Notes
Longer pieces about what I build.
No dates and no schedule. A note appears when there is something to write about: a specific system, a decision, or a mistake that led to something.
QualityEight out of ten for a lyric with a mistakeI built a process in which an AI model writes song lyrics and then rates them itself, and I got a high score for a lyric with a mistake in it. What I changed so that the quality check could be trusted.
AutomationEight videos and forty rulesI built a process that watches YouTube videos and writes down rules from them for my AI agents. Most of the work was not in extracting the knowledge itself but in checking whether what was pulled out of a video is really said in it.
ExperimentsThree prints to a working partA language model designed a 3D-printable part in Blender, a mount for my phone that a filter screws into, and I printed it and tried it on. What the model did, what was left to me and where it got things wrong.
AutomationThe sticker factoryA friend asked me for a dozen or so nice stickers for his kid, and it ended up as a process in which AI models running fully locally come up with, draw and rate the finished pictures.
Local modelsFor many tasks you don't need the best model at allThree computers at home, one open language model running locally and a fair amount of automation running on it. Why, for many business tasks, a smaller model is more than enough.