An interactive slide deck on Token-In / Token-Out, multi-turn drift, safe divergence, and black-box agent RL.

A compact note on Megatron Dynamic-CP, ByteScale Hybrid Data Parallelism, and why DP, CP, and PP become coupled when sequence lengths vary.

An interactive model of post-training, inference and RL for MoE models: per-module attention and MLP rooflines across architectures, parallelism, low precision and speculative decoding on one shared engine, checked against measured and public data points.