"Is the AI useful?" is the wrong question. The real one: "can anyone defend what it produced?" New post on what "just use AI" actually asks of engineers, and what doesn't go away no matter how good the model gets. blog.northerntest.se/using-ai-is-...

The strongest connections in a knowledge base aren't in the content. They're in the behavior. Twelve weeks of observation, Hebbian decay borrowed from neuroscience, and the system surfaced 11 clusters I never designed. blog.northerntest.se/i-built-the-...

Stacking probabilistic systems doesn't make them deterministic. It makes them probabilistic in more places. Should have remembered this from telecom class. PRNG seeded by another PRNG doesn't add entropy. New post: blog.northerntest.se/error-budget...

I spent 4 years building trust with AI assistants. A month ago I went all in on Claude. Then it ignored every instruction I'd written. Repeatedly. The models I trusted broke first. New post on the dangerous in-between state we're in now: blog.northerntest.se/the-model-i-trusted-broke-first/

I mounted a FUSE filesystem to observe how my knowledge base documents are used together. Day one: 690,000 events... oooff.. Four days of detective work later, the signal was clean. Then I discovered the filesystem can see what files were touched, but not why. blog.northerntest.se/i-put-a-file...

I've been building a local AI stack, worrying about vendor lock-in, knowledge loss, and pricing traps. Then my friend Patrick Prill wrote a post that named all eight risks. So I wrote about how I accidentally designed around them, and what I still can't solve. blog.northerntest.se/i-read-patri...

Asked my knowledge base who Fredrik is. Got back 85 documents. Email from 2021 said one company. Meeting note from 2024 another. Both correct, at different points in time. It could find him. It couldn't tell a coherent story. blog.northerntest.se/my-ai-remembered-everything-that-was-the-problem/

I searched for "Fredrik Haard" and got back a conversation about testing philosophy. The knowledge base had 100,000+ chunks indexed and couldn't find someone by name. Five phases of failure later, I finally got search working. Each fix revealed the next problem. blog.northerntest.se/teaching-a-k...

I built an AI assistant that can't do anything without asking first. I called them "suggestions" in the code. Not actions. Not commands. I wrote about why that naming mattered more than any feature I shipped. blog.northerntest.se/i-called-the...

I built a system to make me faster. It worked. Then I had to build another system to make me stop. "I Removed the Friction. That Was the Problem." blog.northerntest.se/i-removed-th...

I got tired of re-explaining my own context to AI every session. So I built a knowledge base that indexes 44,000 files of my digital life (emails, calendar, notes, conversations), all searchable, all private. The AI didn't get smarter. It stopped being amnesiac. blog.northerntest.se/i-gave-my-ai...

I built four systems with AI over Christmas. People call that "vibe coding." I hate that term. The prompting is the easy part. Knowing what to build, what to reject, and what the system needs when you're not looking — that's the work. blog.northerntest.se/beyond-vibe-...

AgileTD is a 4-day conference (Onsite & Online) for international agile software testers and engineers gathered to learn, share, and network. More than 140+ sessions, including inspiring keynotes, talks, hands-on workshops, and full-day tutorials. Tickets: lnkd.in/dv-Ss46W