Vol. 62 Will AI Bring Intellectual Equality? If So, What Comes Next? — Kuangxi Podcast Festival: Er Kuang
Summary
- AI’s key inflection point has moved from “machines can produce text” to “outputs need almost no editing” in some contexts, but a leap in tool usability does not mean equal outcomes. From 姬十三’s first shock at encountering ChatGPT, to his view that o3 is “much stronger than my assistant,” to 潘乱’s sustained use of DeepSeek, o4, and Claude 3.7 after the Lunar New Year, and 曼祺’s account of global memory rolling out on April 11, AI is becoming genuinely usable; prompt skill and hallucinations remain constraints.
- There are 2 related but distinct views on the order of substitution: virtual, rule-bound, data-driven work is easier for models to handle first; 潘乱 argues that highly logical, advanced knowledge work may come under pressure before hairdressing, cleaning, and construction. The analysis he cites covers 1,639 occupations, 19,265 functions, and 23,000 work activities: translation is the easiest to replace, while landscaping, security, cleaning, and construction rank near the bottom. Skills humans acquire through practice are considered already matched or surpassed by AI, while biological instincts formed in childhood remain the hardest to replicate.
- AI is leveling access to capability, not income outcomes: people with a foundation, the motivation, and the habit of asking follow-up questions can compress their learning curves, while everyone else gets standardized answers. 潘乱 stresses that ordinary people can use AI to confront a much larger knowledge base; 姬十三 later summarizes it as “equal access, unequal outcomes.” 曼祺’s F1 analogy is equally stark: the tools may become more widespread, but “there are only 20 F1 drivers in the world.”
- The creative industry’s most vulnerable layer is likely the commercialization middle: a tiny number of geniuses get amplified, masses of enthusiasts create for love, while ordinary professional creators lose bargaining power and jobs. 姬十三 relays 蔡浩宇’s split between the top 0.0001% of talent and amateurs; 潘乱 says “screenwriters’ salaries may have to come down,” while 曼祺 describes the future as “an even more explosive age of creative mediocrity.”
- So far, AI has mainly lifted production efficiency without creating new distribution mechanisms or business models, so the gains are flowing first to incumbent technology giants and more skilled users. Apple, Microsoft, NVIDIA, Meta, Google, and the other “Magnificent Seven” are seen as the biggest beneficiaries of the past 2 years; 纳德拉’s question about global GDP not yet showing a clear lift, with a 10% increase marking the start of a revolution, and Microsoft–OpenAI’s arrangement making $100B in cumulative profit-sharing an AGI milestone, both pull the technology narrative back toward monetizable profit.
- Education will not simply abandon memorization and training because answers are readily available; AI is more likely to cut the cost of “playing a thousand tunes” from 1,000 repetitions to 10. 潘乱 argues that foundational knowledge still has to be acquired through memory; 姬十三 adds that AI may make it unnecessary to actually practice 1,000 times. Children already use AI to make project decks, debate “The Foolish Old Man Removes the Mountains,” and turn physics concepts into songs; the real problem is that neither parents nor schools can explain how far children should use it or where the boundaries lie.
- The next phase is defined by a shift from language to behavior. Whether Agents can execute tasks, start companies, and deploy capital may determine whether AI brings fewer jobs or more concentrated layoffs; these remain projections, and models still cannot reliably make market judgments or bear responsibility. Human moats are narrowing to “generative will,” unique experience, and accountability, while expensive brain–computer interfaces could turn intellectual equality into a paywalled hierarchy.
Deep dive
1. AI Has Crossed from “Producing Text” to “Ready for Delivery” in Some Contexts
姬十三 recalls being stunned the first time he used ChatGPT and believes he hosted a livestream the next day. 潘乱 agrees that the biggest shock was “a machine producing text”—and producing writing that looked plausible enough for the technological breakthrough to become an almost instant global consensus.
姬十三 now uses o3 more often, and his benchmark is no longer whether it can answer a question, but whether “there is very little left to revise” once the result arrives. His direct comparison: “o3 really is much stronger than my assistant.”
潘乱 previously believed that prompt skill and a high hallucination rate would keep most people out; after using DeepSeek, o4, and Claude 3.7 continuously after the Lunar New Year, he changed his mind. 曼祺 shared her experience after OpenAI rolled out global memory to a subset of users on April 11: the model’s description of her identity and writing preferences was especially accurate. 姬十三 adds that DeepSeek still tends to make things up in demanding writing tasks, while Claude 3.7 is much stronger.
2. Writing a Wu Kingdom Drama Outline Overnight Shows the Barrier Falling, Not Creative Equality
潘乱 spent half a night producing a Three Kingdoms drama outline from the perspective of Eastern Wu: starting with the relationship between Sun Wu and the central regime, then weaving the Four Great Commanders, the Five Attacks on Hefei, Taiwan, Liaodong, Southeast Asia, the Shanyue, the Baiyue, and Hainan into a maritime-power narrative.
As they continued unpacking the case, 曼祺 said she was initially unfamiliar with Eastern Wu history herself. She browsed Wikipedia, asked ChatGPT questions through Yuanbao, then passed the material to Claude to keep expanding the structure and connections; she also organized recent papers on Eastern Wu and tried to incorporate different perspectives into the play.
潘乱 asked the model to write the reactions of different parties after Lü Meng killed Guan Yu. He showed the result to the screenwriter and director of The Advisors Alliance, who said it was “workable” and might even be suitable for a test shoot. For someone who had never taken a screenwriting class and had bought a screenwriting textbook but only read the table of contents, that was enough to make him say: “It really feels like a different era.”
姬十三 points out that the core conflict between Eastern Wu’s local gentry and the Sun clan’s outside regime is the vivid dramatic tension in the case. 潘乱 replies that this was a conflict he already knew about; AI did not generate it for him. AI can expand “from different angles, each revealing a different view” into multiple paths of observation, but the initial, sharply defined conflict still came from a person.
3. AI Levels the Entry Point, Not the Outcome
曼祺 first breaks “intelligence” into its components: language, composition, thinking, and planning will not be equalized at the same time. Some capabilities are being rapidly commoditized, while gaps in foundations, judgment, and depth of use are widening.
姬十三 uses an AK as an analogy: giving everyone the same gun does not mean everyone can hit the target, but the technology does create the opportunity to acquire that capability. 潘乱 stresses that people with strong motivation, even in remote areas, can access the entire internet’s knowledge base and quickly build a knowledge graph by going down an interactive “rabbit hole.”
The conversation ultimately separates opportunity from outcome: “Access is equal; results are not.” A model can call on general-purpose skills, but it cannot generate the user’s own motive for expression. Some people ask 2 questions and think they have found the answer; others keep asking until the context window collapses.
4. The Order of Substitution Follows the Virtual and Rule-Bound Before It Follows Pay
潘乱 believes that $20 posters, ordinary portraits, low-level translation, and homework tutoring can already be handled by AI. High-end commercial projects still mainly involve “making skilled people even stronger”; it remains difficult for ordinary people to use AI directly to obtain high-value commercial work.
潘乱 makes the point explicitly: the more logical and “high-level” the office knowledge work, the closer it is to the model’s zone of strength. Headcount in VC analyst roles, financial reporting, and the Big Four could all decline, while physical work such as street-corner hairdressing is harder to automate.
姬十三’s technical framework is not high pay versus low pay, but virtual versus physical, and clearly defined rules versus open-ended interaction. Code exists in a clear rule space, making it suitable for reinforcement learning; natural language is an abstraction of the physical world and can therefore learn patterns from massive datasets.
Embodied intelligence remains constrained by motor control and physical interaction. Autonomous driving only needs to avoid collisions within road rules, while a household robot has to identify and grasp hard, soft, transparent, and black objects. Nannies, hairdressers, and household services may therefore have longer substitution windows.
5. Hiring Data Points to a Counterintuitive Result: Skills Built Through Practice Lose First
潘乱 relays statistics from 陈沁, a Zhihu data-analysis blogger known as “Data Emperor”: the project scraped BOSS Zhipin data covering 1,639 occupations, 19,265 functions, and 23,000 work activities, labeled them, then had GPT assess the substitution efficiency of each task and rank them by weighted score.
The conclusion, as relayed by 潘乱, is: “Humans have already fallen behind AI in skills involving knowledge learned through practice, experience accumulated, and tricks developed over time.” By contrast, perception, movement, and biological instincts acquired at birth or in childhood are the hardest to imitate.
Translation sits at the easiest end of the substitution spectrum. 姬十三 previously estimated that translation work may already have been compressed to just 1% or 0.5% of its former scope. Landscaping workers, security guards, sewing workers, cleaners, renovation workers, laundry workers, crane operators, dancers, and construction workers sit at the harder end.
曼祺 reinterprets something 沈向阳 once said: “The data accumulated over 40 years of the internet seems to have existed for this one AI moment.” The cruel part is that the vast quantity of digital traces accumulated by humanity has become precisely the training material for its replacements.
6. The Premium on Professional Content Retreats to Views, Frameworks, and First-Hand Information
曼祺 introduces AI-generated presentations as a segment with many vendors, where the AR of some projects could reach tens of millions of dollars. But an automatically generated “Overview of the AI Industry in the First Half of 2025” may be adequate for a 10-year-old or a newcomer, yet cannot be presented directly to true industry practitioners.
Professional-grade content still needs views, selection, boundaries, logical frameworks, and first-principles synthesis. No matter how mature the production line becomes, some people will still insist on hand-building a car; the issue is that this work is necessarily inefficient and low-volume, leaving only a small group at the very top.
The full F1 analogy is not motivational fluff: give any driver the fastest tool in the world and the driver still matters, but there are only 20 F1 drivers in the world. Bringing high-level capabilities down to everyone does not mean the existing number of professionals and the existing level of commercial demand can all be preserved.
姬十三 says the editorial team’s internal discussions led them to conclude that first-hand information has become more valuable: before a conversation with a real person, it does not exist online and is not yet data. Chatbots can offer inspiration on an article’s logic and perspectives, but are not necessarily fully accurate. 曼祺 links this to LatePost’s positioning around exclusive reporting: once a report is published, others can process it from different angles.
潘乱 admits that he used to write commentary based only on material found online. 曼祺 bluntly says this school of work will become more dangerous; 潘乱 therefore says he needs to embrace AI more actively, because AI makes him work harder.
7. The Commercialization Middle Layer of the Creative Industry May Be Hollowed Out
姬十三 relays a view 蔡浩宇 posted online: meaningful game development in the future may remain only at 2 ends—the top 0.0001% of geniuses forming elite teams to create unprecedented work, and the overwhelming majority of enthusiasts making things temporarily to satisfy their own ideas.
姬十三’s interpretation is that “game development” itself will be redefined: the peak will continue producing genuine large-scale works, while the other end produces small creations serving only oneself and one’s friends. The large number of professional jobs that currently sit between commercial and indie games may not survive.
曼祺 believes this is not creative equality but an amplification of the gap. People with the strongest ideas are amplified by AI, while people with the most passion can create for love. The middle layer—neither the most capable nor the most passionate and idea-driven—will have the hardest time proving why it should continue treating this as a profession.
潘乱 puts the industry’s outcome more bluntly: “Screenwriters’ salaries may have to come down.” If production capability becomes widespread without a corresponding expansion in social demand, 曼祺 worries that what arrives will be “an even more explosive age of creative mediocrity.”
8. Education Will Not Abandon Cognitive Scaffolding Because Answers Are Readily Available
姬十三 first asks 2 parents with children whether, if standard answers are no longer scarce in the future, children still need to memorize by rote and whether school education is wasting time. 潘乱 replies that in every era, people with the means will believe the current education system is suboptimal, but school education remains an equalizing tool for the broader population.
潘乱 believes people acquire much of their basic framework for understanding the world through rote memorization during adolescence. If they rely entirely on AI for knowledge from age 6 to 20, that may not constitute ideal education. Rote memorization and understanding are not natural opposites.
姬十三 accepts the logic of “playing 1,000 tunes before recognizing music, observing 1,000 swords before recognizing weapons,” but adds that AI may make it unnecessary to practice 1,000 times: “Maybe it only takes 10,” after which the learner can continue deeper along the feedback loop.
9. AI Natives Have Already Bypassed Usage Boundaries Designed by Adults
The first time 姬十三’s 8-year-old son used ChatGPT, he did not search for an answer. He proactively declared a persona: “I’m an 8-year-old child. My mom makes me do homework every day and I’m annoyed. What should I do?” This was native conversational use, not treating the model like Baidu.
For a group project in fifth grade, the children discussed Tesla’s life and theories among themselves, but ultimately “secretly” had AI make the presentation. The group discussion and opinions were produced by humans; the written deck was generated by AI.
曼祺 has seen elementary-school students make Doubao and ChatGPT debate “The Foolish Old Man Removes the Mountains.” She has also seen physics teachers give each chapter’s concepts to a model, turn them into rhythmic lyrics, and then make songs to support revision. These uses were not predesigned by adult curricula.
姬十三 admits that blanket restrictions from schools, competitions, and parents are simply “the most powerless” option. Children cannot realistically avoid using AI, but parents have no god’s-eye view of how far they should go or what boundaries to preserve. They can only “feel their way forward.”
10. Production Efficiency Has Surged, but Distribution and Business Models Have Not Innovated in Step
潘乱 believes AI’s clearest change so far has been production efficiency. Beyond search, recommendation, and following, no new information-distribution paradigm has appeared; creators still monetize through advertising, livestreaming, and paid knowledge, and there is no entirely new internet platform or business model in sight.
If the total pie does not grow, early AI adopters will use lower costs to compete for the same allocation, even “to exploit others more effectively.” People who cannot use AI will be more easily pushed down the value chain by more capable users.
姬十三 observes that the biggest gains in the secondary market still accrue to incumbent technology giants such as Apple, Microsoft, NVIDIA, Meta, and Google. 潘乱 believes that higher individual efficiency from AI could lead to higher income; 曼祺 thinks it is difficult for any individual’s income to increase by an order of magnitude as a result.
姬十三 also relays 纳德拉’s question: global GDP has not shown a clear AI-driven increase, and perhaps a 10% rise would mark the start of the revolution. Microsoft and OpenAI have also set $100B in profit-sharing as an AGI milestone. Compared with “economic outperformance in 90% of jobs,” defining the milestone by profit is more naked.
11. The Technology Dividend Can Mean Less Work—or More Layoffs
潘乱 extrapolates from the evolution of the workweek: China once had only 1 day off each week, then moved through 1.5 days and the 2-day weekend. If AI genuinely lifts productivity across society by a full step, the theoretical path runs from 5 workdays a week to fewer days, perhaps even just 1 day of work.
曼祺 preserves the dystopian version: companies may not convert the efficiency dividend into rest, but instead lay off employees and obtain more profit with less money. Technology can raise output; it does not automatically determine how the new value will be distributed.
Which direction prevails depends on policy, organizational constraints, and entrepreneurs’ choices. The conversation does not directly turn “intellectual equality” into a leisure society; 曼祺 says 潘乱’s vision may be too “simple” and naïve.
12. The Labor Market May Reward Craftsmanship Again, While Identity Anxiety Becomes Harder to Price
潘乱 suggests that his sister have her daughter learn jade carving, because supply and demand matter more than the traditional hierarchy of occupations. “You can’t find a migrant worker for RMB3,000, but you can hire many college graduates for RMB3,000.” Blue-collar scarcity may be revalued before social attitudes catch up.
He cites reports that the number of PhD students currently enrolled in China already exceeds 60% of the total number of PhDs in history, and predicts that humanities programs will continue shrinking because of employment rates and salaries. High-income knowledge jobs in finance and the internet industry also no longer need as many people as they once did.
姬十三 uses 王治郅, who still uses a Nokia phone, as a counterexample: if someone has already achieved internal coherence in his own world, he need not force himself to chase every technology. A more practical strategy is to have one craft, or become relatively top-tier in a sufficiently narrow field.
曼祺 worries that the deeper crisis comes after work. If occupation and income no longer signal identity, people may not actually have a sufficiently distinctive inner life. In a science-fiction novel 姬十三 shared, a painter who does not need to work and has no trouble meeting his material needs remains forever second-rate; once work disappears, it may become even harder to locate one’s dignity.
13. The Human Moat Shrinks to Generative Will, Experience, and Responsibility
姬十三 observes within his team that after models became widespread, the most significant difference was not answer quality but depth of questioning. Some people ask 2 questions and think they have found the answer; others keep digging until the context window collapses. What the latter reveal is interest, motivation, and stamina.
姬十三 relays a self-assessment from ChatGPT: “Every skill will be callable,” but the world will listen only to people “with generative will.” Compassion, anger, sensitivity, and curiosity constitute non-general-purpose capabilities. “You can no longer live by some identity; you can only live by who you are.”
姬十三 believes that the final differences between people will certainly include memory and experience. 曼祺 then cites architect 刘家琨: “Synthesize broadly, present simply, rely on data, and settle on intuition.” Architecture’s nonverbal qualities and intuition are, at least for now, difficult to compute completely.
The most practical form of irreplaceability is responsibility. Academic peer review may become “2 AIs battling with magic, with a human taking the blame”; even if a bank teller’s tasks can be automated, handling money and accountability still requires a responsible human party.
14. Brain–Computer Interfaces Could Turn “Intellectual Equality” into a Paywalled Hierarchy
姬十三 recalls that the 2018 season of the Chinese talk show I Can I BB once debated whether to support a brain chip enabling “one-second knowledge sharing.” The conversation argues that once a large model is built into a chip, the premise is no longer pure fantasy. The extreme version is people lying in nutrient pods while “carbon-based life supplies tokens to silicon-based life.”
姬十三 warns that explosive technologies have never stopped where humans hoped they would; they only keep “flying outward.” He also mentions that Life 3.0 explores 12 possible endings for humanity and AI, with extinction, coexistence, and control all among the branches.
The real stratification comes from price. If a brain–computer interface costs half an individual’s wealth, the rich can purchase different grades of intelligence. This invites comparison with a Cyberpunk 2077-style world in which technology and wealth owners are completely separated from the underclass that contributes bodily function.
曼祺 counters that science fiction may favor this ending simply because it is more frightening and dramatic. Her narrower version is the “Doctor Strange” capability: people with a foundation and agency use AI first to simulate more paths of observation and more possible outcomes.
15. The Right Way to Use AI Is as a Thinking Adversary, Not an Answer Machine
曼祺 points out that many of China’s leading products remain freely available, making equality at the access level fairly substantial. The question has shifted from “Can you get it?” to “Can you use it deeply, in relation to the knowledge and life you already have?”
潘乱’s live demonstration is to add Yuanbao as a WeChat friend and pin it, then forward articles, long chats, and merged group-chat histories for it to interpret, analyze, and expand on. “You can squeeze it,” treating the model as a patient and resilient thinking partner.
For people with little impulse to produce, AI can connect the “fleeting impressions” in their heads into broken phrases, then turn the fragments into an article. 姬十三 believes it can also supply the “different angle revealing a different view,” making people more confident when facing a problem. 潘乱 says sustained use makes him more industrious.
AI is not a prerequisite for acquiring investment know-how. 姬十三 cites an interview with 张新民 for the program Face to Face: without industry experience, he could infer the condition of a listed company using only 3 basic financial statements. AI merely lowers the cost of organizing material and reasoning through multiple angles.
16. The Keyword for the First Half of 2025 Is “From Language to Behavior”
姬十三 summarizes the model advances of the period in 1 phrase: “From language to behavior.” The core of the past 2 years was large language models generating content; progress in the first half of 2025 around Agents, tool use, task execution, and planning began turning language into action.
Software tasks now show a clear gradient: asking a model to act as a calculator is relatively simple, while asking it to build a Douyin account directly is still unrealistic. Even when developers receive a human assignment, they may turn around and use Cursor. The capability frontier will move, but complex systems cannot yet be completed with a single instruction.
曼祺 believes that comprehensive decisions based on market data, company policy, and on-the-ground instinct should still be made by the boss. Further out, if Agents can start companies autonomously, people with more capital and compute could deploy more “startup teams” simultaneously, potentially concentrating resources even further among existing owners.
潘乱 ultimately traces his own “comeback from a second-tier university in northern Jiangsu” to the same principle: a degree is only a ticket through the door; the long-term difference lies in “how you collect information, process information, refine information, and output information.” AI may lower the barriers at each step, but it has not decided for anyone why they should set out.