Pedagogical sites I built while learning robot-learning. Rough but functional.
Standalone walk-through of π₀.5. Six pages, ~30 min, 17 interactive visualizations. Built for readers who haven't seen a transformer paper.
Five load-bearing ideas behind π₀.5, anchored at multi-head attention. ~25 min end-to-end, 14 visualizations. (Earlier version; less polished.)
What changed between releases, anatomy, training, inference, RL-on-top, ecosystem.
Walk-through of the FAST tokenizer paper.