Blog
Notes on compression, systems & AI infrastructure · 中英文随手记
RSS
Random-reward RL for capability probing
Training response can reveal capability that a checkpoint's current accuracy does not show.
Continuous resource attribution for RL training
The problem behind Omni Profiler: connecting CPU/GPU resource use to the code responsible for it during RL training.
Cross-hardware bitwise reproducibility in LLM inference
Integer arithmetic makes reproducibility an execution contract, with explicit precision, overflow and serving trade-offs.
Evaluating agent process claims against execution traces
Checking an agent's result, its claims about execution, and the cost of its trajectory as separate questions.
Storage contracts for hybrid KV cache offloading
A reproducible CPU byte path for hybrid-cache research: variable-sized records, reader-safe storage, honest byte accounting, and unresolved model acceptance.
She Only Told the Beginning
真美啊，美得我只能阿巴阿巴。
Tiny Beautiful Things
The best thing you can possibly do with your life is to tackle the motherfucking shit out of love.
In Between
What lasts when something is gone?