不知不觉两周又过去了，终于打完了《死亡搁浅 2》，可以回到《追忆似水年华》了。摆烂还是挺开心的，下班之后买菜散步然后打游戏。可惜之后可能没法按时下班了，惨啊，不过正好有时间多看看 AI 相关的东西。蹭了一节公司送的拳击体验课，太爽了！！！那种明明一点力气都没了但依然在教练的鼓励与自身的意志下努力挥拳的感觉，非常上瘾，搞得当天晚上直接失眠，but I want more!
听说前司强制项目组所有人月底前必须用 Codex 或者 Cursor 解决不少于一个工作相关的需求，我就 emmmm, 不知道该说啥了，我这种偏向 AI 激进派的都听了会摇头，这不是大家纯给领导的焦虑演戏打工么，如同当年强制打自家游戏段位和绩效绑定一样。当然，想想，我要是还在前司会有多受欢迎啊。。。
Opus 4.7 发布了，官方的技巧介绍：Best practices for using Claude Opus 4.7 with Claude Code，加的 adaptive thinking 看着还是挺有争议的；Boris Cherny 也介绍了一些技巧，这个思路还行：
many of my prompts these days look like "Claude do blah blah /go". /go is a skill that has Claude 1. Test itself end to end using bash, browser, or computer use 2. Run the /simplify skill 3. Put up a PR.Meta 也有新模型，I Tested Meta Muse Spark Against 4 Frontier Models，不知道能不能重新上牌桌呢。另外有个 Gemma4 的介绍：A Visual Guide to Gemma 4，整体看社区反馈似乎还不错的样子。
Claude 的 Context Management 小技巧介绍：Using Claude Code: Session Management & 1M Context；他们的新文章更有趣一些，managed-agents：
We virtualized the components of an agent: a session (the append-only log of everything that happened), a harness (the loop that calls Claude and routes Claude’s tool calls to the relevant infrastructure), and a sandbox (an execution environment where Claude can run code and edit files). This allows the implementation of each to be swapped without disturbing the others. We're opinionated about the shape of these interfaces, not about what runs behind them.Why Isn’t Everything Different Yet，辩护文，如果 AI 真的有 KOL 吹的那么强，那么为啥世界还没发生天翻地覆的变化。有很多理由，我还是挺认同这个的：
Fast by historical standards is still slow by Tuesday standards。想想确实如此，AI 才出来几年。我每次做这个 Curiosity Log，都会感叹，过去两周竟然发生了这么多事情。每周都有更多人在赞美 AI，Eight years of wanting, three months of building with AI，还有在 AI 帮助下用.Net 写推理引擎的，dotLLM,
On decode, dotLLM reaches 66-88% of llama.cpp throughput. Prefill is a different story - dotLLM is roughly 2-5x slower than llama.cpp across the board.当然也有偏悲观主义的文章，The Future of Everything is Lies, I Guess，里面提的一个概念很不错：
The shape of things LLMs are good at seems to be jagged. 不同于人类能力基本是从易到难平滑过渡的，LLM 的能力边界是锯齿状的，LLM 在解多元微积分的同时，却无法理解简单的词语谜题，一边非常聪明，一边又蠢到让人想笑。karpathy 的 2025 LLM Year in Review 也提到了类似的概念，
We're not "evolving/growing animals", we are "summoning ghosts".他最近也有一篇写了相关的，a lot of the capabilities are relatively "peaky" in highly technical areas. Typical queries around search, writing, advice, etc. are *not* the domain that has made the most noticeable and dramatic strides in capability. Partly, this is due to the technical details of reinforcement learning and its use of verifiable rewards. But partly, it's also because these use cases are not sufficiently prioritized by the companies in their hillclimbing because they don't lead to as much $$$ value.还有把 Open web 比作黑暗森林的，The Cognitive Dark Forest，当个思想实验看看还挺有趣的。
当然至少我认同 “Conviction Collapse” and the End of Software as We Know It 里说的：
AI will be really good at making certain processes more efficient. But it won’t be really good at making new processes unless people start to focus on that. And that’s a human creativity thing.紧接着就是这个讨论，The Z/L Continuum - Do AI engineers even need to read code anymore? 里面提到
Ryan Lopopolo from @OpenAI took the stage and said “code is a liability” and that we should all strive to be “token billionaires” - and on the last day, Mario Zechner of Pi got standing ovations after he told folks to “slow the fuck down” and “read every fucking line of critical code”，我大概偏向于自适应吧，靠自己判断要不要看代码，那其实还是偏向 slow down 这一端。里面其他观点也都挺有趣的，比如AI was supposed to make us more productive, but we all just work more，比如Many AIEs are quick to host MCP funerals, while enterprises adopt MCP and MCP Apps faster than ever, due the inherit security risks of agent skills, 比如IDEs are dying, Github is seeing unprecedented traffic levels (15x their biggest year, which was last year), so is @Cloudflare.说到 pi 作者，他有个新的分享 Building pi in a World of Slop，还是那些观点，Agent 的代码能力源于互联网，而互联网上 90% 的代码本身就是老旧的垃圾代码，所以要小心，别放手；人类虽然会犯错但是会铲屎重构；只在任务范围极其明确、代码高度模块化的前提下，才让 Agent 做去处理代码。以及 pi 作者加入公司来维护 pi 了，I’ve sold out，依然是核心部分开源。
正好上面提到了 Cloudflare，他们刚发一篇 blog，Shared Dictionaries: compression that keeps up with the agentic web，为了应对 agent 时代他们网络基础设施压力陡增，提了一个很聪明的解法：
In order to scale with more requests hitting heavier pages that are re-deployed more often, compression has to get smarter.还有这个，The Building Block Economy：
AI is okay at building everything from scratch, but it is *really good* at gluing together high quality, well documented, and proven components. And, AI prefers to do this when it can unless explicitly prompted otherwise.新版的 Codex 也带着 Computer Use 了，更让我好奇的是也带着记忆了：
We’re also releasing a preview of **memory**, which allows Codex to remember useful context from previous experience, including personal preferences, corrections and information that took time to gather.说到记忆，gbrain 挺有意思的，This is my personal opinionated OpenClaw/Hermes Agent setup with full Graph RAG, Vector Search, and retrieval on top of Karpathy's markdown as system-of-record LLM knowledge wiki.然后 Nowledge Labs 也有一篇文章：为 AI Agent 构建记忆系统。说到 Karpathy 这个 knowledge wiki，他自己又发帖说：
it's the tractable form of brain upload. 好奇的是，如果我死了，大家看我的博客所展现的我的脑子，似乎和我自己认为的我的脑子不会很一样。才发现这个评测 Task-Completion Time Horizons of Frontier AI Models - METR，还蛮好的，
The task-completion time horizon is the task duration (measured by human expert completion time) at which an AI agent is predicted to succeed with a given level of reliability.注意默认是 log scale 的，改 linear scale 才能感受到 AI 发展有多快。这个评测是从这篇文章发现的，Dario Says Continual Learning Is Solved. Is It?说到评测，Dawn Song 发了个好玩的帖子：
Our agent Terminator-1 scored ~100% on 8 major AI agent benchmarks, e.g., SWE-bench Verified & Pro, Terminal-Bench, beating Claude Mythos. It solved 0 tasks.提醒大家不能盲目信这些 benchmark 的最终结果。MiniMax cli很聪明的解法：
Instead of calling a Minimax TUI, they give you a skill that you can use to call the "mmx" command line tool from any other harness.How we built a virtual filesystem for our Assistant，
The agent doesn't need a real filesystem; it just needs the illusion of one. Our documentation was already indexed, chunked, and stored in a Chroma database to power our search, so we built ChromaFs: a virtual filesystem that intercepts UNIX commands and translates them into queries against that same database.有点意思的思路，给 agent 的 api 还是文件读写那些，下面换成了数据库的操作。黄仁勋的最新访谈，谈英伟达做什么、不做什么，说
more when necessary, less when possible，就啃着硬骨头干；也从他的角度分析了 TPU 和 GPU 的利弊，想到这个 From SIMT to Systolic: A Foundation for GPU and TPU Architecture，也是雄文：I'll try to convince you that TPU is the better platform even for the thing GPUs were supposed to own outright, which is custom kernel authoring.说到英伟达，他们发了个新模型， Lyra-2 turns an image into a 3D world you can walk through, look back, and drop a robot into for real-time rendering, simulation, and immersive applications. 挺酷的。
围绕 Mythos 和相关的，Vulnerability Research Is Cooked，Why Anthropic believes its latest model is too dangerous to release
关于训练，你不知道的大模型训练：原理、路径与新实践，随便看看；推理的更有意思，Codex 在推理框架上能蹬出什么优化，
然后亮点是它debug修复的过程中除了比较图片的PSNR和MAE之外还会用ffmpeg去截一张生成视频的图然后用多模态的方式读这个图判断画面是正常还是乱码。Codex 确实聪明，我最近在 vibe 网页，它也会自己去调用工具截图看问题。关于推理还有这个，TritonLLM v0.1.1: Agent 时代的大模型推理，
Attention is all you need，attention is also gold。我们的注意力和时间是最稀缺的资源，应该花在更有意义的事情上。以及这个：面向 SGLang 的 Profile Analysis SKILL：3 张表定位 Kernel Fuse 和 Overlap 机会。还不错的画图 skill，diagram-design: Thirteen editorial diagram types for Claude Code. Self-contained HTML + SVG. No shadows, no Mermaid-slop；也有搞笑的 skill，比如这个 caveman，省 token 也不是这么省的吧。
说到基础工具，GitButler 看着也挺有意思的，作者说 github 这种
The old model assumed one person, one branch, one terminal, one linear flow.他们的则是Designed to stack branches, to multitask, to control and organize your changes, to easily undo - to be simple, powerful and intuitive, no matter who (or what) you are.从卡马克的 帖子 的评论区看到了这个 Using a LLM to compress text，还是蛮有意思的想法，LLM 能预测正确的就不压缩了，只记录预测错误的。
Autoresearch 还有后续，Autoscaling Autoresearch: Give your agents elastic GPUs，卖铲子的来啦！
Lemire 的 The Automation of Nonsense，很真实，太多系统创造的 bullshit job 了，对这种东西拿 AI 应付就是极好的。想到《是大臣》S3E1 里 Sarah 的吐槽：
I want a job where I don’t spend endless hours circulating information that isn’t relevant about subjects that don’t matter to people who aren’t interested. I want a job where there is achievement rather than merely activity. I’m tired of pushing paper.Unbound Slop，Unbound 主打 SDF 建模，出了这个加上 AI 的版本，看着挺好玩的。
PufferLib 给强化学习环境加了 Overcooked，有品位。
10x editor 都加入ai相关功能了，
It turns out that AI agents want to use 10x too. They need fast code navigation as much as we do. Stop waiting for AI to grep files and give it access to 10x.朋友做的，unity-repl，挺有趣的，大概流程这样
Agent 发 C# 源码，Mono.CSharp 在 Unity Editor Main Thread 上直接求值，Session 里声明的类型和变量跨调用保留，直到 domain reload。确实如他所说，这样的话整个 Editor / Runtime API 都是可用工具面。我之前想过有没有可能抽象一个 bash 层给 UE，这么想确实不如 repl 来得好，当然，Unity 毕竟有 C#，UE 似乎也不是不行，考虑到有 Live++ 这种神器。AI 融入 Niagara 工作流：从数据理解到制作与优化的全链路实践，可惜我不太懂 Niagara，确实该学学，文章里的思路是对的，导出 json，然后 ai 写 json，然后导回去。AI 因为能看到裸的 hlsl，所以性能优化估计也不错。认同：
Niagara 这套系统，本质上是给人用的。大致扫了一眼，How Microsoft Vaporized a Trillion Dollars，只能说 OpenAI 包括资本市场确实敏感，可千万不能买 MSFT。
看收纳仙人的文章，A Dot a Day Keeps the Clutter Away.
来点传统技术，v8 的分享 How Many Compilers Is Too Many? A Look at V8’s History, Tradeoffs, and Architectural Choices - YouTube；关于 float，It’s OK to compare floating-points for equality，Floating point from scratch: Hard Mode；贴图压缩，The True Size of ASTC Textures; etw，Capture ETW events with C++；Tailslayer is a C++ library that reduces tail latency in RAM reads caused by DRAM refresh stalls.
在给 Seedance2.0 接 UI 的时候体验测试了下其能力，确实真的很不错。之前自己一直不是很关注生图、生视频这块，确实有空也要看看。
港股通终于开了，开始建仓 03441。这波美股反弹也确实是快，没吃满，不过也没事，安全重要。少加班，控制风险，活下去。