2026 年 4 月的 AI 行业正在经历一场深刻的哲学分裂：最强的模型不再是最贵的，最开放的不再是最弱的。

从 24 Fall 开始算，这是第四个学期。前三个学期的路径：打数学和逻辑基础，自学从 Python 到 RISC-V 汇编的完整栈，动手实现 CPU、写 C、做全栈项目。这学期是系统软件层的集中展开：操作系统、数据库、网络三个方向同步推进，配合 MIT 6.S081、Berkeley CS 186、CS 168，加 Redis 工程实践和一个从零上线的 Agent 应用。

上学期（25 Winter）的自学集中在理解层面。这学期开始验证理解和动手之间的差距。三门校内课：EECS 2021 要求用 Verilog 亲手实现一个能跑 RISC-V 指令的 CPU；EECS 2031 用 C 语言处理系统编程任务；EECS 2030 深入到面向对象的实现者视角。课外还做了 Spring Boot + React 的全栈 AI 项目，从零走到 Docker 部署上线。

上学期（24 Fall）用三门课完成热身。这学期开始不一样了：校内三门课 + 自学四门课同步推进，工作量直接翻倍。CS 61A、CS 61B、CS 61C、Nand2Tetris——技术知识开始有了层次感，从 NAND 门到高级语言的完整链路第一次在脑子里贯通。

文科硕士毕业，做了几年数据分析师。流程熟了之后，开始意识到一件事——我会用工具，但不懂工具背后的逻辑。能调 API，但不理解 HTTP 请求的完整链路；会写 Pandas，但遇到性能瓶颈只能靠 Stack Overflow。岗位天花板不是薪资，是你能解决的问题的复杂度上限。

这是决定读 CS 二学位的第一学期复盘。

In the early days of the Generative AI revolution, the industry was obsessed with "Parameters." We measured progress by the billions, then trillions, of weights packed into a model's neural architecture. But by 2026, the consensus has shifted. As we stand in the era of Gemini 3.0 and Claude 4, we’ve realized that raw intelligence is useless without a high-fidelity, low-latency "Working Memory."

Welcome to the age of Context Engineering. If the LLM is the CPU, context is the RAM. And just as in traditional computing, the way we manage this RAM defines the ceiling of what the system can actually accomplish.

In the early days of the Generative AI explosion, the industry was obsessed with the "Brain"—the Large Language Model (LLM) itself. We measured success by parameter counts, context window sizes, and benchmark scores like MMLU or HumanEval. However, as we cross into 2026, the narrative has shifted fundamentally. We have realized a hard truth: The model is not the product.

A raw model, no matter how intelligent, is like a powerful engine without a chassis, steering wheel, or brakes. In a production environment, an engine alone is a liability. The "Product" is the entire system that ensures the engine moves the vehicle safely to its destination. This realization has given birth to the discipline of Harness Engineering—the orchestration, safety, and orchestration layer that transforms a probabilistic model into a deterministic agentic system.

In the early days of Large Language Models (LLMs), prompt engineering was often derisively compared to "alchemy" or "incantations." Developers spent countless hours testing whether "please" improved model accuracy or if threatening the model with a "hypothetical fine" would elicit better code. These were the years of heuristics—vague, trial-and-error patterns that relied on the idiosyncratic behaviors of early transformer architectures.

As we move through 2026, that era is definitively over. The "Magic Spell" has died, replaced by the System Contract. Prompt engineering has matured into a disciplined branch of software engineering where natural language is treated as a high-level orchestration layer, governed by structural integrity, schema enforcement, and rigorous performance optimization. This post explores this transition and the new patterns defining production-grade AI systems.

Welcome to the new documentation site!

This blog will feature technical deep-dives, tutorials, and insights from full-stack and AI engineering projects.

What to Expect​

• Backend architecture patterns
• Frontend best practices
• AI/ML implementation guides
• DevOps workflows

Stay tuned for more content!