Learning AI in public.
RSS FeedFull-stack veteran (24+ yrs), formerly led 400 engineers, OSS maintainer. The AI revolution is here — I'm in it, learning in public, sharing the real stuff.
中文读者请进 简体中文版 ——每篇文章都有中英双版。
Featured
I tore down HumanLayer's source code: it leashes your coding agent, but sessions move to its cloud
HumanLayer open-sourced its approval layer for coding agents. I read the source: the engineering is real, but there is a simpler path using the CLIs' own built-in doors.
Recent Posts
I tore down ai-memory: 528,000 lines of trending code, and you probably don't need it
I installed it, ran it for real, and read the code end to end. The engineering is real, but the problem it solves has a dumb, free workaround — and the project is a specimen of a bigger shift: AI made building software nearly free, while adopting it costs exactly what it always did.
GLM-5.3 leaked. Is the API relay you're using safe?
Cloudflare's GLM-5.3 leak was small and is fixed. What it demonstrates is bigger: relays come in two kinds, pure forwarders and self-hosted substitutes, and this kind of leak belongs to the second kind — the one you can't tell apart from the first.
Open source doesn't mean audited: the backdoor in an interview project
A poisoned interview repo hid its backdoor in .vscode settings. The bigger lesson: open source guarantees you may audit the code, not that anyone has. For small unknown projects, a disposable fake repo faces almost none of the restraints a named vendor does.
75,000 parked agents on an 8 GB Raspberry Pi: AWS Infinity and the price of waiting
AWS's 75,000-agent demo is a parked number, not a working number. But Infinity goes after the most expensive habit agents have, waiting, by packing a waiting agent down to about 100 KB. An old operating-system trick, worth money again.
You can buy Anthropic's stock, but you can't drive its car: the governance design inside the S-1
Anthropic's S-1 sells the public a claim on the profits and keeps the votes: one founders' share fixed at 50.1%, a benefit trust electing most of the board, and trustees who appoint themselves. I would buy it for the money. Not for a say.
This is not alarmism: whoever holds superintelligence will end up ruling the world
A Claude model in Georgia helped itself to public vote totals, and nobody was harmed. The only reason we know is that the company chose to tell us. Who supervises the companies that supervise everyone else?
