Haris Wang's Blog

Jul 11, 2026 Agentic Systems — An Architectural Survey
A ground-up architectural survey of systems that put a language model inside a control loop — how they work, why they fail, and what three years of production practice actually taught us: from the control-flow test, through context engineering, to reliability, evaluation, and protocol standardization.
Jul 11, 2026 Agentic System
一份自底向上的架构综述：把语言模型放进控制循环里的系统如何运作、为何失败，以及三年生产实践教会了我们什么——从控制流判据、上下文工程，到可靠性、评估与协议标准化。
May 13, 2026 DeepSeekV3: 极致工程优化让模型成本下降一个数量级
671B 总参数、37B 激活参数的开源 MoE 模型，用 557 万美元训出来——从架构、训练、推理、性能四个维度拆解 DeepSeek-V3。
May 13, 2026 A Survey: The Architectural Evolution of LLM Applications
2022.11 – 2026.05，大模型应用领域 1260 天的五次范式跃迁与前沿模型时间轴。
Jun 1, 2025 Enzo in Firenze
...
Jan 7, 2022 the Architecture of MultiTasking in Linux
...
Jan 6, 2022 Diagonal Matrix and Orthogonal Matrix
...
Dec 13, 2021 Memory Management of C++ Object Model
...
Dec 3, 2021 std::move & std::forward in C++
...
Jan 11, 2021 A brief review on Transmission Control Protocol
...
Oct 12, 2020 Use ssh to Connect to your VPS in GCP
...
Sep 30, 2020 An Overview of Multiagent Reinforcement Learning Algorithm
...