Notes文章

AI and new software, explained plainly AI 与新软件，用大白话讲

How I think about AI-written code, AI agents and what changes when software gets cheap to build. Each note links to code you can run where there is any. Questions or corrections: hi@byteplain.com. 我怎么看 AI 写的代码、AI Agent，以及软件变便宜之后什么会变。有代码的地方都附了能跑的仓库。有疑问或发现错误，写信到 hi@byteplain.com。

2026-08-08 · How I Work · AI & Judgment

AI extrapolates fluency. I extrapolate emotion. AI 外推流畅，我外推情绪

“AI does the technical part, people do the human part” gets it backwards. The two fail in opposite directions, and quality comes from pairing them so each corrects the other. This is the review process I actually use on client email. “技术交给 AI，与人沟通交给人”这个分工是反的。两者出错的方向正好相反，质量来自让它们互相纠正。这是我处理客户邮件时实际在用的审查流程。

2026-08-03 · AI · Safety

AI-written code always has bugs nobody has found yet. Here is how I use it anyway. AI 写的代码总有没找到的 bug，我照样大量用它，靠的是闸门

Every time I ask AI to review its own code, it finds new bugs. You can't count your way to safety. You can sort mistakes by what they cost and put hard gates in front of the dangerous ones, so a bug can still happen but can't turn into harm. 每次让 AI 审它自己写的代码，它总能再找出新 bug。靠数 bug 数不出安全。能做的是按后果给错误分类，在危险的那几类前面设硬闸门：bug 还会有，但变不成伤害。

2026-07-16 · AI & Judgment

Where value goes when building is free 当“造东西”变免费，价值去了哪

When the cost of something drops to zero, its value moves to the new bottleneck: judgment, verification, and a person who can be held accountable for a machine that can hurt someone. A model can’t provide that last part. 当一件事的成本降到零，价值就转移到新的瓶颈上：判断、验证，以及一个能为可能伤人的机器负责的人。最后这一点，模型给不了。

2026-07-15 · AI Agents · Governance

Governing AI agents that touch money: the agent proposes, it never executes 让碰钱的 AI Agent 受治理：agent 只提议，永不执行

Every write with a material consequence should leave the model’s reach and pass through an authorization layer the agent can’t write or bypass. Governance comes from architecture, not from a carefully worded prompt. An escrow flow serves as the worked example. 每一个会产生实际后果的写操作，都应该脱离模型能直接控制的范围，先经过一层 agent 既不能改写、也绕不过去的授权。治理靠架构，不靠措辞小心的提示词。文中用一个托管支付流程做例子。

2026-07-15 · Self-Hosted · Privacy

A private AI assistant with exactly one way out: designing the outbound boundary 只留一个出口的私有 AI 助手：设计出站边界

“Your data never leaves your machine” only means something if you can point to the single place it does leave and prove nothing else can. Trust comes from drawing that boundary precisely, not from bigger promises. “数据永远不离开你的机器”，只有当你能指出数据唯一离开的那个地方，并证明别处都出不去时，这句话才有意义。信任来自把这条边界画清楚，而不是承诺更多。