2025 年 3 月文摘
纽约时报观点 | 美国政府认为 A.G.I. 即将到来
有人说，这证明了我们美国的 AI 发展模式是错的。我们以为需要海量算力、需要大公司垄断，但 DeepSeek 展示了用更少的算力、更少能源、更少成本也能办到。也有人认为我们被 OpenAI、谷歌、Anthropic 三巨头洗脑了，实际上完全可以走一种更加分散、更加“太阳朋克”的路径。
你同意这个看法吗？
我认为两件事是并行的： 首先，最前沿的研究暂时还是需要极高的计算量和能源投入，这是大公司在做的事情。它们也有很强的动机去提高效率。但它们仍会需要这方面的投入。 其次，除了这些前沿，确实会有一些稍微落后于前沿的扩散式应用，它们所需的计算和能源更少。我们在两方面都要赢。
马克·安德森访谈录音：2024 年 5 月，我和本·霍洛维茨去华盛顿。我们没法见拜登，因为当时谁都见不到拜登。但我们见了白宫内部的高级官员，我们谈到对 AI 的担忧。他们的回应是：“是的，在拜登第二任期内，国家 AI 议程将是：让 AI 只掌握在两三家大企业手中，直接监管并控制这些公司，不会允许创业公司随意写代码并开源。”
Ninety-five theses on AI ⭐
There is a substantial premium on discretion and autonomy in government policymaking whenever events are fast moving and uncertain, as with AI.
Scaling LLMs still has a long way to go, but will not result in superintelligence on its own, as minimizing cross-entropy loss over human-generated data converges to human-level intelligence.
The open vs. closed source debate is mainly a debate about Meta, not deeper philosophical ideals.
e/acc and EA are two sides of the same rationalist coin: EA is rooted in Christian humanism; e/acc in Nietzschean atheism.
High functioning psychopaths demonstrate anti-social behaviors in their youth but learn to compensate in adulthood, becoming adept social manipulators with grandiose visions and a drive to “win” at all cost.
Manus 的护城河在哪里？
有些人开发产品的心态是，我不打磨到 100 分，我就不推出来。 然而，其实产品永远有值得改进的点，难到那就一直改进，一直不推出吗？
要有一个生态占位思路，先推出产品，先占住这个生态位，然后不断迭代，不断提升能力，就能够做得越来越好。 相似的例子是 Cursor，他们最开始发布的产品，效果也没有那么好，没有那么惊艳，但是他们等到了 claude sonnet 3.5 这个模型的出现，于是全球各地开发者都在自来水推荐 Cursor 。 试想一下，如果 Cursor 等到了 claude sonnet 3.5 发布之后，才去开始动手做，那么机会还会是 Cursor 的吗？
《the brutalist》 film
why architecture?
Is it a test?
No, it is not.
Nothing is of its own explanation. Is there a better description of a cube than that of its construction? There was a war on and yet it is my understanding that many of the sites of my projects survived. They remain there still in the city. When the terrible recollections of what happened in Europe cease to humiliate us, I expect them to serve instead as a political stimulus, sparking the upheavals that so frequently occur in the cycles of peoplehood. I already anticipate a communal rhetoric of anger and fear, a whole river of such frivolities may flow undammed. But my buildings were devised to endure such erosion of the Danube’s shoreline.
Ironies of Automation
有所不为
既然我不懂，那我就不写。
个人投资 checklist
https://xueqiu.com/4926075175/325615317
商业模式 checklist:
checklist1.1: 赚的多，要求行业空间巨大，至少得是千亿 rmb 利润规模以上，最好是万亿规模。行业规模会随着时间上升。
checklist1.2：赚得容易，这里用毛利率/roe 几个指标来评估，毛利率：40%+，roe:20%+。
这里后面还会丰富指标，想到了就来更新
checklist1.3: 赚的久，这里讲企业的竞争优势。根据下面的竞争优势进行挑选。
企业竞争优势有品牌优势，网络效应，成本优势，转换成本 （其中政府管制和进入壁垒，专利技术，这两不纳入我处理的范畴）
企业文化 checklist:
是否言行一致，吹的牛是否实现了，是否行事风格谨慎
是否专注
是否乱投资/乱花钱
是否在持续提升企业竞争优势
管理层是否人品靠谱，比如私生活不行，道德败坏
管理层是否偷钱，包括关联企业利益输送，并购事项中是否把企业优质资产剔除，垃圾资产装入。
企业决策方向是否正确，就是做正确的事。是否出现错误的核心决策
企业执行力是否强，是否持续实现了自己的目标，甚至超额完成任务
对待普通员工是否以身作则，一视同仁。对于能力跟不上的管理层是否尽快替换，并且找到合适的人替换
对待上下游合作伙伴是否带着伙伴一起赚钱，当然这里不是说大善人，而是遵循物竞天择，能跟得上企业发展速度的伙伴能一起持续赚钱。
对待股东，这里主要是股东回报，包含常说的分红/回购，还有投资效率，另外对待小股东是否友好。
财报信息披露完善程度
对消费者不利事项
违法事项
估值 checklist:
checklist: 估值比较简单，主要是自己的目标收益率需求，我给自己的目标是至少不能低于 10%的年化收益率（sp500 基准），以及腾讯是机会成本基准比较对象。
估值上要对下限严格，这样确保下限收益率至少年化 10%，要一眼知胖瘦，上限方面，估值不要超过 2-3 年时间才能消耗的到年化收益率的估值范围内，随着才能一直调整估值。
杂项 checklist:
checklist4.1: 不加杠杠，不做空
checklist4.2: 保障自身自由现金流稳健，满足家庭开支
checklist4.3: 政府对企业的生死影响，比如在线教育 A4，对互联网金融业务的影像。
checklist4.4: 战争、关税、技术封锁、合规等的影响
善战者无赫赫之功，聊聊职场中的隐形英雄以及我们该怎么做？
故善者之战，无奇胜，无智名，无勇功。
意思就是真正善于打仗是知道什么该打，什么时候不该打，什么时候该防守，什么时候该进攻。做到了，就很难出现奇异的胜利，都是理所当然的胜利，旁人看不出来什么神奇的地方。也不会有机智多谋的名声，也不会有奋勇无匹的气魄。就是简简单的胜利而已。
一个公司，一个机构，目的到底是持续稳定和有增长的产出，还是看起来虽然危机重重，但是一步步都在解决问题，但是不断自己创造问题，创造性地遇到被人不会遇到的问题，并解决他们。
On Writing #1 ⭐
https://x.com/435hz/status/1901445127995801927
Actual LLM agents are coming. They will be trained
If Manus AI is unable to properly book a plane or advise on fighting a tiger bare-handed, it’s not because it is badly conceived. It has just been bitten by the bitter lesson. Prompts can’t scale. Hardcoded rules can’t scale. You need to design systems, from the ground up, that can search, that can plan and that can act. You need to design actual LLM agents.
LLM agents “dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks”
聊聊 Agent 架构 – Single Agent / MCP / Multi-Agent
如果项目在早期，没有遇到很明显的瓶颈，并不需要用 Multi-Agent 架构，用 Single Agent 简单的架构足够能做好。工程架构越简单，后续基础模型升级带来的增益越大。
隐说 No.6 姥姥的大米粥
这些年我自诩还算通透，写过生死哲学的文章，与人也聊过生死哲学，可到了灵堂前，才惊觉所有理性不过是纸糊的铠甲。这位早年行医救人的老太太，晚年仍能用晒枣子的竹匾接住漏进屋檐的月光。我想起小时候村里的深夜非常的黑，我比较害怕回舅舅家睡觉的那条路，但姥姥会带我到路口，给我说“孩儿，不怕，再黑的天到头了也得亮”，这让我长大后几次碰见困境的时候，想起姥姥，看到姥姥后却总能给我重新审视自己的困境、去勇敢面对的力量。
什么才是软件的关键价值？
软件就是一个框子，装下各家的管理思想和流程。
适度专业吸引客户，过度专业吸引同业
观棋录
在 AlphaGo 的设计中有个重要的细节：训练 AlphaGo 的神经网络时所采用的反馈函数只依赖于输赢，而同输赢的幅度无关。 … 于是我忍不住设想，如果 AlphaGo 在训练时采用不同的反馈函数会是什么结果。不妨假设存在一个 BetaGo，一切都和 AlphaGo 设定相同，只是反馈函数定义为盘面领先的目数。（换言之，从一个正负之间的阶梯函数变成线性函数。）可以猜测 BetaGo 的「棋风」应该比 AlphaGo 凶狠许多，更追求杀着，更希望大赢。
厌蠢的人一定智慧不足？
表达能力的丧失是从大词叙事开始的。
If you’re thinking without writing, you only think you’re thinking.
关于做选择：先找到局部最优，再无限修正。
找到全局最优的能力，是通过找到 n 个局部最优的过程建立起来的。
The Burnout Machine
The Worst Programmer I Know
Tim wasn’t delivering software; Tim was delivering a team that was delivering software. The entire team became more effective, more productive, more aligned, more idiomatic, more fun, because Tim was in the team.
Just don’t try to measure the individual contribution of a unit in a complex adaptive system, because the premise of the question is flawed.
查理芒格的 100 个思维模型研究 ⭐
Measuring personal growth
Every 3-6 years, you become a different person.
the rule of 72 in finance. It’s a simple formula that estimates the number of years it will take for an investment to double in value. If the annual interest rate is 8%, it’ll take 72/8 = 9 years for the value of your investment to double.
Quynh, an old friend who runs a publishing house in Vietnam, believes that there are three big problems in life: career, family, and finance. It usually takes people a decade to figure each out.
For the first decade after graduation, you figure out what you want to do with your life. For the next decade, you get married, buy a house, and have kids. For the next decade, you build out your savings to retire. Her goal is to solve these problems as fast as possible, so she can focus on more interesting problems.
What I learned from looking at 900 most popular open source AI tools
该让机器人交社保吗？
“不是外卖小哥需要社保，而是社保需要外卖小哥。”
小米的汽车工厂，如今年产量已经逼近 40 万台，但整个工厂系统只有 2000 人。整个工厂流水线，几乎完全不需要“智人”手工操作。工厂车间甚至都不用开灯，是名副其实的“黑灯工厂”。
每个人都同时具备“生产者”与“消费者”的双重身份。只有当市场上存在足够的适格消费者时，你作为生产者的努力才有价值。
Rework Book Summary
Planning is guessing Writing a plan makes you feel in control of things you don’t control
Scratch your own itch The easiest, most straightforward way to create a great product or service is to make something you want to use. If you’re solving someone else’s problem, you’re constantly stabbing in the dark. When you build what you need, you can also assess the quality of what you make quickly and directly, instead of by proxy.
什么是 Cancel Culture 取消文化 ⭐
Fun With GPT-4o Image Generation
New ChatGPT image gen can draw sexy men but not sexy women
My black body story (it’s physics).
what we learn through our education has always been filtered by those who came after the events. It can be easier to explain things using current ideas, but it’s easy to forget that those who invented the ideas didn’t have them yet. The act of creating them may only be well understood by stepping back to the time they were working.
2020 letter
为什么有些富人爱吃苦？
无论是璀璨华夏，还是希腊罗马，这个的普遍存在很好地说明了，历史不会重复，但会押韵。
佛教和斯多葛主义有很多的不同，但有个共同的根基，就是按本体论去框定，它们都是「一元论」。 … 我们都会觉得，空气是透明的，身体是不透明的。但是在一元论者看来，它们是同一个东西，只是作为我们身体的那部分表现得不透明，作为空气的那部分表现得透明而已。
他们还据此认为，当我们跑步时，其实我们并没有运动，而是宇宙中之前表现为我们身体的部分变透明了，成为空气；之前是空气的，则变成了我们的身体。
a Stoic is a Buddhist with attitude, one who says fuck you to fate
这大概是权贵特别是今天的技术权贵喜欢斯多葛主义的另一半原因的完美注脚：它认为一切都是最好的安排，算是唯一一种不鼓励他们放弃财富或权力的哲学主张
Notes on MCP
I understand that cosplaying as Java developers (or, worse, TypeScript) is a common affliction in modern Python codebases, but I kept wondering exactly why I would need to create a new server to expose an existing API, …
Implementers, Solvers, and Finders
People want to make decisions rather than execute them Turns out science agrees on this: People want power because they want autonomy. Most of the time, folks desire to move up the career ladder not for pay, better title, or keys to the executive washroom (are those still a thing?) but because they wish to be able to exercise greater autonomy over their lives. Psychologist Daniel Pink agrees - he’s found that the three qualities that contribute most to workplace satisfaction and overall productivity are autonomy, mastery & purpose.
Are you given near-total autonomy in choosing what you work on? Can you tell your boss “That’s an interesting idea but my time would be better spent elsewhere” (and not get fired on the spot)? You’re a Problem Finder.
https://x.com/karpathy/status/1905051558783418370
Introducing 4o Image Generation
Why numbering should start at zero
https://www.cs.utexas.edu/~EWD/welcome.html ⭐
In corporate religions as in others, the heretic must be cast out not because of the probability that he is wrong but because of the possibility that he is right."
The semantic apocalypse ⭐
A funny thing about science fiction is that it never really predicts the future. The details are always off, and it is the details that make all the difference. Now that real AI is actually here driving cars, writing poetry, making music, and filtering your spam, the detail that has made all the difference is that these machines don’t think like us. In fact, these “intelligences” probably don’t think at all. What they definitely do, however, is copy.
Their mimetic nature is the missing detail, the reason our future isn’t already on the shelves of bookstores. Instead the future is trending toward a state that can only be called uniquely weird.
It’s now just a matter of time until our world becomes a Jurassic Park filled with newly-issued work by long-dead creators. Perhaps this means future creators will become judicious about how much work they publish, and therefore how much data they provide, to prevent such “style cloning.” Or perhaps creators will embrace this strange sort of immortality. Maybe they will eagerly train neural networks on their own work; every artist becoming the master of an atelier formed from themselves. Perhaps the future of the creative arts is an assembly line of style clones.
We actually consume the product itself. Art is meant to be imbibed. How much will it really matter if there is a “certified human” sticker on a script or a song or a painting?
Perhaps in the future each of us will have such a replica living on after us, trained on the digital effluvia we’ve left behind.
So then what, exactly, is the semantic content of an AI-produced work of Hildegard von Bingen? It is a “deep fake” of meaning. Such a work points to nothing, signifies nothing, embodies no spiritual longing. It is pure syntax. For art this is the semantic apocalypse. It’s when meaning itself is drained away by the mimetic powers we’ve unleashed.
Imagine a future website where every time you click refresh a new and perfect Shakespeare sonnet is generated on the page in front of you. And you click again and again and again and again. Imagine then, your dread.
transformer poetry
418 I’m a teapot
what to do
And the best kind of thinking, or more precisely the best proof that one has thought well, is to make good new things.
关于金钱、财富
钱最大的功能就是用来买清闲和空间的
能控制自己的时间，能每天醒来的临时决定今天怎么过。赚钱能力远高于开销，以至于不用去想钱。能自信地说“我不知道”而不怕丢面子；在自己的领域可以讲真话而不用怕被报复。财富不等于钱。有钱而没有自己可以自由支配的时间，有钱没人身自由，那也白搭。