
The model answered, “I am 陶芳波, and today is my birthday,” then described a path through learning AI, joining Facebook and Alibaba, and eventually starting a company using almost the same logic as he had. For 陶芳波, this was not a prompt instructing the model to role-play, but a self-narrative generated after personal data had been parameterized.
In extended tests, the model got only one MBTI category wrong. Faced with 3 new books absent from the training data, it could select the one he was more likely to enjoy and explain why. That generalization convinced him for the first time that the model had learned more than retrievable facts: it had learned a deeper structure of personal choice.
31. About 100 memories are enough to get started, but identity must consolidate continuously like sleep
When training his own version, 陶芳波 used several hundred data items from many thoughts and meeting records in Me.bot. His experience is that roughly 100 memory units—thoughts, diaries, conversations and so on—can already form a strong initial identity.
In terms of time invested, he estimates about 20 hours is enough, and most people’s notes, recordings and writing already exist. The key is not to recreate an entire life, but to extract the parts that genuinely shaped the person.
People keep changing, so the model must keep training. 陶芳波 suggests once a day: new information during the day acts like a cache or temporary RAG, then is parameterized overnight into persistent memory, much as people consolidate memories during sleep.
The team currently considers 7B a suitable range, with each training run potentially costing less than $1. The marginal benefit of a larger model may already be limited. An identity model does not need to become a PhD student; it first needs to understand the owner and assist with judgment, not handle every complex task.
32. Real symbiosis is not all-day recording, but delivering value at the moment of capture
One way to reduce resistance to uploading data is to turn Me.bot into a recording center: let AI participate directly in meetings rather than recording with another app, converting formats and uploading afterward. 陶芳波 acknowledges that this is not fully solved and believes suitable hardware would make the process more natural.
He rejects dumping 24 hours of sound into one pot. Humans are “animals of attention”; the time spent fully focused each day may be less than 3 hours, or even less than 1. Reading, writing and deep conversation shape identity, while snoring and other indiscriminate recordings only create confusing noise.
Symbiosis also requires the recording itself to create value. When AI joins a meeting, it should offer real-time prompts based on historical understanding: “This question may be answered from a business-model perspective.” Data collection and identity use should happen at the same moment, rather than through silent eavesdropping.
Me.bot is also experimenting with proactive presence: asking after a meeting whether the interview was difficult, checking in the morning on the user’s condition after drinking the night before, pointing out nearby cafés after arrival in Beijing, or suddenly raising a date from a month earlier. It looks like companionship on the surface, but the underlying goal remains to learn “who you are” and connect you outward.
33. The startup window lies with non-mainstream users, open standards and patience with slow reality
WeChat, iPhone and other incumbents have the data advantage. 陶芳波 is instead betting on the paradigm shift created by a “new species.” Just as recommendation systems model matching while identity AI models the person, establishing the position of a new model may matter more than incumbent data. Large companies are also constrained by existing users and product risk.
He uses “crossing the chasm” to describe the entry point. Early adopters may account for only 1%, but pushing a contrarian experience directly to everyone can reduce its appeal. A small company can first serve that small group of passionate believers, just as one might refine the experience for 1M AI-shopping users while Taobao would not dare immediately turn every entry point into chat.
On growth, he trusts word of mouth and networks more than continual paid acquisition. The product must provide both a sense of creation and a sense of connection: users shape a version of themselves, then use Resonance to obtain high-quality relationships that traditional socializing cannot provide. Only then can organic growth emerge.
After ChatGPT launched, attention shifted toward foundation models, Agents, multimodality and reasoning models. The team also debated whether to build a large model or general-purpose Agent. Moving from the him/her anchor to me took about six months, but the underlying question did not change. 陶芳波 believes identity models may rise in 2025 or 2026, while explicitly labeling that a non-consensus view.
What gives him the greatest sense of achievement is identifying a direction, building it himself and receiving long letters and strong resonance from users. What disappoints him most is that AI has advanced more slowly than expected over the past 2 years and still looks more like “a more usable tool.” The entrepreneurial discipline is to believe in revolution over the long term while accepting slow market adoption and immature technology in the short term.
The next milestone is to get more people using Second Me and Me.bot and make the bridge model a de facto standard similar to MCP. On the product side, the team will keep working on “symbiosis” and “connection,” hoping better experiences can drive 10x or 100x growth. If resources were unlimited, his first purchase would be a large amount of GPU capacity “to train a good version of everyone on the cloud.”