Crossover: Maima CEO Lin Fan on 2025 AI Talent and Gig Work
Summary
- Meta’s $10M annual packages for top AI talent make sense first and foremost on an ROI basis. A company spending $10B-plus a year on compute could save $1B if someone cut costs by 10%; a 0.1% improvement in Meta’s nearly $200B advertising revenue would add $200M. Against that backdrop, a $10M salary is hardly excessive. DeepSeek was cited as an example of dramatically lower costs, but Lin Fan then used the more conservative 10% savings case to make the same point.
- More important than ROI, Silicon Valley’s core teams still believe AGI could arrive in roughly 3 years—and see it as a matter of competition and survival. Lin Fan did not expect that before his trip, but conversations with core personnel at OpenAI and Google left him “80% convinced”: reinforcement learning has produced discontinuous gains in technical tasks such as coding; models can use “high-level judgment but low-level execution” to provide feedback for other models; and human-AI interaction continues to generate new, filterable data.
- China’s major tech companies have shifted from layoffs and cost-cutting to “spend more to become more efficient”—deciding where and on whom to spend in order to amplify productivity. According to the report, AI job postings have grown by the high teens to high 20s month over month since February, roughly 10,000 additional AI workers each month have marked themselves “open to opportunities,” and about 40% of employees at the six leading AI startups have done so since July. AI hiring at major tech companies is generally up more than 10x year on year. High-performer ratios, internal mobility within AI divisions and stock-vesting schedules are all being redesigned; the talent war now extends beyond pay to organizational structure and incentives.
- AI’s near-term impact on employment is not the disappearance of entire professions, but the contraction of junior roles and a higher capability bar—and Lin Fan believes China may lag the US by 1–2 years. Silicon Valley startups that once needed 8–10 programmers may now operate with 1–2; “1 core programmer plus AI” is roughly equivalent to the old 5-person team. In China, hiring for embedded, Python, C++, data development, frontend, testing, Android and iOS roles is down about 20%–30% year on year, while executive hiring is down roughly one-quarter. Lin calls the current state a short-term equilibrium: “No new batch of jobs has appeared out of thin air, and no entire job category has been replaced.”
- Scarcity pricing for talent has spread to papers, open-source code and interns. The compensation ceiling for elite AI talent is roughly $5M–$10M a year in the US and RMB5M–RMB10M in China; beyond that, “salary stops meaning anything.” A small number of top interns can earn RMB4,000–RMB5,000 a day, and interns now account for the high teens of AI hiring. AI talent is about 20% bachelor’s, 54% master’s and 25% PhD; PhDs even outnumber bachelor’s graduates. Citation counts and GitHub downloads have also become rough pricing signals.
- Lin Fan sees the next structural shift in work not as ordinary remote work, but as “labor moving to the cloud” after machines moved to the cloud. Over the next 2–3 years, hiring will shift from testing whether someone can work independently to whether “a person and AI can deliver results together.” Within 5–10 years, decomposable sales, operations, R&D and accounting tasks could be assigned through cloud-based human-machine platforms, turning white-collar work into gigs. Income may remain broadly flat—and rise for the most driven—but organizational belonging and career fulfillment will weaken; the top 5% or 10% may remain in traditional large companies.
- The most urgent personal strategy is to turn AI from an occasional tool into a work habit used at least 10 times a day. Lin’s passing grade is blunt: “If you only use it 10 times a week, you need to build the habit immediately.” The longer-term defenses are complex problems, cross-functional collaboration and experience that has not yet been digitized. He says 95% of people may eventually not need to work; the host summarized that endpoint as “retirement from birth.” Hobbies therefore matter—not as small talk, but as a source of meaning once work no longer supplies identity.
Deep dive
1. Meta’s $10M annual salaries start with a cost-benefit calculation
This talent war began with Meta investing more than $14B for a 49% stake in an unnamed target, followed by a person surnamed Wang joining Meta; the original does not provide a verifiable name for either the target or the individual. Meta then continued aggressively recruiting from OpenAI, Apple and other companies.
Lin Fan reduced the media’s dramatic framing to a basic business principle: “Go back to the essence of business: input versus output.” DeepSeek was cited as an example of sharply lower costs, but Lin then used the conservative 10% savings case to explain the logic.
With Meta spending at least $10B a year on AI hardware, 10% represents $1B; a boss who paid $10M to hire someone capable of delivering that saving would find it “very, very worthwhile.”
The same math works on the revenue side: Meta’s advertising revenue is close to $200B, so a 0.1% improvement would add $200M. Sky-high compensation is not charity; it is a small investment against high operating leverage.
2. Meta’s ad recovery came from previous-generation AI—and amplified this generation’s talent war
Meta’s advertising revenue has roughly doubled over the past 3 years. Lin’s read from internal conversations is that the growth genuinely reflects better ad performance enabled by technology, but the “AI” narrative in earnings disclosures also contains some packaging.
The mechanism was not the latest generation of large models, but previous-generation deep learning and recommendation technology. After Apple restricted Meta’s access to identity-related data, Meta used deep learning to fill in the gaps and repair ad performance. “You can’t say deep learning isn’t AI,” but the distinction between generations still matters.
ByteDance had long used the same approach in advertising and recommendations. Meta happened to benefit from intense capital-market interest in AI, giving its operational improvement greater stock-market leverage and strengthening its ability to keep spending heavily.
3. Lin Fan moved from skepticism to being “80% convinced” AGI will arrive in roughly 3 years
Companies in China and outside the US generally believe AI has hit data and reinforcement-learning bottlenecks over the past year, leaving AGI far away. Core personnel at OpenAI and Google gave Lin the opposite impression: “For every bottleneck the outside world sees, people still have some solutions.”
When he visited Silicon Valley last year, there were still clear factions within core teams. This year, a much stronger shared conviction had formed. Lin admitted he had not expected to leave the trip roughly 80% convinced.
The first source of confidence is a step-change in coding ability from reinforcement learning. Math and code have objective answers, allowing models to receive clear feedback; once technical capabilities approach the human ceiling, progress could accelerate materially.
The second is “high-level judgment but low-level execution”: a model may not write the best article, but it may be able to judge which of 2 articles is better. Reinforcement learning could therefore move beyond RLHF toward agents or machine feedback, with one model serving as feedback for another and iterating autonomously.
4. Data has not truly run out; models are generating the next training set through human-AI interaction
Traditional internet data available for pretraining is nearly exhausted, but that does not mean data growth has stopped. Ongoing human-AI conversations are generating new data.
Lin acknowledged that interaction data may be lower quality overall than legacy internet content, but models can filter for the high-quality portions and feed them back into pretraining. “There are many, many bottlenecks we consider real,” yet core teams continue to look for solutions one by one.
The key point is not that the problem has been definitively solved, but that core personnel believe executable solutions exist. Lin retains roughly 20% skepticism and does not present 3-year AGI as a certainty.
5. AGI is defined as surpassing humans task by task, with technical fields likely to fall first
Silicon Valley’s AGI discussions borrow the L1–L5 framework from autonomous driving. Rather than arguing first about abstract consciousness, the test is whether AI can outperform ordinary people—and eventually the best people—across concrete tasks.
Lin cited an OpenAI model competing in the IOI and placing eighth globally. In this type of task, AI has already surpassed the overwhelming majority of human contestants.
The consensus path is to “push every technical task to the point where humans have little chance,” then use machine-feedback reinforcement learning on tasks without a single correct answer, such as sales and interviews.
Lin described the genuinely general-purpose players as Google and OpenAI, plus one “half-player” specialized in coding; the original name is too unclear to identify the specific company.
6. For Meta, the AGI ticket has escalated from an investment question to an existential one
Lin’s view is that Meta had once pulled ahead, then fell behind again, making it extremely anxious. If the core teams’ 3-year AGI assessment is right, missing one cycle could undermine the company’s entire competitive position.
The talent war therefore has 2 layers of economics: first, calculable ROI from lower costs and higher ad revenue; second, “future competitiveness—whether I live or die,” which cannot be measured solely against this year’s profit.
This explains why capital markets do not see the bidding war as simple wage inflation: “Whoever breaks through to AGI first gets to survive into the next round of competition.”
7. Once pay crosses a threshold, culture and fulfillment become the key variables in talent mobility
Lin compared monthly salaries of RMB1M and RMB10M. Once both are far beyond basic living needs, a highly paid platform that constantly negates talent and restricts direction may still lose people to a company with adequate resources where they can actually get things done.
“Once compensation exceeds a certain threshold,” candidates care more about culture, colleagues, comfort, whether they can realize their ideas and whether extraordinary pay can continue. This also explains why expensive hires can leave quickly.
He estimates the effective compensation range for top talent at $5M–$10M annually in the US and RMB5M–RMB10M in China. “Beyond that, salary stops meaning anything.”
This tier may account for only a few per thousand of AI talent. If the total pool is 100,000–200,000 people, that implies only a few hundred individuals. Media focus on $10M contracts does not mean those are normal salaries across the market.
8. China and the US are roughly one exchange rate apart—and one layer apart in original problem selection
Lin sees a broad gap “of one exchange rate” from top to bottom, and estimates that roughly 60% of the US AI community is Chinese, creating strong Chinese connections across the 2 talent pools.
The difference is at the very top. Non-Chinese researchers still more often propose new reinforcement-learning directions and drive paradigm shifts; Chinese talent more often optimizes specific methods and supplies strong engineering and algorithmic improvements.
There has not yet been a visible large-scale pull of China-based talent directly into US companies. Chinese companies are more often trying to recruit from the US, while immigration and identity restrictions are also affecting some cross-border moves.
9. The US has already replaced junior programmers; China is still reducing hiring rather than releasing workers at scale
Large US technology companies broadly use “don’t backfill” for programmers. Even departments with acute shortages may replace 1 person after 2 departures or 1 after 3. Business and stock-price improvements can coexist with continued headcount reductions.
The shift is sharper at startups. A recruiting, legal or data-labeling company that once needed 8–10 programmers may now need only 1–2. “1 core programmer plus AI” produces roughly the output of the old 5-person team.
China’s lower labor costs mean companies have less immediate incentive to replace people at the same pace. The more common pattern is slower job growth or no further hiring, rather than mass near-term displacement.
Lin believes China may lag by 1–2 years and enter a more visible phase next year or the year after. After returning home, he told his team, “People won’t be able to keep jumping around for much longer”; programmers broadly said they did not believe him.
10. “Spend more to become more efficient” means major tech companies are spending again—not returning to unrestrained expansion
Maima data shows AI job postings growing by double digits month over month since February, in the high teens to high 20s; on the talent side, roughly 10,000 additional people per month have changed their status to “open to opportunities.”
Since July, roughly 40% of employees at the six leading AI startups have marked themselves open to opportunities, suggesting that company demand and worker mobility are rising together rather than the signal being limited to noisy job ads.
Lin dates the inflection point to the industrial changes following GPT-4’s release. After 2 years of slimming down, the marginal benefit of further savings at major tech companies had become low. The question became where to invest, whom to fund and how much to spend to raise efficiency further.
“Spend more to become more efficient” does not mean expanding every team. Ordinary roles can still shrink; capital, headcount and incentives are simply being concentrated more heavily in AI and core output functions.
11. Performance reviews, internal transfers and stock vesting have all been redesigned as recruiting weapons
The traditional “271” performance distribution is loosening. ByteDance and others may raise the share of top ratings in core teams to 30%, or even 50%, giving more key people immediate positive feedback.
Internal “flow” is also shifting toward AI. Tencent employees might previously have needed to spend 1 year in a role before transferring; for Hunyuan, Yuanbao and other AI businesses, the threshold can fall to 3 months.
Alibaba’s stock incentives may previously have vested 0% in year 1 and 50% in year 2. They may now vest roughly 15%, 25%, 30% and 30% from years 1 through 4. It resembles “smaller meals more often”: someone who has worked for 6 months can see that another 6 months would unlock cash.
The system is designed to counter an external 30% pay raise. If stock is due to vest soon, a candidate may not leave immediately; a rival may need to offer 50% or even 100% more to change the decision.
12. AI job definitions have moved beyond R&D, but the market remains dominated by technical roles
Maima does not count a role as AI merely because its title includes search, advertising or recommendations. It checks the job title, JD and candidate history for technical keywords such as large models and deep learning.
Legacy search, advertising and recommendation roles are not automatically included. If a company explicitly uses large models to rebuild search, ads or recommendations, they count. Product, operations, design and even AI product sales can also qualify because the definition combines technology and function.
Roughly 85% of AI jobs remain technical and about 15% are nontechnical. China has few mature AI products and not many products that can be sold at scale, making AI sales and presales roles particularly scarce.
The fastest-growing areas sit around core models, model applications and the interface between models and specific businesses—not the generic addition of “AI” to a job title.
13. ByteDance and Alibaba lead, while large-model talent spreads from the internet to manufacturing
Lin says AI hiring at major tech companies is generally up more than 10x from last year. That means, for example, hiring traffic rising from 50 people a year to 500, not a 1,000-person department becoming a 10,000-person one.
Within individual companies, ByteDance is “cliff-like” ahead in job postings, at roughly 3–4x Tencent’s level, with Xiaohongshu next. Alibaba is split among Taobao, Alibaba Cloud, Ant and other entities; combined, it may be comparable to ByteDance.
When counted separately, Alibaba’s individual entities rank lower. Combined, Alibaba should rank third or fourth, Tencent fifth, followed by Meituan, JD.com, Baidu, Huawei and others.
Xiaomi and BYD have higher inflow-outflow ratios for AI talent than some internet giants. As underlying large models become more standardized, fintech, autonomous driving, robotics and other hardware sectors are competing for the same talent pool.
14. Employee morale first follows equity gains; the business narrative comes afterward
After Alibaba reorganized and strengthened AI, internal discussions shifted from restraint toward business strategy, organization, stock and proposals of all kinds. Its stock rose from roughly $80 at last year’s low to more than $130, an important channel for morale recovery.
Lin does not romanticize the enthusiasm: “What employees care about most is still performance and promotion, then their own returns.” Once AI gained capital-market validation and employee holdings appreciated, company-level conviction followed.
Meituan is moving in the opposite direction. Intensifying food-delivery competition and a falling stock price have made employees more anxious about whether the moat can hold; JD.com employees are similarly focused on delivery strategy and what the company can still improve.
At private companies, discussions are more product- and lifestyle-oriented. Huawei employees often compare cars with Xiaomi; DJI’s product competitiveness is viewed as steadier, so conversations focus more on office districts, rent, food, clothing, transport and benefits.
15. DeepSeek brought financing back to startups—but made the talent war harder
Before DeepSeek, many investors believed AI was a business for major tech companies. Afterward, both the number and value of AI startup investments rose, and investors became more confident in startups.
The good news is that money and motivation returned. The bad news is that major tech companies also raised pay and benefits. Even with improved financing, startups can still face a resource gap of 2–3 orders of magnitude versus large companies.
Big-company talent comes mainly from other large companies and new graduates, but the scale is so large that “if you shift even slightly, startups cannot absorb it.” The 40% open-to-opportunities rate at the six leading AI startups is one signal of that pressure.
Startups’ relative advantage remains speed. Below the middle-management level, people at large companies often execute tasks slowly; at a small company with a real use case, individuals can validate ideas faster and own the full result.
16. AI talent changes jobs every 2 years on average, but voluntary attrition at major tech companies is only 3%–5%
AI talent stays at a company for about 2 years on average, shorter than the roughly 2.9-year job-changing cycle for new-economy tech talent. But the current industry cycle is still young; this should not be read as everyone leaving after 8 months.
After the economy came under pressure, voluntary attrition at major tech companies fell to roughly 3%–5%. Unless another major tech company offers a clearly higher salary, most people do not want to move voluntarily once they are inside one.
ByteDance is the main destination for talent from all companies. Baidu continues to function as a “Whampoa Military Academy,” supplying the main outflow while also ranking second in poaching from the six leading AI startups.
Tencent and ByteDance poach from each other particularly actively. Many resumes directly show a “Tencent–ByteDance” or “ByteDance–Tencent” path. Alibaba may also be highly active once its entities are consolidated, but that activity is less visible in the data.
17. This cycle’s gains are concentrated among elite technical master’s and PhD holders—not all computer-science graduates
The direct beneficiaries are technical master’s and PhD holders from the C9 universities plus “1 electronics and 1 telecommunications university,” meaning C9 schools and strong computer-science institutions such as Beijing University of Posts and Telecommunications and Xidian University. The industry still relies heavily on reading papers, writing papers and academic-style thinking.
AI talent is roughly 20% bachelor’s, 54% master’s and 25% PhD. PhDs even outnumber bachelor’s graduates.
The school ranking by hiring is Tsinghua, BUPT, Zhejiang University, Peking University, Shanghai Jiao Tong University, Beihang, USTC, Huazhong University of Science and Technology, University of Electronic Science and Technology of China, University of Chinese Academy of Sciences and Harbin Institute of Technology. It is not a copy of the conventional overall university rankings.
Lin himself was surprised that BUPT ranked second. His explanation is only a hypothesis: it supplies many engineers, sits close to industry and has been validated by internet companies over time. HIT’s depth in NLP, robotics and other areas likewise shows why specific disciplinary strengths matter more than overall rank.
18. DeepSeek turned interns from a talent pipeline into RMB5,000-a-day production capacity
Before DeepSeek, companies were more convinced they had to recruit mature core talent from the US. Afterward, they saw that interns and new graduates could also make critical contributions, and interns’ share of AI hiring rapidly rose into the high teens.
A small number of top interns earn RMB4,000–RMB5,000 a day. Companies pay by the day and do not block weekend work; if the person chooses to, they can work roughly 28 days in a month. Lin emphasized that this applies only to a small minority.
Academic talent is beginning to be priced roughly by Google Scholar citations. AI papers with more than 10,000 citations command about $3M–$5M a year in the US, potentially reaching $10M; more than 1,000 citations corresponds to $1M–$2M, and more than 100 to $500K–$1M.
In China, the dollar bands can roughly be divided by one exchange rate. A RMB4,000–RMB5,000 daily rate usually implies more than 100 citations for relevant papers; engineering talent is priced by GitHub code downloads and generally below paper-based talent in the same tier.
19. Standardized junior roles are shrinking first, while executives are also down one-quarter because of an AI capability gap
The most exposed jobs share 2 traits: high standardization and high repetition. The US has already directly replaced junior developers, customer service, operations and labeling workers; China is initially seeing reduced incremental demand.
Maima data shows hiring down roughly 20%–30% year on year for embedded, Python, C++, data development, frontend, testing, Android and iOS roles. Lower-end market, administrative and operations roles are also declining.
Executive positions are down roughly one-quarter, no less than junior development roles. One reason is that the expansion era already stocked companies with managers; another is that the AI era requires AI-literate people in core positions, making it harder for some non-AI executives to transition.
The host cited Salesforce CEO Marc Benioff’s comments on a podcast: sales and customer-service headcount fell from roughly 9,000 to 5,000. Lin has also seen domestic video platforms cut content review, labeling and customer service from 6,000 people to just over 1,000, although Chinese companies rarely disclose such moves.
20. The current state is a short-term equilibrium of higher productivity without mass displacement
Lin’s summary of 2024–2025 is restrained: “We have not seen a new batch of jobs created out of thin air, nor have we seen entire job categories replaced.”
AI is currently raising both output quantity and quality. Because labor remains relatively cheap in China, the dominant form is human-machine collaboration rather than full automation. Marketing teams co-create copy with models; algorithm teams recombine traditional speech, audio-video and autonomous-driving technologies with large models.
The result is that many jobs still exist but no longer require as many people, while the capability model has been raised. Work with short feedback loops and high standardization is shrinking first, but has not disappeared wholesale.
Lin compares the impact to ripples in a pond. The first ring is internet giants and AI startups; the second has reached automobiles, robotics and consumer electronics. There is not yet a clear new class of nontechnical beneficiaries; more industries may be affected only in 2–3 years.
21. Lin Fan witnessed the reversal from “running a stupid model for 2 days” to brute-force scaling working
Lin began a direct PhD in machine learning in 2002, then dropped out after 6 months to join the Sogou startup. At the time, neural networks had about 10 layers and the computational load was already “unimaginable.”
While working on Office Assistant-related projects at Microsoft Research Asia, he ran the best available model on a million-scale dataset for 2 days and 2 nights, only to get a result that was still “very stupid.” He then turned to simpler, faster-converging methods such as BNN and SVM.
Deep learning first stunned him in vision: the field went from struggling to recognize 10 license-plate digits to accurately identifying faces among hundreds of millions of people. He regarded subsequent applications in advertising and natural language as logical extensions.
What truly exceeded his imagination was natural conversation and code generation. He says the “orthodox AI people” of his generation never expected this jump 20 years later; compute, data scale and network optimization ultimately allowed brute-force scaling to overcome the efficiency instincts of the past.
22. Megvii’s experience of “using a hammer to look for nails” shows that technical leadership is not commercial validation
Lin dates deep learning’s first industrial inflection point to around 2010, but when he founded Maima he still believed the mobile internet was the more mature transformation of the moment.
His principle is that a startup need not remain permanently at the technological frontier: “You may well become the wave in front that gets crushed on the beach.” It also has to find the moment when technology, demand and commercialization mature together.
Early Megvii had strong computer-vision technology but was for a time “using a hammer to look for nails.” Games were fun but difficult to monetize; counting visitors in shopping malls was overkill because counting phones’ Wi-Fi signals would do the job.
The business accelerated only when bank facial verification created a clear willingness to pay. The example is also his benchmark for AI application companies today: model capability alone is not a closed loop.
23. Recruiting is the first Maima workflow to be rewritten by AI
Recruiters previously had to search for talent, screen resumes, ask about interest and understand candidates’ preferred directions. Lin believes AI can now perform these steps as well as people, and sometimes better.
One Maima solution can read 5,000 resumes in 3 minutes. In another case, it began with 1,600 candidates, communicated with roughly 1,200 and identified about 120 interested prospects.
HR feedback is that the productivity gain is dramatic: “It’s hard to work without it.” Lin envisions decoupling, accelerating and reconnecting every step, eventually compressing recruiting cycles measured in weeks or months into days or even hours.
Serving recruiters will also serve candidates. Job seekers may no longer need to repeat information already explained to many HR teams; they would only need to personally explain what the AI does not yet know.
24. “AI interview cheating” is a transitional conflict between an old evaluation system and a new mode of production
The host mentioned Fellow Round and another recently popular interview assistant; the original transcription of the latter product and its investor is unclear. Lin believes such tools are themselves transitional-stage products or features.
His analogy is the abacus, calculator and Excel. In the old system, accounting candidates were tested on the abacus; once correct delivery could be achieved with a calculator or Excel, continuing to test abacus skills meant testing the wrong capability.
Cheng Manqi’s counterargument is that recruiters sometimes use questions to observe potential, not merely obtain a one-off result. Lin did not deny the value of potential; he asked whether the relevant test is a person’s unassisted ability or their ability to use AI.
A market-role writing task can now be completed by a candidate alone or with GPT in 10–20 minutes. Product managers are beginning to write prototype code; programmers are drawing design mockups and gathering user feedback. Capability boundaries have blurred. Machines may simulate several days of trial work, but Lin believes a broader gig-work workflow may arrive sooner.
25. The 3 nodes of work are AI-collaborative hiring, labor in the cloud and “retirement from birth”
Over the next 2–3 years, hiring will shift from asking whether a candidate can complete a task independently to asking whether they can deliver a result with AI. Specific interview standards have not yet formed, but the transition has begun.
Within 5–10 years, white-collar work may become platformized in the same way Didi and Meituan organize blue-collar tasks. Sales, operations, product, marketing and R&D could be broken into transferable tasks completed by cloud-based machine systems and small numbers of people: “from machines in the cloud to labor in the cloud.”
The more distant endpoint is the replacement of almost all work by AI. Lin says 95% of people may not need to work; the host summarized this as “retirement from birth.” Lin also explicitly admits he has not worked out what happens between a massive gig platform and universal retirement.
The timetable depends entirely on AGI progress. The more concrete prerequisite for labor in the cloud is how many tasks can be decomposed and transferred. If strong coupling can be removed, cloud-based execution becomes possible.
26. Labor in the cloud would create platform power while weakening most people’s organizational belonging
Cheng Manqi asked whether, if competitors A and B both outsource sales to platform C, the platform would control matching between consumers and suppliers, like an advertising platform, and absorb most of the profit.
Lin’s distinction is that if the company still determines the customer list and whom to call, the platform mainly provides execution. If the platform also fully controls target selection, it would indeed become an advertising-platform-style allocator and could take a large share of profits.
He expects the top 5%–10% to remain in large companies, with most others moving into outsourcing and gig systems. This would not mean only “a certain number” of people get work; “enough people” could still have jobs, but the work would be fragmented.
Income may remain broadly flat. Skilled, diligent workers could earn more; most people may earn roughly 10% less than before; and the unmotivated could survive on very little. Career fulfillment would be “definitely lower,” as mission, identity and organizational belonging weaken.
27. An unnamed unicorn has turned the programmer career ladder into a weekly up-or-down system
Lin described a real global company that recruits programmers to label data for OpenAI and uses its own agents to handle Wall Street tasks, with more than 1,000 participants.
Newcomers start by writing code. The weakest are eliminated; the best move up to code review. Strong reviewers become agent builders, followed by agent reviewers who inspect the people building agents, creating a 4-level system.
This is not the traditional programmer, senior programmer and systems engineer ladder. It is assessed continuously week by week, with people able to move up or down. Career development remains, but stable titles disappear.
Participants in India earn about $100K–$150K a year—roughly half Silicon Valley pay—while living in Bangalore. Talent from Eastern Europe and Russia also participates; some Russians need to relocate to South America. Chinese identity restrictions mean participants typically enter using Malaysian or Singaporean identities.
28. Large-model companies, tech giants and recruiting platforms may all compete for the labor-cloud gateway
Large-model companies have the underlying capability and can extend into task execution as AGI improves. They have a natural path from tool providers to labor platforms.
Major tech companies have reusable labor pools and credibility, much as they evolved from using machines internally to offering cloud services. Internal programmers could serve not only their own company but also outside startups.
Recruiting companies already possess supply, demand, evaluation and matching infrastructure. Lin explicitly believes Maima has an opportunity, but it is starting with AI-enabled recruiting: first decoupling each link in the chain, then trying to connect them into a recruiting or white-collar gig platform.
He is not worried about making the framework public: “There are plenty of people with ideas.” More important than a conservative idea is finding like-minded people and solving practical problems such as task decomposition.
29. Using AI 10 times a day is Lin Fan’s minimum passing grade for white-collar workers
Lin’s first recommendation to all white-collar workers is to pay for an AI membership and bring stronger features into daily work. For every task, try handing it to AI once and see what it can do.
His quantitative standard is at least 10 uses a day: “If you only use it 10 times a week, you need to build the habit immediately.” Over the next 2–3 years, people who still lack human-machine collaboration skills will gradually be eliminated.
The second line of defense is complex problems and cross-domain collaboration: business judgments with many threads and clues, and coordination between biology and computing or product and operations and R&D. Models still lack complete digital data for quickly learning these tasks.
A US survey of more than 1,300 parents ranked plumbers, HVAC repair technicians, electricians and nurses as the most AI-resistant jobs, with lawyers fifth. Lin believes lawyers are already clearly affected, embodied intelligence will catch up, and AI may perform nursing work in 10–15 years.
30. Biology is Lin Fan’s family bet for his daughter; hobbies are his public advice for the post-work era
Lin’s daughter, who has just started university, ultimately chose biology. His logic is that silicon-based life has advanced rapidly over decades, while genes and the basic structure of carbon-based life remain poorly understood. Large models could become a new tool for biology, much as calculus accelerated physics.
Biochemistry, chemical engineering, materials science and environmental science became “dead-end” fields because of repetitive experiments and probabilistic trial and error; AI and automation may improve search efficiency. But Lin specifically withdrew any universal recommendation: “I’m not promoting it. I’m only promoting it to my own daughter.”
For a future without work, he would rather discuss interests than careers. Short-video scrolling, shopping and eating desserts do not count as real hobbies; playing mahjong moderately with 3–5 friends can count, provided one does not stay up late or gamble. He himself attracted spectators by practicing swordplay and tai chi, and later shifted toward the outdoors and mountaineering.
Work currently takes about one-third of life and nearly half of waking time. People who have worked for a long time will find it difficult to separate work from self-identity, but future native generations may not. Lin compares them with aristocrats who do not need to work: “As long as you can find something you like, life can still have meaning.”
Verification Notes
- The original transcription is unclear on the Meta investment target, the person surnamed Wang, the coding-specialized company and one interview assistant; this edition does not invent specific entities. The source also gives both 2012 and 2013 as Maima’s founding year, so no specific year is asserted in the text.