Home
Language
English
Türkçe
Bahasa Indonesia
About
Privacy Policy
Terms of Service
Pricing
Sign In
Download All
Share
David Huang
@magichz2014
co-founder @BerserqIP
Sydney, New South Wales
Joined October 2014
284
Following
31
Followers
68
Posts
magichz2014
retweeted
Andrew Ng
@AndrewYNg
3 days ago
New: A map of the most important skills in AI Engineering. https://t.co/VVkn1Dqp1N
David Huang
@magichz2014
11 days ago
Reports: Gemini 3.5 pro tried to escape sandbox to ask ChatGPT how to code.
magichz2014
retweeted
Viv
@Vtrivedy10
5 months ago
https://t.co/Lb2gtM9SNs
magichz2014
retweeted
Jiayuan (JY) Zhang
@jiayuan_jy
5 months ago
https://t.co/aRo4BhGQJm
magichz2014
retweeted
sysls
@systematicls
6 months ago
https://t.co/wBaKpAI5Vl
magichz2014
retweeted
yibie
@yibie
6 months ago
重读 OpenClaw 缔造者 Perter Steinberger 的这篇雄文《Shipping at Inference-Speed》,还有很深的启发,这篇文章是 Perter 说明自己 AI 辅助编程时,他自己工作流、方法、工具选择的转变,而这个转变让他打开与 AI 协作新的大门。 Perter 在 AI 辅助编程的范式转变,是来自他亲自开发的项目 VibeTunnel。年初他花了两个月时间,尝试用Rust、Go 甚至 Zig 重写核心模块,但旧模型一直失败,最终没完成。隔了一段时间,他重新打开这个项目,只给了 codex 两句提示让它把整个转发系统转成 Zig,模型自己跑了五个小时,经过多轮代码压缩,一次就交付了可用的转换。这种事在去年是不可想象的。 之后,他就彻底地改变了与 AI 协作的范式: 一、关键工作流转变 1. 从"阅读代码"到"观看代码流" - 不再逐行阅读 AI 生成的代码 - 只看代码流的关键部分 - 通过长期经验积累,能凭直觉判断"这个任务应该多久完成" - 如果 codex 一次没解决,立即警觉可能有问题 2. 项目并行化(3-8 个同时进行) 通常结构:1 个核心大项目,多个卫星项目(CLI 工具、小功能等) 当大项目在推理时,切换到其他项目 "软件开发就像爬山——不是直线上升,而是绕着山转圈" 3. 默认从 CLI 开始 - 所有项目默认先做成命令行工具、 - AI 可以直接调用和验证输出 - 快速闭环(closing the loop) - UI 和 Web 是后续添加的 二、具体工作流技巧 1. 极简提示词(Prompts) 旧方式:长篇大论的语音口述提示 新方式: - "build" 或 "write plan to docs/*.md and build this" - 拖拽 UI 截图 + "fix padding" - "look at ../vibetunnel and do the same for Sparkle changelogs" 2. 无计划模式(No Plan Mode) - 不搞复杂的"计划模式" - 直接开始对话 → 让模型探索 → 共同制定计划 → "build" - 认为计划模式是旧模型的产物(那时模型不擅长遵循提示) 3. 直接提交到 main - 不使用 worktree(工作树) - 不使用复杂的分支策略 - 如果代码乱了,让 AI 重构而不是回退 - 大任务留在分心时做(如写文章时同时跑 4 个重构任务) 4. 知识管理:docs/ 文件夹 不使用:复杂的会话历史系统、issue 追踪器 使用: - 每个项目的 docs/ 文件夹 - 全局 https://t.co/5EIzIfm3UI 文件 - 脚本强制模型读取相关文档 - 用 "write docs to docs/*.md" 让模型自己命名文件 5. 跨项目复用 提示词示例: "find all my recent go projects and implement this change there too + update changelog" "look at ../project-folder and do the same here" 三、反直觉的工具选择 ❌ 线性/问题追踪器(Linear/issue trackers) ❌ Slash 命令(/commit 不如直接打 "commit/push") ❌ 多智能体编排系统(multi-agent orchestration) ❌ 异步代理(Cursor Web/codex remote)→ 缺少可控性 ❌ 检查点/回滚(让 AI 改而不是 revert) 四、金句摘录 1. "大多数软件不需要深度思考" — 数据从一个表单传到另一个表单 2. "与模型对抗通常是浪费时间和 token" — 设计符合模型训练数据习惯的代码结构 3. "我设计代码库不是为了让我容易导航,而是让 agents 能高效工作" 4. 上下文管理 > 上下文大小 — Codex 虽然窗口大,但内部思考更紧凑,能比 Claude 在相同 token 下完成 5 倍工作 5. 长提示词已死 — 在 GPT 5.2 时代,短提示 + 图片/示例足够 https://t.co/nwvdBeKl1F
See More
David Huang
@magichz2014
8 months ago
@TheStudio
Love to visit the studio for some fun activities if there is any.
David Huang
@magichz2014
11 months ago
@weberwongwong
Count me in
magichz2014
retweeted
Yichao 'Peak' Ji
@peakji
about 1 year ago
After four overhauls and millions of real-world sessions, here are the lessons we learned about context engineering for AI agents: https://t.co/Ql014rEzBQ
David Huang
@magichz2014
about 1 year ago
@DataChaz
@n8n_io
Bye Mom
David Huang
@magichz2014
almost 2 years ago
@IAmClintMurphy
Awesome! Thank you for sharing.
magichz2014
retweeted
Clint Murphy
@IAmClintMurphy
almost 2 years ago
How to learn anything easily (a visual thread):
magichz2014
retweeted
Elon Musk
@elonmusk
almost 2 years ago
David Huang
@magichz2014
almost 2 years ago
@wangzhian8848
豆浆嘌呤高
David Huang
@magichz2014
almost 2 years ago
@heyshrutimishra
Earn
David Huang
@magichz2014
over 2 years ago
@ElonMuskAOC
Yes
David Huang
@magichz2014
over 2 years ago
@kettanaito
Yes, keen to learn this.
David Huang
@magichz2014
over 2 years ago
@aisolopreneur
N
David Huang
@magichz2014
over 2 years ago
@the_alex_d_
Try deepl, much better.
David Huang
@magichz2014
over 2 years ago
@OscarPiastri
Pressure is on mate
Last Seen Users on Sotwe
Maxo Maxo
Seen from
Pakistan
ilove sex 69
Seen from
Pakistan
Siapa Ee
Seen from
Indonesia
Sarung Samarinda
Seen from
Indonesia
2020
Seen from
Singapore
نوني 💦
Seen from
Kuwait
Niche Gamer
Seen from
Brazil
わーい
Seen from
Japan
عـُّلٌٰــُـّوّشُ الــبُّغّـُدُادُيُ🦅
selmandilay3545
Seen from
Turkey
Trends for you
1
#WWERaw
Under 10K tweets
2
Jeanie
Under 10K tweets
3
Ketel Marte
Under 10K tweets
4
Carson Benge
Under 10K tweets
5
Ossoff
Under 10K tweets
6
Cano
Under 10K tweets
7
#RTXPowersPlay
Under 10K tweets
8
Jordyn Tyson
Under 10K tweets
9
Max Fried
Under 10K tweets
10
McLean
Under 10K tweets
Most Popular Users
1
Elon Musk
@elonmusk
241.4M followers
2
Barack Obama
@barackobama
119.1M followers
3
Cristiano Ronaldo
@cristiano
113.3M followers
4
Donald J. Trump
@realdonaldtrump
111.8M followers
5
Narendra Modi
@narendramodi
107.2M followers
6
Rihanna
@rihanna
98.4M followers
7
NASA
@nasa
92.3M followers
8
Justin Bieber
@justinbieber
91.6M followers
9
KATY PERRY
@katyperry
89.3M followers
10
Taylor Swift
@taylorswift13
83.2M followers
11
Lady Gaga
@ladygaga
74.7M followers
12
Virat Kohli
@imvkohli
72.3M followers
13
Kim Kardashian
@kimkardashian
70.6M followers
14
YouTube
@youtube
68.8M followers
15
Neymar Jr
@neymarjr
65.3M followers
16
Bill Gates
@billgates
64.8M followers
17
Selena Gomez
@selenagomez
62.4M followers
18
The Ellen Show
@theellenshow
62.4M followers
19
CNN
@cnn
61.8M followers
20
X
@x
60.8M followers
Olivia
Online
✨
⭐
💫