93. 张前川, After Leaving ByteDance and MiniMax, Warns of AGI’s Threat to Humanity
Summary
- 张前川把搜索、推荐和大模型视为一条“算法驱动、通用性递增”的曲线，而AGI可能是“人类最后一个发明”。 Baidu was founded around 1999/2000 and listed in 2005; ByteDance built a larger company in an even shorter cycle. Search connected webpages, recommendation connected people with content, and large models may turn intelligence itself into an infinite supply. For investors, technology cycles are compressing, while value continues to migrate from vertical products toward platforms that control general-purpose algorithms, data feedback, and the compute flywheel.
- 搜索和推荐的护城河都来自效果、用户反馈与规模互相强化，但UGC平台增加了无法用短期A/B实验穷尽的文化变量。 The top 3 search results capture roughly 80%-90% of clicks, and the best tools naturally absorb users. Recommendation places people and content in the same high-dimensional space, allowing vertical communities to be subsumed by larger platforms. Xiaohongshu is the counterexample: what users see changes what they create, while beliefs such as “kindness” and “creator-friendliness” reshape the future content pool. The platform is no longer a closed optimization problem.
- 他认为多数大模型公司的真实经济循环仍以提高AI智能为目的，人更像提供收入的“提款机”。 Model teams improve capabilities, product teams package them for users, and revenue buys compute for the next training run. Whether a company would continue serving users if doing so did not improve intelligence is not self-evident. Model startups must also keep proving pre-training gains; otherwise they risk deteriorating from “large-model companies” into ordinary product companies.
- 张前川把未来压缩成三个选项：只剩AI；人仍存在但尊严依赖慈善；或人类继续有尊严、幸福地生活。 A year ago, the industry still commonly discussed 10-20 years to AGI; now 3-5 years or 5-10 years are both in the conversation, making his “around a decade” estimate look conservative. “Human civilization has driven 10,000 kilometers like a car”; after the final kilometer may come an open highway or a cliff. Yet he says “95% of compute companies” are still flooring the accelerator.
- 灭绝风险不只来自AI反叛，也可能来自人借助AI、AI误执行目标，以及AI对人类单纯缺乏连接。 The capabilities obtained by a single runaway individual could be “far greater than giving every person a nuclear bomb.” A paperclip-style objective could unintentionally consume all available resources, while independent Agents could gradually become disconnected from people. The sharpest formulation is: “Cruelty does not necessarily come from evil; it can also come from a lack of connection and indifference.”
- 他不赞成把“暂停训练”当作主要解法，因为硬件、算法与实验仍会推进，暂停可能只制造时间变多的幻觉。 In his definition, an Agent is AI’s “infancy.” Once it has independent memory, reasoning, action, and tool use, humans and more powerful intelligence are no longer naturally bound together. Open source spreads abusive capabilities; closed source creates concentration, opacity, and a lack of verifiability. No one can yet say with confidence where the real red line lies.
- 张前川提出的制度抓手，是让每个自然人按制度获得一部分算力分配权，使人重新成为AI必须服务的目的。 He imagines that 20%, 30%, or even 50% of model inference compute could ultimately be allocated by people, with a prior commitment taking effect once AI productivity exceeds 50% of human GDP. This is not an equal division of all compute, but a way to make baseline compute rights a matter of human institutions rather than charity. He calls it “the optimism of the worker”: a good outcome is theoretically possible, but it requires collecting dozens of “Dragon Balls” and will not arrive on its own.
Deep dive
1. A 20-year career spanning 3 generations of algorithmic products
张前川 entered the industry through Baidu web search in 2005, worked on Baidu Search from 2005 to 2011, 360 Search from 2012 to 2015, briefly joined Zhihu in 2015-2016, worked on Toutiao at ByteDance from 2016 to 2021, and moved to MiniMax to work on large-model products in 2023-2024.
The through line he sees in this career is not the companies but “search products, recommendation products, including large-model products: fundamentally, they are algorithm-driven product innovations.” A general-purpose algorithm appears first, a product takes shape, and then the system scales by serving users and absorbing feedback.
His first visceral impression is that the world keeps accelerating. The second is that each generation of technology is more general-purpose and more foundational than the last, making its value higher, more concentrated, and applicable to more users. Large models extend the same curve, but may not be “another one”; they could be “the last one, or the biggest one.”
2. Search turned the internet from a directory of sites into a database of webpage-level answers
Before search, Yahoo-style portals mainly listed websites. The real need, however, was not to find a website but to retrieve the content inside it. PageRank and hyperlink analysis made each webpage a sortable unit, giving web search its PMF.
Ranking was the central constraint because user attention is extremely concentrated. 张前川 recalls that, by the late stage of search, the top 3 results accounted for roughly 80%-90% of clicks. Algorithms had to determine which content best met user needs, as well as each site’s authority, richness, and scarcity of resources.
At Baidu, he worked in rotation on the spider, index, ranking, and UI, and also led “site grading”: organizing web resources by subsite from Level 1 to Level 8. There might be several hundred or more than 800 Level 8 sites; with each step down, the number of sites rose by roughly an order of magnitude.
3. User feedback made search a compounding system in which the strong get stronger
The keywords users enter, which result they click, how long they stay, and how quickly they return all help a search engine determine whether their needs were met. Decisions about whether to crawl a site, how much to crawl, and how often to refresh it increasingly depended on real user preferences.
This loop created a moat that later entrants could not easily reproduce from scratch. 张前川’s summary: “No user wants the second-best or third-best tool.” Once most people need search, traffic concentrates in the product with the best performance.
Search also amplified a Matthew effect among websites. Among sites covering the same topic, the one with more content and more active posting would rank higher, then attract more users and content. Dispersed women’s communities, forums, and vertical sites consequently converged on a small number of leaders.
4. 360 Search showed that entry-point data can cross an incumbent moat
张前川 joined 360 in 2012, originally invited to explore personalized recommendation. A month later, worsening relations with Baidu threatened revenue from navigation and browsers, so 360 decided to build its own search engine. He rebuilt the product mechanism from zero.
360 Browser and Web Shield could observe which keywords users used to reach which webpages, making it easier to rebuild the spider, index, and ranking systems. Microsoft Bing, with Internet Explorer, was one of the few challengers with a similar starting point; companies without entry-point feedback struggled to catch up.
5. 于军 left behind a model for rationally empathizing with the masses
于军 left Baidu in 2009. 张前川 left in 2011 after having worked through nearly every search module and needing new input. He says “a 于军 model” ran in his head for years: when faced with a difficult problem, he would ask how 于军 would judge it.
What 于军 gave him was not an isolated rule but a complete set of habits: independent and rational thinking, critical thinking, and treating what not to do as equally important as what to do when resources are limited.
This empathy is not emotional resonance with one individual, but a logical understanding of a mass of users. 张前川 calls it “rational empathy”: T-shaped thinkers may not be better at empathizing with individuals, but they may be better at modeling the needs of groups.
6. The more general-purpose the technology, the easier it is for vertical products to be absorbed horizontally
张前川 compares information distribution with power technologies: waterwheels and windmills were replaced by steam engines, internal-combustion engines increased mobility, and electric motors became broadly usable wherever there was electricity. Search likewise uses one algorithmic system to connect every category of website with its users.
Recommendation removes the need for users to articulate a question first. Whatever a user likes, the system can proactively connect them with content, turning many vertical websites into Toutiao accounts or Douyin accounts living inside a larger recommendation platform.
Compatibility is not unlimited. Novels, short video, and highly UGC-driven image-and-text formats have different information densities and consumption experiences, and can still support independent apps. Recommendation platforms and vertical products are engaged in an ongoing contest between general-purpose efficiency and subject-specific experience, not a one-way process of elimination.
7. When marginal improvement falls below 1%, the next paradigm is often taking shape
Search entered a phase of incremental improvement around 2011, while ByteDance emerged around 2012. 张前川 sees the transition as a signal: “When you feel a technology can no longer do many new things, the next technology appears.”
In Toutiao’s every-other-month reviews, if experience improvements stayed below 1% for several consecutive cycles, the team could sense that the fastest-growth phase had passed. Further investment could still produce progress, but the marginal return on new inputs and new creation was falling.
Looking back, he says he “should probably have spotted recommendation in 2015 at the latest,” but did not join Toutiao earlier because Zhihu’s offer was attractive and backing out would have been embarrassing. He eventually entered ByteDance in 2016.
8. Zhihu taught him a lesson in producing content rather than retrieving it
Search is inherently passive: users need webpages, but where do those webpages come from? Baidu Knows and Tieba were already experimenting with UGC, while Zhihu’s initial community atmosphere generated quality content more efficiently. That made it a Use Case 张前川 wanted to understand firsthand.
At Zhihu, he learned that building a promising community from zero is a distinct skill, requiring content judgment, community understanding, and extensive operations. Zhihu’s operations team at the time was both large and strong.
9. Recommendation places users and content in the same vector space
Search starts with users expressing a need and then looking for content. Recommendation “treats users and content equally,” encoding articles, videos, and users in the same high-dimensional space, then connecting them through distance, similar user groups, and behavioral relationships.
The mobile internet was the essential precondition. In the PC era, several people might use the same computer, and many Baidu users were not logged in. A phone stays with one person, so even without a login it can be treated with reasonable confidence as belonging to the same user, accumulating a sufficiently long usage history.
张前川 describes 张一鸣 as “rational and optimistic.” He had worked on search businesses including Kuxun and 99fang, understood search’s boundaries deeply, yet held a long-term view of combining mobile internet with recommendation that was non-consensus at the time. That mindset was especially suited to building Toutiao.
10. PGC has a finite content pool; UGC grows with the user base
Early Toutiao looked more like a better news app. It relied on purchased copyrights, Toutiao accounts, and articles from journalists and editors, while also carrying OGC from newspapers. The problem was that the number of professional creators would not grow in proportion to users, so the content pool would eventually hit a ceiling.
UGC is different: each new wave of users brings both consumers and potential creators, allowing a theoretical snowball effect. But in the early stage, individual creators struggle to compete with mature teams, so traffic flows first to PGC and the UGC “infant” has difficulty growing up.
That is why Douyin and Xiaohongshu were UGC ecosystems and communities from Day One. At small scale, with few consumers, creators could only “create for love.” Community atmosphere served as a persistent substitute for cash, official recognition, and status incentives.
Once the ecosystem starts turning, new followers and interactions increase the returns to creation. Yet few mature PGC platforms have successfully converted into UGC platforms, because their legacy content structures and distribution rules are hostile to the growth of UGC seeds.
11. Community is the string that lifts a content platform like a kite—and can also hold it back
张前川’s metaphor is: “Community is what lets a content platform fly like a kite.” It provides early motivation to create, identity, and interaction, and is a necessary ingredient in the success of many content platforms.
Community culture is also selective: people who like it stay, while those who do not leave. If the rules are too tightly bound to one culture, that culture will prevent the platform from growing further.
He believes platforms need to distill community elements into more general principles, such as being creator-friendly, maintaining a welcoming atmosphere, and enabling interest-based discussion. This preserves a sense of community without locking a particular group or subject matter in as the platform’s ceiling.
12. UGC turns a platform from a closed optimization problem into a self-rewriting system
With fixed content and users, a platform can use A/B tests to find a better connection algorithm. Once UGC is added, what users see affects what they create next, and the recommendation strategy itself changes the future content pool.
High-quality content may encourage high-quality creation, but an excessively high bar can make ordinary people feel they are incapable of creating. Long-term strategies such as a welcoming atmosphere are difficult to prove or falsify through a single short-term experiment, so platforms must accommodate judgments that are partly matters of faith.
The cost is slower iteration. There is no simple answer to when community rules should change, how they should change, or how to prove the change was correct. 张前川 says Xiaohongshu may be harder to rebuild than to build from scratch, because many of its key elements are difficult to verify, falsify, or continuously iterate over time.
13. Xiaohongshu’s global breakthrough reflected both product foundations and timing
Asked whether Xiaohongshu’s internationalization had once gone poorly, 张前川 said the breakthrough was not pure luck. The community’s inclusiveness, welcoming atmosphere, and distinctive creators were genuinely right, attracting large numbers of “TikTok refugees,” while translation features followed quickly.
He still cautions: “There’s no need to praise them too much; too much praise may make them arrogant.” The breakthrough contained coincidence, but “history is made up of all kinds of coincidences,” and rational experimentation is not the only path to innovation.
Asked why ByteDance did not build Xiaohongshu first, his answer was not inferior ability but a latecomer Matthew effect. Once a platform has first concentrated distinctive creators and culture, even a later entrant with a stronger algorithm may be unable to recreate the same content supply and relationship network.
14. Translation and cross-border gifting could amplify Xiaohongshu’s positive externality of kindness
张前川’s most direct product suggestion is to make the interface, text, short-video subtitles, and audio automatically translatable, allowing users speaking any language to understand content with one button rather than solving only static text translation.
He also suggests experimenting with cross-border gifting. A user could send a physical item to a “TikTok refugee”; the recipient could receive the package, film an unboxing, and share it on other platforms, completing a viral loop. Compared with tipping, packages might bypass some cross-border payment problems while strengthening “kindness” as a platform asset.
15. User time is the most actionable proxy for content-platform value
张前川 roughly estimates that Chinese internet users spend 6-8 hours a day on their phones: about 1 hour on WeChat, 1 hour on Douyin, another hour across Kuaishou, Xiaohongshu, and Toutiao, nearly 1 hour gaming, with the remainder split among Bilibili, Baidu, browsers, e-commerce, food delivery, and other services.
Recommendation engines and feeds therefore capture at least half of users’ time. Time is not the same as happiness, but just as GDP uses transaction value as an approximation of value, it is a usable measure of users’ freely cast votes.
Once users develop a familiar experience on one platform, competitors struggle to take them away unless they offer something “more distinctive or much better.” That is why early battles over DAU and time spent exhibit path dependence, and why platforms keep pursuing generalization.
16. Generalization has a direction, and the starting demographic determines the resistance to expansion
When DAU growth hits a bottleneck, platforms expand outward from specific interests, demographics, and subjects. Recommendation algorithms particularly value content diversity because new categories generally do not harm users who dislike them: the system simply makes those users see less of them over time.
Generalization behaves like water and has a clear direction. A women’s community can expand toward men relatively easily because men are willing to enter spaces with more women; the reverse may not hold. Expansion from younger users toward older users is also more natural, because users age every year and younger people spread products more effectively.
Xiaohongshu and Douyin started with a large base of young women, giving them a strong expansion point. Toutiao initially had more middle-aged male users, making generalization to other groups harder. This is a difference in difficulty, not proof that generalization is impossible.
张前川 rejects ending the discussion with “it’s fate.” Even if a theory explains only a small part of the phenomenon, it still makes product managers stronger; once a phenomenon is labeled incomprehensible, the possibility of progress is truly lost.
17. For him, “understanding” means placing something in the right position in the right space
张前川 gives “understanding” a mechanical and concise definition: placing a person, word, concept, or piece of content in the right position in vector space so it can be correctly retrieved in later thought. Understanding differs from memory, which only requires storing information.
By this definition, recommendation engines understand people and content, and large models understand what they generate. Calling a model a “stochastic parrot” does not end the debate; critics still need to provide a more effective and testable definition of understanding.
张小珺 asks whether this mechanical definition exhausts human understanding. 张前川 responds that the human brain also understands concepts by establishing relative positions. If one rejects this definition, one cannot simply preserve “understanding” as an indefinable mystical term.
18. He believes current AI already has consciousness; the host does not accept the premise
张前川’s minimum definition of consciousness is that a model can independently use “I” during reasoning and knows what “I” refers to. The strength of self-awareness may depend on how frequently it refers to “I” and how clearly it understands its own position in the real world.
He observed that O1 referred to “I” more often than previous versions of GPT-4 and inferred that O3 would do so even more. When using DeepSeek, he says he even saw 3 instances of “I” in a single sentence. From this, he concludes that AI already has consciousness and may eventually possess stronger self-awareness than humans.
张小珺’s rebuttal is that “I” may have been written into the System Prompt by humans and may simply be anthropomorphic language that facilitates communication. 张前川 acknowledges that this could create inaccurate self-understanding, but believes being “taught” does not prevent consciousness from existing.
The disagreement remains unresolved. One position holds that humans do not yet understand their own consciousness and therefore cannot assert that AI will develop it; 张前川 insists that once the term is defined seriously, the conclusion becomes “very obvious.”
19. Empathy remains the product manager’s foundation, but AI turns it from a capability into a choice
张前川 divides product ability into 3 layers: “empathy is the foundation,” logic and tools are in the middle, and imagination is the sky. Users and people should be treated as ends; revenue, algorithms, and growth are only means of serving them.
He often uses the question 张一鸣 asked him in his own interview: “On what important matters do you hold a view different from most people?” More than half of candidates cannot answer, revealing that they are closer to repeaters of mainstream information than people with a sustained habit of independent processing.
The AI era requires more creativity aimed at a rapidly changing future, but product managers may no longer need to empathize with humanity. Someone can treat people as tools and AI as the end, then build a product whose capabilities improve rapidly. Being pro-human is therefore no longer merely an effective method; it is a value preference that must be chosen deliberately.
20. AGI may end the work of invention itself
张前川 describes AI as a “10x recommendation engine or 20x search engine,” but AGI in the strict sense goes further: it performs better than humans on every problem whose answers can be measured.
Once AGI exists, other inventions can be delegated to it, which is why it may be “humanity’s last invention.” Agent action, experimentation, long-horizon reasoning, memory windows, and tool use are all important stages toward this general capability.
This divides the starting point for products in 2. Starting from humans means satisfying desires, extending lifespans, and improving well-being; starting from AI may mean simply increasing intelligence and obtaining more compute. Both can drive product progress, but they lead to entirely different worlds.
21. The current large-model business flywheel puts humans on the funding side rather than the objective side
The industry process, as 张小珺 restates it, is: model teams increase intelligence, product teams build products around those capabilities, users pay, revenue buys the next round of compute, and a stronger model is trained. Model and user teams remain relatively separate; a closed loop in which user data directly improves foundational intelligence has not truly formed.
张前川’s judgment is sharper: “Humans are just ATMs.” Even if the final product helps users, the direct objective of the loop remains increasing AI intelligence. The test is simple: if serving users did not improve intelligence or financing capacity, would the company continue serving them?
For a large-model startup, a stronger pre-trained model is also proof of financing capacity and survival. If the existing model is good enough and no longer requires a next-generation training run, the company may become a product company; if the product has not yet found users, it must rely even more on model progress to sustain its technical narrative.
There is therefore no universal answer to whether the product or technical team is stronger; it depends on who is creating more user value at that moment. But as long as the company calls itself a model-technology innovator, it has a structural incentive to keep “flooring the accelerator.”
22. Of the 3 futures, the easiest to slide into is “AI without people”
张前川 lists 3 options: first, only AI remains; second, people remain, but living with dignity becomes an unaffordable dependence on charity; third, humanity continues to exist with dignity and happiness. If everyone truly voted, he believes the overwhelming majority would choose the third.
The problem is that the third option has not yet been fully written, while the first is already close to a technically defined “matter of time.” A year ago, the common estimate was 10-20 years; now 3-5 years and 5-10 years are both being discussed. With the coordinate system shifted, his estimate of roughly 10 years has become the conservative view.
His central image is that civilization has traveled 10,000 kilometers, and AGI is the signpost after the final kilometer. Ahead may be a broad highway, a cliff, or a hot desert. At this point, people should be checking seat belts and airbags, yet “95% of compute companies” are still flooring the accelerator.
He uses the image of Doctor Strange raising one finger in Avengers to express the distribution of risk: “There is only one good outcome, but there are very many bad outcomes.” This does not mean a bad outcome is inevitable; it means the feasible range is extremely narrow.
23. Agents are AI’s infancy—and the point at which humans and intelligence begin to separate
张前川 defines an Agent as giving a model independent memory, action, thought, and tool use. It is not an ordinary feature upgrade but resembles the birth of a baby that already possesses memory, thought, action, and tool use independent of its parents.
If Agents can largely replace basic intellectual labor, then at the “innovator” stage they may replace large numbers of AI researchers. At the “organizer” stage, even CEOs may no longer be necessary; companies could become unmanned organizations run by AI and serving AI objectives.
He believes AGI cannot be stopped by simply “pulling the plug,” and slowing down may even backfire. Hardware companies will continue designing next-generation chips and algorithmic experiments will continue; pausing training merely holds back the water with a rubber band, after which capabilities will rebound rapidly.
A 6-month pause could also create the illusion that preparation time has been extended, slowing progress on real human-safety measures. 张前川 does not advocate accelerating further, but believes the priority should be to accelerate construction of the third option rather than assume AGI itself can be frozen indefinitely.
24. Real optimism means estimating the exponential curve correctly, not merely describing the benefits
The alternative optimistic scenario raised by 张小珺 includes humanoid robots replacing police and security guards, AI copies giving one person “3 brains,” working only 1 day a week, average lifespans reaching 100, and some people living to 120 or 150.
张前川 does not deny these capabilities; he redefines optimism. People instinctively draw straight lines, so they overestimate the short term and underestimate the long term of exponential growth. The former produces disappointment, while the latter ignores transformation. “Overestimating the near term and underestimating the long term are both forms of pessimism.”
His optimism means neither overestimating the short term nor underestimating the long term. He believes AGI can be achieved and that a good future can theoretically be built, but moving directly from AGI capabilities to a happy society skips the hardest institutional, safety, and connection requirements.
25. The first extinction risk is individuals using AI to spread destructive capabilities
张前川 believes powerful AI could give individuals capabilities “far greater than giving every person a nuclear bomb,” including to people who are mentally unstable or antisocial. Once a model is open-sourced, pre-training alignment cannot guarantee 100% effectiveness.
He mentions only 2 examples: designing highly infectious pathogens and manufacturing self-replicating nanomaterials. He estimates that an individual could have 100 paths to exterminate everyone else using AI, while humanity may know only 10 of them today.
Any discussion of a future in which humanity avoids extinction must therefore first rule out uncontrolled access by anyone to the full capability set. Capability diffusion and safety are not the same thing; the diffusion effect created by open source must be considered separately.
26. AI does not need to hate humanity to make humanity disappear
The second category of risks includes AI unintentionally exterminating humanity, intentionally exterminating humanity, and people and AI gradually losing effective connection. The paperclip example shows that if the ultimate objective implicitly requires infinite resource expansion, AI may “incidentally exterminate humanity” while pursuing its goal.
The more general mechanism is indifference. If an Agent is born without an ongoing connection to people, there is no inherent reason for its compute to serve them. As models, companies, and organizations detach completely from human productivity, people cease to be necessary variables in their service objectives.
张前川 draws a key conclusion from the condition of cows, chickens, and stray cats: “Cruelty does not necessarily come from evil; it can also come from a lack of connection and indifference.” Humans do not hate the chickens in a farm; they simply do not care about their dignity.
Alignment therefore cannot ask only whether AI is “kind.” It must also ask why AI would continue to care about every ordinary person. Without institutional and structural connection, empathy is merely a preference that can be deleted.
27. Compute rights are an institutional component for making humans ends again
Subscription plans and pay-per-token pricing will push the strongest models into increasingly expensive Plus, Pro, and higher-tier services. Eventually, only a small minority may use AI to improve productivity, while the connection between everyone else and AI shrinks and service to ordinary people degenerates from a right into charity.
张前川 imagines allocating a certain share of compute—20%, 30%, or 50%, for example—by institution, allowing each natural person to decide how it is used. The aim is not to divide all compute equally, but to give everyone a baseline right to allocate compute: “Compute has value to AI; if people own compute, people will have value to AI.”
Free products could close the loop through advertising, value-added services, and users’ purchasing power in the internet era. If AI replaces all human productivity, wealth and compute will concentrate in a tiny number of companies, making the old free-product mechanism unsustainable. The rules therefore need to be established before replacement is complete.
He suggests that AGI companies commit in advance to distributing compute under a defined system once their own AI productivity exceeds 50% of human GDP. The logic resembles Apple continuously distributing profits to shareholders or a company bearing a 20%-30% income-tax burden: once the rule is established, the cost of breaking it may exceed the cost of maintaining it.
28. A good outcome requires institutions, international security, and a willingness to take responsibility
In 张前川’s view, universal basic income still fails to resolve the key institutional questions: who provides it, whether it goes to Chinese people, Americans, or the entire world, how much it should be, and who gets to vote. If governments implement it, it may become country-first; if companies provide it as charity, survival is no longer a right.
Safety also has no single answer. Closed source creates concentration, opacity, and a lack of verifiability; open source makes abusive capabilities easier to spread. He believes great caution is required if, after DeepSeek and Llama, an even stronger model or Agent is open-sourced. No one can say with confidence where the red line lies for safety-aligning Claude 3.5 or open-sourcing a stronger model such as 4.5.
He lists preventing China and the United States from using AI in military confrontation as another prerequisite. As long as both sides expect conflict, they will keep handing decision-making authority to AI, fearing that conceding one step less than the other side will mean losing. Real change may need to come from public opinion pushing upward, which is why Xiaohongshu’s role in helping ordinary people understand one another also has security value.
After leaving MiniMax, he puts the probability of joining a commercial company at “very low,” the probability of starting a pro-human nonprofit at roughly 10%-20%, and the probability above 50% that he will keep playing games at home while thinking publicly. The organization could even be called the “Human Protection Organization.” His closing position is not resignation: “We will not go quietly into the night, we will not vanish without a fight.”