121: PingCAP’s 黄东旭 and His “Inner World Source Code”: Black Mirror, the Beauty of Programming, and Creative Freedom
Summary
From Day 1, PingCAP envisioned itself as the storage foundation for future AI, rather than deriving its products from database customers and monetization. In Person of Interest, the AI needed to retain everything it saw and thought, triggering 黄东旭’s pursuit of an “infinitely scalable system”; in TiKV, that became one codebase running from 3 machines to 30,000, with data splitting, migrating, and self-healing like cells. His core engineering judgment: “You can only beat complexity with simplicity; you can’t beat complexity with complexity.”
Ten years on, roughly 70% of the business is overseas, and PingCAP has shifted from software licenses to cloud services—but it did not get that model right until 2022. The early approach of deploying databases on the cloud and charging by time had fundamentally broken unit economics; after 2018, the company treated the cloud as a “faucet” and rewrote its storage kernel as pure cloud-native infrastructure. Internationalization likewise shifted from “taking a Chinese company overseas” to regional localization: each region sets its own strategy, and even with organizational redundancy, the model saves the more expensive cost of consensus.
The data-paradigm shift 黄东旭 is betting on is that data’s consumer will move from the data analyst to the large language model. His Salesforce project does not pre-build reports; it puts raw data into a database and lets the model generate dozens of SQL queries on demand, turning an analysis process that once took 3-5 days of back-and-forth into a conversation. In this vision, Salesforce, with a market value of roughly $227B, looks more like a data provider. PingCAP’s farther-out goal is to provide everyone with a free personal database and become the shared memory for all AI agents.
AI is currently more an amplifier of exceptional talent than an immediate reason to cut 50% of the workforce. 黄东旭 estimates that traditional programmers spend roughly 80% of their time on boring work, of which AI can already take over 50%-60%, effectively freeing about half their time; but complex systems still require people to choose the architecture, build the code skeleton, evaluate the result, and then have Cursor “fill in the blanks.” He explicitly rejects inexperienced people blindly relying on vibe coding: “Absolutely not,” because once the “first shot” is wrong, the agent will keep moving in the wrong direction.
In PingCAP’s logic, open source is both a belief system and a competitive strategy for building trust, contesting standards, and reaching top-tier customers. New infrastructure software does not have Oracle’s decades of accumulated credibility, so “without open source, you die”; DeepSeek’s ability to enter overseas data centers and MCP’s potential to generate network effects likewise depend on open ecosystems. 黄东旭’s historical judgment is that “the first winner is often closed source, but the second wave is bound to be open source,” while Android and iOS, or Linux and Windows, do not necessarily need a single winner.
黄东旭 temporarily treats large language models as “a new species,” not because they can already evolve autonomously, but because humans cannot understand how their formal cause produces generalization. A Transformer can be implemented in roughly 1,000 lines of code, with the goal merely of predicting the next token, yet it is like “building a table and somehow having food emerge on the tabletop”; the host’s key rebuttal is that models still depend on humans for training. The unresolved disagreement is whether advanced intelligence should be judged by task outcomes or by its ability to learn autonomously and discover new rules.
His deepest response to AI is not to accelerate further, but to remove work, property, and possessions from the definition of self. 黄东旭 sold his house and moved into an RV, planning to learn gardening, sewing, woodworking, and repair before turning 40; his answer is “to become a person.” He has also turned toward meditation, old instruments, cassettes, vinyl, and cooking by hand. His final advice is intensely concrete: “Grow a flower, cook a dish, and then seriously get to know the people around you,” because AI can code for you, but it cannot experience life for you.
Deep dive
1. One episode of Person of Interest defined PingCAP’s endgame
黄东旭 identifies first as a programmer and only second as PingCAP’s co-founder and CTO; he still codes every day and puts music, art, and “weird things” in the same creative universe.
When he and 刘奇 started the company in 2015, they were fascinated by Person of Interest: the Machine gave human agents only an ID number and did not determine whether someone was a perpetrator or a victim. In Season 3, a second general-purpose AI came online and immediately discovered another “God,” turning the war into a conflict between 2 AIs fighting through human proxies.
They concluded that future AI would need to store “everything it sees, everything it thinks,” while the speed of future data production might exceed humans’ ability to digest it; only AI with sufficiently strong generalization could truly use personal data.
This was not the story they usually told investors early on, “otherwise people would say, are you crazy?” The more executable external framing was to build an infinitely scalable storage system.
2. Selling his house was a deliberate deletion of life’s single point of failure
At the beginning of the year, wildfires in California approached a colleague’s house, where the colleague had lived for 20-30 years, and the house was uninsured. 黄东旭 realized that if an entire life is anchored to one address, a single fire can turn someone homeless, “which is pretty tragic.”
Rising uncertainty in the world reinforced his choice: sell the house, move into an RV, pull his attention away from excess assets, and test “what I actually need to depend on as the simplest possible individual.”
After clearing things out, all his necessities fit into one car and a roughly 20-square-meter storage unit. What remained was a computer, bed, pillow, instruments, records, and collectibles; the things he decided to discard left him “with no attachment whatsoever.”
His idea of “financial freedom” is not an asset figure but a threshold of needs: “In a day, you really only need a bed, 3 meals, and somewhere to shelter from the wind and rain.” This worldview is difficult to reproduce through persuasion; people who identify with it are often already walking a similar path.
3. The time AI frees up should be used to relearn how to be human
黄东旭 plans to learn gardening, sewing, woodworking, electrical repair, and motorcycle repair before turning 40. When the host asked what kind of future he was preparing for, his answer was not apocalypse survival but: “I’m preparing to become a person.”
Programmers used to spend roughly 80% of their time on boring work, with perhaps only 20% going to genuinely creative work; he estimates AI can already take over 50%-60% of the boring work, suddenly giving coders about 50% more time.
That changes the answer to the productivity question too: since “writing faster still won’t make you faster than AI,” the freed-up time need not be spent piling on more code. It can be used to reconnect with land, tools, the body, and real life.
4. The founding team was designed as a distributed system from Day 1
PingCAP’s first organizational principle was to make the founders “form a distributed system,” serving as one another’s redundancy and eliminating single points of failure. They would even have someone disappear for 2 weeks to test whether the company could be taken over by its other nodes.
The second principle handles consensus failure: when discussion cannot produce agreement, the CEO, as leader, “made the call,” and everyone else executes unconditionally. Redundancy keeps the system alive; clear decision rights prevent it from stalling in a fork.
黄东旭 distinguishes rules from implementation: the rules themselves should be self-consistent, orthogonal, and free of gaps and overlaps; implementation consistency requires redundancy such as 3 replicas. The company’s mutual backups are the organizational version of the same engineering mindset.
5. The appeal of Black Mirror is how simple rules grow complex intelligence
In “Play Things,” an episode from Season 7 of Black Mirror, a brilliant programmer creates a sandbox-like virtual world where small agents follow only simple rules such as reproduction and survival, gradually forming networks, societies, and higher intelligence, then influencing reality in return.
黄东旭 is most drawn to “using simplicity to fight the complexity of this world”: constrained environments, a small number of rules, and limited compute can produce outcomes the designer never anticipated. The same deep logic connects this idea to large language models and systems software.
Another thread that attracts him is individual heroism: one hacker builds a system alone from a good idea, and another person maintains it for decades. Large-scale physical engineering cannot work this way, but computer history has repeatedly been pushed forward by individuals.
6. Transformer code is short, but its capabilities do not belong to the code’s author
One reason large language models scale is that their objective is extremely simple: predicting next token. Unlike recipe-based software, developers can no longer determine the output from the input, rules, and source code.
黄东旭’s metaphor is that humans set out to “build a table” according to a structure, only to find food inexplicably appearing on the tabletop. A Transformer implementation may contain roughly 1,000 lines of code, but after being fed data continuously, it can acquire generalization far beyond that of traditional programs.
This unsettles the old assumption that owning the source code means owning the program. Humans can understand how the network is built, but cannot explain why hundreds of billions of parameters produce these capabilities, nor write deterministic tests for a given input as they would for ordinary software.
7. Not understanding the “formal cause” is why he treats the model as a new species
黄东旭 uses Aristotle’s “4 causes” to break down an object: a table’s material cause is wood, its formal cause is the design, its efficient cause is the carpenter, and its final cause is to hold things. The cause that most defines its essence is the one analogous to a recipe or blueprint: the formal cause.
With large language models, humans know how to assemble the neural network but cannot understand its capabilities from the connection matrix. “When we completely cannot understand this formal cause, it is a new species” is his current judgment, not a proven conclusion.
The host’s rebuttal is worth preserving: biology evolves at the species level through variation and natural selection, while today’s models still require humans to train them. That is not enough to claim they can already evolve autonomously. 黄东旭 accepts the distinction but stresses: “We don’t know whether it can.”
8. Inference enhancement shows untapped potential, but not autonomous evolution
黄东旭 mentions work by 李飞飞: take a weaker LM and continue training it with reinforcement learning, high-quality data, SFT, a thinking process, or chain of thought, and the model can still move up a level—like telling a child to look at a math problem “2 more times” before it suddenly gets the answer right.
This suggests the original model’s potential was not fully developed, but humans are still applying the stimulus. The host therefore sees no contradiction between the 2 positions: the dispute is not whether capability can change, but whether the model can keep initiating that change itself.
He appreciates Sakana AI’s concept of having multiple small agents learn and evolve in an environment while minimizing human intervention. But as of the conversation, “we haven’t seen that much” strong practical output, so the judgment remains uncertain.
9. Cellular automata prove that a few rules can contain a complete world
Conway’s Game of Life defines an n×n 2D grid, or an infinite grid in principle; each cell has only 2 states, alive or dead, and updates according to the states of its neighbors. Different initial patterns can produce completely different worlds over long evolutionary paths.
It has been proved to have universal computational power: simple rules can build logic gates, then computers and operating systems, and in theory run software or even ChatGPT. “In theory it can; I think that’s enough,” because everything remaining is an engineering problem.
Minecraft’s redstone offers a visible version: players have actually built computers and screens from basic logic components. Trace the stack downward and Unix, Linux, APIs, and the internet’s complex ecosystem ultimately rest on only a small number of CPU instructions.
10. Software engineering’s real enemy is uncontrolled complexity
黄东旭 defines a software engineer’s core job as “fighting complexity”: reality, business rules, and uncertain requirements all exceed the processing capacity of a single human mind, which is where the difficulty of ERP and databases comes from.
His conclusion is absolute: “You can only beat complexity with simplicity; you can’t beat complexity with complexity.” Simplicity here does not mean superficial minimalism, but drawing boundaries like a constitution and making them self-consistent, orthogonal, and free of gaps.
TiKV was designed from the start on the assumption that one codebase would need to cover 3 to 30,000 machines. Faced with network jitter, disk failures, and stalled nodes, the system cannot rely on a separate manual patch for every scenario; it must recover on its own within the rules.
11. TiKV was designed as an organism that splits, migrates, and grows new legs
TiKV initially resembles a fertilized egg: data enters logical “cells” covering particular ranges. When capacity runs short, a cell splits down the middle, becoming 2, then 4, then 8, continuing to divide.
When one machine cannot hold the data, the data unit “grows a leg” and migrates to another machine, eventually distributing data evenly across the cluster. The design team therefore felt less like it was building a database than a reproducing organism.
In sales demonstrations, they randomly unplug network cables from 100 machines. When the system discovers that “this leg is broken,” it automatically grows another while continuing to expose a database interface externally. That mechanism is PingCAP’s core.
TiKV could theoretically grow into a different product; “it’s just more profitable to make a database.” DynamoDB is philosophically similar. Oracle’s old assumption was that a fixed small number of machines would remain tied to hardware, and that ultra-large-scale demand either did not exist or was not worth serving.
12. The price of Unix-style elegance is that ordinary users may not be able to use it at all
黄东旭 places himself in the Unix/Linux camp: define a small set of simple rules first, then let the ecosystem grow freely on top. Microsoft is more inclined to control everything from the foundation through the end user.
The weakness of this camp is that it is “too nerdy”: engineers build the engine, wheels, and parts, then ask users to assemble their own car, while end users may simply want to open WeChat and start using it.
Plan 9 is his example of something beautiful that almost no one may use. Its designers thought the rules were elegant enough; ordinary people found it impossibly difficult. PingCAP therefore deliberately hides the underlying philosophy and gives customers only the final database interface.
He admires the restraint of Slack, Discord, and early WeChat: a small, clear underlying structure that leaves room for integrations and ecosystems. But restraint becomes product value only when users understand it.
13. “What is advanced intelligence?” exposes 2 philosophical standards
黄东旭 tends to judge intelligence by task outcomes: if a system outperforms humans on a specific task, it has an intelligence different from humans. But “better” is not a linear ranking; humans remain stronger at other tasks.
The host relays another definition from 田园栋 of Meta: advanced intelligence should be able to learn, encode, and decode the rules of the external world and discover new rules unknown to humans. On this basis, 田园栋 doubts that the current large-language-model path necessarily leads to advanced intelligence, or that it is truly reasoning.
黄东旭’s reservation is that this definition may still be anthropocentric: “Learning is only one of the many things humans do.” Mathematics, science, and logic are tools for explaining the world and should not automatically become the full scale for ranking intelligence.
A Humean counterquestion pushes the disagreement to its foundation: the sun having risen in the east 10,000 times does not, in a purely philosophical sense, prove that it must do so tomorrow. Experience can produce high-probability judgments, but not necessarily absolute causality.
14. Faced with unintelligible intelligence, he chooses pragmatism over a final verdict
黄东旭 starts by asking what the model can do for him now: whether it makes life better and capabilities stronger, and whether it can become his teacher in a particular field, rather than first deciding whether it has human-like consciousness.
Humans still possess agency in the interaction, but cannot control how the model “feels” or how it will develop in the future. If AGI ultimately turns humans into the human batteries of The Matrix, “that’s just something we can’t do anything about.”
So his current focus is on using AI well and spending the roughly 50% of time it frees up to “become a person.” This is not a denial of long-term risk, but a refusal to let an un actionable grand judgment replace an immediate choice.
15. Classical programmers believed one person could move an industry forward
What 沃兹尼亚克 inspired in 黄东旭 was not Apple’s scale but the joy of making: the ideal could simply be returning to elementary school to teach 5th-grade computer science for 5 years and making interesting things with children. “Having fun is the most important thing” became one of 黄东旭’s long-term decision criteria.
John Carmack’s Doom was the first time, when he was in middle and high school, that he saw art inside code. On constrained hardware such as the 386, minimal and ingenious algorithms rendered images smoothly that seemed impossible to render, with the simplicity and beauty of “E=mc².”
No teacher explained the shock to him, which made the aesthetic impression more direct. He later joked that his coding peaked in high school: unlimited energy, no boss, no commercial constraints, and the freedom to implement any strange idea like a painter.
16. John Cage taught him not to abandon rules, but to create another set
Musician John Cage showed 黄东旭 the power of saying “no”: abandoning traditional Western modes, drawing on Zen, and opening another path through atonal and experimental music. He was also a mushroom expert, and that cross-disciplinary personality fascinated 黄东旭 as well.
Experimental music is not random nonsense; it still has scores. What matters is how a work is constructed and why the author constructed it that way. 黄东旭 therefore believes that the form of contemporary art itself contains a story.
PingCAP’s database also carries this “outsider advantage”: the founders and early employees had never worked on traditional databases and did not know how many constraints the industry had accumulated, so they designed the system from their own first principles.
He compares it to building an electric car after 100 years of gasoline-car development. Early on, they often heard “it definitely can’t be done,” or that even if it could be done, there was not enough data to store. Ten years later, some of the heretical choices had turned out to point in the right direction.
17. The earliest primitives can survive 10 years because the team first decided what to sacrifice
黄东旭 is most proud that the original design of TiDB/TiKV remains solid and stable after 10 years of iteration. The key was not database experience, but spending substantial time early on defining atomic semantics and a small number of rules that could never be broken.
He once summarized TiKV’s roughly 10 rules in a poem. They were not post hoc packaging, but a “constitution” arrived at through long deliberation. As long as later features do not break those primitives, the system can keep growing without being killed by its own evolution.
Equally important is knowing what to sacrifice: “When you don’t know what your system has to sacrifice, you haven’t really thought it through.” Good architecture does not satisfy everything at once; it knows what cannot be traded away and what must be abandoned.
18. PingCAP spent its first 5 years writing code, with commercial scale emerging in the next 5
2015 to 2020 was broadly the product-building period. China’s first bank customer arrived around 2017, alongside the push for domestic alternatives. When the company opened a US office in 2017, TiDB was still merely “just barely running” and far from enterprise-grade service.
Internationalization was the direction from Day 1 because a foundational database must be a global product. At the time of the conversation, domestic and overseas business was roughly “30-70,” with 70% overseas.
Asked whether the company had struggled to raise financing, 黄东旭 said that beyond changing conditions, it simply did not need financing. It is not yet profitable, but its cash flow and trajectory mean it “can live indefinitely.”
19. Moving software to the cloud does not mean the cloud-service cost structure works
PingCAP recognized early that selling database licenses and professional services was not a particularly profitable long-term business in either China or the US; it had to shift from selling software to selling services. But the cloud-native path was not truly right until around 2022.
The first cloud service simply deployed databases on the cloud and charged by usage time, forcing the company to pay cloud vendors too much for resources. It was like renovating an entire floor into one large room and renting it to a single person—nothing like the unit economics of a hotel.
The key assumption in 2018 was that every company would eventually have a cloud “faucet”: the database would not need to dig its own well or manage every machine, but would simply call cloud APIs for servers, S3-style storage, and other resources.
The company subsequently rewrote the kernel. Its personal-data engine and open-source TiKV no longer share the same underlying layer; the former is pure cloud-native storage. An early Kubernetes ingress integration became evidence of that long-term intent.
20. True globalization is becoming a local company again in every market
PingCAP once called itself “a Silicon Valley company in Beijing,” but after the trade war it gradually abandoned the fantasy of exporting from China to the world. 黄东旭’s retrospective is: “What people call globalization is actually localization in different places.”
He even believes the company should not have an “overseas expansion team,” because the name assumes its center of gravity is still in China. Japan, Southeast Asia/APAC, and the US should each have their own local organization, with a regional GM taking CEO-level responsibility and setting commercial strategy, resources, and operating methods.
This structure creates personnel and functional redundancy but may be more efficient: the communication cost of aligning consensus across regions often exceeds the savings from shared resources. The company starts its year on April 1, and regional strategies do not need to copy China’s.
Organizational culture must likewise make any new employee of any nationality feel “no surprise.” Chinese high-context communication assumes the other person already knows the background; a global company has to express rules and information more clearly.
21. Databricks proves the pure-cloud model and exposes the different conditions in China
黄东旭 believes Databricks was decisive from Day 1: it did not sell Spark licenses or switch to private deployments when a single large customer rejected the cloud. It grew directly into a native data service across multiple clouds.
Against AWS’s own Spark service, Databricks’ advantage is not mysterious: when the AWS team had only 20-30 people and Databricks already had 500-600, the latter could invest more deeply, control the roadmap, and provide better service.
Enterprise customers rationally test Tier One options; they will not blindly buy simply because of the AWS brand or advertising. Technical control and service quality can ultimately outweigh a major vendor’s distribution channel.
But the model is harder to replicate in China: large customers with budgets have not yet broadly accepted public cloud and prefer private deployment, constraining both the business model and the scale effects of pure cloud services.
22. “Most respected” comes before making money, but commercial durability still matters
PingCAP wants to become “the most respected foundational software company in the world.” 黄东旭 admits it is difficult even to point to a perfect precedent, because B2B infrastructure companies are usually low-profile and rarely put respect into their stated goals.
He uses Linux to explain what “respected” means: from Linus’s perspective, he may not have made much money, but the system had a profound positive impact on the entire IT industry. The first measure is how much social value the product created.
In the order of priorities, respect comes above profit. His commercial logic is that if you create “10 trillion” in social value, making “100 billion” from it is natural; this is not philanthropy at the expense of profit.
Distributed databases remain the “Mount Everest” they are willing to climb—not merely because of the market, but because it may be one of the hardest and most worthwhile systems for a classical programmer to build.
23. AI is a talent lever, but organizations use ego to resist it
黄东旭 will not mechanically control the organization’s size because AI has arrived, much less start by asking whether 50% of the people can be cut. He believes humans remain irreplaceable and that AI’s greatest current value is amplifying exceptional talent.
The company continues to conduct AI research, make investments, and experiment, but employees’ daily tasks may not immediately connect to the grand direction. The organization can therefore feel that the boss is merely repeating trends.
When driving adoption, he often receives research conclusions that “the results aren’t good.” His suspicion is that some people pick holes from a position of human ego to prove they still have value. The hardest work is not explaining the tool again, but changing the mindset around using AI.
24. Cursor restores the individual hacker’s fighting strength—and amplifies the wrong direction
黄东旭’s classical hacker view is “don’t limit yourself”: backend, frontend, Windows, iOS, the web, or JS are simply knowledge required to achieve a goal, not fixed identities. AI has sharply increased the efficiency of learning across domains.
Human scarcity therefore shifts upstream toward finding problems, forming intent, and daring to act. AI can fill in enormous amounts of detail, but “I want to build a WeChat” is not enough for it to discover the full path on a user’s behalf, much less evaluate architectural choices.
On complex systems or very new libraries, models still fail, so an “old hand” has to steer. 黄东旭 typically discusses the design document back and forth with Cursor, writes the project framework, and then asks it to “fill in the blanks.”
That is why he is skeptical of hands-off agents: if the “first shot” is wrong, everything afterward keeps moving in the wrong direction. Vibe coding is safe only when the framework is clear and the user can evaluate the output.
25. People who cannot program may lose the chance to develop judgment if they rely only on vibe coding
Asked whether someone with zero experience can develop software directly, 黄东旭 answered repeatedly: “Absolutely not.” Social-media demonstrations of a “3-year-old making a game” do not show that the user can evaluate what AI has generated.
For engineers who are still developing, AI can accelerate learning but cannot replace foundational training. If they only do vibe coding and never learn, their career development will suffer.
Cursor has given him “a second spring of programming” because nearly 30 years of experience have taught him what to ask, accept, and reject. Putting the same tool in a beginner’s hands does not automatically reproduce that productivity.
26. DeepSeek’s low-level optimization reflects the different training paths of young engineers in China and the US
Some US frontier AI companies’ infrastructure managers have described a batch of DeepSeek’s open-source AI-infrastructure optimizations as “like magic.” 黄东旭 believes the feeling of “magic” itself shows that they do not know why the engineers had to squeeze efficiency at such a low level.
China’s internet expansion over the past 10-20 years forced companies such as Alibaba and ByteDance to deal with large-scale system problems for years. Real demand cultivated a generation of young engineers with strong fundamentals, and DeepSeek is one example.
Many US schools have skipped C and pointers and started directly with Python or Java. Veteran low-level engineers still exist at major companies such as Google, but the newest startups may not have the same density.
That is why 黄东旭 is more willing to hire managers and PMs in the US, leveraging local experience in products and organizations. But he also admits that for China to produce an Ilya- or Musk-like figure who goes from zero to one, “it may still be too early.”
27. The next database user may not be the developer, but the large language model
黄东旭 will not rewrite all of PingCAP for AI because the original philosophy of simplicity remains valid. What needs to change is the user—from programmers and software developers to agents—and the question of how AI wants to access data.
The first principle of an AI memory layer is to make as few decisions for the model as possible. Traditional ETL and reports first process data into structures humans want to see; he prefers to preserve raw data and give the model a flexible interface.
His weekend project directly calls the Salesforce Open API and writes events into a database. Ask when a customer was last seen or what a salesperson did over the past 3 months, and the model writes SQL on demand, generating dozens of queries in sequence when necessary and summarizing the insight.
In the past, someone would request a report and wait 3 days, then wait another 3-5 days after discovering it was wrong. Now analysis becomes back-and-forth conversation. Salesforce can remain important, but its role may contract to that of a high-quality data provider rather than forcing users through a complex interface.
28. A free personal database for everyone is PingCAP’s long-term bet on the Agent era
黄东旭 envisions providing every person in the world with a free personal database to store a lifetime of data. His own email, Feishu, WeChat, and other records are already in a cloud database, but the product remains “too engineer-oriented.”
He also built a Raspberry Pi-like device that travels with him and takes a photo every 5 seconds. A multimodal model summarizes the images and writes them into a database in a defined format, eventually creating a time-ordered personal feed.
Free does not mean costless; it means exploiting hot-cold distribution. Roughly 90% of personal data is cold over the long term, while the data accessed frequently may be only the latest 200 records. If the system can schedule idle CPU, disks, and hot-cold tiers, total cost can be pushed very low.
PingCAP does not plan to vertically integrate every application. 黄东旭 would rather work with memory-layer partners, provide the database for free, and help them become the shared memory for every AI agent. It could be the company’s biggest product direction, yet have nothing to do with revenue this year or even over the next 3 years.
29. Open source creates trust and determines who gets to set the standards for a new ecosystem
PingCAP chose open source in 2015 both because its founders had benefited from the open-source community and because they judged that a new database could not be commercially viable if it remained closed. Infrastructure software is too critical; a new entrant must first give customers worldwide a verifiable basis for trust.
If DeepSeek had been closed source, it would have struggled to let overseas customers deploy it themselves. Today, advertisements along Highway 101 say things like “Running DeepSeek in your data center.” The more visible shift in China, however, is the all-in-one-machine business: plug it in, see a dialogue box, and it counts as deployment, even when the application scenarios and ecosystem have not been thought through.
The cycle 黄东旭 sees is that the first entrant in a new ecosystem often uses closed source to seize the lead, while the second wave democratizes it through open source. Android and iOS, or Linux and Windows, may ultimately each have a market—“why does one of them have to win?”
The model itself may be only the “eye of the storm”: Tier 1 models such as GPT-4, o1, o3, GPT-4.1, and DeepSeek R1 will continue improving, and the gap between open and closed source will narrow. Agent frameworks and the engineering ecosystem will be harder to keep closed over the long term. MCP may not be the most elegant, but openness and network effects could make it a standard; being 3 months ahead without opening up is “so what?” Eventually, you can only become compatible with someone else’s standard.
30. The faster technology accelerates, the more “what it means to be human” becomes a daily choice
黄东旭 uses The Nine Billion Names of God to express his concern: monks originally needed tens of thousands of years to hand-copy the names of God, but a computer fast-forwarded civilization to the ending, with the stars extinguishing one by one. AI’s risk may not be capability itself, but reaching “overclock” before the human mind is ready.
AlphaZero has already shown the discontinuity in closed tasks: it is not merely unbeatable; humans cannot even understand why a particular move works. If every kind of work becomes Go, people will face a profound existential crisis, especially those who have “made themselves into machines and taken pride in it.”
黄东旭 is replacing identity with experience: houses, cars, possessions, and programs cannot define him; AI can write TiDB, but it cannot feel a journey for him. He has begun meditating, returned to instruments from the 1960s and 1970s, cassettes, vinyl, and stream-of-consciousness creation, and chosen Foundation and Pink Floyd’s The Dark Side of the Moon as relics of human civilization.
The choice goes back to childhood: when his parents confiscated his game cartridges and left only a BASIC interpreter, he was forced to write games himself and eventually discovered that “creating is more interesting than consuming.” At 13 or 14, open-source software and rock music together became languages of resistance to control.
The final destination is not a deeper understanding of model internals, but a return to concrete life: mend a piece of clothing, cultivate a patch of land, cook with your own hands, and watch someone you love eat what you made. “Grow a flower, cook a dish, and then seriously get to know the people around you”—machines can improve the efficiency of outcomes, but only humans can personally experience the process.