Articles

Everything published so far, newest first. Pieces marked as drafts are honest works in progress.

• You and AI

  What AI does to us

  What does using AI do to us — to our self-worth, our self-reliance, our loneliness? A sober look at the evidence: what's established, what's hype, and how to stay whole.

  10. July 2026 · 8 min read

• Workshop notes

  Agents on a long leash

  AI agents work on my code autonomously for hours — and do remarkably little damage. That's not down to trust, but to four guardrails: isolation, permissions, verification, and the rule "ask instead of improvising".

  6. July 2026 · 4 min read

• Workshop notes

  My AI keeps notes on me

  Language models forget everything between two conversations. The remedy: a folder of plain text files where the AI records what it has learned about me, my projects, and its own mistakes. How that works — and where it goes stale.

  6. July 2026 · 4 min read

• Basics

  When a model runs out of context

  What a context window is, why the middle of long chats suffers first — and three simple things you can do about it.

  6. July 2026 · 3 min read

• Workshop notes

  Convincingly wrong

  The AI explained, in publishable prose, why my server was serving stale code. The explanation was plausible, detailed — and wrong. The "fix" derived from it took the real website down. On the most dangerous property of AI answers.

  6. July 2026 · 4 min read

• Workshop notes

  The AI that looked like malware

  My antivirus reported a trojan: "This program is dangerous. It executes commands from an attacker." The attacker was my AI — and all it wanted was to start a test server. Why harmless AI code can set off alarms.

  6. July 2026 · 3 min read

• Basics

  The EU AI Act for small operators

  Europe's AI law isn't just for tech giants: even businesses that merely use AI — in operations, on their website, with customers — pick up obligations. What applies in practice, without the legalese — as of July 2026.

  6. July 2026 · 7 min read

• Workshop notes

  Fourteen languages, one AI — and why humans still get the last word

  A one-person website in fourteen languages used to be unaffordable. AI translation makes it possible — but only native speakers make it right. What their corrections revealed surprised me.

  6. July 2026 · 4 min read

• Basics

  llms.txt — my websites introduce themselves to AIs

  A small text file is meant to make websites more readable for AI systems. All my sites serve one — including this one. What llms.txt is, what it looks like, and what it realistically does (and doesn't do).

  6. July 2026 · 3 min read

• Basics

  MCP explained: the USB port for AI tools

  For an AI to do more than talk, it needs tools — and a plug that fits them all. That is exactly what the Model Context Protocol provides. What's behind it, shown on our real setup.

  6. July 2026 · 4 min read

• Workshop notes

  AI culls my wedding photos — and invents what it can't see

  Sifting, rating and pre-developing hundreds of frames: AI is a great assistant for that. But where the visual signal gets thin, it guesses — and a pony became two sheep. My rules for AI in the photo archive.

  6. July 2026 · 4 min read

• Basics

  Which model for what? The three dials

  Model class, thinking budget, number of attempts: three dials determine how good an AI answer gets — and what it costs. Know them, and you'll also understand why answers sometimes quietly get worse.

  6. July 2026 · 4 min read

• Workshop notes

  The ugly cat

  All I wanted was a comfortable tomcat for a website. What I got explains two things at once: why AI can't draw — and why quality is a budget decision.

  6. July 2026 · 4 min read

• Behind the site

  Why lostcontext.ai?

  The name has two meanings — and both are on purpose. On the mission, building from the inside, and what should grow here.

  6. July 2026 · 2 min read