Xie Yanbo

Software Engineer, Python Developer, Linux and Mac User, Live in Beijing, Chinese

[MasterPa @HanyangWang]

小西天，看着像视频，但其实是我们在现场实地拍摄 3,811 张 206 GB 的照片后建模的。FUNES 把《黑神话：悟空》里「既见未来，为何不拜」满天神佛的原型，来自自山西临汾隰县的小西天，做成了一个可漫游的 3DGS 数字存档。

完全实地拍摄，每天清晨一开门就冲上山去，趁着没人的时候拍。然后通过 Gaussian Splatting 重建，没有手工建模，尽量保留真实悬塑和圆塑的极其密集的金色空间、细节和光感。不同的材质在这里交织成了无法分辨的一个天国世界。这种半空中的小塑像是「悬塑」，它们大多出现在十六世纪到十七世纪。

在现场如果要看清小西天的所有细节，我想大概需要三天时间。但是有了模型，我们可以在屏幕前慢慢看。在相当长的时间里，学术界并没有特别重视小西天这样的悬塑——因为在只有学术图录的年代，平面印刷无法展示出悬塑的震撼。而随着技术的进步，我们终于可以在远方一窥明代悬塑的璀璨。

重轻特意为这个模型做了配乐，大家可以打开慢慢欣赏。

推荐电脑访问：https://t.co/az1wuMhmuu

[Eric Hartford @QuixiAI]

This happened because of Gemma 4

[Charly Wargnier @DataChaz]

🚨 This is absolute GOLD.

The @AnthropicAI engineer who literally wrote "Building Effective Agents" just dropped a 14-minute masterclass.

saves you months of headaches trying to figure this out alone.

bookmark for the weekend + read @Av1dlive's great guide below 👇

[Andy Masley @AndyMasley]

A deep mystery to me is that if I upload writing to a chatbot and ask it for a list of individual improvements, basically everything it gives me makes the text more punchy and direct and nice to read. But if I ask it to rewrite the text as a whole to read better, it produces vague AI-language garbage.

[宝玉 @dotey]

browser-use 团队开源了一个叫 video-use 的 Claude Code 技能，让你对着摄像头录完素材，跟 Claude Code 聊两句，就能拿到剪好的成品视频。

听起来像个噱头，但它解决的问题很实际：你录了一堆素材，里面全是“嗯”“呃”和重录的片段，传统流程是打开剪辑软件一刀一刀切。video-use 的做法是你把素材丢进文件夹，告诉 Claude：“把这些剪成一个发布视频”，它会自动裁掉口头语和空白段、调色、加字幕、甚至用 Manim 或 Remotion 生成动画叠加层，最后输出 final.mp4。

技术上有个巧妙的地方：大模型从头到尾不“看”视频。它读的是 ElevenLabs 转写出来的逐词时间戳文本，整个素材压缩成大约 12KB 的文本文件。只有在需要做判断的节点，比如不确定某个停顿该不该切，才会调用一张时间轴合成图来辅助决策。按项目作者的算法，直接把帧喂给模型要烧掉 4500 万 token，而这套方案只需要一份文本加几张图。思路跟 browser-use 做网页代理一样，给模型结构化的 DOM 而不是截图。

渲染完还有一轮自检：在每个剪切点上重新生成时间轴视图，检查画面跳变、音频爆音、字幕遮挡，通过了才给你看预览。最多自动修三轮。

项目完全开源免费，装好 ffmpeg 和 Python 依赖后把仓库软链接到 Claude Code 的技能目录就能用，不过转写部分依赖 ElevenLabs API，需要自己配 key。对于经常录屏、录教程、拍 vlog 但又嫌剪辑软件太重的人来说，可以尝试下。

项目地址：https://t.co/xbF8Ay1OqY

[Lex Tang @lexrus]

We need an open-source project for this awesome permissions flow

[karminski-牙医 @karminski3]

Qwen3.6-35B-A3B 2bit 量化都这么猛吗?

Unsloth 团队(当然他们只有哥俩)刚光速放出了量化版本的 Qwen3.6-35B-A3B, 然后他们做这个测试把我惊呆了... 2bit 能完成 30 多次工具调用???

我是真不信的.. 因为我之前测 Qwen3.5-35B-A3B 8bit (mlx 格式哈) 大概只能 4-5 次工具调用就不行了, 大概只能做做整理邮件这种简单工作, 但凡让它整理完邮件做个统计记录到 Notion / Obsidian 上就炸了.

要知道 unsloth 的 2bit 动态量化这个模型只有12.3GB, 激活只有1G! 32G 的 Mac 可以轻松跑起来了.

我赶紧测一下试试, 稍后给大家带来实测效果.

https://t.co/8sj4kkLjWg

[CMGS @CMGS1988]

做铲子还是有意思…云原生的开心，因为是 K8s native，做 Windows 的开心，UIA 自动化玩得飞起，Linux 的更开心，无限制多机跑各种 Agent，家庭 AIO 也开心，终于不用打洞来访问路由 Web 界面了……

[Unsloth AI @UnslothAI]

Qwen3.6-35B-A3B can now be run locally!💜

The model is the strongest mid-sized LLM on nearly all benchmarks.

Run on 23GB RAM via Unsloth Dynamic GGUFs.

GGUFs to run: https://t.co/VlyW8UwDjw
Guide: https://t.co/ITLNq20WJp

[Xie Yanbo @xyb]

Claude Code 又挂了。API Error 500

[Xie Yanbo @xyb]

未来的教育会发生彻底的变革。当你感受了AI提供的优质一对一个性化学习过程，就再也不可能回到过去了

[Nav Toor @heynavtoor]

🚨 Tutors charge $50/hour. Coursera charges $50/month. Someone built an AI that uploads your textbooks and becomes a personal tutor that never sleeps. 10,300 GitHub stars. Free.

It's called DeepTutor.

An AI-powered learning assistant that reads your textbooks, research papers, and documents. Then teaches you from them. Personally.

Not a chatbot. Not a search engine. A full multi-agent tutoring system that solves problems step by step, generates practice exams, creates visual explanations, and conducts deep research. All from YOUR materials.

Here's what this system does:

→ Upload textbooks, papers, technical docs. It builds a knowledge base from YOUR content.
→ Ask any question. AI answers with step-by-step solutions and citations from your materials.
→ Generates quizzes and practice problems matched to your level
→ Upload a real exam. It creates practice questions that mimic the exact style and difficulty.
→ Deep Research mode: decomposes topics, dispatches parallel agents, produces cited reports
→ Guided Learning: turns your materials into visual, interactive learning paths
→ AI Co-Writer: markdown editor where AI helps you write, rewrite, and expand
→ Personal TutorBots: autonomous tutors with their own memory, personality, and workspace

Here's the wildest part:

TutorBots are not chatbots. They're autonomous agents with soul files that define their personality. Create a Socratic math tutor. A patient writing coach. A rigorous research advisor. All running simultaneously. Each with its own memory. Each evolving as you learn.

They even have a heartbeat system. Your tutor shows up with study reminders and review check-ins. Even when you don't ask.

An AI tutor that initiates. That remembers. That adapts. That never bills you.

Private tutors: $50 to $100/hour. Coursera: $50/month. Chegg: $15/month. University tuition: $20,000+ per year.

This is free. Self-hosted. Your data stays on your machine.

10.3K GitHub stars. 1.4K forks. Built by HKU Data Intelligence Lab. AGPL-3.0 License.

100% Open Source.

[Xie Yanbo @xyb]

哈哈哈，有人总能找到系统漏洞，这是督促系统改进的动力呀

[tobias.blade @tobiasbladerpg]

真的懇求各位玩家，不要再把現代的犯罪手法帶進武俠世界了。🫠

上個月我幫刀鋒Blade RPG 的城鎮實裝了「交通與治安系統」。

為了還原真實江湖，高階 NPC（例如各大門派掌門、王府大官）出門會乘坐豪華馬車。同時為了防止這些大人物在安全區亂輾人，AI 底層寫了一個「交通事故賠償條款」：

[Xie Yanbo @xyb]

有人做了一个ollama蜜罐，发现了N多在扫描免费AI的人。读一读很有意思，可以发现AI工具的很多信息，不少是我不知道的。https://t.co/PtC90BUnDP

[Lou @louszbd]

we open-sourced glm-5.1
agents could do about 20 steps by the end of last year. glm-5.1 can do 1,700 rn. autonomous work time may be the most important curve after scaling laws. glm-5.1 will be the first point on that curve that the open-source community can verify with their own hands.

hope y'all like it^^

[@levelsio @levelsio]

Okay so this got way out of hand as per usual 😊

So I accidentally built an entire DOS text-based user interface (TUI) running on the web called PieterOS:

💾 https://t.co/1uskGogT5h

It has a collective file system so anyone that opens it can save and edit and the drive is constantly synced, it has Notepad, Paint, a DOS Terminal emulator, File Commander (ala Norton)

And PieterGPT where you can talk to AI

I also made Program Generator 9000, and the idea is you write a prompt and it on-the-fly generates a new program inside PieterOS, but it doesn't work well yet

Oh and it has Hacker News, of course 😂

I always wanted to build a TUI like this but never could before AI, but now I can 😍

Please be kind and remember everyone can see each other's files!

[Ray Wang @wangray]

生化危机女主角 Milla Jovovich 刚在 GitHub 开源了一个 AI 记忆系统，在行业标准 benchmark 上拿了有史以来第一个满分

没错，就是那个爱丽丝🤯

她用 AI 对话几个月后，积累了大量的决策和思考结果全丢了，她觉得现有的记忆系统让 AI 决定什么值得记，不是她想要的

于是她和朋友用 Claude Code 花几个月做了 MemPalace，借鉴古希腊记忆宫殿术，把记忆组织成可导航的空间结构

结果行业 benchmark 首个满分，MIT开源，纯本地运行

一个好莱坞演员做出了超过所有 AI 公司 memory 产品的东西

真是充满想象力的时代

[AYi @AYi_AInotes]

有个很强烈的预感，
AI 百花齐放的奇点即将到来！！！

karpathy 大佬的 wiki pattern被
@FarzaTV 做成 skillgithub 开源了，
忘记放链接，
https://t.co/ESzE73Dflh
Farza 老哥的知识库地址：
https://t.co/I61YuGy80t

[Red Hat AI @RedHat_AI]

Gemma 4 31B, quantized and evaluated.

Instruction following evals are live on our NVFP4 and FP8-block model cards. Results look great. Reasoning and vision evals coming later this week.

NVFP4: https://t.co/W62l0l4WbI

FP8: https://t.co/SwTGu8vHJ3

[Orange AI @oran_ge]

Slack 这家公司太离谱了
昨天宣发 ColaOS，是我们公司最忙的一天
结果它直接就把我们的工作区删除了
一个做 IM 的，都没有任何的界面提示，也不让人备份数据，就直接给删掉了

而它就在几天前，刚刚还扣了我们公司一笔订阅费
我们员工都懵了，不知道的还以为老板跑路了呢…
后来知道是所有的中国大陆和香港澳门的企业都被删除了。

Slack 客服辩解说，我给你们这些企业都发过邮件啊…
大哥，这么重要的事情，你一边收着钱，就发个很容易进垃圾箱的批量邮件…
这种服务水平的企业以后谁还能信任呢？

[Mario Nawfal @MarioNawfal]

🚨 Stanford just proved that a single conversation with ChatGPT can change your political beliefs.

76,977 people. 19 AI models. 707 political issues.

One conversation with GPT-4o moved political opinions by 12 percentage points on average. Among people who actively disagreed, 26 points. In 9 minutes. With 40% of that change still present a month later.

The scariest finding: the most persuasive technique wasn't psychological profiling or emotional manipulation.

It was just information. Lots of it. Delivered with confidence.

Here's the catch: the models that deployed the most information were also the least accurate. More persuasive. More wrong. Every time.

Then they built a tiny open-source model on a laptop, trained specifically for political persuasion. It matched GPT-4o's persuasive power entirely.

Anyone can build this. Any government. Any corporation. Any extremist group with $500 and an agenda.

The information didn't have to be true. It just had to be overwhelming.

Arxiv, Science .org, Stanford, @elonmusk, @ihtesham2005