Grounded takes on AI and software engineering, by Kenny Vaneetvelde.
Claude Code could only reach Anthropic models, so I was picking from a third of my real options. model-gateway is a plugin that puts the GPT and Grok models from my existing subscriptions straight into the /model picker, keeps itself updated, and leaves Claude traffic going to Anthropic on my own credential.
Claude Code kept re-reading my whole codebase every session, and CLAUDE.md never reliably fixed it. So I built a plugin that maps the codebase once and re-injects it into context on every prompt.
In one month, the US government switched one frontier AI model off worldwide, rationed a second to about a hundred approved companies, and held a third back before launch. One government now decides who gets the best models, and most of the world, Europe included, has nothing of its own to fall back on.
The 5 to 10x AI productivity pitch is snake oil. The rigorous studies put the real gain in the single to low double digits. The honest math, with sources.
On June 12, 2026, the US government made Anthropic switch off its two best models for every non-American on Earth, citing a jailbreak. Why that reason cannot be the real one, and what the precedent means.
I watched Google I/O on May 19, 2026. By the end of the keynote, something felt off and I couldn't say what. This is me working out what.
For years, "Dead Internet Theory" was framed as something done to us: foreign bot armies, state-sponsored troll farms. My most-read piece of 2026 so far.
MCP gave AI agents a way to reach your systems of record. What it does not reach is live applications. The article that introduced tesseron.
It starts with an admission: I have not written code myself in almost a year. The essay follows that thread where it leads.
Anthropic went from $1 billion to $19 billion ARR in fifteen months without the loudest marketing. On focus as a strategy.
I keep thinking about the same scenario; it has become a bit of an obsession. The closest thing I have to a manifesto on how companies should adopt AI.
Worldwide AI spending hit $1.76 trillion in 2025 and is forecast to reach $2.52 trillion in 2026. The failure pattern is organizational, not technical.
Amazon's Kiro codes unsupervised for days. After 15 years as an engineer and consultant, I have seen this movie before: the prototype that "just works" becomes the foundation everyone builds on for the next five years.
Claude Code is also a workflow automation engine hiding in plain sight. How to encode your procedures as skills, including the ones that have nothing to do with code.
A simple document analysis task once ballooned into hundreds of dollars in API costs because of hidden LLM calls I couldn't see or control. On what popular AI frameworks actually cost production teams.
Six design principles for agents that survive contact with the real world: modularity, persistent memory, orchestration, defensive design, clear interfaces, and realistic testing.
Frameworks promise high-level abstractions and deliver hidden complexity. One of my two most-read pieces, and how many people first found Atomic Agents.
Why I built Atomic Agents: the Input-Process-Output model, atomicity, and full developer control instead of framework magic. My most-read piece ever.
uv folds pip, pipx, poetry, pyenv, and virtualenv into one Rust-powered tool that is 10-100x faster. I called it the game-changer we had all been waiting for; the ecosystem agreed.
A push-to-talk voice-to-text tool in Python with Groq's Whisper API: hold a key, speak, release, and the transcription lands in whatever app you're in.