Vol.60 Information and Knowledge Management and Learning Efficiency in the AI Era — Kuangxi Podcast Festival: Kuang 2
Summary
- AI’s real breakthrough in knowledge management is not letting people “store more,” but making years of dormant material searchable by meaning, synthesizable across documents, and usable in production for the first time. Traditional search required users to remember exact keywords, while Notion, Yuque and similar tools demanded costly manual classification; now information can remain scattered across documents, group chats, meetings and articles, with users asking questions directly and AI aggregating answers along with their sources. Longtian’s minimum bar is simple: “Starting today, put your knowledge and information online. AI can handle the rest.”
- Knowledge management has long been a notorious startup trap because everyone’s workflow is different, while the cost of organizing information often exceeds its reuse value. Liu Fei groups it with maternal-and-child and pet communities: every new technology cycle produces people who rebuild the category because existing products do not fit their needs. He also points to the millions of words piled up in his Evernote and asks what they are supposed to be used for—without a retrieval mechanism, a knowledge base is merely “a graveyard of knowledge,” and its users are “cyber hamsters.”
- Large-model hallucinations cannot be reduced to zero architecturally in the near term, but product constraints can already push them into a usable range. Liu Fei advocates choosing among web search, Deep Research and designated publications by task, while forcing the model to rely only on supplied materials. Longtian calls hallucination a “feature” of this generation of models, saying multiple rounds of verification and citation constraints can lift the “evidence rate” above 90% in some scenarios and above 98% in isolated cases; direct numerical calculations should still be delegated to computational tools.
- AI is reshuffling the value chain for content and white-collar work: standardized tasks such as research organization and art execution are being hit first, while topic selection, experience and judgment are temporarily becoming scarcer. After using Deep Research, Liu Fei says the research work for Half-Coffee was “roughly cut in half.” Zhuang Minghao says AI already handles around 80%–90% of the front-end production of some 2D scenes, characters, icons, logos and other art assets in games. Once vocal tone, pauses and verbal tics can also be imitated, the core value of podcast creators may shift from content generation to topic selection—but Liu Fei admits, “I don’t have an answer to this question.”
- VC is a form of knowledge work especially well suited to AI; the real unresolved question is whether machine-led decisions can survive five-, seven- or even ten-year return cycles. Standard questions about a founder’s background, product, competitors and funding rounds may yield a roughly 60-point diligence result from a single prompt. Liu Fei says small US funds are already experimenting with “one-person companies,” AI employees and data-driven investment decisions with almost no founder interaction. If these funds perform well over many years, the industry may have to turn the page.
- Zhuang Minghao’s trading framework is straightforward: as long as the long-term AI trend remains intact, Nvidia is still a hold; below $100 is a buy zone, while above $130 is a level for swing trades. He stressed that this was only his personal view and not investment advice. His Bitcoin analogy was to sell at $100,000 and buy back around $80,000–$90,000, with the same condition that the AI trend remains in place.
- The key to education and cognition is not continuing to hoard information that can be memorized, but keeping questions, agency and real-world experience in human hands. As students become “AI natives” and professions such as translation face direct disruption, memory and rote learning will matter less, while the ability to combine and create knowledge will matter more. But Guan Yadi warns that beyond efficiency lie the irreducible complexities of parenting and offline conversation. The conclusion is not to force oneself to “embrace AI,” but to “first become an expert in living, then put AI to work for you.”
Deep dive
1. AI Turns Knowledge Management from an Archiving Problem into a Retrieval Problem
Liu Fei has been writing for roughly ten years. His most direct pain point is not a lack of records, but the fact that once old articles, diaries and work materials accumulate to a certain scale, it becomes nearly impossible to reread them one by one—let alone remember what he wrote and when.
Traditional search also required users to remember fairly precise keywords. If they remembered only the idea but not the original wording, the material was effectively disconnected. AI changes that by finding content through meaning and relevance, then extracting and synthesizing the results.
Liu Fei summarizes the issue this way: “Recording is actually easy. You can just clip, copy and paste as you go. But once you’ve built the library, how are you supposed to use it?” The missing piece was always the final reuse loop.
2. Knowledge Management Keeps Becoming a Startup Trap Because Personal Needs Cannot Be Standardized
Looking back over the fifteen years since around 2010, Liu Fei puts maternal-and-child communities, pet communities and knowledge management in the same category of notorious internet startup traps: once people have a child, get a pet or need to organize information, they often decide existing products are “not usable” and start building their own.
Knowledge management is especially difficult. Everyone consumes different information, follows different classification habits, produces different kinds of output and works within different relationships. A single fixed workflow can rarely cover all of that.
That is why every new technology cycle brings another group proposing to “rebuild knowledge management with this technology.” AI is not the first startup craze in the category; it is simply the latest variable aimed at lowering the cost of organization and retrieval.
3. A Knowledge Base Is Not a Folder but a Production Chain from Input to Output
Liu Fei’s plain definition is that the public-account articles, reports, saved items, group chats and documents encountered every day are stored, classified, summarized, processed and coordinated into reports, copy or presentations. That information flow is knowledge management.
His summary at the end of last year ran to 132 pages. The underlying materials may have included hundreds of reports, hundreds of charts and hundreds of articles. The human brain cannot retain all of that, so an external system must handle storage, filtering, recombination and updating.
The problem is that information is usually scattered across WeChat public accounts, WeChat groups, chat histories, Feishu documents and other databases, while processing it requires constant tool switching. Everyone wants an integrated product, but actual work habits make a true “all-in-one” system difficult.
The value of a knowledge base should therefore not be measured by how many files it stores, but by whether the material can be converted back into results the next time someone writes, makes a decision or collaborates.
4. The Cost of Manual Classification Turns a Beautiful Knowledge Base into a “Graveyard of Knowledge”
Liu Fei says products such as Yuque and Notion are well built, but the old problem remains the barrier to use: every time someone adds an article, they must manually judge its relationship to existing material, then classify, format and tag it. For ordinary users, that cost is unsustainable.
The best-organized system he has seen belongs to Liu Shaonan, who works on knowledge organization and runs a paid subscription product. That gives him a sustained incentive ordinary users do not have.
Liu Fei himself kept throwing articles into Evernote until the archive reached several million words. The problem then became not a lack of content, but not knowing how to use it again. He appeared to possess a vast body of material, but the sense of possession was merely “illusory knowledge.”
Liu Fei also uses the term “cyber hamster”: when the cost of acquiring public information approaches zero, people keep circling things in and piling them up. If technology cannot rebuild the relationships among those materials, the result is not a knowledge base but a “graveyard of knowledge”—or simply a folder for storing garbage.
5. Semantic Search Lets Information Stay Scattered While Users Retain Only the Question
AI’s advantage is not limited to finding the word “e-commerce.” It can also connect related experience from two-sided platforms, ride-hailing, food delivery and transaction platforms, allowing users to discover categories they never created themselves.
Longtian argues that people should first put their knowledge and information online. The more information they generate, the less they need to worry about over-organizing it; classification, extraction and linkage can later be delegated to smarter tools.
He uses this podcast festival as an example: the event’s progress was scattered across dozens of documents and dozens of group chats. He only had to ask how far the overall plan had advanced, and AI produced a structured summary with cited sources.
Online meetings have also gone through two waves of demand growth. The first came during the unusual period; the second came with AI. In the past, recording a meeting still left someone with the work of producing minutes. Now a conversation can directly generate a reasonably reliable summary.
6. Feishu Is Trying to Make AI Infrastructure, but Complexity Remains Product Debt
Longtian sees Feishu not merely as a tool for office workers, but as a knowledge-production and management platform ordinary people can use. Feishu Minutes, Documents and Knowledge Base together cover conversation records, collaboration and subsequent retrieval.
Asked whether the interface is too complex, he acknowledged that more powerful functionality inevitably makes a product more complex, and that new users do face a higher comprehension cost. Feishu will continue improving the user experience and offer simpler, clearer ways to use the product for different user groups.
On training and case studies, the recommendation at the event was to follow the available courses and materials. The product direction is to help users find functions relevant to them, rather than requiring everyone to understand the full capability set.
In international markets, the domestic version is Feishu and the international version is called Lucky. The two differ in deployment, local laws and compliance requirements. Multinational companies can use the corresponding global collaboration solution.
7. Hallucinations Cannot Be Eliminated, but Evidence, Boundaries and Verification Can Constrain Them
Liu Fei stresses that AI products and configurations differ. The web search, deep-thinking and Deep Research modes of DeepSeek, Yuanbao, Doubao, Feishu and GPT produce different results, while the Chinese or English sources searched also affect the output.
While producing Half-Coffee, he encountered models that simply invented business stories. But when he supplied a complete biography or formally published material and required the model to follow the source, errors fell sharply—though human review was still necessary.
Zhuang Minghao’s view is that current architectures offer no visible way to eliminate hallucinations 100% in the near term. The workable approach is to narrow the task, restrict the sources and constrain the output format, creating a small environment in which the model’s work can be verified.
Longtian takes a more aggressive view: “Hallucination is not a problem with this generation of AI. It is a feature of it—it is a feature.” Products can perform double and multiple checks much as humans do. He says engineering can push the “evidence rate” above 90% in some scenarios, with isolated cases exceeding 98%.
8. AI Is Like an Excellent Liberal-Arts Student; Task Allocation Matters More Than Blind Replacement
Liu Fei’s simple boundary is that generative AI resembles “an exceptionally good liberal-arts student.” Creation, summarization and some forms of reasoning can be handed to it boldly, but direct numerical calculation is generally not its strength.
Mature tools already exist for computation. The better approach is to have a human or an Agent call those tools, rather than asking a language model to perform every calculation through its text capabilities.
Zhuang Minghao believes workers will spend a long time deciding “which tasks to give to AI.” AI’s arrival does not make existing tools obsolete. In some cases, current solutions are more accurate and stable, and a general-purpose Agent may not be the optimal choice.
9. VC Due Diligence Is Highly Standardized; a 60-Point Result Can Already Be Generated by a Prompt
Zhuang Minghao participates in early-stage investing. He says the dimensions VC cares about have remained relatively stable: founder background, product, competitors, funding rounds and so on. At base, the work is repetitive information collection, organization and judgment.
He cites a short prompt from a US investor and argues that, given to a model today, it would probably produce an initial result that is “pretty decent, maybe 60 points,” without first building a complex workflow. The prerequisite is that the investor has already distilled the dimensions that matter.
More localized elements of the Chinese market remain constrained by data availability. Information on company registration, patents and similar matters requires specialized databases, and China still lacks a particularly good workflow product that can solve these problems directly.
Zhuang Minghao’s sharper judgment is that for collecting, organizing and outputting knowledge-based information, “humanity has been beaten to a pulp.” The craft of traditional VC does not mean every step contains a high technical barrier.
10. Fully AI-Run Funds Are Testing Whether VC Can Still Be Built Around “Feel”
Liu Fei says small new funds in the US are moving to the opposite end of the spectrum: they minimize human intervention, do not meet founders in person, and use data and AI for analysis, scoring and tracking. AI employees may even handle materials and the signing of investment agreements.
These funds may have only one person, with every other employee being an AI employee. They represent an extreme contrast with traditional investing, which depends on meetings, preferences and a founder’s personal feel. Liu Fei does not claim either side will win; he only says the experiment has begun.
Early-stage investments cannot be verified immediately. Projects typically need five, seven or even ten years to show results. Only after this batch of fully technical funds delivers long-term performance will the industry have an answer. If they perform very well, “this industry may have to turn the page.”
11. Nvidia’s Long-Term Logic Remains Intact; the Short-Term Framework Is Buy Below $100 and Trade Above $130
On the premise that this was “only my personal view and not investment advice,” Zhuang Minghao said Nvidia’s long-term logic remains unchanged and the stock is still a hold as long as the AI wave maintains its current trend. The sharp volatility of the past period, especially the last few months, has also reflected policy, tariffs and other factors.
His short-term price framework is simple: Nvidia is buyable below $100; above $130, investors can take some swing trades and buy back below $100.
He compares the strategy with Bitcoin: “Sell at $100,000, then buy back around $80,000 or $90,000.” The experience is explicitly conditioned on the trend continuing and labeled as a personal view.
12. Podcast Execution Costs Are Collapsing; the Creator Moat May Retreat to Topic Selection
Liu Fei believes conversational podcasts are temporarily harder to replace because the host must understand the guest in real time and ask follow-up questions. Half-Coffee-style business stories rely mainly on public materials and published books, so in theory they are easier for models to replicate.
More than 100 episodes already provide enough training data for a model to learn the show’s style. Tone, emotion and verbal tics were once thought to be protective layers, but podcast-generation tools can now imitate rhythm, pauses and even speech habits.
His honest answer is: “I don’t have an answer to this question.” The core value may eventually shift from content generation to topic selection. But if another creator has strong topic-selection ability and uses AI for research, polishing and voice production, that creator could also replace an existing show.
AI has already materially changed the workload. Liu Fei uses Deep Research extensively, having the model first organize books and source material running into hundreds of thousands or even millions of words, then manually verifying and polishing it. The research-organization stage takes “roughly half the time.”
13. As Learning Efficiency Rises, Some Traditionally “Important Processes” Will Lose Their Status
Guan Yadi recalls that preparing a course used to require reading many books. AI can quickly produce a draft, but it may also replace part of the training the brain receives through reading and distillation.
The question at the event therefore became whether the pleasure of instant feedback would reduce people’s tolerance for the pain of learning and weaken deep thinking. Liu Fei believes the shift is “inevitable.”
His view is that “a lot of work we once thought was important will become unimportant.” The large amount of memorization and fact retention in nine-year compulsory education was once necessary to build a basic knowledge system, but AI and the internet are now providing an external backstop.
The liberal arts will not disappear; the location of competitiveness will shift. Memory and rote learning will matter less, while the ability to combine and create knowledge will matter more. Education will undergo a major transformation, but the event offered no mature replacement curriculum.
14. This AI Cycle Is Unsettling Because It Is the First to Directly Do Ordinary People’s Jobs
Guan Yadi points out that earlier AI could play Go, chess and StarCraft, drive cars and even defeat world champions, but most people do not make a living from those activities.
Generative AI writes emails, performs data analysis and organizes information every day—and “that is our work.” Ordinary people are seeing machines competently perform tasks they have repeatedly trained themselves to do, creating a more visceral experience than a world champion losing to a machine.
Guan Yadi also says that in one ranking of professions by replacement risk, she and another guest independently put translation first. Many major international conferences already use AI rather than humans for simultaneous interpretation. The question is no longer merely whether this will happen someday, but what existing practitioners are facing now.
15. AI Natives Have Arrived; a General-Purpose 60- or 70-Point Answer Is Enough to Expand the Capability Frontier
At the final session of a campus-recruiting process, Longtian asked outstanding students from around the world who used GPT in daily life. Nearly 100% raised their hands. His view is that students of this generation who do not use GPT to write papers may simply lose to those who do. They are “a generation of AI natives.”
Younger children now speak directly with Doubao, repeatedly asking “why” questions. The default interface for obtaining information has shifted from the search box and the book to a conversational model.
After DeepSeek displayed its chain of thought, many ordinary users saw for the first time how a problem could be broken into steps. Longtian gives the example of a front-desk operations colleague with a budget of several thousand yuan who did not know how to organize an event. After handing the task to DeepSeek, the model listed the first, second and third steps. The result may have been better than what the employee could have done alone—and the employee learned from it.
Experts may find the reasoning inadequate, but for a completely unfamiliar task, a 60- or 70-point framework is enough to get started. Longtian’s optimism comes from the fact that “the world has always been pretty makeshift”; general knowledge can fill a large amount of the basic guidance gap.
16. Technological Revolutions Create New Jobs, but the Adaptation Cycle Is Accelerating Beyond a Human Lifetime
Longtian notes that every major industrial and technological revolution has produced new occupations. Ten years ago, few people could have predicted that short-video editing would become a job. New technology will also change employment, but it is unclear whether the number of new roles will exceed the number of roles disrupted.
Humans can adapt to new changes over time, but no one knows exactly how long the process will take.
His timeline is as follows: the First Industrial Revolution may have taken roughly 100 years to truly raise productivity and generate a massive increase in global GDP; the second may have taken 50–70 years; the third, from computers and chips to the internet, may have taken 30–40 years. Whether AI takes 20 years, ten years or five, no one knows.
From God’s-eye perspective, 30 or 50 years is just one page of history, “but for many people it is their entire life.” Massive change brings opportunity and possibility, but individuals are also bearing the cost of the transition.
17. Purely Efficiency-Driven Work Is Easier to Replace; Experience and Complexity Preserve a Human Role
Liu Fei divides demand into efficiency-oriented and experience-oriented categories. Textiles are judged mainly by the end result, and machines can match or outperform humans in most scenarios, which is why industrialization could replace textile workers at scale.
Television did not eliminate film, and film did not eliminate theater, because being in a cinema with other people or watching a real person perform on stage is itself part of the product experience. It cannot be perfectly substituted by a more efficient medium.
To assess occupational risk, ask whether the job is merely about completing Excel sheets, formulas and standardized results, or whether it also includes topic selection, content structure, human interaction and non-replicable experience. The replacement logic is different in the latter case.
Guan Yadi calls this value “complexity”: taking a child out to sit in the sun, fixing smoke backdraft in the home or sitting down for a long offline conversation are not activities optimized first for efficiency. Handwriting may likewise move from the mainstream recording method into the realm of experience and expression, much like calligraphy.
18. AI Checking AI Is Producing a “Magic Duel,” While the Education System Makes Students Bear the Consequences
Zhuang Minghao cites problems with AI-detection systems for academic papers. Feeding AI the “Preface to the Pavilion of King Teng” may produce a 74% similarity score; feeding it The Three-Body Problem may produce a score above 60%. AI detection itself can therefore hallucinate.
At the same time, students may be told that AI use cannot exceed 20% in liberal-arts subjects and 15% in science subjects. Students use AI to write, schools use AI to check, and the absurdity is that “two AIs are fighting each other, while humans ultimately bear the consequences.”
Primary and secondary schools are already trying to establish more detailed rules. At Zhuang Minghao’s fifth-grader’s school, teachers and students have separate guidelines: AI cannot be used for major Chinese- or English-language essays, and it cannot be used for the ideas section of English project work, though it may be used for the PPT design.
Enforcement still depends on teachers’ experience. A teacher may judge that an English essay was AI-generated based on the language level normally expected of a fifth-grader and require a rewrite. As models improve, these judgments and boundaries will only become harder.
19. The Best Personal Knowledge Base Is Often Not the Most Advanced but the One with the Shortest Path
Zhuang Minghao calls himself lazy. His eventual methodology is to manage knowledge in as few steps as possible. He uses a WeChat fan group as his knowledge base because the group is not very active, allowing him to continuously send in reports, charts, links and PDFs.
When he actually needs something, he searches the images, links and files in the group chat and starts working once he finds the relevant material. The system is not elegant, but it uses a familiar tool and requires no additional organization.
Liu Fei’s equivalent tools are Jike and Flomo. He posts thoughts suitable for public viewing on Jike and keeps private ones in Flomo. The key is not to save the original text, but to leave behind his own judgment; after turning 30, a morning insight may be forgotten by evening.
Zhuang Minghao adds that everyone on the panel has a habit of producing content, which makes it easy to overestimate ordinary people’s need to publish. “Output drives input” works for creators, but most people may not have a clear expression scenario and do not need to copy a content producer’s heavyweight system.
20. Knowledge Has Value Only When Connected to a Concrete Problem; Otherwise It Is Just Entertainment in New Packaging
Liu Fei advises people to “manage their knowledge with a question in mind.” AI is hot, but that does not mean everyone needs to read all of OpenAI’s papers from start to finish. First understand the parts related to one’s field, thesis or current work problem.
Studying something simply because a brand is popular or an industry is hot does not necessarily create knowledge. If someone buys a company’s stock and actually loses money, the materials become knowledge only when they understand its founding team, business and price movements.
His definition is direct: “Only what is useful to us is knowledge.” Random reading without a connection to reality may be entertainment, even when the subject is a classic.
This is not a criticism of entertainment. Guan Yadi can browse books at random and collect more than 10,000 volumes as a personal interest. The real danger is mistaking collection for mastery, or treating the sense of ownership produced by digital hoarding as cognitive growth.
21. Trust Determines Whether AI Can Enter a Live Content Setting; Showy Digital Humans Still Have No Value
The event used Teacher Lanxi’s exceptionally long post-stream summary, “Lanxi in One Sentence,” as a benchmark. Readers trust it as a reliable summary, while the first hurdle for AI is trust. Trust comes from repeated communication, depth of dialogue, repeated returns of results and whether verification gradually narrows the gap with the original text.
Longtian recalls that Feishu once had AI host a live guest-chat event internally. The result was “a complete mess”: the long prompt ran slowly, responses were not fluid and no real dialogue emerged. The experiment showed that generating answers is not the same as carrying the rhythm of a live setting.
He believes real-time conversation is no longer especially difficult as long as the system is not a long-reasoning model. But the principle remains not to “use AI for the sake of using AI.” A digital human installed purely for novelty has no meaning.
Guan Yadi proposes an experiment: for the first 45 minutes, let the guest’s AI avatar speak with the audience, then reveal during the break that the real guest has been present all along. Longtian thinks the idea has theatrical value, but stresses that when AI occurs naturally, no one will be able to stop it; there is no point forcing an AI gimmick. Guan Yadi says to give the technology some time to iterate—it may already be completely different by November.
22. AI Will Blend into Life Like the Internet; Our Cognitive Framework Does Not Need to Be Rewritten Around the Tool
Liu Fei’s closing point is: “There’s no need to force yourself to embrace AI just because this is the AI era.” Ten years ago, no one declared every day that they were embracing the internet. Word, Excel and smartphones ultimately entered work and life through real needs.
Zhuang Minghao compares this with the 1999 “Internet Survival Challenge,” in which people were locked in a hotel with only a computer and money and asked to survive through the internet. More than 20 years later, if someone took today’s smartphone back to 1999, it would be difficult to explain where food delivery, flights, navigation, contacts and virtual worlds came from.
Longtian observes his parents, who were born in the 1960s. In the early internet era, they needed their children’s help to register accounts; now, when he visits, he sees packages they ordered online themselves. Mature technology eventually disappears into infrastructure, without requiring everyone to first take a course in the system.
Guan Yadi insists that agency must remain with people: “Let life guide me on what knowledge to learn, what experience to accumulate and what tools to borrow.” Her final formulation is not anti-efficiency, but: “First become an expert in living, then put AI to work for you.”