Translated from the Chinese original. This post is the development-environment installment of the “LLM Eats Everything” series. For how to build the hardware foundation, see the previous post, The Agent’s Home: The Individual’s Hardware Foundation in the AI Era; for how to build the software layer, see The Tools I Grew With AI. This one is about what comes next: once the hardware and software are in place, how do you keep your Agent online 24 hours a day, while you connect anytime, anywhere, from any device? ...

Translated from the Chinese original. A couple of days ago in Nanjing, I finally finished all 234 episodes of the podcast Feihua Lianpian (肥话连篇), start to finish. The show drops recommendations from time to time: a restaurant in some city, a series the hosts are bingeing, a little gadget they like. While listening I think “worth trying”, make a mental note to check it out later, and then… nothing happens. Later, when I actually want to find that one restaurant again, all I can do is rummage through my fuzzy memory, and I basically never find it. Xiaoyuzhou (the podcast app) currently has no ability to search the text inside the audio. ...

Translated from the Chinese original. An empty Nanjing city wall, deep and quiet stretching into the distance I wanted to write a travelogue but was too lazy to do it properly, so I first rambled through everything I wanted to say into Feishu’s Recording Bean, then slowly sorted it out. 1. How It Started It started simply: a friend said there was a music festival in Nanjing, and Chen Jingfei happened to be on the lineup, so they asked me to come along. I said yes instantly. ...

Translated from the Chinese original. In Kazakhstan: Planning with AI, a Drone Blown Away by the Wind, and the Afterglow of a Golden Age 1. How It Started A couple I know were going to Kazakhstan and asked what my plans were for the May Day holiday. I happened to have none, so I happily said yes and dragged my college roommate along. The whole trip ran from April 25 to May 4. I took 4 days of annual leave, so it came to about eight or nine days. ...

Translated from the Chinese original. My Hardware, My Rules: Customizing Your Phone on Demand in the AI Era In the previous post, The Agent’s Home: A Hardware Foundation for Individuals in the AI Era, I mentioned one idea: in the AI era, it’s worth owning a rooted phone. I gave four reasons at the time, but didn’t expand on them. ...

Translated from the Chinese original. Yesterday Claude Code’s source code leaked, and the hottest move in the Chinese-speaking community has been following tutorials to turn off telemetry. Please don’t do this. This article explains why. Main image What happened yesterday The source code of the current version of Claude Code (cli.js.map) was leaked through reverse engineering. The leak includes all of the telemetry reporting logic: which fields are reported, how many channels they go through, and how to turn telemetry off via environment variables. ...

Translated from the Chinese original. This is the hardware installment of the “LLM Eats Everything” series. The previous post, The Tools I Grew with AI, covered how I built things at the software layer. This one covers the hardware foundation underneath: how I built it, why I built it this way, and my read on where the hardware market is heading. This post gets a bit geeky. I actually wanted to write it a long time ago and put it off for almost a month, and in that month hardware prices went up quite a bit more. ...

Claude Code: Avoiding Bans, Choosing Models, and Design Philosophy Translated from the Chinese original. Claude Code paired with Opus 4.6, and Codex paired with GPT 5.4 — these are the two products running far ahead of the pack in the coding agent space. My own setup is the Claude Code Max plan plus Codex’s $20 basic plan; I do most of my project work in Claude Code and turn to Codex for the really hard debugging. ...

Translated from the Chinese original. Description: The AI starts acting dumb mid-conversation, a Skill’s success rate swings up and down, you copy someone else’s setup and can’t reproduce their results—these problems often share one root cause. Understanding the context window is the first lesson in using any AI tool well. Treat your AI’s context the way you treat your most precious attention. Preface: You’ve Probably Run Into These Problems After twenty-odd turns with the AI, it starts contradicting itself, and even forgets what you said five minutes ago You have an Agent run the same workflow: sometimes it passes on the first try, other times it fails for no obvious reason You throw a big chunk of data at the AI for analysis, and it answers off-topic, or the analysis is complete nonsense The same Skill gives completely different results on a different model; the same model succeeds this time and fails the next A Skill that used to work fine starts failing after you install more Skills You copied someone’s Agent configuration and project setup in full, yet the actual results are miles away from the demo The AI assistant of Meta’s director of safety deleted all of their emails—because the conversation was so long that the safety constraints were lost when the context was compressed If you’ve had similar experiences (or worry something similar could happen to you), this article may help you find the root cause. ...

Translated from the Chinese original. Last month I wrote a piece called LLMs Devour Everything: The Tools I’ve Grown with AI, about the information-processing system I’ve built for myself with AI over the past year or so. The response was good, and I got quite a bit of feedback. But after finishing it, I still felt there was more to say. That piece was mostly about “what I built,” while what’s actually been turning over in my head lately are some more abstract things: how AI affects our lives, our social relationships, even how we define human value. These thoughts are fragmented and tangled up with each other. I’ve talked through some of them with friends, and I’ll write them all down here as a handful of loose musings. ...