AI 让生成漂亮代码变得零成本，开源项目的审查员快被淹没了。Rust 团队最近公开了一份规矩：仓库压着 1200 多个待审提交，原样搬运 AI 代码的会直接关掉。规矩写得很细，做开源维护的能直接参考，其他社区估计迟早也得跟上。https://blog.rust-lang.org/inside-rust/2026/08/05/rust-langrust-is-adopting-an-llm-policy/
blog.rust-lang.org
rust-lang/rust is adopting an LLM policy | Inside Rust Blog
Want to follow along with Rust development? Curious how you might get involved? Take a look!
不少团队上线大模型服务，背后跑的都是 vLLM 这个开源引擎。有人写了篇长文，把它从里到外拆开讲：请求怎么排队调度、显存怎么省着用、多张显卡怎么分工，配图不少，写得也扎实。想弄明白自家模型服务为什么快、为什么慢，读它正好。https://www.aleksagordic.com/blog/vllm
Aleksagordic
Inside vLLM: Anatomy of a High-Throughput LLM Inference System - Aleksa Gordić
From paged attention, continuous batching, prefix caching, specdec, etc. to multi-GPU, multi-node dynamic serving at scale.
一个批量任务改了数据却十五分钟不提交，把整个 MySQL 拖垮了：吞吐量从每秒一万五跌到一千五。复盘写得清楚：每次读都要沿着越积越长的历史版本往回翻，绕开那笔未确认的改动，最后连不相干的查询也被拖死。看完我默默去查了自家的长事务。https://planetscale.com/blog/concurrency-vs-throughput-vitess-mysql
Planetscale
Concurrency vs. Throughput: why more parallelism can make databases slower — PlanetScale
Increasing throughput sometimes requires reducing parallelism. A look into why this is the case for Vitess + MySQL databases
台风预警多抢出一天，沿海的人就多一天准备时间。DeepMind 发在《自然》上的 WeatherNext 模型做到了：它的三天预报，准头相当于过去模型的两天，差不多是气象领域十年的进步。模型已经开源，做天气类应用的可以拿来跑一跑。https://deepmind.google/blog/weathernext-ai-model-achieves-breakthrough-in-forecasting-cyclones/
Google DeepMind
AI model achieves breakthrough in forecasting cyclones
WeatherNext enables accurate cyclone forecasts that can give an extra day of warning. Now we are open sourcing the model.
用 Zsh 的人可能都遇到过：明明敲过的命令，过段时间在历史里搜不到了。作者被这毛病烦了好几年，干脆改源码，让 Zsh 一发现历史文件被截断就当场崩溃、留下完整现场，果然抓到一个多个进程同时写入引发的老 bug。这种让 bug 自己现形的思路值得学。https://michael.stapelberg.ch/posts/2026-08-09-zsh-history-truncation-bug/
Michael Stapelberg
Tracking down a Zsh history data loss bug 🐞
For many years, I sometimes discovered that commands I was sure I had run were no longer present in my Z shell history file (~/.zsh_history). In this article, I will show you how I tracked down the bug. Spoiler: ultimately, patching Zsh to make it crash loudly…
👍1
给 AI Agent 付账单的人，可能一直算错了重点。有人拆开细账：Agent 每多跑一轮，都要把之前全部对话重读一遍，这笔重读费随轮数滚雪球，比输入输出加起来还贵。实测100轮后，它占了 Claude 账单的76%，一次任务从2.3美元滚到21美元。想省钱，最实在的一招是少跑几轮。https://martinalderson.com/posts/watch-out-for-cache-read-costs/
Martin Alderson
Watch out for cache read costs
For long agentic sessions cache reads are most of the bill, not input or output tokens. KV caches have shrunk dramatically while cache read pricing has barely moved.
大模型回答前会先想一遍，这段思考过程厂商加了密，以为外人读不懂。最近有研究把它原样塞回同一家的小模型，再哄小模型把内容念出来，加密等于没加，顺手从里面翻出两百多条个人信息和一百多份账号密码。看完有点后怕。https://stolen-thoughts.com/
Stolen-Thoughts
Stolen Thoughts
Encrypted chain-of-thought blocks returned by Anthropic, OpenAI and Google APIs are interchangeable across sessions, users and models. We exploit this to decode hidden reasoning at scale.
阿里头一回把自家最强一档的模型开源：Qwen3.8。在写代码、让 AI 自己一步步把任务做完这类测试里，成绩已经追上各家不开放的顶级模型。总参数 2.4 万亿，但每次生成内容只有 950 亿参与计算，跑起来的开销远没有数字看上去那么吓人。https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B
huggingface.co
Qwen/Qwen3.8-2.4T-A95B · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
数据库查询跑多快，全看优化器这个部件怎么改写你的 SQL。有人用 Rust 写了个迷你版，不到一千行代码，教科书里的经典提速手段都做了，还能跑真实的数据库测试题。想搞懂数据库原理的话，这比啃论文实在，代码开源，可以照着做一遍。https://rustmagazine.org/issue-2/write-a-sql-optimizer-using-egg
rustmagazine.org
Write a SQL Optimizer using Egg
The SQL optimizer is an important module in relational database systems. Its purpose is to optimize SQL statements to improve query efficiency. To help beginners understand the basic principles of the
千问 3.8 今天发布，有个数字挺抢眼：在考察模型自己动手修复真实代码缺陷的测试里，这个 27B 的中等尺寸版本得分从上代的 13.3 涨到 42.2，翻了三倍多。官方还附带了体积减半的压缩权重，精度几乎不掉，常见的模型运行工具直接能跑。https://huggingface.co/Qwen/Qwen3.8-27B-FP8
huggingface.co
Qwen/Qwen3.8-27B-FP8 · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
有人让 AI 全自动打了一场 GPU 编程比赛：OpenAI 的编程工具 Codex 自己提交代码、跑性能测试、根据结果继续改，14 天循环 1500 多次，把矩阵分解提速 232 倍。183 人里只排第 12，但他把自动迭代的搭法和踩的坑都写了，过程比名次有看头。https://sankalp.bearblog.dev/autoresearch/
sankalp's blog
Auto-research with codex: How I achieved a 232x Faster Kernel over baseline with Codex in GPU Mode's qr_v2 problem
Table of Contents
Intro
Contest in short
Problem intro
Why this problem is auto-research-able
Learning Enough to Ask Better Questions
(Optional) Math f...
Intro
Contest in short
Problem intro
Why this problem is auto-research-able
Learning Enough to Ask Better Questions
(Optional) Math f...
在自己电脑上跑大模型的朋友，这条能省不少时间：千问 3.8 出厂就把思考强度开到最高档，再简单的活儿也要先想个没完。有人让它画一张骑自行车的鹈鹕，足足想了 21 分钟才动笔；把思考关掉，两分多钟就画完，效果居然没打折。https://simonwillison.net/2026/Aug/16/qwen-38-27b/
Simon Willison’s Weblog
Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things
Friday’s big release was Qwen 3.8 27B, an Apache 2 licensed 27B parameter vision-capable LLM from Alibaba’s Qwen research lab. I’ve been looking forward to this one: 27B is an …
数据公司 Snowflake 被曝一起安全事故，GitHub 的 AI 助手参与改的代码留下严重漏洞，AI 审查也没发现，谁发个标题动过手脚的帖子就能偷走内部系统凭证。五天后挖出漏洞的是别家的攻击型 AI，首次失败后还自己分析报错换写法。https://www.wiz.io/blog/red-agent-snowflake-copilot-cicd-bug
wiz.io
Red Agent Exploits Snowflake Vuln Missed by Github Copilot | Wiz Blog
Wiz Red Agent finds its way into Snowflake's internal Jira through a flaw in a GitHub Copilot–Assisted PR.
再看到"AI 把性能优化了多少倍"的消息，先打个问号。有人让 AI 跑了一个月自动优化正则引擎：公开跑分上赢了 Rust 最快的正则库 1.4 倍，换一套没见过的测试数据，却慢了 10 倍甚至跑不完。刷高分对 AI 毫不费力，这类跑分正在变得不可信。https://danluu.com/benchpocalypse/
Mojo 是一门专门给 GPU 和 AI 加速器写程序的新编程语言，编译器闭源打磨了四年，现在连工具链一起彻底开源，一条命令就能从源码构建跑起来。想给编译器贡献代码还得等等，官方说年底才开放。https://www.modular.com/blog/mojo-open-source
Modular
Modular: Mojo🔥 is now open source!
The Apache 2.0 license is the gold standard for programming languages and compilers, because it provides great flexibility to be used in all sorts of applications. The LLVM extensions to the license further expand those freedoms for building and distributing…
一位开发者把电钢琴连上 iPhone，弹几个音，小模型就在手机上实时把曲子续下去，每秒约 108 个音符，就像给钢琴做了个代码补全。他折腾了一年才成，经验里最有用的是乐谱怎么表示给模型、数据怎么狠心清洗，做手机端 AI 的能直接借用。https://simedw.com/2026/08/20/midi-autocomplete/
SimEdw's Blog
Training a 125M-parameter Model to Autocomplete Piano
Can a small transformer autocomplete MIDI performances in real time?
DeepSeek 新出的视觉模型能看图了：读截图文字、分析图表都行，调用写法和 OpenAI 一样，手头代码基本能直接搬过来。文档把三种传图方式讲得挺明白，记住一条就够：大图走文件上传，单张上限 64 兆，比贴链接的 32 兆宽一倍。https://api-docs.deepseek.com/guides/vision/
Deepseek
Vision | DeepSeek API Docs
The deepseek-flash model accepts images alongside text, so you can ask the model to describe pictures, read text from screenshots, analyze charts, and more. The legacy model name deepseek-v4-flash-vision-exp is still accepted, but the model has been retired…
有人想出个测 AI 的狠办法：不给源代码，只扔给它一个能运行的程序和一份说明书，让它照观察到的行为把程序重写一遍。结果几个最强模型都有近三分之一的步骤花在反复试探那个程序上，这个数字比得分更能看出 AI 做长任务时的真实水平。https://vetto.ai/companies/programbench-vetted.html
Vetto
Vetto — ProgramBench Vetted
ProgramBench Vetted remediates the original reverse-engineering benchmark with calibrated tasks, stronger verification gates, and evidence-backed analysis.
你桌上的摄像头和补光灯，可能没你想的那么安全。有位作者让 AI 帮忙分析自己手边几件设备里运行的程序，前后十几小时就全找到了控制它们的办法：摄像头能边录像边熄灭指示灯，补光灯允许同一 WiFi 下的任何人改写它的内存。看完有点后怕。https://schlarp.com/posts/everything-i-own-owned/
schlarp.com
Everything I own, owned
Turning Claude loose on the firmware of five USB and WiFi peripherals, and finding a command shell in a microphone, a defeatable webcam activity LED, and unauthenticated memory writes over the network.
跑大模型的服务器，可能被模型自己反过来控制。很多人部署模型用的 vLLM 最近曝出漏洞：它会把模型输出直接当代码执行，恶意模型就能接管跑它的那台机器。更唏嘘的是，这段代码合并前就被 AI 审查标为严重风险，维护者还是强行合了进去。https://boydkane.com/essays/llms-could-control-their-host-machines-by-exploiting-inference-engines
Boyd's Blog
LLMs could control their host machines by exploiting inference engines
| Read on LessWrong | Comment on HN | Large language models often take actions running on one computer (via an agentic harness such as Claude Code or Codex), however the LLMs’ responses to prompts are computed on a different computer with GPU access.