写给自己的随手记 · a running set of notes
Notes随想
一些观察和想法,先写给自己 —— 去过的聚会、冒出来的念头、用 AI 做东西时学到的东西。谁读到觉得有用,那就更好。
Bits I jot down for myself — meetups I show up to, ideas I'm chewing on, what I'm learning building with AI. If it's useful to you too, even better.
暂时没有邮件订阅 —— 收藏这页,或者用 RSS 跟着读。/ No email list yet — bookmark this page, or follow along by RSS.
The Cloud Thinks, the Mac Plumbs云端想事情,Mac 接水管
八月一个周末,我的一串定时任务漏跑,而且没有一个告诉我。我花了一段时间才想明白这不是一个故事,然后把整套系统拆开:要动脑的放云端,只有 Mac 能做的留在 Mac,不需要 AI 的就写成普通脚本。九月,它还是黑了十天 —— 发现这件事的,是同一套系统里的另一部分。
One August weekend a string of my scheduled jobs quietly missed their runs, and none of them said so. Untangling it took a while; then I split the whole thing — the thinking goes to the cloud, the Mac keeps only what needs the Mac, and anything that doesn't need AI becomes a plain script. In September it still went dark for ten days. What noticed was another part of the same system.
读这一篇 / Read →Rent vs buy in NYC — how I think about the question纽约租还是买 · 我怎么想这个问题
别拿月租和月供比,要比的是五年后两个世界里你一共有多少钱。这一篇用一个虚构的例子把框架走一遍:盈亏平衡价、机会成本、纽约买卖的全部成本、租房的时机。不是理财建议。
Don't compare rent to the mortgage payment — compare what you'd have in two worlds five years out. A plain-words framework worked through on a fictional example: the break-even price, opportunity cost, the full cost of buying and selling in New York, and timing if you rent. Not financial advice.
读这一篇 / Read →The Question Bank提问手册
我在练习聊天时更松弛一点 —— 少一点技巧,多一点次数,感觉比表现重要。这些年攒下的问题里,能让人愿意多说两句的那些,整理在这里:原则、追问的手艺、轻松的开场题、认识新朋友、和老朋友聊深一点,还有几个最好别问的。
I've been practicing being more at ease in conversation — reps over technique, the feeling is the metric, not the performance. These are the questions from my notebook that actually open people up: principles, the craft of following up, light openers, networking, going deeper with close friends, and a few to never ask.
读这一篇 / Read →How I built a live health monitor for my 14-year-old dog给暖暖做一个实时健康监测
暖暖戴着一个 PetPace 项圈。这一篇讲它的数据怎么变成安静的提醒、一张私密的每日页面和一张公开的趋势卡 —— 以及第一版错在哪里:不看活动量就读心率、把一次 HRV 下降当成警报、把没有数据显示得像一切正常。
How 暖暖's collar data becomes quiet alerts, a private daily page and a public trends card — and what the first version got wrong: reading pulse without activity, treating a single HRV dip as a warning, and letting a gap look healthy. Tracking, not veterinary advice.
读这一篇 / Read →Claude Code worktrees — when to work in an isolated copyClaude Code 工作树 · 什么时候让它在副本里干活
Claude Code 有一个"工作树"开关:打开,它就在一份独立的副本里干活;关掉,它直接改你的文件。我用一个厨房的比喻讲清楚什么时候该开、什么时候该关,还有绝对路径那个最容易踩的坑,以及怎么清理堆起来的旧副本。
Claude Code's worktree toggle decides whether a session works in an isolated copy or directly in your files. A kitchen analogy for when to turn it on and when to leave it off, the absolute-path trap that makes a test pass when it shouldn't, and how to clean up the copies that pile up. Not official documentation.
读这一篇 / Read →The Scoreboard That Wouldn't Flatter Me我做了一块不肯讨好我的记分板
我这几个月一直在给自己的生活写自动化 —— 每天的简报、体检数据、投资、待办。这周我做了最后一块:一个每周给自己打分的系统。真正难的地方不在收数据,在于怎么让它不学会讨好我。
I've spent months automating my own life — daily briefs, health data, portfolio, to-dos. This week I built the last piece: something that grades the week. The hard part wasn't collecting the data. It was stopping it from learning to flatter me.
读这一篇 / Read →The Whispering Dome会传话的穹顶
九月底我去了拉合尔和卡拉奇见同事。拉合尔有一座 1674 年的清真寺,穹顶是双层的,在一个角落说话,对角的人听得清清楚楚。同一周的晚饭桌上,有人问我:你的名字是什么意思?
At the end of September I went to Lahore and Karachi to meet the team. In Lahore there's a mosque from 1674 with a double-layered dome — speak in one corner and someone in the opposite corner hears you clearly. That same week, at dinner, someone asked me what my name means.
读这一篇 / Read →The Gate I Didn't Add我没装的那道闸门
一份周报里的一句话说,agent 系统缺的不是更聪明的模型,是一道能拒绝的闸门。同一周,我自己的消息系统正好在示范这句话是什么意思 —— 而我对着这个证据的第一反应,是想装更多自动化,不是装闸门。
A line meant for AI agents in general said the missing piece isn't a smarter model — it's a gate that can refuse. The same week, my own message system was quietly demonstrating exactly that. My first reaction to the evidence was to add more automation, not a gate.
读这一篇 / Read →The System Got Me Wrong系统写错了我
我给自己的生活装了一套 AI 系统,其中一个每周三晚上写一篇帖子,我只要点一下就能发。上周它交上来的稿子里有一句:「力量训练还没开始。一次都没有。」那一周我练了六天。它没有说谎 —— 它只是读错了地方。
One of the automations I run over my own life drafts a post every Wednesday night, ready for me to approve with a tap. Last week's draft said: "the strength rebuild hasn't started. Zero sessions." I had trained six of the previous seven days. It wasn't lying. It was reading the wrong layer.
读这一篇 / Read →Finger fracture — choosing an NYC hand surgeon手指骨折 · 纽约手外科医生的选择
从急诊或 urgent care 的临时夹板,到手外科医生正式接手,通常只有 7–10 天。这一份指南是给这一段的:今晚准备什么、明早打给谁、电话里必须问到哪两个日期、门诊当天用英文问医生什么、以及术后怎么恢复。
A step-by-step guide to the 7–10 day window between the ER or urgent-care splint and the specialist consult — what to gather tonight, who to call in the morning, the two dates you must get on the phone, the English questions to ask the surgeon, and how recovery actually goes.
读这一篇 / Read →The Bug They Didn't Fix那个没被修掉的 Bug
用了十多年 Spotify,最近打开 Daily Mix,整串放下来一首都不想跳过。于是我去查了他们到底做了什么 —— 结果最关键的一条,是 2015 年一个没被修掉的 bug。
Ten-plus years on Spotify, and Daily Mix suddenly runs end to end with nothing I want to skip. So I went looking for what they changed — and the most interesting answer is a bug from 2015 that nobody fixed.
读这一篇 / Read →June PM × AI Meet Up纽约 PM 聚会
纽约一场每月一次的 PM × AI 聚会后的随记:比起模型,更难的是 context 和治理;build vs. buy 的反复纠结;为什么 domain 深度才是真正的护城河;以及 AI 到底替代不了什么。
Notes from a monthly New York PM × AI meetup — context over models, the build-vs-buy loop, why domain depth is the real moat, and what AI still can't replace.
读这一篇 / Read →More lands here as I write it — roughly whenever I have something worth keeping.