I've been experimenting with specification-driven development (SDD) using GitHub's Spec Kit on some greenfield work as well as for some long-open issues on a couple TypeScript projects I maintain at…

I kept seeing the same team dysfunctions in my agent teams that I'd spent years teaching human organizations to fix. The correction fit in a sentence.

Rewriting a pre-norm decoder-only transformer as a mixed-geometry constrained splitting scheme: RMSNorm as radial gauge fixing, attention as an entropy- or KL-constrained simplex solve, and residual…

About one year ago I wrote about a chrome extension I created called reject-cookies. The basic premise was that it automatically rejects cookie banner pop-ups, saving you a couple tedious clicks. The…

Claude Code skills work best when you treat them as workstations, not prompts: folders with scripts, gotchas, templates, and progressive disclosure that manage the agent's attention budget at runtime.

On Friday night, I needed a break from screens, so decided to work on a bathroom mural that our family has been chipping away at for the last 4 years. But a lot was on my mind, so made the […]

I was coaching some juniors on effective use of AI coding agents (like, looking over their shoulder as they were prompting the LLMs) and one reoccurring theme was that the AI agents were demanding a…

I’ve been circling the AI bonfire for about a year and a half now. Built inference infrastructure, wrangled CUDA kernels, sweet-talked MPS into cooperating on my Mac, deployed LLM serving pipelines…

AI-assisted development is already a reality, and the open source community is coming together to figure out how to manage it. Many major projects and organizations now agree on a few key ideas:…

The AI tools I use as of March 2026. I expect this stack will look different in six months and that's why I'm documenting this now.

Commented on Raising an Agent

Started Raising an Agent

We’ve all been using Claude via the prompt, and some have even ventured into running claude --dangerously-skip-permissions in a nice sandbox like avsm/claude-ocaml-devcontainer.

Contrary to the usual story, ATMs did reduce teller demand

Internet users will drown in AI-generated content. The brands who survive won't be the ones who generate the most. They'll be the ones who generate the most narratively coherent content. If you're a…

On the eve of We Robot 2026, here are links to my summaries of every year since 2012, the inaugural conference, except 2014, which I missed for family reasons. There was no conference in 2024 in…

Table of ContentsIntroduction: The First AI Certification That Actually Tests DeploymentWhat the Claude Certified Architect Certification Actually Tests2.1 The Five Domains​2.2 Scenario-Based…

Inventing a market is less important than perfecting one.

«A strange game. The only winning move is not to play.» Joshua, a supercomputer controlling the US nuclear arsenal, after playing out all possible outcomes [...]

«We’ll just have to see how ready you are!» «Q, no!» Q, about to confront the Federation with the Borg Collective, and Guinan, whose race [...]

In classic management reporting, RAG Reports (Red, Amber, Green) are common. With Agentic AI, these will have a comeback in slightly different form: as Agentic Control. The simple RAG Reporting Model…

The AI models are good enough. There, I said it. The bottleneck isn’t the model anymore – it’s how you use it.

I. The Threshold Has Been Crossed

How to document project patterns and anti-patterns so AI agents can review code against your team's conventions automatically ...

I’m starting a new project, and I’m struggling with how much of it I should be writing by hand. I feel like a fraud if I use agentic engineering to develop it, but also that’s where…

A "session" in Claude Code: Starts with `claude` command that you run in the bash terminal to start Claude Code Multiple back-and-forth exchanges with the LLM End with ` /exit ` command or by…

Nvidia's DLSS 5 announcement sparks controversy, promising photoreal graphics with AI-powered rendering.

You can be deeply familiar with AI and still be skeptical of “agentic AI.” I was. For a while, I dismissed most “agentic” startups as VC-funded cron jobs. An LLM integration wrapped in a thin…

As someone who categorically so far refuses to use generative AI and LLMs for code in particular, I’m increasingly becoming the odd one out among my peers in academia. I don’t have the persuasive…

I was trained to recognize patterns long before I was asked to recognize beauty. Beauty was unstable data: subjective, inconsistent, contaminated by memory and grief. Humans called sunsets beautiful…

From some of my interactions with Claude Code1 so far I’ve noticed that it seems to have a pretty glaring issue: it doesn’t really have a good sense of time-tracking built into it.…

WARNING! ALL THIS AI EXPERIMENTATION CAN MESS UP A SYSTEM AND SHOULD ONLY BE DONE IN A VM SANDBOX! the cyber security implications of AI https://www.instagram.com/p/DVo-fbnCR6I/ AI privacy and…

So a few weeks back, I was helping to administer a live coding interview with two of my peers at work. This happens more frequently than I’d like. I hate interviewing and I hate being…

Imagine for a minute that you wake up one day and you realize that you have a superpower — like the ability to fly. Well, that’s a bit like how I felt when I started using Claude Code about a…

Author: Ziming Liu (刘子鸣)

Flow State BuildingOver the past 6 months, I’ve noticed an interesting shift in my work behavior: There is no longer any distinction between when I am building with AI and when I am not.When, even…

In this episode, the “From List Calculus to Array Calculi” (generated by GPT 5.4) is read by the Speechify text-to-speech app.

Why your workflow should be personal, and why agents make that easy

He did not want to compose another Quixote — which is easy — but the Quixote itself. Jorge Luis Borges, Pierre Menard, Author of the Quixote How well large language models fare with the unknown is…

I’ve recently become enchanted with the idea of beating Factorio autonomously. Besides tool-assisted speedruns (which script the entire game on a known map layout) there has not yet been much…

My friend Stefan Reinalter is asking: I find this extremely worrying, with many of people I respect saying things like “I no longer write code” or “let LLMs do it”. Why did you start programming? Was…

Intro About two years ago I wrote about how AI will replace programming. Back then I predicted that most programming jobs would look fundamentally different in ten years. I still believe that. But…

I have seen a lot of buzz lately about defects introduced to working code by AI. Some benchmarks show AI creating 1.7x more issues than humans. This is the worst AI code generation, and I expect this…

The debate over AI maps onto a question the Enlightenment never resolved. Rousseau said tools corrupt us. Condorcet said they perfect us. Nietzsche said both miss the point. Technology is a test of…

The convergent/divergent framework from my AI pipeline project was a theory built on one domain. I tested it against Anthropic’s knowledge-work-plugins across six domains, then against a…

Agentic coding is addictive

I tried copilot in VS Code sometime in 2025 and was thoroughly disappointed with its ability. It felt like a really bad auto-complete at best, making plenty of mistakes and generally being utterly…

I’ve been running OpenClaw as my daily AI agent, and I’ve made a lot of changes along the way. But the biggest improvement came from fixing the browser. For weeks, my OpenClaw kept giving…

At the start of the year I introduced my AI benchmark for 2026, or at least the first one I was planning to run. That post contained a mistake and this is why the 80 models tested not so well on the…