Software Engineering Research

Reading, Writing, and Hacking#

阅读、写作、和黑客式编程#

“A week of coding can save you an hour of reading the documentation.” What is software anyway?

What is software’s engineering?

What’s Up with Our Software?

“Adding manpower to a late software project makes it later.(The Mythical Man-Month)”

向一个已经延期的软件项目添加人力，只会让它更延期。

“Software engineering: making life easier for code peasants,given that we can’t seem to build good software quickly.”

软件工程：就是让代码农奴的日子好过一点——鉴于我们似乎就是没法又快又好地造出软件。

“Faster(productivity) and better (quality).The ultimate dream.”

更快（生产力）和更好（质量）。终极梦想。

“Kill ourselves in software development.”

在软件开发中把自己卷死/累死/耗死

The Only Thing We Can Still Do:

我们唯一还能做的：

Seek fun and profits

追求乐趣和利润。

Slove really interesting (and hard) problems!

解决真正有趣（且困难）的问题！

Find 100 bugs in GCC/LLVM; figure out why Android apps lag;…

Build useful tools!

在GCC/LLVM里找100bug;搞清楚Android应用为什么会卡顿；…

构建有用的工具！

Quick test: Would real software engineers (like those at Google or Meta) care about your work?

快速测试：真正的软件工程师（比如Google或Meta的那些人）会在乎你的工作吗？

How to Read a Paper in Three Steps 如何三步读一篇论文

S. Keshav. How to read a paper, ACM SIGCOMM Computer Communication Review, 3(37), 2007.

• Quick scan 快速扫描

• category, context, correntness, contributions, clarity 分类，上下文，正确性，贡献，清晰度，记下关键点

• Jot down the key points 

• read with care, but ignore details as proofs 仔细阅读，但忽略细节（如证明）

• Virtually re-implement 虚拟复现

• re-create the work as if you were the author 以作者的身份重新创作这项工作。

How to Read a Paper(cont’d) 如何读一篇论文（续）

A “machine-learning” approach 一种“机器学习”式的阅读方法

This paper identifies a new problem! 这篇论文提出了一个新问题！

Back-prop didn’t I notice this important peoble? 反向传播：我为什么没注意到这个更重要的问题？

This paper tackles an existing problem! 这篇论文解决了一个现有问题！

Hide the solution and try to slove it yourself (like auto-regressive models) 隐藏答案，自己尝试去解决（就像自回归模型那样）

Back-prop didn’t I come up with this solution? 反向传播：我为什么没想到这个解决方案？

We are “transformers” 我们就是“Transformer”

Traps and Pitfalls in Paper Reading 阅读论文时的陷阱与误区

Only papers solving my problem are relevant 只有解决我问题的论文才相关

• Good problem solvers have lots of tools in their toolbox 好的问题解决者工具箱里有很多工具

This paper is limited in XXXX. It’s useless! 这篇论文在XXXX方面有限制，所以没用！

• Every paper has limitations, be constructive in your criticism 每篇论文都有局限性，请建设性的批评

I give up！ 我放弃了

 - Don’t give up easily — get some background from ChatGPT 不要轻易的放弃——先向ChatGPT了解一些背景知识。

Quick Recap 快速回顾

1. Read a ton of papers  海量阅读论文

2. Read really good papers (to develop a good taste in research) 阅读真正好的论文（以培养对研究的好品味）

  - The training-data quality matters 训练数据的质量至关重要

  - “Textbooks are all you need” but… “教科书就是你所需要的一切”，但是…

    90% of “top-conference”papers have almost no “real” contributions (maybe doing well,but not groundbreaking)  90%的顶会论文几乎没有真正的贡献（也许做得不错，但并非突破性）

   - Lots of papers can mislead junior researchers 大量论文可能会误导初级研究人员。

3. Use tools wisely (citation manager, symbolic engines ,GPT,…) 明智地使用工具（文件管理工具，符号引擎，GPT…）

More Comments on Reading…

Manuel Blum’s advice to a beginning graduate student

Just don’t panic  不要慌