-
MemoryTax: The Bitter Lesson of Agent Memory
Does agent memory still help once the model can search the raw history itself? Memory systems and coding agents on LoCoMo with three models.
-
Qwen3.8-27B on One RTX 4090
245 tokens per second on the official weights, 322 with better ones, against 200 for vLLM and 102 for llama.cpp on the same card: batch-1 speculative decoding with interactive figures from real traces.
-
Make Sand Think
A riddle, and an honest answer: how a handful of sand becomes a machine that talks.
-
我实现了 GPT-5.6-sol Ultra 自由
为什么我坚持用最强的模型，以及我如何用一台小服务器和 Sub2API 搭出自己的 AI 中转站。
-
LoRA as Parametric Memory: a failed experiment
I tried writing conversation facts into model weights via LoRA fine-tuning. The accuracy was dismal and temporal questions were completely unanswerable.
-
Where Is Linear Attention Actually Slow on GPUs?
Poor GPU performance for linear attention is an intuition, not a conclusion. We ran experiments, then had half of our interpretation overturned. This post records that process.
-
ForgeHLS and DiffHLS: why I built them, and why HLS still feels like a trap
A personal note on building a large-scale HLS dataset, pushing DiffHLS forward, and why I remain skeptical about HLS as a long-term research direction in 2026.
-
探索更大的世界
留学申请路上的记录
-
Notes on Proofs, Arguments, and Zero-Knowledge
Notes on machine learning, security, privacy, and mobile computing. Guided by Liyao Xiang.