This side of the river is the old world. The other side is AI. I want to ferry us all across.
RSS FeedFull-stack developer for 24+ years. I formerly led a 400-person engineering team, and I'm still a core maintainer of several large open-source projects. Everything here is something I've actually built, run, or broken: AI engineering practice, open-source dynamics, infrastructure, and security.
ä¸æč¯ģč č¯ˇčŋ įŽäŊä¸æį ââæ¯į¯æįĢ éŊæä¸čąåįã
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
Can CrowdStrike's kill switch actually stop the Codex running on your laptop?
CrowdStrike's AI Kill Switch promises that a human can stop a rogue agent. Walking it through with Codex: the switch sees an agent through three layers â hooks inside the agent, the Falcon sensor on the machine, and the AI Gateway â and it misses whatever runs outside them. The only kill action live today is cutting model access at the gateway; the human presses the switch, but the machine does the filtering first.
OpenAI bought Pierre: your local git repo stays yours â what moves is where agent code lives in the cloud
OpenAI is acquiring Pierre, maker of Code Storage. Local development does not change: you or your agent still create the git repository on your machine, and the local form of Codex never touches Pierre. What moves is the cloud side â which form puts your code on a platform's servers, who Code Storage is actually for, and what that means for an individual user.
A backdoored Claude Code extension under a verified name: how to check your machine, and how to protect yourself
Two builds of the official Claude Code extension in Cursor's marketplace were tampered with, live for about 11 hours, reaching at most 5,000 users by Cursor's own statement on X. They plant a key in your authorized_keys, open a tunnel, and leave persistence that updating does not remove. Here is how to check your own machine, what to do if you find something, and the habits that keep it from working next time.
GitHub put agents in Actions. Can a small project finally afford proper CI?
GitHub Agentic Workflows are generally available: describe the work in a Markdown file, in ordinary words, and an agent sorts issues, reads failed CI runs, updates documentation and drafts tests inside Actions. I judge it on what a small project actually cares about: the upkeep burden, the habits it builds, ease of use, and safety, and end with a one-month test you can run on your own evenings.
I rebuilt a 1.47 MB, 761,000-parameter intent router and raced it against if statements. The if statements won
I rebuilt the 1.47 MB, 761,000-parameter intent router from a small account's post to its exact published specs and raced it against a set of keyword rules. It lost on every phrasing it had never seen, and all 12 off-list sentences were sorted somewhere wrong. Here is why it lost, and how a router this small should actually be trained and used.
DeepTutor, taken apart: its core trick fits in one skill, so why did 54,000 people star the platform around it?
DeepTutor's core â a structured record of what a learner got wrong, computed into quizzes and spaced review â is the one layer a single agent skill can copy. The 54,000 stars belong to the platform around it: parsing pipelines, output surfaces, an always-on tutor inside chat apps, and the institutional layer. I read the code end to end and checked how much of the crowd is actually using it.
