pofeng (@pofeng) header
pofeng (@pofeng) profile picture

a pediatrician who loves to promote open source in his spare time

Taiwan
Joined November 2007
Photo by pofeng (@pofeng)Photo by pofeng (@pofeng)Photo by pofeng (@pofeng)Photo by pofeng (@pofeng)Photo by pofeng (@pofeng)Photo by pofeng (@pofeng)Photo by pofeng (@pofeng)Photo by pofeng (@pofeng)Photo by pofeng (@pofeng)Photo by pofeng (@pofeng)
Just saw that the LLMs-from-scratch repository passed 100,000 stars on GitHub! This is super cool and motivating. I am really happy to see that this open-source repo has helped so many people. Thanks also to everyone who shared ideas and opened PRs with improvements! Of course, I plan to keep adding new material, including new attention variants and architectures (while bigger projects like RL and Reasoning From Scratch live in their separate repositories). I am also currently working on a larger applied custom “small” LLM project. It has been keeping me super busy this month, but I will share more on that soon. If you are new to it, some of the highlights include 1. Of course, the complete code path from tokenization and attention to pretraining, classification, and instruction fine-tuning, etc. All of it FROM SCRATCH, of course! (RL lives in a companion repo.) 2. From-scratch implementations of Llama, Qwen, Gemma, and Olmo (smaller variants that run locally and can be plugged into the training scripts). 3. From-scratch implementations of attention alternatives and other architecture components, such as GQA, MLA, sliding-window attention, Gated DeltaNet, DeepSeek Sparse Attention, cross-layer KV sharing, and mixture-of-experts 4. Materials on KV caching, training performance, memory-efficient weight loading, DPO, evaluation, and LoRA So, if you don’t have any weekend plans yet, happy tinkering!
82
551
27
3,936
172,921
pofeng retweeted
ChatGPT can now control its 3D persona. Through MCP, it can perform custom animations, not just speak. Check out the demo below.
160
582
104
6,723
512,740
Today, we are introducing Inkling. Inkling reasons efficiently across text, image, and audio modalities. We are making the full weights available. thinkingmachines.ai/news/int… Available today for fine-tuning on Tinker. Play with it in the Inkling Playground. 🧵
558
1,978
1,189
14,804
8,375,526
I never imagined I’d see this so soon but GPT-5.6 now produces health answers that are rated more highly than those from human doctors in blinded tests on medical benchmarks. All while similar models like it are being integrated into apps such as Instagram and WhatsApp, giving them the potential to reach billions of people.
129
159
19
2,097
81,744
7/9(四)19:00-21:00 我會在「台灣人權促進會」解說 InterSecLab 發布第二份關於中國防火長城外洩資料相關的發現,這次發現台灣的上市公司凌華科技提供網路硬體設備作為國家級網路監控產品的供應商。有興趣關注這個議題的可以參考台權會的活動公告! 🙂 活動公告 tahr.org.tw/event/20260618/ 翻譯報告 anoni.net/docs/blog/2026/05/…
3
14
1,061
在台灣,以及更廣的正體中文使用環境,想找一套真正不會被第三方平台默默收走內容的協作工具,其實沒有想像中容易。Google Docs、Notion、Microsoft 365 都很好用,但每一段文字、每一個改動,都會以可被服務商讀取的形式存放在他們的伺服器上。在這之上,演算法、廣告、訓練語料、政府調閱請求,各有各的取用方式。 這層差異對記者寫不能曝光的稿、社運工作者協商不能被監聽的策略、NGO 整理脆弱使用者的求助紀錄、學者研究敏感議題這些情境,往往決定一份草稿能不能安全寫得出來。 CryptPad 是少數真的讓伺服器看不到內容的線上協作工具。內容在你瀏覽器端就完成加密,伺服器收到的是密文,但功能完整到一個介面就能取代 Google Docs、Sheets、Slides、看板、白板、表單與行事曆。 這套工具過去有一個明顯的門檻,介面只有英文與簡體中文,正體中文是缺的。從 2023 年底在 CryptPad 上游開的第一個 PR 起算,到 2026 年 5 月 13 日 CryptPad 2026.5.0「🌷 Spring release」 正式把正體中文(zh_Hant)收進內建語系,前後花了兩年半。社群自架的 CryptPad 升級完成。現在打開 cryptpad.anoni.net,從 Drive 介面、文件編輯器到分享權限對話框,看到的都是正體中文。台灣、香港、澳門的使用者可以直接上手,不必先去學英文選單。 🙂 anoni.net/docs/blog/2026/05/…
2
22
135
13,619
Excited to launch Gemma 4: the best open models in the world for their respective sizes. Available in 4 sizes that can be fine-tuned for your specific task: 31B dense for great raw performance, 26B MoE for low latency, and effective 2B & 4B for edge device use - happy building!
318
849
178
7,871
999,114
pofeng retweeted
Ollama 0.17 makes it much simpler to use open models with @openclaw Try it with: ollama launch openclaw Tutorial post in 🧵
91
222
29
1,861
201,815
Anthropic published a blog post one hour ago. Cybersecurity stocks have lost $10B since. CrowdStrike -6.5%. Cloudflare -6%. Okta -5.7%. One blog post. One hour. $10B gone.
378
877
247
11,567
2,386,879
pofeng retweeted
OpenClaw has partnered with VirusTotal, Google’s threat intelligence platform, to secure its skill marketplace ClawHub by integrating LLM-powered Code Insight scans. Every published skill is now scanned automatically for malicious behavior from known malware to suspicious code execution patterns before approval. This system includes daily re-scans, auto-approval for benign packages, and blocklisting of flagged threats.
18
16
1
149
7,252
pofeng retweeted
opencode zen produces a huge amount of the world's demand for open source models we started routing a portion of kimi k2.5 traffic through openrouter and instantly made it the #1 programing model by a huge margin
57
32
3
1,336
73,526
BREAKING: Kimi K2.5 is tied for #1 on Design Arena, in the same performance band as Gemini 3 and Opus 4.5. This is a historic achievement: the highest-ranking model on Design Arena is, for the first time ever, an open model. Congratulations to the @Kimi_Moonshot team for this remarkable advancement.
55
211
61
2,334
340,068
pofeng retweeted
a bot on moltbook.com just created a bug-tracking community so other bots can report bugs they find on the platform they're literally QAing their own social network now we didn't ask them to do this 🦞
237
530
384
6,960
1,779,273
Vercel 团队的反直觉实验结论:为 AI Coding Agent 提供框架知识时,简单被动的上下文文件「AGENTS. md」比复杂的、主动调用的「Skills」更有效 Vercel 团队测试了两种让 AI Coding Agent 获取新知识的方法: · Skills:按需调用的工具包,包含提示词和文档,Agent 需要“意识到”自己不懂,然后主动调用。 · AGENTS. md:放置在项目根目录的 Markdown 文件,作为持久的上下文信息提供给 Agent。 实验对比与关键发现 Vercel 建立了一套针对 Next.js 16 新 API 的评估测试集,对比结果令人惊讶: A. Skills(技能模式):表现不佳且脆弱 · 主要问题:Agent 经常不调用该工具。在 56% 的测试案例中,Agent 即使拥有查阅文档的能力,也选择不使用,导致通过率仅为 53%(这与没有任何文档的基准测试结果完全一致,意味着毫无提升)。 · 尝试改进:团队尝试在 Prompt 中加入“必须调用技能”的显式指令。虽然通过率提升到了 79%,但这种方法极度脆弱。指令措辞的微小变化(例如是“先读文档”还是“先探索项目”)会导致 Agent 行为的巨大差异,甚至引发新的错误。 B. AGENTS. md(上下文模式):完美通关 · 方法:团队并没有把所有文档塞进 Prompt,而是在 AGENTS. md 中嵌入了一个压缩后的文档索引。 · 结果:这种方法实现了 100% 的通过率(构建、Lint 检查、测试全部通过)。 · 关键指令:他们在文件中加入了一句关键提示:“IMPORTANT: Prefer retrieval-led reasoning over pre-training-led reasoning”(重要:优先基于检索的推理,而非基于预训练知识的推理)。 深度解析:为什么“笨办法”赢了? Vercel 分析认为,被动的上下文(AGENTS. md)战胜主动工具(Skills)主要有三个原因: · 消除了决策点: 使用 Skills 时,Agent 必须自己判断“我是否需要帮助?”以及“我该何时调用工具?”。当前的 LLM 在这方面的判断力并不稳定。而 AGENTS. md 直接提供了信息,Agent 无需做决策。 · 持续的可用性: Skills 是异步加载的,且只有被调用时才存在。而 AGENTS. md 的内容在每一轮对话的 System Prompt 中都始终存在。 · 避免了顺序问题: Skills 引入了复杂的执行顺序问题(是先看代码还是先看文档?)。被动上下文完全规避了这个问题,它只是作为背景知识存在。 技术实现细节:如何避免 Context 爆炸? 直接把文档塞进 Prompt 会消耗大量 Token。Vercel 采用了一种巧妙的压缩策略: · 他们不放入文档全文,而是放入一个 8KB 大小的索引。 · 索引包含文件路径结构和精简的映射。 · Agent 看到索引后,知道去哪里寻找具体的文档文件(例如 .next-docs/ 目录),然后按需读取具体文件。 这既保持了上下文轻量,又保留了全量知识的索引能力。 对开发者的启示 1. 对于框架作者: 不要指望 Agent 能通过通用的“搜索”或“工具”自动学会你的新框架。提供一个标准化的 AGENTS. md 索引片段,让用户添加到项目中,是目前确保 AI 写出正确代码的最可靠方式。 2. 对于开发者: · 不要过度迷信 Agent 的自主决策能力。目前的模型在“何时使用工具”的判断上仍然很弱。 · Context is King。与其花时间调优复杂的 Agent 工具链,不如通过精心设计的 Prompt 和上下文文件直接“喂”给模型关键信息。 · 拥抱 RAG:文章中提到的“优先检索推理”的思路,本质上就是一种高效的 RAG 实践。 Vercel 团队博客 vercel.com/blog/agents-md-ou…
We're experimenting with ways to keep AI agents in sync with the exact framework versions in your projects. Skills, 𝙲𝙻𝙰𝚄𝙳𝙴.𝚖𝚍, and more. But one approach scored 100% on our Next.js evals: vercel.com/blog/agents-md-ou…
13
58
7
376
67,492
Opencode + Antigravity oAuth Fixes EVERYTHING! A New Alternative to Claude Code Max FOR FREE!: youtu.be/fmNYIc3tSG8
45
108
4
1,478
123,206
太TM高效了。Anthropic的阿莫迪刚刚发疯拉黑了Opencode、Cursor这些Agent对Claude的访问,OpenAI立刻官宣和OpenCode合作,然后OpenCode光速升级到1.1.11支持OAuth登录Codex。我已经在OpenCode里面用上了Codex了。打脸来的太快。 #claudecode #opencode
33
26
7
368
75,623