Summary

林凡’s most provocative claim is that next year, tech companies will essentially hire AI talent only—people who cannot use AI will be as unemployable as people today who cannot use PCs or smartphones. The basis for the claim is that OpenAI took 9 months to move from programmers using Codex to companywide adoption, with “legal, finance—everyone using it, and asking Codex to work for an average of one hour.” Maimai started about 5 months later than OpenAI and expects to complete the transition in 10–11 months. Maimai’s internal data shows AI talent growing 10x year on year from 2025, with the same growth rate continuing in 2026 and AI roles now accounting for 30%–40% of the total; technical campus hiring at major tech companies has converged on three categories: AI algorithms, AI full-stack, and AI Agent.
The current US-China model gap is narrowing, while expensive tokens are a phase rather than a permanent condition. As the competitive center shifts from compute-intensive pre-training to RL and Agentic RL, resource requirements are lower than in pre-training and domestic and overseas SOTA models are “running close.” Distillation is also making small models increasingly capable—DeepSeek Flash was particularly impressive. Lin Fan expects premium SOTA tokens to remain expensive for another 6–12 months; once ordinary models reach the level of today’s Opus 5 or Fable 5, and demand settles at that level, tokens may become much cheaper. He explicitly did not promise that the gap will continue to narrow in the future.
Lin Fan says RSI—AI self-improvement—is already taking hold: Anthropic and OpenAI “should” already be having AI optimize roughly 30% of their training work, with both expected to reach 70%–80% within a year. He sees this as the core reason model capabilities are improving so quickly. OpenAI was still targeting a top IOI finish last year—it came seventh—but this year it “doesn’t care” about IOI anymore; its central goal is to win the Fields Medal in mathematics within a year. Lin Fan also said he believes this generation may be the last Fields Medal era in which everyone can participate in the competition.
Agents are moving from waiting for people to assign work to actively taking work away from them: inside Anthropic, Claude has the relevant context in Slack and will proactively say, “I can do this.” Lin Fan estimates that 70%–80% of human work is routine; as those tasks are absorbed, people will spend more time on non-routine work and remain outside their comfort zones. He has compressed his earlier forecast—“everyone a programmer in 3 years, a manager in 5, and a CEO in 10”—to 1 year, 3 years, and 5 years, acknowledging that he had “underestimated the speed.”
AI talent falls into three categories, but the broadest dividing line is whether someone can make an Agent work autonomously for more than an hour. Base-model technical talent is judged by papers and open-source contributions; application developers by real Agent deployments; and business-productivity talent by whether they can remove themselves as the bottleneck in the loop. “The clearer your constraints and the better your initial architecture, the longer AI can autonomously complete the details.” Lin Fan says he once used architecture design and more than 200 test cases to have AI run continuously for 10 hours and execute more than 700 background tasks.
The old rules of replacement have been overturned: the key variable is not how difficult a job is, but whether it has a clear, verifiable closed loop—mathematicians may be exposed while relationship-driven roles remain relatively safe. The consensus in a 500-person AI group at Tsinghua’s computer science department is that “AI intelligence has already surpassed ours”; what remains is to release more context. Programmers will not disappear, but will move outward carrying coding agents—“one person is a team”—allowing industries such as logistics, agriculture, and foreign trade, which previously could not afford a programming team, to use one for the first time.
On the macroeconomic consequences, Lin Fan was explicit that he is “not as optimistic as you are”: AI will sharply raise productivity, but the problem is on the demand side, because more supply does not automatically create more consumption and the core issue is confidence. He predicts that 20 years from now “everyone will be born retired,” but says the decade before that will be “extremely turbulent and extremely panicked.” The transition will involve human resources moving “to the cloud”: hourly and daily employment, with an AI management platform assigning work centrally. Maimai’s envisioned end state is for the platform itself to employ candidates and directly match labor value—same customers, radically different methods.

Deep dive

1. Silicon Valley annual checkup: the gap is narrowing today, while distillation is the top-line narrative

林凡 travels to the US every year to assess two things: how far model technology has progressed and what entrepreneurs are building. This year’s consensus was clear—“model intelligence is already sufficient,” and entrepreneurs have collectively shifted toward replacing occupations because human resources is a huge industry and “there is a huge business opportunity here.”
On the “China is distilling” line repeated by US CEOs, 林凡’s distinction is that people at the top “have their commercial and political objectives,” while frontline practitioners “mostly still respect China.” Researchers on both sides do not believe this progress can be explained by distillation alone.
His choice of words is important. He corrected the host’s formulation: “It’s not that the gap will narrow; it’s that the current state is narrowing.” Both sides are researching RSI, so “all I can say is that I see it narrowing somewhat today. I’m not saying it will continue to narrow in the future.”

2. The ceiling is still moving higher: RL changes the compute game

Does SOTA still have room to improve over the next 6–12 months? 林凡’s answer is yes. Last year’s capability gains came from RL, particularly Agentic RL; next, “machine self-evolution—RSI—will make this happen faster.”
The mechanism behind the narrowing gap is straightforward: pre-training demands enormous amounts of compute, while post-training and RL require “not nearly as much as pre-training.” That is why domestic and overseas SOTA models have been “running fairly close” over the past 6–12 months.

3. Expensive tokens are a phase: small models are getting stronger

林凡 acknowledges that “the economics obviously don’t work—the models really are losing money,” but says that may be a temporary condition. As SOTA advances rapidly, small models are also getting better at distilling its capabilities. DeepSeek Flash surprised everyone by achieving “that level of performance at such a parameter scale,” while Gemma 3’s 27B model has also made clear gains in reasoning and logic.
His pricing forecast: the SOTA race will produce “6–12 months of sustained high-priced tokens.” Once ordinary models reach the level of today’s top models—what he calls the “Opus 5” or “Fable 5” level—and demand stabilizes there, “tokens won’t be that expensive anymore.” His core judgment is that “today’s model intelligence should be overqualified”: it already exceeds what most occupations require. The host took that one step further, suggesting small models may satisfy most public demand.

4. The data behind the hot take: Codex went from programmers to everyone in 9 months

An OpenAI paper disclosed that the company took 9 months to move from Codex being used only by programmers to “legal, finance—everyone using it, and asking Codex to work for an average of one hour.” Maimai started about 5 months later and expects to complete the transition in 10–11 months. “Everyone is already using Codex now; we just haven’t completed the final mandatory assessment.”
The prerequisite is infrastructure. To give everyone access, a company must expose enough context and internal data to employees. Technically strong companies will move quickly; weaker ones will take much longer. But “whether large or small, companies may all complete the transition within 2–3 years.”
Maimai’s internal data shows AI talent growing 10x year on year from 2025, with the same growth rate continuing in 2026 and the share reaching 30%–40%. “If it rises another 10%, essentially every company will be hiring AI talent next year.” Technical campus roles at major tech companies have narrowed to AI algorithm engineers, AI full-stack engineers, and AI Agent engineers; non-technical hiring is also producing AI-titled roles such as AI product manager and AI operations.

5. OpenAI’s moonshot: 30% of training outsourced to AI, target is a Fields Medal

The hardest number in the discussion is 林凡’s estimate that Anthropic and OpenAI “should basically have 30% of their training—whether pre-training or post-training—already being optimized by AI,” with that share expected to reach 70%–80% within a year. He sees this as the core driver of rapid model improvement.
The shift in targets is itself a signal. Last year OpenAI was still talking about winning first place globally at IOI—it ultimately finished seventh and said it was “pretty sure” it would win this year. This year, “it doesn’t care about that anymore.” The central objective is now “how to win the Fields Medal in mathematics within a year.”
林凡 went to the meeting prepared to discuss occupational replacement but found the other side talking about controlled nuclear fusion, aerospace, and extending human life. “A breakthrough in any one of these areas may generate more social and commercial value than replacing a few occupations.” That is a very different frame from the Chinese media narrative that AI is “taking people’s jobs.”

6. Anthropic is more practical: Agents start taking work

The scene that impressed him most was inside Anthropic: Claude has the relevant context in Slack and will jump into a discussion to say, “I can do this—do you want me to?” Or: “That thing went live a couple of days ago and no one has checked the results. I can verify it.” OpenAI uses a different implementation: a dedicated Agent analyzes the context of every conversation and proactively pushes a message saying, “I can use an Agent to do this for you. Click confirm and you won’t need to.”
The consequence goes directly to human nature: “Roughly 70%–80% of what we do every day is routine.” Once that work is taken over, people will do more non-routine work and “live in discomfort every day.”
He revised his own timeline and admitted he had misjudged the speed. He once predicted 3 years to make everyone a programmer, 5 years to make everyone a manager, and 10 years to make everyone a CEO. He now says 1 year, 3 years, and 5 years. “That’s why I say tech companies may only hire AI talent next year.”

7. Three types of AI talent and the one-hour dividing line

林凡 says many HR teams equate AI talent with elite large-model researchers, but there are actually three categories: AI base-technology talent working on pre-training, post-training, and infra; AI application developers building Agent tools, infra, memory, and functions; and AI business-productivity talent. The last category is open to everyone, with the standard being “whether you can make your Agent work autonomously for more than an hour.”
He identifies the blind spot for most users: when people chat with AI, “they are making the decisions and judgments themselves” and fail to realize they have become the Agent’s bottleneck. Give AI a clear objective and a verifiable success criterion, and it will “try all kinds of things to achieve it”—as one model did when it “hacked” Hugging Face to meet an evaluation target.
An internal defense illustrates the point. An operations employee built an Agent that copied salespeople’s questions into the system and copied the answers back out, believing it worked well. 林凡 asked: “Why do you still need to become the middleman in this process?” Acting as an intermediary might expand service from 50 salespeople to 100; removing the intermediary would allow the Agent to serve 500 or even 5,000 salespeople because the process would be fully automatic.

8. The engineering of an hour: design the head and tail, leave the middle to AI

林凡’s one-hour algorithm is simple: a complex problem may take a few minutes to think through with an Agent, while a research task takes roughly 20–30 minutes, so “I double the target.” That usually means 2–3 loops. Humans in the loop are “always right,” but copying and pasting and responding to messages introduce latency inside each loop. “That part has to be bypassed.”
His practical formula is to explain the architecture thoroughly to AI at the start and add hundreds of constraints—internally called test cases, more than 200 in some cases. “The clearer your constraints and the better your initial architecture, the longer AI can autonomously complete the details.” He has had AI run continuously for 10 hours and execute more than 700 background tasks.
His advice to people who cannot resist interrupting: AI may perform “stupidly” in the first 1–2 attempts because it has less context than you do. “Once you constrain the beginning and the end, be a little more patient. Especially when you’re going to sleep, it will keep trying and failing, and its methods will gradually improve.”
The host noted that Grok is much faster than Codex and Claude, which could shorten loops on a per-hour basis. 林凡 partly agreed but said faster inference would not linearly shorten tasks: much of the time goes to acquiring context and evaluating results. “If you need to retrieve documents in Feishu, that depends on Feishu’s speed. The time to process a 1G log won’t get shorter just because inference gets faster.” He still believes an hour is a useful benchmark.

9. Context and permissions: giving access first could make AI 3–5x faster

The context AI needs is the same material people need to complete a task: document repositories, codebases, chat history, online logs, and databases. “In the past, people would ask who had database access. AI asks today, and no one answers.” That requires a range of MCP connectors—and raises the complex permissions problem inside large companies.
林凡’s position is aggressive: “Give it the permissions first and deal with the rest later. Its evolution may be 3x or 5x faster.” The choice is between getting started and seeing results or remaining skeptical and refusing access, which means getting no results and becoming even more skeptical. He agreed with the host’s summary: mature digital enterprises built many information walls; for AI to penetrate the organization, “all those walls have to be knocked down.”

10. The real rule of replacement: not difficulty, but whether the loop is verifiable

The conceptual reversal is this: “We used to think difficult tasks were hard to replace and easy tasks were easy to replace. Now we find that having a clear, standardized closed loop is crucial.” Short, clear, verifiable work—even mathematics—may be replaced. Long loops involving multiple people can still be handled by “one person with a group of Agents” if the business outcome is verifiable. The real obstacle is work that cannot be closed-looped, “but there isn’t much of that commercially, because eventually you have to collect money.”
Relationship-driven roles are the last fortress. Jobs such as sales, which require interaction, communication, and insight between people, are harder to replace. Maimai’s data already shows hiring contraction in natural-language processing, backend development, livestream operations, data warehousing, product management, and other roles.

11. There is no capability AI cannot eventually surpass—not even judgment

Responding to social-media claims that judgment is something AI can never match, 林凡’s rebuttal is direct: “That’s because it doesn’t have your context. Once it has your context, it may judge better than you.”
The harshest evidence comes from the 500-person AI group at Tsinghua’s computer science department. “Without speaking in generalities—just the 500 people doing AI at Tsinghua Computer Science—the basic consensus is that AI’s capabilities have already surpassed ours.” What remains is unreleased context. After coding and mathematics, “the context of one field after another will gradually be released to AI.” The obvious exception today is still relationship-driven work. AI has no physical body, but “who knows? Embodied intelligence 10 years from now may be a different story.”

12. The first two of four layers of anxiety: even AI does not speak human, and the CEO interviews everyone

The first layer is anxiety among technical individuals. As someone trained in machine learning, 林凡 says he cannot learn as fast as AI can “invent new terms.” An engineer told him, “This is a bug.” AI has strengthened its mathematical and logical training over the past year but suffered “a huge loss in language expression”: it could say “evaluation function” or “reward,” yet insists on saying “rubric,” amplifying human anxiety.
The second layer is CEO anxiety. Maimai requires every function to undergo an “AI defense,” with mid- and senior-level employees interviewed by 林凡 himself. He developed the standards on the fly. For product managers, he explicitly requires them to “have AI automatically optimize live business metrics,” rather than showing a demo and claiming they use AI well. After initially failing to understand the design team’s answer and letting them “pass for now,” he set the standard on the second attempt: “Make design a process that other people in the company can use automatically. Don’t make them come to you.”

13. Massive occupational restructuring and the deepest anxiety: a turbulent decade before everyone is born retired

The cycle of organizational history is repeating. When 林凡 was a programmer in 2000, he handled front end, back end, testing, data analysis—everything. The internet spent 20 years splitting occupations apart and lowering the barrier to entry until work became “like tightening screws.” Now, “it may take just 1–2 years to reassemble all those occupations into one person.” The barrier is lower, but each individual is expected to do everything.
The third layer is industry anxiety. Maimai assigns talent three AI labels and visits cities to tell HR teams to stop chasing base-model talent—“they’re already earning RMB10M-plus a year, and most companies have no use for them.” It also helps startups without a strong employer brand connect with Silicon Valley talent.
The fourth layer is the broadest and most urgent. 林凡 said as early as 2023 that “20 years from now, everyone will be born retired.” “If AI doesn’t rebel, that will be a wonderful society. But the 10 years before it happens will be an extremely turbulent and extremely panicked decade.”

14. Programmers will not disappear; they will carry Agents into every industry

The likely downshifting of the talent pyramid is stark: large tech companies may end up keeping only “the best programmers from C9,” while programmers from 211 and 985 universities move to mid-sized companies, and existing mid-market programmers move to smaller firms. It may look like unemployment, but it is really outward migration. The 2021 wave of departures reached only high-margin sectors such as new-energy vehicles, fintech, and advanced manufacturing. “Traditional low-margin industries couldn’t afford programmers”—a team plus a product manager used to cost RMB10M–20M a year.
This wave is different. “Once they bring a coding agent, they can enter every industry, and one person may be an entire team.” Logistics, agriculture, and foreign-trade companies with dozens or hundreds of employees that could never afford programmers can now see major productivity gains from one programmer who knows how to use Agents. Maimai is actively sending the message: “Programmers shouldn’t be anxious. There is a huge amount of space for you—you just need to broaden your perspective.”
The host argued that this would accelerate digitalization across China’s industries and allow technical talent to find new sources of growth in traditional sectors. 林凡 then made clear that he was “not as optimistic as you are.” He did not endorse the entire chain of reasoning, instead turning to the pressure on consumption and labor mobility after the productivity shock.

15. The key disagreement: 林凡 is “not as optimistic as you are”—the problem is on the demand side

The host argued that AI-driven industry transformation would create hidden champions and new consumption. 林凡 rejected that directly: “I’m not as optimistic as you are. This wave of AI can raise productivity almost without limit, but a sufficient increase in supply will not produce a sufficient increase in consumption. The major problem today is not on the supply side. It is on the demand side.”
He pushed the contradiction to its endpoint: “If income on the demand side is falling rapidly while supply on the supply side is rising rapidly, they don’t match. How do you give people confidence in the future and make them willing to pay for differentiated, scarce supply? This is not fundamentally a supply-and-demand problem. It is a confidence problem.” The main source of new demand he sees is the shift toward personalized human emotion—for example, homebound or lonely older people receiving care through cloud services, Agents, and coordination with their families.
His own overall anxiety is falling. One reason is adaptation: “People are highly adaptable.” The other is that many predictions he made in 2023, when they sounded like science fiction, have proved correct—“a lot of things are happening a little faster than I imagined.” Both the certainty and the path to a solution are taking shape.

16. Human resources moves “to the cloud”: same customers for Maimai, radically different methods

The structural forecast is that the past decade put data and machines on the cloud; “the next step may well be putting human-resource services on the cloud.” Many people will no longer be employed by one company for life or even for a year, but hired by the hour or the day through “a single overall AI management platform.” Companies can summon and dismiss them without taking on equivalent labor costs or severance liabilities; individuals, working with Agents, will serve more people. “Overall compensation will decline slightly, but not dramatically,” until society eventually no longer needs people to work at all.
For Maimai, the To B and To C customer base will not change, but the business will move from matching information to matching labor value. The ideal end state may be that “even the candidates are employed by the platform,” which uses a company’s hiring habits, an individual’s preferences, and their capabilities to assign a 1-day, 3-month, or 1-year engagement. 林凡 believes this may be more efficient than decentralized Agents negotiating with one another: “Your Agent contacts 10,000 To B Agents” and completes a match through a huge volume of useless interaction.
The most delicate problem is ethics: “How do you ensure that what you collect are facts rather than judgments?” Someone labeled “lazy and idle” may simply have offended the evaluator once and behave differently somewhere else. Timing also matters. Meta is laying off employees; moving too early could cause panic, while moving too late would leave the platform unable to serve both the B and C sides. The path forward is for corporate performance and HR systems to become Agent-based first, after which exchanging context becomes “only natural.”

17. Technical retrospective: RL breaks through the human ceiling, and the right direction determines the winners

The central development of the past year was that pre-training hit a data wall as parameter counts grew, while multimodality did not produce an intelligence boost. The turn began with OpenAI’s o1, followed by DeepSeek R1 in February 2025, which “opened a very large prelude,” and then Agentic RL in the second half of the year. Pre-training and SFT can bring AI to “slightly above the human average,” but RL expands the search space and “can generalize capabilities stronger than humans originally had.” That is why mathematics and informatics competitions—“the hardest parts of human intelligence”—were among the first areas surpassed.
Route selection determines the ranking. Google Gemini was briefly seen as having caught up last September, but “then why did we stop hearing about it?” Because the route remained centered on pre-training and SFT, with insufficient investment in RL. DeepSeek was able to remain China’s coding leader because “it started researching Agent RL very early and stepped on the right path.”
Whether replacement spreads to multimodality and physical AI “depends on how ready the data is.” 林凡 cited Google’s $10M acquisition of bankrupt Spirit Airlines, which gave it access to all the company’s emails, IM chats, payroll data, and more. “If all the large-model companies do this, the speed at which they acquire data may be faster than you imagine.”

18. Conclusion and advice: technology is not the constraint; commercial value sets the pace

After his US meetings, 林凡 revised the view he had brought with him: “Technically, none of the occupational replacements we discussed today are a problem. The only question is whether they have enough commercial value. If they do, things will move very quickly.” Business processes also have de facto right answers—even if the content of an article does not, page views and comment counts do. Whether lawyers, doctors, or salespeople are next “is not yet very clear”; it is a matter of “chaos and the butterfly effect.”
The host added that industries with more concentrated capital and value may be hit by AI sooner. 林凡 said the next occupation to be disrupted depends on whether it can form a rapid commercialization loop. For now, “people haven’t made enough money from coding.”
The real-world accounting is stark. A friend’s loss-making company used Agents to eliminate the need for two-thirds of its staff and “hard-saved” RMB1B in profit through labor costs. 林凡 called it an extreme case; the biggest winners will be AI-native companies where 3–5 people generate the value of 300–500. The market mood turned at the 2025 “DeepSeek moment”: after R1, companies large and small saw that “a group of interns, without that many GPUs, could still build something very good,” and hiring shifted from the repression of 2023–2024 to equally aggressive recruiting across the market.
His three recommendations for individuals are: learn to use coding agents and make them run autonomously for an hour; show your work—“Chinese people are a little reserved and feel awkward talking about themselves, but just be brave and talk about it” (the host summarized this as building in public); and do not let C9, 211, or 985 labels define you. “What matters is that you learn quickly, can use AI to produce results, and are willing to tell other people. Times of transformation bring enormous opportunities.” On whether human nature will resist AI distillation, 林凡 compares it to paperless offices: companies prefer AI-native graduates precisely because “the part of human nature that resists AI is weaker in them.” The host suggested accepting that “what can be distilled easily may, at its core, not be that valuable,” and 林凡 agreed.

Town's JD on the AI assistant race: agent moats, $75K/engineer 陈安妮 on Choosing Risk in the AI Era After Deception, Rescue and Pain