Scaling laws are often shown as retrospective plots, but the more useful framing is prospective: what should I train next, how much data should I use, and what loss should I expect if the trend holds?
The visualizer project came out of wanting a more concrete way to reason about these tradeoffs. Instead of treating a paper equation as a static object, the interface turns compute, model size, and data into controls.
This post will expand into a design note on how research tools can make theoretical relationships operational.