Summary

ByteDance and Meituan’s seemingly opposite management styles are rooted in 2 different business economics. 魏小康 characterizes ByteDance as “short-cycle, short-chain, high-gross-margin” and Meituan as “long-cycle, long-chain, low-gross-margin”: the former is closer to Google, broadly recruiting strong people, while the latter resembles Amazon, using operating plans and long-term iteration to build retail fundamentals. In his view, “only Meituan and Pinduoduo really understand how to do retail”; Alibaba and ByteDance understand advertising better.
Once a founder’s intelligence clears the threshold, curiosity, low ego, motivation and energy are what create separation. By low ego, 魏小康 means being “highly confident yet highly humble”: hard to knock down, but able to take in feedback and recognize what one does not know. Founders of multibillion-dollar companies do not necessarily need Olympiad-level intelligence; seeing the industry clearly and catching the wave matters more, because an entrepreneur is “just a wave on the tide of the times.”
The core of organizational building at a startup is not elaborate制度, but getting the business and its people to operate around the same goal. Short-cycle businesses can use public OKRs, while long-chain businesses are better served by Amazon- or Meituan-style operating plans. Even in a 10-person company, people may be more than “10% apart in their understanding” of the week’s objective. Spending a few hours each month on alignment can save the entire team 10% of its time.
Recruiting should take up 80%–90% of the people function, with 60%–80% of recruiting work moved upstream into defining demand and building candidate supply. Strong hires typically take 3-4 months from search kickoff to start date; waiting until the business is already in trouble means “not even an immortal could fix it.” Founders should continuously answer 3 questions: what kind of people the business problem requires, how many, and when they are needed—and build the candidate pool in advance.
Interviews are far less reliable than managers think; reference checks are the highest-leverage step in judging talent. 魏小康 estimates that even many top executives get at least 10%–20% of interview judgments wrong, while a one-third miss rate is hardly unusual for founders. Strong candidates are easy to identify, but a carefully prepared 50- or 60-point candidate can disguise themselves as a 70- or 80-point candidate. The most effective approach is to find genuinely credible former managers and collaborators and “figure out what kind of person they really are.”
AI is reshaping organizations from functional assembly lines into lean squads organized around business modules, but 魏小康 remains skeptical of OPC. “100% AI coding” is already a hard requirement for programmers in his book; once one person can take a product from definition through testing, and sales and growth staff can also code, team efficiency could rise 5x or 10x. He still believes each direction needs 1 or 2 strong operators working together: “OPC is a waste of this era and of oneself.”
This AI startup cycle favors 90s and 95s founders with battle experience, resource networks and current frontline involvement—not youth for its own sake. 魏小康 sees the window lasting 2 years at most, and perhaps only 6-12 months at the short end; asking people born in the 2000s to handle fundraising, direction-finding, team formation and brutal competition simultaneously is unfair. 曲凯 argued that the battlefield does not divide by age. 魏小康’s response: doing R&D well inside a mature organization and taking the top job while facing strong rivals are 2 different things. The ideal founder is “mature, but still youthful”（成熟，但是同时又有少年感）.
魏小康’s startup direction is to use AI to rebuild the supply-demand match in recruiting and reduce deception in blue-collar job searches. He wants to minimize manual work in the matching chain and make business managers and recruiters many times more productive. He also says that “99%” of Chinese blue-collar jobseekers—including workers and service staff—have been deceived; the team’s simple goal is to help hundreds of millions of jobseekers get scammed less, find suitable work faster and earn a little more.

Deep dive

1. Business economics determine organizational form; different management methods are not necessarily contradictory

魏小康’s firsthand sample is unusual: he handled recruiting at ByteDance from 2017 through December 2020, during Douyin’s rapid growth and international expansion; from December 2020 through January 2026, he handled recruiting at Meituan before moving into an AI product-manager role.
In his view, China’s mobile internet was still growing from 2021 through 2026, but at a slower pace—precisely when companies began to be tested on their organizational fundamentals. Meituan spent years mastering the “hard, unglamorous work” of retail, while ByteDance excels at businesses with shorter cycles and higher gross margins.
Many compensation, performance-management and recruiting policies at the 2 companies “look as if they operate in reverse,” but a closer look reveals highly similar founder philosophies and cultures. The difference is that each is studying what kind of organization can build durable strength.

2. Once intelligence clears the bar, curiosity and choosing the right era matter more

魏小康 breaks the first common trait of strong entrepreneurs into learning ability and curiosity: some people learn quickly but lack the urge to explore. What really matters is continuing to interrogate a problem—“Why this way? Why not that way? And why must it be this way?”
His view is that the pure-intelligence requirement for building a multibillion-dollar company is simply the top 1% of the 985 population. 曲凯 suggested the threshold might be lower; 魏小康 reiterated that this level of raw intelligence is enough because running a business is not the same as solving the hardest problem in physics.
He recalls a story 贝索斯 shared about studying physics: he and a classmate spent 3 hours unable to solve a partial differential equation, while another friend glanced at it and immediately produced the answer. The point was not that entrepreneurs must be the smartest people in the room, but that elite academic intelligence and understanding an industry are different requirements.
He relayed the core message of a 2021 internal presentation: “We are just a wave on the tide of the times”（我们只是时代浪潮上的一朵浪花）. Recognizing that one has been lifted by the era is what keeps a person clear-eyed about luck and cycles.

3. Low ego means being hard to knock down while admitting what you do not know

魏小康 does not define low ego as a lack of confidence, but as being “highly confident yet highly humble”: holding to one’s judgment through setbacks while genuinely listening to feedback and knowing where one has failed or made mistakes.
曲凯 pressed him on how easily confidence and ego get conflated. 魏小康 added that, beyond rationality and a commitment to facts, low ego may also come from repeated exposure to setbacks. Some people know the facts but still cannot take them in because their pride gets in the way.
Motivation must be strong enough to carry a person through prolonged hardship, but it does not have to come from poverty. 魏小康 stressed that entrepreneurs bear more psychological pain: no one can solve the hardest problems, key subordinates get poached, and competitors launch public attacks. It “has little to do with material desire.”
The final personal requirement is high energy, but individual ability must ultimately translate into organizational building: “You definitely can’t do it through one person alone.” Building a sustainable, battle-ready team is the most important founder-level test.

4. The substance of organization is getting people and the business to operate toward the same goal

Established frameworks simply use different vocabularies: Alibaba talks about the Six Veins of the Sword and the Political Commissar system; Tencent’s “Yang Triangle” asks whether employees have the ability, are willing to do the work, and whether the company’s mechanisms make execution easier.
魏小康 divides organizational work into 2 parts: first, the “select, deploy, develop, retain and exit” cycle for people; second, combining people with the business by setting goals, breaking down tasks, creating collective force, distinguishing strong from weak performance and iterating continuously.
曲凯 questioned whether a startup with simple goals really needs to break them down in detail. 魏小康’s answer was that even with just 3 people, they may not understand the same thing. By 10 people, they will “get at least 10% of the information wrong,” making weekly or monthly rolling alignment more important.
Short, fast-moving businesses can use ByteDance-style public OKRs; businesses with long chains and online-offline coordination can look to Amazon- or Meituan-style operating plans. Spending a few hours in discussion each month may seem slow, but it can save the team 10% of its time.

5. Culture spreads through the core team; complex performance systems and job ladders slow startups down

The cultural language at leading companies is remarkably similar: ByteDance says “candid and clear” and “pursue the extreme”; Meituan says “dare to speak the truth, have the courage to hear ugly truths” and “pursue excellence.” 魏小康’s explanation is that the scientific methods required to get things done are fundamentally shared.
Startups do not need to rush into writing a complete culture manual. At a company with dozens of people, employees watch how the founder works; at a company with hundreds, they watch how 10-plus or 20 core members work. If those people lead by example and explicitly stand against certain behaviors, the culture will replicate downward.
A performance system should not come first. Early-stage bosses must be able to see the frontline, quickly judge who is doing well and decide how much to award. “If you need to spend 1 or 2 months discussing it like a large company, then don’t do it.” When a company’s valuation rises 5x or 10x, the difference between a 10-month and 12-month bonus cycle is not the central issue.
He is equally opposed to building job grades and promotion systems too early. No one has yet explained what jobs should look like in the AI era; fixed standards instead encourage employees to manufacture achievements, write documents and attend review panels for promotion. “With that time, just focus on the business.”

6. Training contributes only 10%; real growth comes from battlefield selection

Both ByteDance and Meituan place a high value on development, but 魏小康’s underlying judgment is blunt: “People cannot be developed through training.” More precisely, people grow, get eliminated and are selected through real work. The company’s first responsibility is to provide a real battlefield.
Meituan’s “721” framework assigns 70% of growth to fighting battles, 20% to learning from strong seniors and 10% to company training. Especially in the new AI cycle, someone with 10-plus years of experience may not understand many things much better than someone with 3-5 years. “Give people a battlefield, and the good ones will sort themselves out.”

7. Recruiting starts with defining demand in advance; interviews come second

魏小康 summarizes the recruiting chain as defining demand, finding supply, assessing candidates, negotiating and following through to onboarding. Most managers spend a great deal of time interviewing without first thinking through the problem at the front end.
His recommended allocation is for demand definition and supply-building to account for 60%–80% of recruiting work, while 80%–90% of the people function’s time should go to recruiting. The core questions are not how many resumes arrive, but what kind of person the business problem requires, how many people are needed and when they must start.
Recruiting often requires 3-4 months of preparation, so demand should be reviewed on a monthly or otherwise appropriate rolling basis. Waiting until the business already has a gap before starting is “wrong—no immortal could solve that problem for you.”
Before building supply, the team should map the market: which companies work in the relevant area, how well they are doing and how many strong operators they have. The team must also face reality: it may currently need people who are merely above the industry average, rather than competing for the top 3% or top 5%.

8. Talent standards depend on how mature the battlefield is; abstract high-potential cannot be pursued apart from the business

If the business and offline supply chain are already mature, 魏小康 believes the company should find strong operators who know the area and can quickly assemble the necessary resources. But when exploring entirely new questions such as how search should work in the AI era, the gap in accumulated experience between newcomers and veterans may be smaller.
魏小康 believes candidates earning no more than RMB1M–RMB2M a year are generally sufficient if they have no obvious weaknesses across the 4 or 5 qualities discussed earlier and possess 1 or 2 clear strengths. Above RMB2M in annual compensation, the difference comes more from long-term motivation and energy than from intelligence.
The choice of where to work can quickly multiply the payoff. In recent years, choosing ByteDance, ByteDance’s Doubao, an AI application company or several large insurers could have produced a 10x or even 100x difference in returns over 2-3 years. Some people dropped out along the way, however, so the outcome cannot be measured purely in financial terms.
曲凯 proposed “putting a batch of people on the battlefield and keeping whoever fights their way through.” 魏小康 explicitly disagreed: the organization serves the business, not the elimination process. A startup with limited resources should first assemble the team with the highest probability of success, using experienced operators where necessary. If someone later cannot keep up, the company should grant the options they are due and part ways amicably.

9. Trusted networks and credible reference checks are worth more than adding interview rounds

The natural disadvantage for a startup expanding its talent supply is a weak brand, making the founder’s network, former colleagues, employee referrals and social-media reach most important; headhunters and job sites come later. “Find a few reliable people around you, then figure out how to bring the reliable people around them into the company.”
Strong candidates are often easy to identify. The hard part is that carefully prepared unreliable candidates can hide their flaws. 曲凯 added that people scoring 90-plus and people scoring in the teens can sometimes both appear “amazing”; distinguishing them in a 1- or 2-hour interview is inherently difficult.
魏小康 estimates that many senior executives get at least 10%–20% of their interview judgments wrong, and that a one-third miss rate can occur during the startup phase. Even after working with someone for 1-2 years, 360-degree evaluations can still diverge dramatically. That is why he puts credible reference checks first in the assessment process.
A 3- or 4-person team should have the founder do sourcing first rather than hire an HR person immediately. The founder can use AI to search and screen GitHub, tech communities, Zhihu, Xiaohongshu and job sites; HR can then find contact details and handle outreach, serving as “the connection between AI and the real world.”

10. The core of recruiting negotiation is a 3- to 5-year outlook, not a few thousand yuan of salary difference

The people a startup genuinely wants are usually not moving just for a 30%–50% raise. 魏小康 cares most about what a candidate wants to achieve over the next 3-5 years; if the company can solve that underlying need, money is easier to negotiate.
曲凯 pointed out that many people do not even know what they want. 魏小康 agreed, saying 80%–90% of employees at large companies cannot clearly answer “Why are you leaving?” or “What problem do you want your next job to solve?” The negotiation can resemble career planning or “a therapy session.”
The interviewer needs to ask what genuinely made the candidate happy over the past several years: professional achievement, influencing others, recognition or money. For experienced people who already earn well, “a little more money won’t make them happy.” The company must find the intrinsic drive that can sustain long-term commitment.
Compensation can be generous, but the work-hours model needs to be assessed separately. ByteDance might once have raised pay from 100 to 140 or 150 while introducing a 6-day week every other week; Pinduoduo might have offered 170 or 180 with only 1 day off each week. The hourly economics still worked, but 魏小康 reserved judgment: “Whether the AI era still requires this is open to debate.”

11. Google and Amazon represent 2 recruiting machines worth studying but not copying

The Google-style model serves short-cycle, high-gross-margin businesses: an independent sourcing team first brings strong people into the funnel; an interview committee not tied to a specific role makes a back-to-back judgment on whether the candidate meets the company standard; after hiring, the candidate and 8 or 10 internal teams choose each other.
The Amazon-style model serves long-cycle, low-gross-margin businesses: the company spends 1-2 months preparing an annual operating plan that lays out goals, budgets, headcount, levels and costs, then has independent interviewers and the Bar Raiser jointly conduct the debrief.
As 魏小康 summarizes it, Amazon requires each new hire to rank in the top 5% of the existing team while eliminating 10% every year; after 3-4 years, the team is systematically refreshed. Both machines are too heavy to install directly in a startup, but back-to-back interviews are worth retaining.
Google now deploys a recruiting organization numbering several thousand; ByteDance at its peak also had several thousand recruiters, including employees, interns and outsourced staff. For R&D recruiting, having 1 person manage 2 or 3 interns and complete roughly 40 hires a year is considered ideal. A startup should cut that target in half, while an HR person who lands 8-10 genuinely strong hires a year is “an outstanding HR.”

12. AI is turning functional assembly lines into business squads and sharply reducing coordination costs

魏小康 is not addressing foundation models that have already entered the knockout rounds. But even an application company valued at a few hundred million or several billion dollars may need only a few people or a few dozen people. He remains skeptical of OPC and still wants 1 or 2 strong operators working together in each critical area.
His analogy is that today is “at worst, the beginning of the PC internet era”: people who understood the internet then had opportunities to build social networks, portals and search engines, rather than limiting themselves to personal websites. “OPC is a waste of this era and of oneself.”
For programmers, “100% AI coding” is already a hard requirement. The boundaries between product, frontend, backend, testing and algorithms are disappearing; one person can take a business module from product definition through testing, greatly reducing communication and coordination.
Once sales, growth, audit and other non-R&D teams can code, their own productivity may rise 5x or 10x, and they no longer need to wait continuously for support from product and engineering. The product and engineering team only needs to build some basic infrastructure, after which the company’s operating tempo changes fundamentally.

13. The AI startup window favors experienced, focused leaders, and recruiting itself will be reinvented

魏小康 differs from the many VCs who favor people born in the 2000s: the commercial window lasts “at most 2 years at the long end, and perhaps only 6-12 months at the short end.” The top leader must quickly complete fundraising, explore a direction and build a team, while navigating reasonable and unreasonable terms as well as legal and illegal competition.
曲凯’s rebuttal was that real battlefields do not divide people by age and that people born in the 2000s can also emerge quickly. 魏小康 responded that following a mature founder and doing R&D well is not the same difficulty as leading an incomplete team into a direct confrontation with experienced, well-funded rivals.
This wave began on November 30, 2022, and 3.5 years have now passed. He is more bullish on 90s and 95s founders who have already started companies or recruited teams and possess networks of former employees, classmates and friends. He also expects people born in the 2000s to break through in the next wave, 4 or 5 years from now. The current wave requires founders who have already worked out roughly 70%–80% of the direction, have a decent ship and can raise money. The ideal state is not being “slick,” but having resources, remaining on the frontline and genuinely wanting to change the industry: “mature, but still youthful.”
His own startup direction is AI recruiting: improving supply-demand matching, minimizing manual intervention and addressing deception in the blue-collar labor market. 魏小康 says that “99%” of Chinese blue-collar jobseekers—including workers and service staff—have been deceived. The longer-term question remains unresolved: once AI makes cross-functional collaboration much simpler, entrepreneurs will still need to work out the “organizational model of the AI era” together.

AI:AM #4: Cameron on Model Consciousness, Duvenaud's Gradual Disempowerment, swyx's AI-Eng Alpha How We Got Fred Wilson, Benchmark and Index to Invest $94M | Why Robinhood's Strategy is Wrong