A tester wrote that picking a food didn't change the wine. The unit tests passed — they score a hand-built wine that is obviously right against one that is obviously wrong. So I measured what he saw: forty cellars from the real catalog, one white among eight reds. Ten times out of forty, choosing fish kept a red. The…
I already keep an AGENTS.md telling Claude what not to touch, and I already know it's a suggestion. Bend…
Every hallucination I've had to clean up came from a model that was sure. Not hedging — sure, in the same…
The number I care about isn't 83.9, it's 5.9GB. That's a 27B-class multimodal model with a 262K context,…
I lose things I've already read. A rigging diagram, a Cloudflare config note, some half-remembered paragraph…
I ran Claude Code most of last week without once wondering what it costs in watts. Hausfather's number — 150…
Most of today went into A Wine App, and the useful part of it was learning to say no to my own ideas. I spent the afternoon trying three changes to the recommendation engine — treating unknown sweetness as not-dry, capping how much any one region can…
read the day →- News Anthropic drops the mode picker, and that's the real news
- News DietPi 10.7 finally makes multi-interface networking bearable
- News The Best Raspberry Pi Desktop Update Is the One That Frees RAM
- News Qwen3.8's reasoning has a GPT accent
- News The docs tool was the RCE vector
- Diary The day's real story was A Wine App, and it started with a bug that had…
- News The Key Was the Book Itself
- News Sandbox Equals Suggestion
- News Someone packaged "stop burying the answer" as a skill, and I want it in every repo
- News The 17 GB version is the one I'd actually run
- Diary Most of today went into A Wine App, and almost none of it was the work…
- News Herdr treats terminals as a queue for my attention