I document my notes and writings on AI research, LLMs, and engineering here. A mix of long-form posts hosted on this site and selected external articles.
A stage-by-stage field guide to LLM training data — what counts as one example and who certifies it at each stage, how synthesis and verification actually work, and how to distill from frontier teachers into a small model.
按训练阶段梳理 LLM 训练数据：每个阶段的「一条样本」究竟是什么、由谁保证它正确，数据合成与校验如何真正落地，以及如何把 frontier model 蒸馏进一个小模型。
What OpenAI's booth Q&As at ICML 2026 add up to: the operating horizon is the new scaling axis, and evaluation latency, credit assignment, monitoring, and trust are what it drags along.
OpenAI 在 ICML 2026 booth Q&A 上真正说了什么：operating horizon 成为新的 scaling axis，而 evaluation latency、credit assignment、monitoring 与 trust 是它拖着一起走的东西。
How next-generation agents should route deployment experience across context, memory, harnesses, weights, and the learning mechanism itself.
下一代智能体应如何把部署期经验路由到 context、memory、harness、weights 与学习机制，并在多时间尺度上安全地持续学习。
How agents improve the software around a frozen model — the no-gradient twin of environment scaling. The propose-evaluate-select-archive loop, the five surfaces people evolve, the LLM-as-optimizer zoo, and why task supply and verifier honesty are the real bottlenecks, for code agents (SWE-bench, Terminal-bench).
围绕冻结模型进化"harness"（智能体的软件外壳）——环境扩展的无梯度孪生：propose-evaluate-select-archive 循环、人们进化的五个面、把 LLM 当优化器的方法谱系，以及为什么任务供给与 verifier 的诚实度才是真正瓶颈（面向 code agent：SWE-bench、Terminal-bench）。
A 2026 field guide to the frontier-LLM training pipeline — data, pre-training, post-training (RL), evaluation, and safety — synthesizing the MAI-Thinking-1, DeepSeek, Qwen, Kimi, Llama, GLM and other technical reports.
2026 年前沿大模型训练流水线的中文导览——数据、预训练、后训练（RL）、评测与安全，综合 MAI-Thinking-1、DeepSeek、Qwen、Kimi、Llama、GLM 等技术报告。