Is curing disease really a software problem?
Moderna's personalised cancer vaccine just passed its first Phase 3 trial, and Elon Musk called mRNA a software problem. As an engineer, I find the analogy half right, and the half that is wrong is the interesting part.
How do you evaluate an AI employee?
You can't watch an agent work, and its report sounds equally confident whether it is right or wrong. So stop grading the report and start grading the artifact.
You can't predict how hard a bug will be
Features can be estimated. Bugs can't, because a bug's difficulty is the depth of a wrong assumption you haven't found yet. So stop estimating and start drilling.
Your grader is rejecting right answers
A paper on step-level math verification found that LLM graders reject most valid solutions that don't look like the textbook one. If you use a model to check another model, measure the checker on its own.
谁牵涉你的注意力，谁就是在要你的命
注意力是唯一不可再生的资源。守住它的办法不是拒绝所有人，而是承认大部分事情做到三成就够了。
Debugging Whole-House Bluetooth: What I Learned from OpenComm, Multipoint, and a Raspberry Pi
My open-ear headset kept dropping when I walked away from the Mac. After testing power classes, multipoint, and codecs, the fix was architectural: a Raspberry Pi as a remote Bluetooth endpoint on the home network.
LLMs are too positive. Ask for the veto.
Language models are trained to agree with you, so stop asking them for approval and start asking them to kill your idea.
Solve the cold-start problem first
Most products die from empty rooms, not missing features. Answer the day-one question before you write any code.
The math of 500K a year as an indie developer
500K a year sounds like a dream until you divide it. Work the number backwards and it becomes a table of four businesses — and only one of them is probably yours.
写作是一种角色扮演
写作不是自我暴露，而是扮演一个你自己选定的叙述者，想明白这一点，下笔就没那么难了。