A QUESTION STILL OPEN
GRPO efficiency starts before the kernel
A practical investigation of sparse rewards, loss normalization, and what a faster training step actually proves.
Open this story ↗A SKY THAT TAKES ITS TIME
Every new question leaves a little light. Start remembering your journey, then follow the spaces you haven’t explored.
Opt-in and local: this feature remembers only selected site paths, visit counts, and opening times on this device. It sends no exploration data to a server. No query strings, selections, or browsing outside this site are recorded. Pausing keeps existing stars; clearing removes this observatory’s records.
A QUESTION STILL OPEN
A practical investigation of sparse rewards, loss normalization, and what a faster training step actually proves.
Open this story ↗RECENT ORBITS
Your constellation belongs to this browser. Export a copy if you want to keep a record.