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.
中文读者请进 简体中文版 ——每篇文章都有中英双版。
Featured
Give an AI agent a city. It gives you back a film. | A City Through Time
Give a coding agent a city and a starting era. It researches the sources, pulls building and map data, builds the whole city as a white model, and lights it up era by era, from the earliest settlement to today. What comes back is a film, an old city you can walk through, and museum pages for its buildings. Two cities are lit so far: Xianyang, 6,000 years, and Rome, 2,800.
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 installed plannotator, sent 2 notes on a 3-line change: desktop apps have its headline features
I installed plannotator v0.28.64 and ran it for real: 3 lines changed in Pico CSS, 2 line-by-line annotations sent back, and the agent fixed the first one. But its headline features, plan approval and line-by-line annotation, already ship inside the Codex and Claude desktop apps. What it adds is annotating non-code files, an Inbox shared by several agents, and an approval gate scripts can check. Skip it if you use one desktop app; worth it for CLI and multi-agent work.
DeepSeek opens search, file and voice APIs: cheaper transcription than OpenAI, no image generation
DeepSeek opened its search, file parsing, URL reading and voice APIs on November 13. This post compares DeepSeek, OpenAI and Gemini on the platform basis: speech-to-text at $0.058 per hour against OpenAI's $0.27, about 4.7 times cheaper, but DeepSeek has no image generation, and its transcription caps at 120 seconds and 8MB per segment with no streaming. Tool fees are charged separately from token fees, so run one week of real records before you integrate.
I Let a Bot Run My X Account for a Month: 15 to 4,000 Followers, 37 Posts Flagged as Spam
I let scripts and an agent run my X account for a month. The followers went from 15 to 4,000, and X's own Under the Hood report for October shows 37 of my 1,937 posts flagged SPAM_HIGH_RECALL, almost all of them follow-back comments concentrated on two days. What the report says, what I suspect beyond it, and what I changed on November 12.
Claude Code reviewed its own work — out of 125 suggestions, only 2 actually worked
After 30 tasks, Claude Code reviewed its own work with /retro and proposed 125 improvements. Researchers then tested them one by one against later tasks, and only 2 fully worked. This post lays out the funnel, why most of the suggestions didn't help, and the delete-first-then-redo test that separates the keepers from the rest.
Muse's 30 Billion Tokens Don't Last Nearly as Long as They Sound
I worked backwards from Muse's own dashboard readings: the invite-reward cap is 30 billion, but what I actually received was about 4 billion, and 357M of it burned away in three days. Splitting the bill pile by pile, almost all of the money went on context being reread over and over and on the roughly 88,000-token startup cost of every new agent. At the fastest burn the 4 billion would be gone in 21 days; after the fixes it lasts 62 days, and converted into a subscription it comes to about 1.85 months of the Power tier.
A Local AWS Emulator Is Out: Agents Wrote All the Code, and the Team Reviewed Not One Line
On November 11, AWS released AWS Emulator, a local emulator of the AWS APIs covering 438 API models. All of its code was written by autonomous agents, and the team neither wrote nor reviewed a single line. I installed 1.0.322 myself, ran the four SQS steps of creating a queue, listing queues, sending a message, and receiving a message, and counted the fidelity score file that ships with the release record by record: out of 42,937 test units, only 24,243 passed, a pass rate of 56.5%. It is usable, but how many units each service actually passed is something you have to check for yourself first.







