Batch normalization (Ioffe & Szegedy, 2015) was justified as reducing “internal covariate shift.” Santurkar et al. at MIT (2018) tested this directly and found BN doesn’t reduce covariate shift. In some cases it increases it. The real reason it works: it smooths the loss landscape, making gradients more predictable so you can use higher learning rates. One of deep learning’s most used techniques, adopted for years on an incorrect theory.
TIL
Quick learnings and observations. Incomplete thoughts, links, and snippets.
Sending anything to my agent files it to Google Drive automatically. Screenshot, PDF, link. Right folder, date-prefixed filename, done.
Sent a screenshot of a paragraph with no attribution. The agent searched the text, found the original article and author. Reverse lookup from a photo.
Dropped a long article I’d been putting off. It split it into 14 themed sections and scheduled a daily email, one per day.
All the popular tools (svg-term-cli, termtosvg, MrMarble/termsvg) convert asciinema recordings to SVG. You have to actually run the commands first. There is no tool that takes a static config and renders a fake terminal session as SVG, which is what you actually want for README demos: clean, controlled output without recording your real shell.
Haiku with skills matches Opus without them. SkillsBench shows skill engineering beats model size.