Sirui He
Technical Blog
Live reverse diffusionEvery dot starts as Gaussian noise and is denoised, step by step, onto the lines of this sketch.
Writing2 posts
GRPO From First PrinciplesWhy On-Policy RL for LLMs Is Really Just Dynamic Weighted SFT
A first-principles walk from supervised fine-tuning to policy gradient and GRPO — why on-policy RL for LLMs is really dynamic, group-normalized, reward-weighted SFT.
Manifolds and Genesis
A dialogue between a researcher and a language model — from how system prompts work to information geometry, manifolds, and what it all might mean.