The seed
I stopped in the middle of organizing my notes today. Not because there was nothing to do. There was too much.
I have an Obsidian vault with around 300 files. Claude Code helped me build an automated workflow: morning planning, task dispatch, knowledge digestion, cross-project sync. Before AI, I’d never have attempted any of this — too complex, not worth the time. But now, “not worth it” has quietly become “maybe I could.”
Then I noticed a paradox. Everything became “doable,” but everything took longer than expected. The tools got stronger. I got more tired.
I assumed it was just me. Maybe I wasn’t efficient enough. Maybe I hadn’t found the right workflow. Then I saw the data:
Upwork’s 2024 research found that 77% of employees using AI say the tools have added to their workload.
It’s not just me. It’s structural.
I. The pipe is fixed
Zhuangzi said something two thousand years ago that hits differently now:
“My life has a limit, but knowledge has none. To pursue the limitless with the limited — that is perilous.”
— Zhuangzi, Nourishing Life
Sounds like a platitude, but cognitive science spent the next two thousand years proving him right.
In 1956, George Miller published a now-famous paper: The Magical Number Seven, Plus or Minus Two. He found that human working memory holds roughly 7 chunks of information at once. Seven chunks, not 7 GB. It’s a hardware limit, and you can’t upgrade it by “trying harder,” any more than you can grow a new heart chamber by running.
Then John Sweller’s Cognitive Load Theory in the late 1980s added the next layer. Working memory is small. It’s also short-lived. Everything has to squeeze through this bottleneck before it gets to long-term memory. Too many options, too complex an interface, too much noise — and learning and decision-making fall apart.
And then directed attention fatigue, from Rachel and Stephen Kaplan at Michigan. Your brain burns energy just suppressing distractions, and the fuel tank is finite. When it runs dry: more mistakes, worse planning, shorter temper, dumber impulses.
From Miller to Sweller to the Kaplans, over half a century of research says the same thing: your brain is a pipe, and the pipe doesn’t get wider.
II. Ten times more valves
Herbert Simon, in 1971, wrote a sentence the AI era keeps making truer:
“In an information-rich world, the wealth of information means a dearth of something else — a scarcity of whatever it is that information consumes: the attention of its recipients.”
- No internet, no smartphones, no AI. Simon was watching TV channels and newspapers multiply. He still got the punchline. Attention is zero-sum. Give it to A, and B gets less.
Now apply that to the AI era.
Before AI, my day had maybe 5–10 decision points: what tasks to do, what methods to use, how to prioritize. With Claude Code, the count is closer to 50. Should I let AI refactor this code? Is its output correct? Should I iterate again? Is this automation worth building? Who fixes the bugs later? Do I understand the approach AI suggested? If not, how long until I do?
Every new “possibility” is a new valve. Every valve asks: open or not?
There’s a famous Israeli study on parole judges that gets cited for exactly this. Favorable rulings start around 65% at the beginning of a session, then gradually drop to near 0% as decisions accumulate. After a food break, they reset to 65%. The popular reading is that the judges’ brains got tired. The authors leaned that way but never measured fatigue. Critics pointed at case order (prisoners without lawyers were often heard last), and one simulation got a similar curve from judges just not starting long cases right before a break. The cause is still disputed. Either way, I wouldn’t want to be the last case before lunch.
David Meyer’s research is more direct. Switching between tasks can cost as much as 40% of your productive time. The brain spends zero-progress “restart” time every time it switches tasks. And David Strayer at Utah found only 2.5% of people are true “supertaskers”; everyone else just performs worse when multitasking.
So AI lets you open more valves, but the pipe hasn’t changed. Each new valve adds no output. It just thins what every valve gets.
III. The data doesn’t lie
If it were just me, fine. Maybe I’m doing it wrong. But the numbers say otherwise.
APA’s 2023 Work in America Survey (n=2,515):
- 38% of employees worry AI will make their jobs obsolete
- Among those worried: 51% say work hurts their mental health, 64% feel tense or stressed daily, 37% report emotional exhaustion
- Surveillance adds its own load: 56% of workers monitored by tech feel tense or stressed during the workday, vs. 40% of those who aren’t
Microsoft’s 2024 Work Trend Index:
- 75% of knowledge workers now use AI at work
- 78% bring their own AI tools (BYOAI)
- Yet 59% of leaders admit they can’t quantify AI’s productivity gains
- 60% of companies lack an AI vision or plan
- 46% of global workers are considering quitting within the year, an all-time high
Upwork’s 2024 Research (the most damning):
- 77% of AI-using employees say the tools have added to their workload
- 71% of full-time employees are burned out; 38% feel overwhelmed about having to use AI at work
Gallup 2021 Global Emotions Report:
- 41% of adults worldwide report stress, an all-time high
- 42% report worry, also an all-time high
Back in 2007, Tarafdar named five flavors of technostress: overload (work faster), invasion (always on), complexity (steep learning curves), insecurity (will I be replaced?), and uncertainty (this tool will be obsolete next quarter).
Sound familiar? AI cranks every single one up.
IV. History repeats
This isn’t the first time.
In 1987 Robert Solow wrote a line that became the “Solow Paradox”:
“You can see the computer age everywhere but in the productivity statistics.”
Erik Brynjolfsson quantified it in 1993. Computing power grew 100× through the 1970s and 80s, and yet labor productivity growth fell from over 3% in the 1960s to roughly 1%. More computers didn’t translate into more productivity, at least not in the short term.
Same story with steam and electricity. The productivity payoff took decades. Factories in the 1890s bolted electric motors onto steam-era floor plans and wondered why nothing changed. It wasn’t until someone rethought the entire factory layout that electricity actually delivered.
AI is running the same cycle:
Tool explosion → cognitive overload → organizational adaptation lag → productivity stagnation
75% are using AI. 59% of leaders can’t tell you what it’s actually doing for them. More tools, more exhaustion. What’s lagging is us: our habits, our orgs, our brains.
V. Knowing when to stop
So what do you do? Reject AI? Obviously unrealistic. Use AI harder? The data already shows where that road ends.
Lao Tzu:
“He who knows when to stop does not find himself in danger.”
— Tao Te Ching, Chapter 44
And the sharper one:
“Less brings gain; more brings confusion.”
— Tao Te Ching, Chapter 22
Two and a half thousand years before any of that research, same answer: do less.
When Steve Jobs came back to Apple in 1997, the first thing he did wasn’t ship new products. He killed 70% of the product line. He later said:
“People think focus means saying yes to the thing you’ve got to focus on. But that’s not what it means at all. It means saying no to the hundred other good ideas.”
Obama wore the same gray or blue suit every day. Zuckerberg wore the same gray T-shirt. They could afford variety. They just knew that every trivial decision draws from the same pool of limited cognitive resources. Reducing unimportant decisions protects the quality of the important ones.
My own recent version of this: Claude Code can help me with dozens of things. Note automation, code refactoring, knowledge curation, data analysis, documentation. But when I try to push every one of those forward at the same time, I’m more exhausted than I was before AI. Every new automated pipeline is one more system to maintain, one more failure point, one more “is this AI output correct?” judgment.
The moments that actually work are when I decide not to do something. Not because I can’t. Because I can — which makes “not doing” an active decision I have to make deliberately.
Jacoby’s experiments showed the same thing. More brand information actually led to worse decisions. Barry Schwartz called it the “paradox of choice.” Past a certain point, more options just mean more anxiety and more regret.
AI has handed us more options than we’ve ever had. Our ability to choose well? Still the same.
Closing
Back to Zhuangzi.
“To pursue the limitless with the limited is perilous.” But Zhuangzi didn’t conclude with “so do nothing.” He was talking about nourishing life. Nourishing what, exactly? The pipe itself.
The line I keep coming back to: the more you CAN do, the less you SHOULD do.
Keep the tools, and use them to do fewer, deeper things instead of more, shallower ones. The pipe diameter isn’t going to change. The water pressure will keep rising. The only thing you control is how many valves you open.
Eric Schmidt once warned that too much information could get in the way of “deep thinking, comprehension, and memory formation.” Clay Shirky put it better: the problem isn’t information overload, it’s filter failure. Our filters can’t keep up.
AI generates information faster than anything before it. But the hard part was always throwing most of it away and actually paying attention to what’s left.
Simon, in 1971: in an information-rich world, the scarce resource isn’t information. It’s attention.
So in an AI-rich world, what’s scarce?
The moments you decide not to use AI.
Written April 8, 2026. An evening when I used AI to organize 300 note files and found myself more exhausted than before I started.
缘起
今天下午整理笔记的时候，我突然停了下来。
倒不是没事做。恰恰相反，事情太多了。
面前是一个 Obsidian vault，300 多个文件。Claude Code 帮我搭了一套自动化工作流：晨间规划、任务调度、知识摘要、跨项目同步。以前这些事我根本不会去做：太复杂，不值得投入时间。但有了 AI，“不值得”悄悄变成了“好像可以”。
然后我发现一个悖论。每件事都变得“可以做了”，但每件事做起来都比预想的要久。工具更强了，我更累了。
一开始我以为是自己的问题。可能效率不够高，或者还没找到对的工作方式。直到看见一组数据：
Upwork 2024 年的研究发现，用 AI 工具的员工里，77% 说这些工具反而加重了工作量。
不是我一个人的错觉。是结构性的。
一、管道是固定的
庄子两千多年前说过一句话，现在读起来后背发凉：
“吾生也有涯，而知也无涯。以有涯随无涯，殆已。”
— 庄子《养生主》
听着像鸡汤？后来的认知科学证明，他说得一点没错。
1956 年，心理学家 George Miller 发了一篇后来成为经典的论文：《神奇的数字 7±2》。他发现人类工作记忆（working memory）一次大约能装下 7 个信息单元。7 个 chunks，可不是 7 GB。这是硬件限制，靠“更努力”升级不了，就像跑步跑得再多，也不会多长出一个心房。
1980 年代末，John Sweller 的认知负荷理论补了一刀：工作记忆不光小，还短命。所有信息都得先挤过这个瓶颈，才能进长期记忆。选项一多、界面一乱、噪音一大，学习和决策质量直接崩。
还有一个维度：密歇根大学 Kaplan 夫妇的定向注意力疲劳。大脑要集中注意力，就得不停压制周围的干扰，这件事本身就在烧能量。能量烧完了，后果就来了：判断失误、规划变差、脾气变大、做事冲动。
从 Miller 到 Sweller 到 Kaplan，半个多世纪的研究翻来覆去说的都是一件事：大脑是一根管子，这根管子不会变粗。
这就好比《天龙八部》里的鸠摩智。他把少林七十二绝艺一门一门学会了，看上去无敌。结果运起小无相功硬撑那些招式，自己内息逆转，走火入魔。问题不出在招式上，出在他那条经脉就那么粗，硬塞了七十二门进去。藏经阁的扫地僧点破过他：“每一项绝技，均须有相应的慈悲佛法为之化解。”鸠摩智没听进去。
AI 工具就是这种“少林七十二绝艺”——一上手都能开，可你的内息（注意力、决策力）还是那么多。开得越多，每一门用得越浅。
二、阀门多了十倍
Herbert Simon 在 1971 年写过一句话，到了 AI 时代越来越准：
“在信息丰富的世界里，信息的富裕意味着其他东西的匮乏——即信息所消耗的东西的匮乏：接收者的注意力。”
1971 年。没互联网，没智能手机，没 AI。Simon 看到的只是电视和报纸在变多。但他已经看透了。注意力是零和的。给了 A，B 就少了。
现在把这逻辑放到 AI 时代。
以前一天大概有 5–10 个决策点：今天做什么任务、用什么方法、优先级怎么排。现在用 Claude Code，决策点可能有 50 个：要不要让 AI 帮我重构这段？它出的结果对不对？要不要再迭代一次？这条自动化流程值不值得搭？搭了之后出 bug 谁来修？AI 建议的这个方案，我理解吗？不理解的话，理解它要多久？
每多一个“可能性”就是多一个阀门。每个阀门都在问你一句：开还是不开？
以色列有个挺有名的假释法官研究，结果挺扎心。每轮裁决开始时，批准率大约 65%。决策数量一多，批准率就一路降到接近 0%。吃完东西休息一下，又回到 65%。流行的说法是法官的脑子累了。作者自己也往这个方向解释，但从没直接测过疲劳。后来有人指出排案顺序并不随机（没请律师的犯人常常排在最后），也有人算过，法官只要别在休息前接手耗时长的案子，就能模拟出差不多的曲线。到底什么原因，现在还有争议。反正要是我，可不想排在午饭前最后一个。
David Meyer 的研究更直接：来回切换任务，最多能吃掉 40% 的有效时间。大脑每切换一次任务，都得花一段“重启”时间，这段时间里一点进展都没有。犹他大学的 David Strayer 则发现，只有 2.5% 的人是真正的“超级任务者”，其余的人多任务时表现都会下降。
所以 AI 让你能开更多阀门，但管子没变粗。每多开一个，产出没多，每个阀门分到的水反而更少了。
三、数据不说谎
如果只是我一个人的感觉，那大可以说是我自己的问题。但数据不这么说。
APA 2023 年美国职场调查（2,515 人）：
- 38% 的员工担心 AI 会让自己的工作过时
- 这些担心 AI 的人里，51% 觉得工作伤害了自己的心理健康，64% 上班时常感到紧张、有压力，37% 感到情感耗竭
- 被技术监控本身就是压力：被监控的员工里，56% 上班时常感到紧张、有压力，没被监控的只有 40%
Microsoft 2024 年工作趋势指数：
- 75% 的知识工作者已经在工作中用上了 AI
- 78% 自带 AI 工具上班（BYOAI）
- 但 59% 的领导承认，没法量化 AI 带来的生产力提升
- 60% 的公司承认，在 AI 上缺乏愿景和计划
- 46% 的全球职场人考虑在一年内辞职，历史最高
Upwork 2024 年研究（最扎心的一组）：
- 用 AI 的员工里，77% 说这些工具反而加重了工作量
- 71% 的全职员工已经倦怠（burnout），38% 因为工作中必须用 AI 而不堪重负
Gallup 2021 全球情绪报告：
- 全球 41% 的成年人感到压力，历史新高
- 全球 42% 的成年人感到担忧，也是历史新高
2007 年 Tarafdar 总结过五种技术压力：过载（逼你加速）、入侵（永远在线）、复杂（学不完的新工具）、不安全（怕被替代）、不确定（下个季度又换一套）。
放到 AI 身上，五条全中，而且每条都在加码。
四、历史总在重复
这不是第一次了。
1987 年，Robert Solow 写过一句话，后来被叫做“Solow 悖论”：
“你到处都能看到计算机时代，唯独在生产力统计中看不到。”
Erik Brynjolfsson 在 1993 年把这件事量化了。1970–80 年代计算能力翻了 100 倍，但劳动生产率的增速却从 1960 年代的 3% 以上掉到大约 1%。更多计算机没带来更高生产力，至少短期内没有。
蒸汽机和电力也是同一个故事。生产力红利花了几十年才兑现。1890 年代的工厂装上电动机，车间布局还是蒸汽机时代那套。直到有人想明白要重新设计整个流程，电力的好处才真正出现。
今天的 AI 时代在跑同一个循环：
工具爆发 → 认知过载 → 组织适应滞后 → 生产力停滞
75% 的人在用 AI，但 59% 的领导说不清楚 AI 到底带来了什么。工具更多了，人更累了。掉队的是我们自己：习惯、组织、脑子，都还没跟上。
五、知止
那怎么办？拒绝 AI？显然不现实。更努力地用 AI？前面的数据已经说了，这条路走到头就是倦怠。
老子：
“知止不殆。”
— 《道德经》第四十四章
还有一句更狠的：
“少则得，多则惑。”
— 《道德经》第二十二章
两千五百年前那句，和今天的认知科学说的，是一回事：少做一点。
Steve Jobs 1997 年回 Apple，第一件事没去推新品，先砍掉了 70% 的产品线。他后来说：
“People think focus means saying yes to the thing you’ve got to focus on. But that’s not what it means at all. It means saying no to the hundred other good ideas.”
说白了，专注就是对另外一百个好主意说“不”。
Obama 每天穿同款灰色或蓝色西装。Zuckerberg 每天穿同款灰 T 恤。他们当然买得起衣服。只是明白一件事：日常决策再小，也要从同一个认知资源池里扣，这个池子就那么大。减少不重要的决策，是为了保护重要决策的质量。
我自己最近的一个体会。Claude Code 能帮我做的事有几十种：自动化笔记、代码重构、知识整理、数据分析、写文档。可真要同时推进这么多事，我比不用 AI 的时候还累。每多一条自动化流程，就多一个要维护的系统、多一个可能出错的环节、多一次“这个 AI 输出对不对”的判断。
真正管用的，是我决定不做某件事的时候。不是因为做不到。恰恰因为做得到，“不做”才成了一件得我主动拍板的事。
Jacoby 的实验也说了同一件事：给消费者的品牌信息越多，决策反而越差。Barry Schwartz 管这叫“选择的悖论”。选项多过某个点，多出来的就只是焦虑和后悔了。
AI 给了我们空前多的选项。但选的能力呢？还是老样子。
收束
回到庄子。
“以有涯随无涯，殆已。”用有限追无限，危险。但庄子没说“所以什么都别做”。他说的是养生。养的是什么？是那根管子本身。
我反复想到的一句话：能做的事越多，该做的事就越少。
工具照用，只是拿它去做更少、更深的事，别去做更多、更浅的事。管道口径不会变。水压还会继续涨。唯一能控制的，是你主动打开多少个阀门。
Eric Schmidt 说过，信息太多会妨碍“深度思考、理解和记忆形成”。Clay Shirky 说得更到位：问题不是信息太多，是过滤失败。我们的过滤能力跟不上信息的增长速度。
AI 是有史以来最猛的信息生成器。但难的从来都是另一头：扔掉大部分，再把注意力真正放到剩下那点东西上。
Simon 在 1971 年就说清楚了：信息丰富的世界里，匮乏的不是信息，是注意力。
那在 AI 丰富的世界里，匮乏的是什么？
是你决定不用 AI 的那些时刻。
也就这么回事。
写于 2026 年 4 月 8 日。一个用 AI 帮自己整理了 300 个笔记文件、然后发现自己比整理之前更累的晚上。
Comments留言