唉,这次更新又迟到了,这次是因为五一去广西玩了,内感洞、弄拉、通灵大峡谷、德天瀑布、蓝洞咖啡、明仕田园、南宁城里转转,喀斯特地貌一次看够,粉也确实好吃。我真还挺喜欢的,反正不懂的还可以问问 Gemini。我最喜欢的风景反而是司机带我们走一条小路去德天瀑布的时候,路在半山腰上,时上时下,看着下方的村落与上方的山头,绿意盎然。德天瀑布那边也很有趣,毕竟对面就是越南,中国这边人哗啦哗啦多,越南那边淅淅沥沥的几个,这对比也很有趣。
看到标题就笑了,Opus 4.7 isn’t dumb, it’s just lazy,Amp 的评测也不错 Opus 4.7;说到 Claude,Claude code 降智原因调查也出来了,An update on recent Claude Code quality reports;还有 GPT5.5,又是 Amp:GPT-5.5 In Deep,希望每个模型都有一个这样的 model card:GPT 5.5 - Amp.
Kimi 2.6 发了,Meet Kimi K2.6;deepseek v4 发了,很多文章,DeepSeek V4 预览版本上线并同步开源,哪些亮点值得关注,我用自家公司的 BIOS 二进制文件,考了一次 DeepSeek V4 Pro——结果让我沉默了
Two chips for the agentic era,谷歌芯片分 TPU 8t: The training powerhouse 和 TPU 8i: The reasoning engine 了,有意思。
说到训练,How GPT, Claude, and Gemini are actually trained and served;还有这个,Where the goblins came from,真的好好玩;Fine-tuning 教程:Fine-tuning LFM2.5-1.2B-Instruct with GRPO
ntroducing ProgramBench: 200 rigorous, whole-repo generation tasks where models design, build, and ship a working program end to end,这个有意思,目前通过率全是 0,但确实从这里可以看到,Opus 4.7 确实强。
tison 哥的好文 夜天之书 #119 Agentic Coding 的边界,
软件质量缺少可量化回归指标,隐性知识需要真人提供,说得多好啊;这个也是,Code is free, technical debt isn’t,Software Fundamentals Matter More Than Ever.非常不错的 harness 总结,Agent Harness Engineering,以及 pi 作者的分享 Building pi in a World of Slop,这个也不错,Harness Engineering 时代下有哪些优秀样例;以及 ykiko 的新文章,agent 时代的 clice。
学到了新名词,Mechanical sympathy.
The peril of laziness lost:
The problem is that LLMs inherently **lack the virtue of laziness**. Work costs nothing to an LLM. LLMs do not feel a need to optimize for their own (or anyone’s) future time, and will happily dump more and more onto a layercake of garbage.Agents Are Better Testers Than We Are,那确实;Agents can now create Cloudflare accounts, buy domains, and deploy,好啊,Cloudflare 确实很先进。
T$^2$PO: Uncertainty-Guided Exploration Control for Stable Multi-Turn Agentic Reinforcement Learning,好友崔崔的 paper:
我们发现多轮 agentic RL 不稳定的根源之一是“犹豫”——模型在反复生成低信息量 token,看似在思考,实则原地打转。T²PO 用不确定性来约束探索,让 agent 少走弯路,训练更稳定。非常棒!还有后续,The Bitter Lesson of Agent Harnesses:
The bitter lesson of agent harnesses: your helpers are abstractions too. Delete them. Let the agent write what it needs.Jina AI 的分享很不错,2026 年做搜索就是做 Agent Memory,顺藤摸瓜发现了 MiroThinker;说到记忆,吹了几年的 AI 个人知识库,为什么还是那么难用,这里的分层框架还不错。
谈带宽的显存容量比,大模型时代的新 roofline,
agentic推理已经逐渐滑向严重的memory capacity bound,早点看到我就满仓海力士了。Daniel Lemire on X:The trend is clear: faster and faster memory.The Supply and Demand of AI Tokens,由于复杂的强化学习环境和代码部署任务极度依赖通用计算,CPU 也处于售罄状态。Why fat tailed costs emerge at scale,唉,Long-context, agentic workloads, and more users compound tail risk; tail risk here refers not just to profit loss, but overcommitting resources and crashing systems.学学,独家对话罗福莉:AI 范式已然巨变:
要去做好Agent的Post-train。更具体说,是在Agent上怎么做好RL的scaling...至少在Chat时代,for研究、for Pre-train和for Post-train的用卡比例非常夸张,比如3:5:1,现在一个非常合理的用卡比例可能是3:1:1。OpenClaw 我能欣赏它的产品,但确实没给我带来巨震倒是,虽然我确实觉得越来越有一个一直跑着 AI 的 vps 的需求了,难道是我比较后知后觉?Getting Into AI Infra,非常好的文章,其中的练习我想了想我大概没法马上画出来:
Another good exercise is to draw a cartoon systems diagram of a gaming PC and annotate the rough bandwidths between the components.每周都会看待 agent 帮助下的算子优化,如何让 Claude Opus 4.6 写一个 100% CUBLAS 性能的 GEMM 算子;也有这种 How we got 207 tok/s with Qwen3.5-27B on an RTX 3090,和这种 llama.cpp-deepseek-v4-flash: Experimental implementation of DeepSeek v4 flaash in llama.cpp.
刷到了这本书,看着也不错,《AI Systems Performance Engineering》略读小记。
Context Rot: How Increasing Input Tokens Impacts LLM Performance,不是新文章,做 ppt 的时候发现的,写得很好。
Random thoughts while gazing at the misty AI Frontier:
AI will first automate away the things that are easier to form a closed loop learning system on. This is why code and AI research may be accelerated and then displaced quickly - you can have testable closed loop systems so machines can learn and iterate quickly. The tighter the closed loop, the faster the AI can learn.Deep Research 进化了,但我似乎很久没用了,Introducing Deep Research and Deep Research Max.
可以对照看:在 ChatGPT 中推出工作空间智能体 ,Introducing Gemini Enterprise Agent Platform.
感觉谈 multi agent 的变多了,Multi-Agents: What’s Actually Working,How A2A and MCP work together: five integration patterns for building multi-agent systems,wanman,One Developer, Two Dozen Agents, Zero Alignment
How AI is reshaping developer choice:
developer choice is shifting toward technologies that work best with the tools we’re already using.如何评价 Claude Code 核心工程师「Bash 即一切」的观点,
LLM可能永远都无法非常擅长写Bash。Bash中,对引号、括号进行匹配是个 Dyck-k 问题。然而,Transformer的电路复杂度类别是TC0,因此,它非常不擅长处理这种需要在内部维持状态的工作,理论上就无法完成任意深度的配对任务。更不要说里面还常常会碰到带转义的引号,又要分门别类处理了。刷到一篇 Computer Use 的原理,The internals of Computer Use in Claude Cowork & OpenAI’s Codex,我就说为啥都在 Mac 上先出,Mac 的 Accessibility 做得好,所以 AI 也能更方便地看到那些 UI 控件,真有趣。
有大神在帮 OpenClaw 优化 OpenRouter 的 token 消耗:we intentionally cut aggregate OpenRouter token usage by ~35%, down to ~400B tokens.
Yansu App,挺有趣的想法,让 AI 主动观察用户的行为,默默帮助你去 build 你需要的提效工具;Kami,纸张排版 skill.
Warp is an agentic development environment, born out of the terminal 开源了。
openclaw/mcporter: Call MCPs via TypeScript, masquerading as simple TypeScript API. Or package them as cli,xgrammar: Fast, Flexible and Portable Structured Generation,我竟然才发现这些。
好分析,if-i-could-make-my-own-github,虽然没啥关系,但最近 github 确实不够稳定。
Meet the New Visual Studio Debugger Agent Workflow,很需要,crash 有 windbg-mcp 了,debug 也需要。
麦老师的 Agent,KimiX:Agent Swarm 与更加高效的工具
谈 Claude Design 和 Figma 的文章 Thoughts and Feelings around Claude Design:
There’s an Arts and Crafts principle called truth to materials — the idea that a thing should be honest about what it is and how it’s made, rather than masquerading as something else. Figma ended up being the opposite of this: a set of extremely rigid schemas with a free-form “just vibes, man” costume over the top.Repairing the Ruins: Why AI Can’t Replace Education:
We tend to celebrate knowledge: facts accumulated, results confirmed, information stored. But as the biologist Stuart Firestein has argued, discovery begins not only with what we know but with a disciplined sense of what we do not yet understand. That frontier is where large language models reach their limit.当然总有一些偏悲观的文章,The West Forgot How to Build. Now It’s Forgetting Code
C vs Python & LLMs,说实话我觉得和训练语料有关,但是给 agent 以 profile 工具以及一些提示,应该让它自己优化没啥大问题。
游戏上没看到啥,零散几篇,游戏运行时 AI Native Debug 工程,米哈游坦白局:AI 全面升级游戏管线,崩坏 IP 正在做什么?
来点传统技术,这个有意思:Launch WSL Applications from Windows with WslLaunch,能想象出一种混合平台编程;虽然很久不看图形学了,但是 Metal Lossy Compression Format;Mike Acton’s Expectations of Professional Software Engineers;Daniel Lemire on X: “You can beat the binary search” / X
Hamming, “You and Your Research” (June 6, 1995) ,我也需要 Great Thoughts Time,时常觉得自己的工作太 trivial 了,一点也不重要,所以很没意思,也许我也该学着“每隔七年左右更换研究领域,可以防止思维僵化”,至少要一直学点新东西,不论是投资还是拳击还是做饭,不然太无聊了。
我自己在做个分享的 ppt,地址在这 ai_pre_for_yx. 用的是 slidev 的 skill + Claude Opus/Sonnet + cdp 截图反馈。Opus 太慢了我后面换成 sonnet,但确实体感 sonnet 的理解能力不如 opus. Claude Code 的很多小功能确实不错,btw, recap 都挺实用的。这次我还是让 AI 用 vue 组件或者 css 直接画样式包括图表的,还没上 AI 直接生成图之类,也许下次可以试试。说实话习惯了之前的所见即所得的编辑,有时候还是会被 AI 气到,即使 AI 可以自己截图来调整样式,很多视觉的东西还是自己控制比较安心。可能 slidev 确实就是不适合太精细,主打一个差不多就行,还好这次本也就是差不多就行。内容本身因为面向公司的,删了很多,反而体现自己的思考,到底要讲啥不讲啥。