The curve that broke the textbook

Train 128 models of growing width and plot their test error: it drops, spikes exactly where parameters match the data count, then drops again beyond. That second descent, described around 2019, is where today's giant models live.