Yuanming Hu started this list when Meshy, the 3D AI company he founded, was at $80M in annual recurring revenue, by his own figures. He kept adding to it for about a month, until the number passed $100M. Then, on a business trip over China’s National Day holiday, he spent five days cutting two-thirds of it. The 48 notes that survived, in five sections, were published on WeChat on October 7, 2026.
This is one I strongly agree with.
Last year he wrote a review of Meshy’s first stretch after product-market fit. This one is for the “10 to 100” stage, and it circles three questions: what he wishes someone had told him before this stage, why he became the person he is and how a founder should treat a team, and what he still hasn’t done well.
I. Warmth and principles
Hu traces his habit of putting others first to kind teachers. There was an informatics olympiad coach who always cheered him on, and his MSRA mentor Steve Lin, who patiently fixed his papers. His MIT co-advisors, Fredo and Bill, let him spend his PhD on Taichi and wrote the grant proposals themselves instead of handing that chore to students. His parents told him that losing out a little is a blessing. Putting others first looks like a loss, he argues, but it builds trust and lets people do their best work. That’s why empathy went into Meshy’s culture code at the start.
But a company, unlike a school, has to win to survive. Kindness without principles turns into people-pleasing, and a pushover can’t lead, can’t work with principled people, and drives the best ones away. He wants kindness measured across the whole team and over a long time span.
Fairness is the principle he spends the most time on, and he calls it hard systems engineering. Lazy management makes the dependable people pay. A colleague who never complains gets the dirty work, does it well, and so gets more of it. Without compensation elsewhere, that punishes conscientiousness. Paying people who make noise more than people who quietly deliver, or rewarding overtime over output, is the same mistake. His test before a raise or an equity grant: if everyone’s pay leaked one day, would this look fair?
Two more principles fill out the section. Be candid: call bad work bad, soften the delivery but not the facts, and judge people by whether they solved the problem. If it isn’t working, an exit today beats a slow one later. And weigh the whole picture. Games are addictive because the rewards are instant and the logic is black and white, and real life is neither. A manager trades off short term against long term, individual against team, fairness against efficiency. A mine planted today may go off months or years later, and digging it back out takes twice the effort.
II. Time and founder mode
The section starts from a premise. The person at the top sets the ceiling on what the team understands, and time can’t be replaced, so a founder’s judgment shows in who they spend time with, on what, and in what order. That puts a limit even on kindness. A founder who drains themselves on small things has nothing left for the big decisions.
He models a startup as a system of differential equations in time, where internal and external factors keep acting on each other. The job is to work out those interactions and separate leading indicators from lagging ones (he points to Ray Dalio’s The Changing World Order for this style of thinking). Revenue is the tip of an iceberg. Underneath are talent density, clarity of vision, technical and product strength, culture, morale, shareholder alignment, even the founder’s health. Two companies with similar numbers can be heading in opposite directions.
A friend in Silicon Valley gave him a way to price an hour. Take next year’s valuation, subtract today’s, then divide by 365 and by the working hours in a day. That’s roughly the company’s opportunity cost per hour. Assume half of it comes from the founders’ decisions and a founder can work out what each second of theirs is worth. Once you know the number, a lot of small things stop being worth agonizing over, and a lot of meetings stop making sense. Family and friends are explicitly exempt.
On health, he observes that the best founders have no work-life balance at all. He cites Jensen Huang saying that work is when he relaxes, and prefers “work-life unification.” He sees a tradeoff between health and output and tries to avoid needless harm. In his case that means fewer spicy snacks and McDonald’s spicy wings, less staying up late, some exercise.
A founder needs ego, as long as it serves the team and the vision instead of the need to be right. You have all the context, your reports perhaps half, outsiders maybe 1%, so an outsider’s criticism is aimed at a version of you they imagined. His programmer’s image: the CEO is the try {...} catch {...} in main. There’s no caller above to raise an exception to, and if bad outside advice ends in exit(-1), that’s on you.
Founders are also lonely, and the trap is looking to employees for the emotional support they miss. If a founder needs their approval, it leaks out. Flatterers get rewarded, more people start managing up, and the founder ends up inside an information bubble, sometimes manipulated by the people below. Hu gets his support from elsewhere. He promised himself a Hasselblad if the Series B closed (the photos in the post were taken with it), and he knows that’s lucky, since most startups return nothing. He learns from working with stronger people. And he is glad when colleagues who struggled alongside him grow and do well.
He separates two Chinese words here. The opposite of 快乐, joy, is pain. The opposite of 幸福, closer to fulfillment, is boredom, at least for him. The two are independent. He has had painful fulfillment as well as the joyful kind, but while he’s growing he is never bored.
III. AI and people
Sometimes AI is the army he commands, he writes, and sometimes he is AI’s “carbon-based pet,” his curiosity fed with neatly organized knowledge.
Without fundamentals, AI amplifies confusion. He has asked colleagues to rewrite AI-generated reports by hand because he couldn’t follow them. He mocks the “not this, but that” constructions, the rhyming metaphors, and the “one honest caveat” that makes a model look thoughtful. Before AI, form and content were loosely linked. AI broke that link, and confident nonsense became normal.
He recalls the grading rule in Tsinghua’s Yao class: a correct proof earned full marks, a wrong one zero, and “I don’t know” earned half. That killed the strategy of writing something long enough that the grader gives up looking for the flaw. Meshy’s version is that if you can’t say it clearly, don’t say it. Muddled writing usually means muddled thinking.
Everyone at Meshy has close to unlimited tokens (an earlier post of his was titled “I Work for 10 Claude Codes”), yet interviews still test what he jokingly calls old-method programming, which means writing code by hand. Over several hundred interviews, people with clear hand-written logic were the better problem-solvers, and AI hasn’t changed that. Built on weak fundamentals, AI output is a castle in the air. So in his view AI has made the strong stronger and the weak weaker. It hasn’t lowered the bar for building software.
He also admits that in some fields AI has gone beyond what he imagined. Those fields close the loop in software with no atoms to move (coding), have checkable results (math, verified in Lean), and don’t need specialized knowledge the way law does. The anxiety follows him daily, though it doesn’t cost him sleep. Steam, electricity, computers, and the internet all produced obedient tools. This time the thing people were proudest of, intelligence, has been overtaken in many areas.
Mathematics is his main example. He expects AI to become the main source of proofs soon. Because Lean can check a proof mechanically, mathematicians may know a result is correct without understanding it or learning anything from it. He recommends Scott Aaronson’s “The Age of Wonders and Terrors” and Logan Graves’s “Priest, Monk, and Mathematician,” and quotes the idea that mathematicians will become priests interpreting AI’s oracles, until AI interprets its own proofs better. On a smaller scale that’s already his life. He asks AI about whatever he doesn’t understand and has it build visualizations to learn from.
Do we still need to understand what AI is doing? With coding agents, the practical failure is that you can no longer write the next prompt that makes progress. For production code some understanding is still needed, and he doesn’t know for how long. He sets aside whether carbon-based life should make way for silicon, but says AGI has arrived in a sense, just unevenly distributed across fields.
He ends the section with one thing AI is good for. Take your problems to someone with no stake in the answer, since the person who says you’re sick may just want to sell you medicine. AI comes close to having no stake.
IV. Team and management
A neutron weighs about 940 MeV/c², he notes, while the rest masses of its three quarks add up to about 12 MeV/c². Roughly 99% comes from the motion and field energy of quarks and gluons under the strong force.
Teams are similar. How people interact shapes output as much as what each can do, so Meshy hires for both fundamentals and culture fit. A handful of people can run on friendship. Dozens need a leader who knows the business in detail. Hundreds need principles people actually follow, plus a deep bench of talent.
Meshy pays at the top of the market, partly out of the same kindness and partly because high gross margins allow it. He is blunt that pay can’t outrun what a company earns. With Zhang Yiming, he prefers to call people an investment rather than a cost, because in generative AI almost every result comes from strong employees.
Worse than being short-staffed is over-expansion that dilutes talent density, which he calls an irreversible cancer. Strong people attract strong people, and weak leaders hire weaker ones. His rules: each hire must be stronger than the people already there, if you hesitate don’t hire, and hire only self-driven people, because nobody wants to micromanage.
Vision is what moves a team. People building a ship should long for the sea, or the work is just hammering nails. A founder has to think the vision through personally, since a borrowed one collapses under the team’s questions.
Managers shouldn’t manipulate anyone either, whether reports, bosses, or peers, for instance by negging them. Be fair, work out what each person wants, and fit their role, rewards, and growth to it. He recommends Robin Stern’s book on gaslighting. An early-stage company has to be an extension of its founder, or the founder ends up working against themselves. He quotes Jensen Huang’s advice that you’re building an F1 car you’ll drive yourself, so build one you can drive.
Leave the team room, and don’t kill ideas unless the logic is broken. He uses Andy Grove’s task-relevant maturity, from High Output Management, to decide when restraint works and when to get closer. Because AI companies tend to be led by young technical people, he teaches an internal course called “Management for Good Students.” He has found that strong individual performers often struggle most with managing.
He wants founder mode to be how the whole team works, with leaders able to see two or three levels down and get their hands dirty. The best people want to learn from their leader. Leaders can’t judge work they’ve never done. Technical leaders who don’t know the details turn insecure and start meddling. And real tradeoffs across talent, money, brand, product, and market need the details.
Some calls are like the office thermostat, where somebody is always too hot and somebody too cold. He suggests every new manager play Frostpunk (the first game, not the second). Low performers drive out high performers: after a mess they lower the bar together and blame other departments. So part ways decisively, and kindly if your margins allow.
Then he takes on 老登 (laodeng), internet slang for, roughly, an older person who lectures and acts superior. He turned thirty this year and wonders how younger readers see him. His examples of aging without it are Nintendo, founded 137 years ago, Studio Ghibli, where Hayao Miyazaki is still making films, and Ray Dalio, now 77. They put players, audiences, and readers first. Shigeru Miyamoto would pull a cleaner in from the hallway to play his game and watch silently from behind, which Satoru Iwata called “Miyamoto-san’s over-the-shoulder view.” In this era, Hu writes, not creating new value is the original sin.
He thinks laodeng behavior grows out of insecurity, the fear that the times have moved on and young people no longer listen. Every generation has its own job. When he can no longer do something at the state of the art, he’ll hand it over and set goals instead, the way the Zelda games place towers, shrines, Koroks, and Lynels and let players work out the rest. (Nobody rescues Zelda that way, he jokes. Fine, as long as the players are having fun.) But setting goals for the smartest young people means keeping up with them. Once you stop learning fast it creeps in, whatever your age.
V. Strategy and execution
Most effort is wasted, in startups as anywhere, so time spent deciding what matters beats more heads-down execution. He wants big bets made and losses accepted. He credits his success to luck, and says that isn’t modesty. He’s just used to failing. When the goal is innovation, most attempts fail, so he commits fully while expecting to. Low expectations keep him alert and make him happier. They also make him more conservative.
Resources can stand in for each other: talent density, culture, morale, brand, cash, vision, a founder’s stamina. A weak spot can be covered by strengths elsewhere, as long as it isn’t unacceptably weak.
From Professor Zeng Ming he took the habit of starting from the end. Instead of planning from the cards in your hand, decide what you want to achieve and work out which cards you still need. He believes that lets you do something ten times bigger, and not necessarily harder. “Why can’t I do this yet?” is the basic question. “What do I do today so that I can later?” is the better one.
Tsinghua and MIT taught him that the smartest people around you often work several times harder, and the market only rewards the best result. He left with two firm beliefs: nobody succeeds casually, and whatever luck hands you, your actual ability will take back in the long run. The people there who became leaders in their fields also loved their work, and smart and hardworking loses to smart, hardworking, and in love with it. They were rigorous in experiments and got every punctuation mark right in papers. Hu once thought that level of fussiness was controlling. Then he saw it in top people everywhere, from Miyazaki and Miyamoto to Steve Jobs and his own writing mentors. The condition is that they get priorities right first. Fussing over details is something you earn after the basics are done, and they accept that demands the team doesn’t yet understand will bring pressure, even insults.
The world is a giant makeshift operation, he writes, so doing the basics well puts you in the top 1%. The basics are just hard: know what you want, keep your word, sleep on time, think for yourself, make sure readers can follow what you write, ask only good questions, use your own product daily, pay the biggest contributors more, put facts before face, skip the correct-but-empty statements. So do fewer things. As the company grows he counts each “no” as a gain. Fewer projects leave more time for strategy, fewer events more time with family. Fewer new acquaintances allow deeper friendships, and fewer mismatched hires mean higher talent density.
His last note is that deep thinking can be copied without much risk. It combines market demand, the founder’s and team’s abilities, timing, and the company’s DNA and culture, so only you can execute it at 100. A copier might reach 80, and the market picks the best. If you can’t be the best at something, better not to do it.
The essay closes with a dinner from his childhood. His father was eating with a former student who had left to start a company, and the two of them said that a scientist treats science as home and an entrepreneur treats the enterprise as home (in Chinese, 科学家 and 企业家 both end in 家, “home”). Hu still remembers it more than twenty years later. His own answer to what science and business share is to find a problem important and interesting enough, then do the most basic things properly.
胡渊鸣这份清单是从 Meshy 的 ARR 还在 8000 万美金时开始写的。想到什么记什么，记了一个月左右，ARR 按他文中的统计口径过了一亿美金。国庆出差期间，他又花五天删掉三分之二，最后剩下五个板块、48 条，10 月 7 日发在公众号上。
这篇我特别认可。
去年他写过一篇复盘，讲 Meshy 刚找到 PMF、进入高速增长时的得失。这次公司进入“从 10 到 100”的阶段，他想回答三个问题：进入这个阶段前，希望有人早点告诉自己什么；自己为什么成了现在这种性格，founder 又该怎么和团队相处；还有哪些地方没做好。
一、温度与原则
他“先人后己”的习惯，来自一路上遇到的好老师：中学搞 OI 时一直给他打气的教练，在 MSRA 耐心帮他改 paper 的 Steve Lin，还有 MIT 的两位 co-advisor Fredo 和 Bill。两位导师让他放手做 Taichi，写 grant proposal 这种苦差事也没推给学生。再加上父母常说的“吃亏是福”。他觉得先考虑别人，表面上吃亏，换来的却是很深的信任。Meshy 一成立就把“同理心”写进文化守则，原因就在这儿。
可公司跟学校不一样，得打胜仗才能活下去。善良少了原则，人就成了讨好型的滥好人：管不住团队，跟有原则的人合作不来，还会把优秀的人逼走。所以他说，善要放到整个团队、更长的时间里去看。
他着墨最多的是“公正”，说这是件很难的系统工程。管理者一偷懒，老实人就吃亏。典型的情况是：某人任劳任怨，脏活就都派给他；他做得好，脏活就更多。别处没有补偿的话，等于在“惩罚责任心”。会闹的比踏实干的拿得多，奖励加班而不奖励产出，都是一个毛病。
他给自己的检验是：给谁涨薪、加期权之前先想一下，要是哪天大家的薪资都意外公开了，这个安排还站得住吗？
另外两条原则也很实在。一是客观真诚：不好就说不好，话可以委婉，事实不能改，看人就看他有没有真把问题解决。不合适就早说，今天分开总比拖很久再分开好。
二是全局权衡。游戏让人上瘾，是因为回报即时、规则非黑即白，现实可没这么省事。短期和长期、个人和团队、公平和效率，都得放在一起掂量。雷埋下去，可能几个月甚至几年后才炸；真埋了，就得花双倍力气挖出来。
二、时间与 Founder mode
这一节的前提是：一号位的认知决定团队的上限，时间又不可再生，所以一号位的判断力，最后都落在跟谁、按什么顺序、做什么事上。这也给善良划了一条线。Founder 把精力耗在无谓的地方，到了大决策就没心力了，这对团队一样不负责。
他把创业看成一组随时间 t 演化的常微分方程，内外部因素互相作用。一号位要搞清这些作用关系，分清 leading indicator 和 lagging indicator（他推荐用 Ray Dalio 的《The Changing World Order》来理解这种思路）。营收只是冰山一角，水面下是人才密度、愿景清晰度、技术和产品、文化、士气、股东是否一条心，甚至创始人的身体。只盯营收和估值，判断永远是滞后的。指标差不多的两家公司，可能一家正在加速，另一家已经从根上开始烂了。
硅谷一位朋友教他算时间账：一年后的估值减去今天的估值，除以 365，再除以每天的工作小时数，大致就是全公司每小时的机会成本。假设一半来自创始团队的决策，创始人每秒值多少钱就算得出来了。这个数一清楚，很多事不值得纠结，很多人也不必见了。陪家人朋友除外，那不是能用钱算的。
关于健康，他的观察是：创业阶段最好的 founder 根本没有 WLB。他提到黄仁勋说自己工作的时候就在放松，觉得更贴切的说法是 work-life unification。他在健康与产出之间的取舍，是尽量少做无谓的消耗：少吃辣条和麦辣鸡翅，少熬夜，挤点时间锻炼。
Founder 必须有 ego，但出发点得是团队和愿景，不是证明自己对。你掌握 100% 的 context，下属大概一半，外人可能只有 1%，所以外人的批评，针对的往往是他想象中的那个你。他用了个很程序员的比喻：CEO 是 main 函数里那层 try {...} catch {...}，上面没有人可以 raise exception。听信不靠谱的建议把公司带进沟里，exit(-1) 还得自己扛。
Founder 也很孤独，很多话没法说，能聊到一个层面的大概只有别的 founder。危险在于想从下属那里找情绪价值。一旦需要下属认可，就会不自觉地奖励奉承，大家开始向上管理，founder 最后被包进信息茧房，甚至被下属 PUA。他自己的情绪价值来自别处：答应过自己 B 轮 close 了就买台哈苏，文中的照片就是用它拍的（他也说这很幸运，大多数创业的回报是零）；和更厉害的人共事，能学到东西；还有看到一起打拼的同事成长、过上好日子。
他还区分了“快乐”和“幸福”：快乐的反面是痛苦，幸福的反面是无聊，两者是正交的。这两年他有过快乐的幸福，也有过痛苦的幸福（他说是“酸爽”），但只要还在成长，就不会无聊。
三、AI 与人
他说 AI 有时是他指挥的千军万马，有时他倒成了 AI 养的“碳基小宠物”，被整理好的知识喂着好奇心。
基本功不扎实，AI 只会放大困惑。他会直接请同事把看不懂的 AI 报告手动重写，也吐槽 AI 文章里“不是……而是……”的绕弯逻辑、押韵却不知所云的比喻，还有结尾那句显得很深刻的 “One honest caveat”。以前形式和内容多少挂钩，现在彻底解耦了，一本正经地胡说八道成了常态。
他还记得姚班的考试规则：证明题答对满分，答错零分，写“我不会”给一半。这样一来，“写一大串让阅卷人找不出破绽”的策略就没用了。放到 Meshy，就是说不清楚不如不说。在他看来，表达不清多半是因为没想清。
Meshy 全员几乎无限量用 token（他之前写过《我给 10 个 Claude Code 打工》），可面试照样看重“古法编程”。几百场面试下来，能手写出清晰逻辑的人，解决问题的能力往往更强，有了 AI 也没变。地基不牢，AI 搭出来的就是空中楼阁。他的结论是：AI 没有拉平门槛，反倒让强者更强、弱者更弱。
他也承认，有些领域 AI 已经超出想象了。这些领域能在软件里闭环、不用搬原子（比如写代码），结果可以验证（比如用 Lean 验数学证明），也不需要法律那样的行业知识。这让他每天都有一种对人类未来的焦虑，虽然还不至于失眠。以前几次产业革命造出来的，都是听人使唤的工具；这一次被超越的，是人类最引以为傲的智力。
数学是他举的主要例子。他认为 AI 很快会成为证明定理的主力。Lean 能机械地验证证明，于是人类可能知道结论一定对，却看不懂，也学不到东西。他推荐了 Scott Aaronson 的《The Age of Wonders and Terrors》和 Logan Graves 的《Priest, Monk, and Mathematician》，还引了一个说法：数学家会变成解读 AI 神谕的“祭司”，再往后连这活儿也会被 AI 自己接走。他自己某种程度上已经这样了：不懂就问 AI，让 AI 写可视化工具帮自己学。
那还需不需要看懂 AI 在干嘛？他说在 coding agent 里，完全不懂的代价是写不出下一个能推进问题的 prompt。至少今天写生产代码还得懂一些，能持续多久就不好说了。碳基是否该给硅基让位，他这次不展开，只说某种意义上 AGI 已经实现了，只是在各领域分布不均。
这一节收在一条建议上：遇到问题，找利益无关的人问。说你“有病”的人，可能只是想卖药给你。这么看，AI 至少基本做到了利益无关。
四、团队与管理
他从物理讲起：中子质量约 940 MeV/c²，三个夸克的静质量加起来只有约 12 MeV/c²，剩下约 99% 来自强相互作用下夸克和胶子的运动与场能。
团队也一样，人和人之间的相互作用，跟个人能力一样影响产出。所以 Meshy 招人既看基本功，也看 culture fit。几个人的团队靠兄弟情就行；几十个人，要靠一号位吃透业务细节；到了几百人，就得有真正落地的原则和价值观，外加一支够厚的人才梯队。
Meshy 愿意给市场最高位的薪资。一是与人为善，二是业务毛利高，撑得住。他也说得很直接：薪酬不能超过公司的赚钱能力。他认同张一鸣的说法，把“人力成本”改叫“人力投资”，因为在生成式 AI 这种新行业，成绩几乎都来自优秀的人。
比缺人更可怕的，是过度扩张把人才密度稀释了，他称之为组织不可逆的癌症。牛人招来牛人，弱 leader 只会招更弱的人。所以他的招聘底线是：新人要比现有的人强；让你犹豫的人就别招，后面大概率出问题；只招自驱的人，因为没人想做微观管理。
愿景得创始人自己想透、讲准。造船的团队要先向往大海，不然每天只是在砸钉子。借来的愿景经不起团队追问，甚至过不了自己那关。
前面说 founder 要防着被下属 PUA，反过来，管理者也别去 PUA 别人，比如用贬低（negging）去打压对方。对人公正，同时弄清每个人想要什么，放到合适的位置，给他需要的回报和成长。他推荐了 Robin Stern 讲煤气灯效应的书。创业公司本来就是创始人的延伸，他引用黄仁勋的比喻：你在造一辆自己要开的 F1 赛车，那就造一辆你开得了的。
给团队留空间，除非逻辑明显有问题，别一上来就否定。什么时候克制、什么时候抓细，他用 Andy Grove 在《High Output Management》里讲的 task-relevant maturity 来判断。AI 公司多由年轻的技术型人才带队，他们普遍缺管理经验，所以他在公司里亲自讲了一套《给好学生的管理课》。他发现，个人战斗力越强的人，做管理时心理障碍有时反而越大。
他希望 founder mode 是全队的工作方式。Leader 要能往下看两三层，必要时自己上手：厉害的人想从 leader 身上学东西；自己没做过，就没法判断下属干得好不好；不懂细节的技术 leader 容易没安全感，怕被取代，于是乱插手；要在人才、资金、品牌、产品、市场之间找全局最优解，也离不开细节。
有些决定就像办公室空调的温度，总有人嫌热、有人嫌冷，得接受。他建议新任管理者都去玩一下《冰汽时代》，记得玩一代。Low performer 会把 high performer 逼走：搞砸之后互相安慰、一起降低标准，再把问题推给别的部门。想留住最好的人，就果断和不合适的人分开；毛利高的话，好聚好散也做得体面些。
接着他聊到“老登”。他今年满三十，也会想自己在年轻读者眼里算不算“中登”了。
他举的“老而不登”的例子，是成立 137 年的任天堂、宫崎骏还在创作的吉卜力，还有 77 岁的 Ray Dalio。他们都把玩家、观众、读者放在第一位。宫本茂会从走廊里拉一位清洁工来试玩，自己站在背后默默看反应，岩田聪称之为“宫本先生的越肩视线”。所以他说，这个时代不创造新价值才是原罪。
在他看来，登味的根子是不安全感：怕被时代甩下，怕年轻人不听自己。可一代人有一代人的事，被时代甩下也正常。哪天他做不到 state of the art 了，就交给别人，自己去设目标，像塞尔达那样把鸟望台、神庙、呀哈哈、莱尼尔摆在合适的位置，让玩家自己去解。（他还开玩笑说，这样就没人去救公主了，不过玩家开心就好吧。）
麻烦在于，要给最聪明的年轻人设目标，自己得跟得上。一旦不再快速学习，登味就会找上门，跟年龄关系不大。
五、战略与执行
大多数努力是白费的，创业也一样，所以多花时间想清楚什么重要，比一直埋头干更有用。要敢下大注，输了认账。他习惯把成功归于运气，说这跟谦虚关系不大，主要是太习惯失败了：搞创新，大多数尝试注定失败，那就抱着大概率失败的心态全力投入。低预期让他更有危机感，也更容易觉得幸福，代价是有点保守。
资源之间能互相置换：人才密度、文化、士气、品牌、现金、愿景、founder 的体力。哪块弱一点，只要没弱到不可接受，都能靠别的长处补上。
他从曾鸣教授那学到“以终为始”：别看手里有什么牌再想能做什么，先想将来要做成什么，再去补牌。他相信这样能做成大十倍的事，还不一定更难。“我今天为什么做不到”是基本功，更好的问题是“我今天做什么，将来才做得到”。
清华和 MIT 让他明白，身边最聪明的人往往还比你努力得多，而市场只奖励结果最好的人。他从那里带走两个信念：没有人能随随便便成功；靠运气得到的，长远看都会凭实力还回去。那里后来成了行业顶尖的人，都热爱自己做的事；光是聪明加努力，比不过聪明、努力又热爱的人。他们做实验严谨，写 paper 连标点都要对。
他以前觉得这种较真有点控制欲，后来发现各行顶尖的人都这样，宫崎骏、宫本茂、乔布斯，还有当年教他英文写作、带他写 SIGGRAPH paper 的老师。前提是先把优先级排对，重要的事先做。基本问题都解决了，吹毛求疵才是本事。他们也清楚，提出团队一时理解不了的高要求，眼下是要挨骂的。
他说世界是个巨大的草台班子，把基本的事做扎实，就能进前 1%。问题是基本的事并不简单：想清楚自己要什么，说到做到，按时睡觉，独立思考，确保读者看得懂自己写的东西，只提高质量的问题，每天用自己的产品，让贡献大的人多拿，把事实放在面子前面，不说正确的废话。
既然这么难，就少做几件吧。公司越大，选择不做什么越重要，而且不做反倒有得：少开项目，多出时间想战略；少参加活动，多陪家人；少认识些新朋友，才能和志同道合的人深交；少招不合适的人，人才密度和文化浓度才上得去。
最后一条是，深入的思考被别人拿走也无妨。它揉进了市场、创始人和团队的能力、时代、公司基因和文化，只有你自己能落地到 100 分。别人照搬做到 80 分，也竞争不过，因为市场只选最好的。做不到最好，还不如不做。
结尾他讲了小时候的一顿饭。父亲带着他和一个出去创业的学生吃饭，两人聊到：科学家是把科学当成家的人，企业家是把企业当成家的人。这句话他记了二十多年。至于搞科学和做企业有什么相同，他的答案是：找到一个足够重要而有趣的问题，然后把最基础的事情扎扎实实做好。
Comments留言