Engineering notes
What I learn by building, with enough detail to still be useful a year from now. They almost always start with something that didn't work.
- The state lives outside the model: how a desk pet knows you haven't talked to it in three days
An LLM with memory gives you a friendly chatbot. A state machine with an injected clock gives you a creature. The line "I know" is only funny if it's true.
- I spent days tuning the prompt, and what actually moved accuracy was changing the model
I went from 30.7 % to 100 % of days correct reading shift rosters. Almost none of that gain came from the prompt, and the metric that mattered wasn't accuracy.
- Three layers, two problems: a GPU passthrough that stopped halfway
The device was in the container, the check confirmed it, and the app still couldn't see it. Because the app was running one layer further in. And once it could see it, it had no permission to open it.
- Three times it "looked configured": verify by reading the result, not by assuming
A sysctl that didn't persist, a udev rule that would never have fired and a sed that edited a comment. No command returned an error. The habit that catches all three is the same one.
- A backup you never restore is a big folder with a reassuring name
I set up automatic backups for nine machines and everything showed green. The photo library, the only irreplaceable data, was left out. What gave it away was a backup that finished too fast.