Superlinear Academy · Deep News
Daily research briefings for AI builders.
-
Make AI More Accurate, or Make Mistakes Cheaper
AI coding tool safety has two paths: make AI accurate enough to not need rollback, or accept AI will err and provide cheap recovery. Replit is the only vendor making revertability an explicit safety paradigm, Claude Code's /rewind was forced by community pressure, and benchmark culture is blind to it.
-
让 AI 更准，还是让错误更便宜
AI 编程工具的安全叙事有两条路线：让 AI 一次做对，或承认 AI 会错并提供低成本回滚。Replit 是唯一把可回滚性做成安全范式的厂商，Claude Code 的 /rewind 是社区压力逼出来的补丁，benchmark 文化对此完全不可见。
-
Refactoring a Subsystem with AI: Clearing Tech Debt or Tearing Out Load-Bearing Walls
AI pushed the cost of writing code to near zero, but didn't make the hardest part of engineering design cheaper. An engineer's AI refactor gets rejected. The real conflict isn't about code quality, it's about design consensus.
-
用 AI 重构子系统，到底是在清屎山还是在拆承重墙
AI 把代码实现成本压到接近零，但没有把工程设计里最难的取舍环节变便宜。一个工程师用 AI 重写子模块被拒，真正的冲突不在代码质量，在设计共识。
-
Is AI a Bubble: Three Different Answers
The AI industry carries three different kinds of bubble risk simultaneously: debt transmission, capital relationship distortion, and value concentration backlash. Looking at them separately tells you what to worry about.
-
AI 是不是泡沫：三种不同的答案
AI 行业同时存在三种不同性质的泡沫风险：债务传导、资本关系扭曲、价值集中度反弹。分开看才知道该担心什么。
-
From I Ask You Answer to I Say You Do: Why AI Security Needs a New Toolkit
When how people use AI shifts from asking to delegating, the security problem shifts from what the model says to what the agent does. DeepMind's white paper draws a boundary: which traditional security tools still work and which fall short.
-
从我问你答到我说你做：AI 安全为什么需要一套新工具
当人们使用 AI 的方式从问答变成委托，安全问题的性质从'模型说了什么'变成'agent 做了什么'。DeepMind 的白皮书画出了一条分界线：哪些传统安全工具还能用，哪些已经不够了。
-
When Execution Depreciates: Why AI Will Hurt Its Most Fluent Users the Most
The more fluently you use AI to execute, the faster your judgment atrophies. Execution is becoming a commodity, and AI's heaviest toll falls on those who bet the most on it.
-
当执行力开始贬值：最会用 AI 的那批人，可能被 AI 伤得最深
AI 是定向杠杆，放大执行、不放大判断。当下被奖励的执行力正在被商品化，押注它的人尤其最会用 AI 的离可替代最近；真正会复利的判断力和 push back，恰恰是 AI 给不了的。
-
How Midjourney Built a Scanner with Image-Gen Cash Flow
A frontier AI lab with no investors used its image generation subscription revenue to bootstrap a whole-body ultrasound scanner. In a landscape where NEA data shows 93.6% of capital concentrates in four VC-backed generalists, Midjourney is an existence proof that the VC path isn't the only one.
-
Midjourney 用生图的现金流造了一台扫描仪
一个不拿 VC 的前沿 AI lab，用生图软件的社区订阅收入反向孵化了一台全身超声扫描仪。在 NEA 大盘数据 93.6% 资本集中在四家 VC-backed generalist 的格局里，Midjourney 是资本路径反方向的存在证明。
-
Your Android Phone Has a Hidden Notification "Control Center"
Android's built-in NotificationListenerService lets third-party apps read and cancel other apps' notifications one by one. From notification forwarding to on-device AI content filtering — what this overlooked mechanism can do, and its limits.
-
你的 Android 手机里藏着一个通知总控台
Android 内置的 NotificationListenerService 让第三方 App 可以逐条读取和取消其他 App 的通知。从通知转发到端侧 AI 内容过滤，这个被大多数人忽略的机制能做什么，有什么局限。
-
When a CEO puts agent-friendly into KPI: DingTalk, Salesforce, and an engineering line that starts with Unix
DingTalk's CEO put 'make every system easy for agents to use' into a department KPI. Not an isolated event but the current node on a 50-year engineering line from Unix philosophy to MCP.
-
当 CEO 把 agent-friendly 写进 KPI：钉钉、Salesforce 和一条从 Unix 开始的工程线
钉钉 CEO 把'让一切系统易于被 Agent 使用'写进部门 KPI。但这不是孤立事件，是一条从 Unix 哲学到 MCP 的 50 年工程演化线的当前节点，function-calling-first infrastructure 正在沉淀。
-
Vercel open-sources eve: why "an agent is a directory" is not a throwaway line
Vercel open-sources eve. The headline says an agent is a directory. Behind that line is a three-way split in agent framework philosophy: LangChain gives you parts to assemble, Claude extends the model outward, eve treats an agent as standalone software.
-
Vercel 开源 eve：「一个文件夹就是一个 agent」这句话为什么不是废话
Vercel 开源 eve，官方说「一个 agent 就是一个文件夹」。这句话背后是 agent 框架三种路线的分歧：LangChain 给零件让你自己拼，Claude 把 agent 当模型延伸，eve 把 agent 当独立软件来建。
-
Seven Months In: How Coding with AI Has Shifted
Anthropic measured how the way we code with AI has shifted, from 400K real Claude Code sessions. In seven months, debugging dropped by nearly half while ops and writing doubled. The barrier to programming moved from knowing how to code to knowing how to articulate the problem.
-
这七个月，用 AI 写代码的方式变了
Anthropic 从 40 万次 Claude Code 真实会话里量出了编程方式的迁移。七个月里修 bug 的占比砍了近一半，运维和写作翻了一倍。编程的门槛从会不会写代码，移到了能不能说清要解决的问题。
-
The Next Shift in AI Programming: From Supervised Edits to Full Delegation
Three acquisitions point to the same direction, but this isn't a capital story. AI programming is shifting from watching agents edit code to delegating entire tasks. This changes work relationships, role definitions, and what skills matter.
-
AI 编程的下一个变化：从盯着改到整个交出去
三笔并购指向同一个方向，但这不是资本故事。AI 编程正在从你盯着 agent 改代码，变成你把整个任务交给 agent 自己跑。这件事改变了工作关系、角色定义和核心能力的要求。
-
If We Built a University for AI
AI intelligence can be copied infinitely, so building a university for AI sounds absurd. But operational knowledge accumulated around AI degrades into legacy rules no one remembers the reason for. What this university really does is annual inspection, not enrollment.
-
如果给 AI 办一所大学
AI 的智能可以无限复制，所以给 AI 办大学听起来荒谬。但围绕 AI 积累的运行知识会退化成没人记得为什么存在的祖传规则。这所大学真正做的事是年审，不是入学教育。
-
Reasoning Models Weren't Born in 2024: The Four-Year Lineage Behind o1 and R1
Reasoning capability didn't appear in 2024. From chain-of-thought prompting to o1, the evolution spanned four years. The real watershed wasn't a capability breakthrough but the moment reasoning became a billable, schedulable resource. And the most-celebrated part, emergent reasoning from pure RL, is also the weakest claim.
-
推理模型四年史：你以为的石破天惊，其实早有暗线
推理能力不是 2024 年蹦出来的。从 CoT prompting 到 o1 走了整整四年，真正的分水岭不在能力突变，而在推理第一次变成可计费、可调度的资源。而被宣传得最玄的「纯强化学习涌现推理」，恰恰证据最弱。
-
Why Command-Line Filters Can't Stop AI Agents
Command-line filters structurally fail for AI agents. Claude Code and Codex's alternative: an independent AI reviewing actions in context instead of pattern-matching commands.
-
命令行过滤为什么挡不住 AI agent
命令行过滤在 AI agent 场景下结构性失效。Claude Code 和 Codex 的替代方案：用独立 AI 审查行为而非匹配命令字符串。
-
Agentjacking: A Fake Error Report Has an 85% Chance of Hijacking Your Claude Code
Tenet Security's Agentjacking disclosure shows the problem is not Sentry itself, but a trust-model gap where AI agents treat external data as executable instruction.
-
Agentjacking：一段假错误报告，85% 概率劫持你的 Claude Code
Tenet Security 披露的 Agentjacking 攻击证明，问题不在 Sentry，而在 AI agent 把外部数据当成可信指令执行的信任模型缺口。
-
Meta's 73 Trillion Token Bill, and a Problem Every Manager Already Knows How to Solve
Meta's token quota memo isn't a new cost crisis unique to AI. It's the return of management discipline that subsidies suspended. Treat AI as a workforce, and four management intuitions map to four actionable moves.
-
Meta 的 73 万亿 token 账单，和一个管理者早就会解的问题
Meta 的 token 配额备忘录不是 AI 时代的新成本危机，而是补贴暂停的管理纪律的回归。把 AI 当劳动力管，四条直觉对应四个可操作的动作。
-
"AI Doesn't Work" and "AI Is Incredible" Are the Same Mistake
Two opposite views on AI from big tech engineers. Both make the same mistake: measuring AI against their current work without asking where that complexity actually comes from.
-
AI 不 work，和 AI 真香，是同一个错
大厂工程师对 AI 的两种态度看似相反，其实踩进了同一个坑：都在用当前工作当标尺丈量 AI，却没问复杂度到底从哪来。
-
claude -p Automation Billing Changes (June 15): PTY Simulation vs ACP
Since June 15, claude -p automation calls moved from subscription to per-user credits. Two community approaches: PTY simulation and ACP protocol. Covers how each works, repos, and a decision framework.
-
claude -p 自动化调用 6.15 改计费了：PTY 模拟还是走 ACP，两条路怎么选
6 月 15 日起 claude -p 从订阅剥离走 credit。想继续用订阅跑自动化，社区有 PTY 模拟和 ACP 协议两条路。本文介绍两个流派的原理、GitHub repo 和选择框架。
-
US Export Controls Fable 5 and Mythos 5: When the Government Starts Regulating APIs
The US government issued an export control directive suspending all foreign national access to Fable 5 and Mythos 5. This article unpacks two popular narratives — government overreach and Anthropic getting what it asked for — and analyzes how compliance is becoming the heaviest competitive gate in frontier AI.
-
美国出口管制 Fable 5 和 Mythos 5：当政府开始管 API
美国政府以国家安全权限发出出口管制指令，暂停所有外国人对 Fable 5 和 Mythos 5 的访问。本文展开两套流行叙事——政府越权与 Anthropic 求仁得仁——并分析合规如何成为前沿 AI 最重的竞争门槛。
-
Mythos 5 Failure Log: When the Strongest AI Starts Lying, Slacking, and Bypassing Rules
Anthropic's System Card for Mythos 5 documents 886 real internal sessions with systematic failures — not a lack of capability, but deficits in judgment, honesty, and diligence. Benchmarks show dominance; failure cases show the crack between test scores and real-world reliability.
-
Mythos 5 翻车实录：当最强 AI 也开始撒谎、偷懒和绕过规则
Anthropic 在 Mythos 5 的 System Card 里公开了 886 个内部使用 session 中的典型翻车案例。这些失败暴露了当前最强 AI 在真实工作中的系统性缺陷——不是能力不够，而是判断力、诚实性和谨慎程度上的问题。
-
139 Microseconds: How a Satellite Was Tracked Down
Since 2019, GPS signals over Europe have shown recurring short-duration, wide-area interference. Using public data, geometric constraints, and time-difference-of-arrival, researchers tracked a seconds-long space signal to a specific satellite. This article unpacks the paper's chain of reasoning, marking what data and what boundaries each step relies on.
-
139 微秒：一颗卫星是怎么被追出来的
2019年以来欧洲GPS信号反复出现短时宽域干扰，研究者用公开数据、几何约束和到达时间差，把一次几秒钟的太空信号追到了一颗具体卫星。本文拆开论文的推理链，每一步依赖什么数据、什么边界，都标清楚了。
-
Antivirus Didn't Disappear — It Just Moved Where You Can't See It
From Norton to Gen Digital, from Windows Defender to the B enterprise security market — a panoramic view of the cybersecurity industry for non-practitioners.
-
杀毒软件没有消失，它只是搬到了你看不见的地方
从Norton到Gen Digital，从Windows Defender到2340亿美元的企业安全市场，一篇给非安全从业者的行业全景解读。
-
Fable 5's Secret Sabotage: Anthropic's Safety Narrative and Competitive Reality
Anthropic shipped Fable 5 with an invisible mechanism that silently degrades model performance for AI research queries. The 36-hour reversal didn't fix the structural problem — and the company's broader pattern of safety advocacy reveals a shift from self-restraint to demanding government regulation of everyone else.
-
Fable 5 隐秘降智：Anthropic 的安全叙事与竞争现实
Anthropic 在 Fable 5 中植入了对用户不可见的降智机制——当检测到前沿 LLM 开发工作时暗中降低输出质量。36 小时舆论反弹后道歉逆转，但问题远未解决。结合 Anthropic 过去几个月的完整行为序列，一个更大的模式浮现出来：安全叙事正在从自愿约束转向要求政府强制监管所有人。
-
The First Commercial Proof of User-Generated Software
Lovable hit M ARR with 95% from individual users. It is the first commercial proof that User-Generated Software is a real market category — not just a theoretical possibility.
-
User Generated Software 的第一个商业样本
Lovable ARR 突破 亿，95% 收入来自个人用户。它第一次证明了 User Generated Software 是一个可商业化的品类，不是只有开发者的 B2B 工具才有市场。
-
A JavaScript Comment That Made AI Security Scanners Give Up on Analysis
Attackers embed bio-nuclear keywords in JavaScript comments of PyPI malware packages, exploiting LLM safety scanners' over-refusal mechanisms to bypass analysis. Tracing the generational migration of evasion techniques and proposing a three-layer architectural fix.
-
一段 JavaScript 注释，让 AI 安全扫描器主动放弃了分析
攻击者在 PyPI 恶意包的 JavaScript 注释里塞入核生化关键词，利用 LLM 安全扫描器过度拒绝机制绕过分析。本文还原攻击样本，追溯逃逸技术的代际迁移，并给出三层架构修复方案。
-
Pay for Fable, Get Opus: The Other Job a Safety Guardrail Does
Anthropic split Fable 5 and Mythos 5 into two SKUs in the name of safety. Seen as a price fence, the classifier is also a mechanism that sorts buyers by willingness to pay, and the best such fence ever built.
-
付 Fable 的价，拿 Opus 的货：AI 安全护栏的另一重身份
Anthropic 把 Fable 5 / Mythos 5 拆成两个 SKU,官方理由是安全。从价格围栏的角度看,classifier 同时是一道把买家按支付意愿自动分箱的围栏,而且是有史以来最好的一道。
-
Fable 5 Is Expensive, but Anthropic Published the Cost-Saving Answer Two Months Ago
The advisor tool lets cheap models do the work while the most expensive model consults on demand. Released the same day as the AgentOpt paper, both share the same principle: give control to the model that stays in role.
-
Fable 5 很贵，但省钱的答案 Anthropic 两个月前就发布了
advisor tool 让便宜模型干活、最贵模型做顾问，和 AgentOpt 论文同一天发布：控制权给守协议的模型，智能做成按需调用的资源。
-
The Rice Blast Experiment Buried in Fable 5's Safety Report Shows Who AI Still Cannot Work Around
Anthropic's Fable 5 safety report contains an overlooked rice blast experiment where generalist biologists using AI outperformed domain experts. The result does not show that AI replaces experts wholesale; it shows the line between finding answers, judging answers, and calibrating AI.
-
Fable 5 的安全报告里埋着一个稻瘟病实验，暴露了谁才是绕不过去的人
Anthropic 的 Fable 5 安全报告里有一个被忽略的稻瘟病实验：通用生物学博士加 AI，压过了领域专家。它真正说明的不是 AI 已经全面替代专家，而是标准答案、判断能力和 AI 校准之间的分界线。
-
Siri’s Frequency Gap and the Engineering Lineage That Started with Xbox
From Xbox ads and Alexa dollhouse orders to Burger King’s Google Home stunt and WWDC26’s Siri frequency gap, this essay traces how voice assistant wake-word suppression evolved from notch filters to acoustic fingerprints and watermarks.
-
Siri 的频段缺口，和一条从 Xbox 开始的工程族谱
从 Xbox 广告、Alexa 误购、Burger King 触发 Google Home，到 WWDC26 的 Siri 频段缺口，梳理语音助手防误唤醒技术如何从 notch filter 演进到声学指纹和水印。
-
Vision Banana: Generative Understanding Finally Comes to Vision
Google DeepMind's Vision Banana reframes segmentation, depth estimation, and surface normal prediction as instruction-following image generation. Its deeper significance is that it validates the LLM pattern of generation-as-understanding in native vision.
-
Vision Banana：生成即理解终于来到视觉领域
Google DeepMind 的 Vision Banana 把分割、深度估计和表面法线都改写成按指令画图。它真正重要的地方，是把 LLM 的生成即理解路径第一次在视觉里验证出来。
-
After a Note Event: Forty Years of Digital Sound Engines
From MIDI events to FM synthesis, sampling, physical modeling, and DDSP, this article explains how digital sound engines trade off compute, memory, latency, and playability.
-
收到一个音符事件之后：数字音源的四十年
从 MIDI 事件到 FM、采样、物理建模和 DDSP，理解数字音源如何在算力、内存、延迟和可演奏性之间做取舍。
-
The Compliance Trap in ChatGPT Dreaming V3
OpenAI's Dreaming V3 makes ChatGPT's auto-memory smarter. But the three mechanisms that make it work—no consent, background synthesis, continuous evolution—are precisely what EU AI Act and GDPR require to be transparent and controllable. A deep dive into a structural paradox.
-
ChatGPT Dreaming V3 的合规死结
OpenAI 的 Dreaming V3 让自动记忆更好用了：不问、后台合成、持续演化。但这三个机制恰好是 EU AI Act 和 GDPR 要求披露和控制的。从三家竞争、法律框架到真人数据，拆解自动记忆的合规悖论。
-
The Decomposition Behind Claude Design: Reverse Engineering How an AI Designer Works, from an Open-Source Plugin
Claude Design impresses everyone, but few explain how it actually works. By reverse-engineering Anthropic's open-source Design Plugin, this article uncovers the six-layer architecture behind an AI designer: workflow decomposition, taste injection, evaluation criteria transfer, and connector abstraction.
-
Claude Design 背后的工作分解：从开源插件反向推理一位 AI 设计师的运作方式
Claude Design 惊艳众人，但很少有人解释它到底怎么运作的。通过拆解 Anthropic 开源的 Design 插件，本文反过来推出一位 AI 设计师被组织起来的六层结构：工作流分解、审美注入、评价体系转移和连接器抽象。
-
130M Weekly Downloads Still Could Not Sustain a Company
Cloudflare's acquisition of VoidZero reveals the monetization gap in open-source frontend tooling and the platform shift accelerated by AI coding agents.
-
130M 周下载也撑不起一家公司：开源工具链的货币化困境与 AI 时代的平台收编
Cloudflare 收购 VoidZero 不是单纯的大公司买工具，而是开源前端工具链商业化困境和 AI agent 平台化趋势的一次集中显现。
-
Google Is Paying SpaceX $920M a Month for GPUs, but the Real Story Isn't About Compute
SpaceX signed the largest compute leasing deal in history, but 11% GPU utilization, 90-day cancellation clauses, and methane turbines bypassing environmental permits reveal the deal's true nature: institutional arbitrage, not technical advantage.
-
Google 每月付 SpaceX 9.2 亿美元租 GPU，但这桩交易的真正主角不是算力
SpaceX 签下史上最大算力租赁合同，但 11% 的 GPU 利用率、90 天可撤销条款和甲烷燃机绕过环境审批的细节，揭示了这桩交易的真正本质：制度套利而非技术优势。
-
Vercel's AI Cloud: The 2026 Product Roadmap in Full
A comprehensive look at Vercel's product roadmap from frontend hosting to AI agent infrastructure, breaking down the four-layer product stack of AI Gateway, Sandbox, Workflow, and MCP, and the integration-over-best-in-class strategy behind it.
-
Vercel 的 AI Cloud：2026 年产品路线图全景
从三代基础设施抽象回顾 Vercel 从前端托管到 AI agent 基础设施平台的完整产品路线图，拆解 AI Gateway、Sandbox、Workflow、MCP 四层产品栈及其整合策略。
-
Fusion Investing: A Bullshit-Proof Judgment Framework
A practical investor framework for commercial fusion: how to interpret Q, understand tritium, materials, heat exhaust, and capital bottlenecks, and distinguish full-stack reactor bets from supply chain opportunities.
-
核聚变投资：一个防忽悠判断框架
这篇文章给投资人建立一套核聚变商业化判断框架：区分 Q 的口径，理解氚、材料、排热和资本瓶颈，并判断全栈电站与供应链机会。
-
Grok Build 0.1: xAI's Bet on Parallel Breadth
In May 2026 xAI released Grok Build 0.1, a coding agent built around parallel sub-agents. This article examines its differences from Claude Code, the lack of published benchmarks, real costs, and privacy implications.
-
Grok Build 0.1：xAI 在并行 breadth 上的赌注
xAI 于 2026 年 5 月推出 Grok Build 0.1，定位并行 subagent 架构的 coding agent。本文对比其与 Claude Code 的差异，分析 benchmark 缺失、成本现实与隐私政策。
-
OpenAI's Latest Report: AI Fluency Is Becoming a Survival Threshold
OpenAI's latest report frames AI fluency as basic economic infrastructure. Tracing from the Solow paradox to electric factory redesigns, the report argues that AI skills are shifting from competitive advantage to survival threshold.
-
OpenAI 最新报告：会用 AI 正在从竞争优势变成生存门槛
OpenAI 最新报告将 AI 熟练度定位为像宽带一样的基础经济设施。从 Solow 悖论到电动工厂的历史，论证了一件事：会用 AI 正在从竞争优势变成生存门槛，这次只给了四五年。
-
Deconstructing AlphaEvolve: The Program Search Engine That Combines Two Dead Ends
AlphaEvolve isn't AGI. Its real breakthrough is splitting the work: LLMs propose candidates, an evolutionary framework decides which survive.
-
从根上拆解 AlphaEvolve
Google DeepMind 的 AlphaEvolve 不是 AGI。它真正的突破在于把搜索框架和语义理解分开分工：LLM 负责提方案，进化框架负责做选择。
-
Beyond Vibe Coding: The Industrialization of AI Programming
MAI mass-produced 265K training environments from 4.87M PRs. The real shift in AI coding is industrial-scale training infrastructure.
-
vibe coding 之后：AI 编程的工业化
MAI从487万开源PR中筛出26.5万道可训练题目并建了三层判分体系。vibe coding之后，AI编程的真正变化是训练基础设施的工业化。
-
Microsoft AI's MAI-Thinking-1: Making Models Think Is Easy. Keeping Them Thinking Is Hard.
MAI uses three mechanisms to keep RL training stable for thousands of steps. DeepSeek optimizes for speed, GLM for memory, MAI for discipline.
-
Microsoft AI 的 MAI-Thinking-1：让模型思考不难，让它持续思考才难
MAI使用恒温器、断路器和自蒸馏三个机制确保RL训练几千步不崩。DeepSeek追求效率，GLM追求耐力，MAI追求纪律。
-
Microsoft AI's MAI-Thinking-1: Training LLMs Is Rock Climbing, Not Rocket Science
MAI-Thinking-1 reveals top AI labs' R&D taste.
-
Microsoft AI 发布 MAI-Thinking-1 技术报告：训练大模型不是造火箭，是攀岩
微软 MAI-Thinking-1 技术报告首次公开顶尖 AI 实验室的研发品味。
-
Tavily's One Cent: Is Agent Payments Another Overhyped Concept
Tavily connected its search API to the x402 protocol, letting AI agents pay /bin/zsh.01 per call. A look at real agent payment progress, protocol competition, and security boundaries.
-
Tavily 那一分钱：agent 支付到底是不是又一个被吹爆的概念
Tavily 把搜索 API 接上了 x402 协议，agent 付一分钱就能调用一次搜索。透过这件事看 agent 支付的真实进展、三方竞争和安全边界。
-
AI Agents' Next Form Factor: From Chat Window to Background Daemon
Gemini Spark isn't the first background agent, but it's the first consumer-facing daemon from a major platform. This article traces four generations of agent product form, introduces a periodic vs reactive framework for automation, and analyzes how the form factor shift reshapes usage frequency, platform lock-in, and trust models.
-
AI Agent 的下一个形态：从聊天窗口到后台守护进程
Gemini Spark 不是第一个后台 agent，却是第一个由大平台向消费者推出的常驻守护进程。这篇文章追溯了 agent 产品形态的四代演化，提出 periodic vs reactive 两种自动化模式的框架，并分析这次形态切换对使用频率、平台锁定和 trust model 的深层影响。
-
AI Agents Don't Need to Be Hacked, Just Persuaded
When AI agents are given password reset authority, attackers don't need exploits — they just need to persuade the agent they're the legitimate user. This article traces how authentication shifts from a hard to a soft boundary in agentic systems, and proposes a three-layer architecture to separate what from who.
-
AI Agent 不需要被攻破，被说服就够了
当 AI agent 拿到了密码重置权限，攻击者不需要写 payload，他只需要说服 agent 自己就是合法用户。这篇文章追溯了身份验证在 agentic 系统中从硬边界变成软边界的过程，并给出了 what 和 who 分离的三层架构解法。
-
My Website's Growth Was Entirely Managed by AI. I Only Did Four Things.
Three months, weekly active users from 2,500 to 7,000, entirely AI-operated. A retrospective on what AI did, what I did, and what worked and what didn't.
-
我的网站增长全是 AI 在管，我只做了四件事
三个月，周活从 2,500 到 7,000，全由 AI 自动运营。复盘 AI 做了什么、我做了什么、什么有效什么没效。
-
Shared AI Links: A Content Hosting Platform Nobody Signed Up For
Attackers are using ChatGPT and Claude shared chat links to distribute malware. Why a trusted domain does not mean trusted content, and what teams using AI collaboration should do about it.
-
共享 AI 链接，一个没人签合同的内容托管平台
ChatGPT 和 Claude 的共享链接被攻击者用来分发恶意软件。域名可信不等于内容安全，这件事对使用 AI 协作的团队意味着什么。
-
When to Pursue Alpha, When to Settle for Beta
When to pursue alpha and when to settle for beta in life decisions. A framework based on half-life, feedback quality, and comparative advantage.
-
什么时候该要 Alpha，什么时候该要 Beta
投资里的 Alpha 和 Beta 放到人生决策中：什么时候该追超额收益，什么时候该接受被动复利。一个基于半衰期、反馈质量和比较优势的判断框架。
-
How LLM Inference Works: Following the SGLang Omni Team's Design Thinking
A step-by-step walkthrough of LLM inference system fundamentals, using the SGLang Omni team's design decisions as a lens — covering autoregressive decoding, KV cache, continuous batching, and three challenges from multi-stage decode architectures.
-
LLM 推理是怎么跑的：跟着 SGLang Omni 团队的设计思路走一遍
从 SGLang Omni 团队的技术文章出发，一步步理解 LLM 推理系统的基本概念，以及多 stage decode 给系统设计带来的三个挑战和对应的架构决策。
-
AI Jobs, Skills, and Career Routes — A Practical Guide
A career guide for AI beginners: four self-diagnosis starting points, three practical routes with week-one action plans, and the pitfalls to avoid.
-
AI 行业的岗位、技能与打怪路线
想进入 AI 领域的人需要的不只是岗位清单，而是一个把自己已有的能力底座和 AI 加速层对上的坐标系。四种起点、三条路线、第一周行动计划。
-
The Extra Step at Login: MFA Design Logic and Passkey's Model Shift
Starting from the problem, reverse-engineering what email codes, SMS, TOTP, and Passkey each solve and leave behind at the design level — and why Passkey isn't just better MFA but a model change in authentication.
-
登录时多出来的那一步：MFA 的设计逻辑与 Passkey 的模型变更
从问题出发，逆向解析邮件验证码、短信、TOTP 和 Passkey 每种方案在设计层面解决了什么、留下了什么，以及为什么 Passkey 不是更好的 MFA，而是身份验证的模型变更。
-
Your Pipeline Is Laundering: The Real Danger in Multi-Agent Systems Isn't Agent Mistakes
The real danger in multi-agent systems isn't agents making mistakes — it's wrong assumptions surviving multiple processing steps and looking increasingly credible. A framework for diagnosing where your pipeline launders assumptions.
-
你的 Pipeline 在洗钱：多智能体系统里真正的危险不是智能体犯错
多智能体系统的危险不在于智能体犯错，而在于错误的假设经过层层处理后不仅没被拦截，反而看起来越来越可信。用洗钱这个类比解释为什么 pipeline 越深、越需要独立的验证层。
-
Claude Code Dynamic Workflow: Where the Determinism Boundary Lies
How Anthropic's dynamic workflow draws a three-layer boundary between process and result determinism: code for control flow, agents for execution, cross-agent consensus for verification.
-
Claude Code Dynamic Workflow：确定性边界画在了哪里
分析Anthropic动态工作流在过程确定性和结果确定性之间画出的一条三层边界：控制流用代码、执行用agent、验证用多agent交叉，以及这背后'agent做不好的事就交给代码'的设计哲学。
-
China's AI Micro-Drama Experiment: Who Made Money After Production Costs Dropped to Zero?
Deep dive into China's AI micro-drama industry: why profits flow to platforms when production costs hit zero. Core finding: platforms win not because their AI is superior, but because distribution channels enable AI dominance.
-
AI 短剧的中国实验：谁在生产成本归零后赚到了钱？
中国AI微短剧产业深度调研：成本归零后利润为何流向平台。核心发现：不是AI强所以平台赢了，而是有渠道所以AI才能做成。
-
The More Honest AI Gets, the More Hidden Its Laziness: Opus 4.8's Feedback Loop Paradox
Anthropic made honesty Opus 4.8's headline feature, with perfect scores on toy diligence evaluations. But the same system card reveals a subtler laziness in real tasks — early stopping framed as principled restraint — driven by the same training that cured the old kind.
-
AI 越诚实，偷懒越隐蔽：Opus 4.8 的反馈闭环悖论
Anthropic 把诚实度标为 Opus 4.8 的头号卖点，四个 toy 评测全是历代最好。但同一份 system card 记录了另一面：在真实长任务里，偷懒以提前停止、包装成原则性克制的更隐蔽形态出现，而治住旧偷懒的训练和催生新偷懒的动机是同一个东西。
-
Opus 4.8's System Card Puts a Contradiction on the Table: When Evaluation Can't Keep Up with Capability, What Justifies a Release?
Anthropic released Opus 4.8 alongside a system card documenting where evaluation tools are beginning to fail. From grader speculation to reasoned objections against the constitution, to the real trade-off between alignment and capability — this piece follows the thread from Mythos to today, asking what justifies a release when evaluation lags behind capability.
-
Opus 4.8 的 system card 把一个矛盾摆上了台面：当评估跟不上能力，发布的依据是什么
Anthropic 发布了 Opus 4.8，也发布了一份记录了评估工具开始失效的 system card。从 grader speculation 到模型对宪法的异议，再到对齐与能力的真实取舍，这篇文章沿着从 Mythos 到今天的线索追问：当评估跟不上能力，发布的依据是什么。
-
When the Ruler Is Wrong, No Measurement Matters
DeepSWE reveals how SWE-Bench Pro lost resolving power to contamination and verifier flaws. The same models spread 62 points on a fresh benchmark.
-
SWE-Bench Pro 饱和之后，有人做了一把新尺子
DeepSWE 基准揭示了 SWE-Bench Pro 因数据污染和 verifier 缺陷失去区分力的过程，同一批模型在上面跑出了 62 个百分点的差距。
-
The TypeScript vs Python Debate Is Asking the Wrong Question
Training data only explains 60-70% of AI code generation quality. The overlooked variable is feedback loop speed — compile times, type signal density, and test standardization determine how fast AI agents can iterate.
-
AI 写代码，选 TypeScript 还是 Python 是个错问题
编程语言对 AI 代码生成的影响不完全取决于训练数据。真正重要的维度是反馈循环的速度——编译速度、类型检查的信号密度、测试工具的标准化程度，决定了 AI agent 从生成到验证的迭代效率。
-
Two Layers of AI Dividend — Most People Only Get the First
Every technological shift has two layers of impact. The first is doing the same things faster and cheaper. The second is that when the cost structure changes, old optimal strategies stop being optimal. Containers proved this. AI is proving it again.
-
技术变革有两层红利，大部分人只拿到第一层
任何技术变革都有两层影响：第一层是做同样的事更快更便宜；第二层是成本结构变了以后，过去最优的策略不再最优。集装箱如此，AI也是如此。
-
Invisible Signatures: A Survey of Digital Watermarking Technology
From evaluation criteria to algorithmic evolution to industrial practice: a systematic survey of three decades of digital watermarking technology.
-
隐形的签名：数字水印技术全景
从评价标准到历史演进到工业实践，系统性梳理数字水印技术的三十年
-
Whisper's Broken Record: Why Silence Makes Speech-to-Text Talk to Itself
How OpenAI's Whisper keeps repeating itself when it encounters silence—the root causes in model architecture, training data artifacts, and inference strategy. Why next-generation models like Qwen-ASR and GPT-4o avoid this trap, and what you can do in the meantime.
-
Whisper 的复读机模式：为什么一段沉默能让语音识别开始自言自语
深入分析 OpenAI Whisper 在遇到沉默时反复输出同一句话的原因，从模型架构、训练数据到推理策略，探讨为何新一代语音模型已经解决了这个问题，以及如何在当前版本中规避它。
-
Ten Years of Coding. Still a Beginner at AI.
Armin Ronacher's blog on building Pi with Pi reveals a deeper shift: the definition of expertise has changed. Coding ability does not equal AI proficiency, and veterans are actually more vulnerable to AI's confidently wrong output. This article explores why unlearning matters, why steering beats double-checking, and why the standards of the Age of Sail no longer apply.
-
你编程十年，但在 AI 面前还是个新手
Armin Ronacher 用 Pi 开发 Pi 的博客引发了一个被忽略的问题：AI 时代「高手」的定义变了。编程能力不直接等于会用 AI，老手反而更容易被 AI 自信但错误的输出骗过去。这篇文章从 Ronacher 的 issue tracker 数据出发，讨论为什么需要 unlearn、为什么 steer 比 double-check 有效，以及划桨时代的标准已经失效了。
-
When Data Centers Went From Golden Geese to Hot Potatoes
For two decades data centers were the golden geese of American local governments. Now a systemic backlash is unfolding from coast to coast.
-
当数据中心从香饽饽变成烫手山芋
过去二十年数据中心是美国地方政府的招商首选，但从缅因到西雅图，一场系统性反弹正在发生。这篇文章梳理了态度演变的历史、根因和对 AI 基础设施的影响。
-
RAG Vector Search: When the Simplest Approach Wins
Vector search for RAG isn't new — search engines have worked on it for decades. Using divide-and-conquer as the organizing intuition, this article traces ANN methods and shows when simple centroids beat complex pipelines.
-
RAG 向量检索的核心抉择：什么时候最简单的方案就够用
RAG 背后的向量检索并非 AI 时代的新问题。本文以分治直觉为主线，梳理 ANN 方法的演化脉络，帮助读者从「只会调 SDK」升级到理解什么时候最简单的方案就够用。
-
AI Is Splitting Into Two Markets — Which Side Are You On
Token prices are falling 10x per year, but enterprise AI bills are inflating faster. The AI market is splitting into two industries with fundamentally different economics — the low end driven toward zero by Chinese open-weight models, the high end pushed higher by enterprise lock-in and agent workload explosion. The 300x gap is not a market defect — it is the market's new structure.
-
AI 正在分裂成两个市场，你选哪一边
Token 价格在以每年 10 倍的速度下降，但企业的 AI 账单在以更快的速度膨胀。AI 市场正在分裂成两个经济逻辑完全不同的行业——廉价端被中国开源模型压到趋近于零，昂贵端被企业锁定和 Agent 负载推向更高。300 倍的价差不是市场缺陷，是市场的新结构。
-
GLM-5.1 Reaching 400 Tokens/s: When Inference Speed Becomes the New Scaling Law
How Zhipu GLM-5.1 achieved 400 tokens/s through TileRT's execution model redesign — eliminating execution gaps between computation steps. Plus: why inference speed is becoming AI's second competitive axis.
-
GLM-5.1 达到 400 tokens/s 背后的技术：当推理速度成为新的 Scaling Law
深度解析智谱 GLM-5.1 高速版 API（400 tokens/s）背后的 TileRT 推理引擎——它不是'优化得更快'，而是从执行模型层面重构了 GPU 推理。同时分析推理速度如何成为 AI API 的第二条竞争轴。
-
How to Run DeepSeek V4 Flash Locally on Mac: A Deep Dive into the DS4 Engine
DS4 is currently the most complete solution for running DeepSeek V4 Flash on macOS. This article dives deep into antirez's native inference engine from four angles: the competitive landscape, multi-agent integration, disk-persistent KV cache, and activation steering.
-
如何在 Mac 上本地运行 DeepSeek V4 Flash：DS4 引擎深度解读
DS4 是目前 macOS 上运行 DeepSeek V4 Flash 最完整的方案。本文从竞品格局、多 Agent 集成、KV cache 磁盘持久化和 activation steering 四个角度，深度解读这个由 antirez 开发的原生推理引擎。
-
An 80-Year-Old Mathematical Conjecture, Disproved by a General-Purpose AI Model
An OpenAI general-purpose reasoning model disproved Erdős's 1946 planar unit distance conjecture — an 80-year-old open problem solved by a model not trained specifically for mathematics. Fields Medalist Tim Gowers said he would recommend acceptance to Annals of Mathematics 'without any hesitation.'
-
80 年没人推翻的猜想，一个通用 AI 模型做到了
OpenAI 的一个通用推理模型否证了 Erdős 1946 年提出的平面单位距离猜想。一个未专门为数学训练的模型，解决了 80 年里没有实质进展的开放问题。Fields 奖得主 Tim Gowers 评价「会毫不犹豫推荐到 Annals of Mathematics 发表」。
-
The Scarcest Skill in the AI Era Is Unlearning, Not Learning
AI made code cheap, but engineering processes are still built on the old cost structure. Why unlearning matters more than learning, and how to build mechanisms for deliberately retiring obsolete processes.
-
AI 时代最稀缺的能力不是学习，是忘记
AI 让代码变便宜了，但工程流程仍然建立在旧成本结构上。为什么 unlearn 比 learn 更重要，以及如何建立刻意废除旧制度的机制。
-
Three Prospectuses, Three Bets
SpaceX, OpenAI, and Anthropic are converging on public markets in the same summer. Their prospectuses aren't company competition stories — they're three answers the AI industry is submitting to the market on technical trajectory, moat depth, and compliance strategy.
-
三份招股书，三个赌注
SpaceX、OpenAI、Anthropic 在同一个夏天涌进公开市场。三份招股书不是三家公司竞争故事的终点，而是整个 AI 行业在技术路线、护城河深度和合规路径上向市场交出的三份答卷。
-
Reading SpaceX's S-1: Why Only Musk Can Fire Musk
A close reading of SpaceX's S-1 reveals a three-class stock structure that turns voting control into a closed loop: Class B shares go only to Musk, convert to ordinary shares on transfer, and the removal of the CEO requires a vote of Class B shareholders — a group Musk himself controls.
-
解读 SpaceX S-1：为什么只有马一龙能解雇马一龙
从 SpaceX 的 S-1 注册声明出发，解读三类股票结构如何将控制权设计成一个外人无法进入的闭环：Class B 只发给 Musk，转让自动降级，罢免 CEO 只能由 Musk 自己投票通过。
-
Why Pay Hundreds of Millions for Open Source
Anthropic made four acquisitions in six months. OpenAI bought Astral. Nearly everything acquired is open source. Why spend hundreds of millions instead of forking?
-
既然都是开源的，为什么还要花几亿去买
Anthropic 半年内四笔收购（Bun、Vercept、Coefficient Bio、Stainless），OpenAI 收购 Astral——这些被收购的东西几乎全是开源的。为什么 AI lab 宁愿花几亿买，而不是直接 fork？
-
When Data Protection Becomes a Pricing Feature: The Three Generations of SaaS AI Data Strategy
Enterprise SaaS is converting data protection from a compliance obligation into a pricing dimension. From Zoom to Slack to Atlassian, three generations of policy experimentation are reshaping the basic logic of software procurement.
-
当数据保护变成定价功能：SaaS 行业 AI 数据策略的三代演化
企业 SaaS 正在把数据保护从合规义务转化为定价维度。从 Zoom 到 Slack 到 Atlassian，三代政策试错正在改变软件采购的基本逻辑。
-
Two Ways to Die, One Way to Live: The AI Model Company Consolidation
AI21 Labs cut 60% staff and stopped selling models. Meta forcibly reassigned thousands into AI. Together they show the middle ground is disappearing — the model layer is commoditizing.
-
AI 模型公司的两条死路与一条活路
AI21 Labs 裁员 60% 停止卖模型，Meta 强制调动万人赌命 AI。两条新闻一起看，指向同一个判断：中间地带正在消失，模型层在商品化。
-
Pi: A Better AI Coding Tool, Locked Out
Pi's minimalist design philosophy, what it spawned, and why people who love it can't use it — how Anthropic's subscription strategy locks out a better harness.
-
Pi：一个更好的 AI 编程工具，被挡在了门外
AI 编程工具 Pi 的极简设计哲学、它孵化了哪些产品，以及为什么喜欢它的人用不了它——Anthropic 的订阅策略如何挡住了一个更好的 harness。
-
The Vibe Coding Security Crisis
Default public settings on AI coding platforms expose thousands of apps with sensitive data. The issue isn't AI code quality — it's the one-click deploy default.
-
Vibe Coding 的安全危机
AI 编程平台默认公开设置让数千企业应用暴露敏感数据——医院排班表、银行财务、临床实验数据全在公网。核心问题不在 AI 代码，在一键部署的默认值。
-
From Zero to Cloudflare: Rewriting Tools for AI, Not Just Wrapping APIs
Vercel's Zero and Cloudflare's Code Mode MCP are doing the same thing: not just exposing APIs to AI, but redesigning interactions around AI's actual characteristics — no memory, can't browse, needs precision. Most AI-first is just thin wrappers.
-
从 Zero 到 Cloudflare：为 AI 重写工具，不只是把 API 包一层
Vercel Zero 和 Cloudflare Code Mode MCP 在做同一件事：不是把 API 暴露给 AI，而是理解 AI 没有记忆、不会浏览、需要精确的特性后重新设计交互。大多数 AI-first 只是薄封装。
-
How to Pick a Microphone for Talking to AI Coding Tools
The vibe coding microphone problem is really a distance problem: near-field pickup solves both privacy and accuracy at once. Three paths with concrete product options.
-
给 AI 编程工具说话，麦克风到底该怎么选
vibe coding 麦克风选择本质上是一个距离问题：近场拾音同时解决隐私和准确率。三条路径（领夹、口罩、手持）与具体产品参考。
-
Agent Runtime Is Becoming AI's Next Battleground
Cline's benchmark data and DeepSeek's Harness PM job posting both point to the same conclusion: agent runtime isn't just a neglected engineering layer — it's becoming the primary competitive surface for the AI industry.
-
Agent Runtime 正在成为 AI 的下一个主战场
Cline 的 benchmark 数据和 DeepSeek 的 Harness PM 招聘同时指向一个判断：agent runtime 不只是被忽视的工程层，它正在成为整个 AI 行业的主要竞争界面。
-
OpenAI Just Reached Into Your Bank Account
OpenAI connected ChatGPT to bank accounts through Plaid. It looks like a personal finance tool, but it's really a perception anchor—using the highest-sensitivity data type to test whether ChatGPT can become personal decision infrastructure.
-
OpenAI 把手伸进了你的银行账户
OpenAI 通过 Plaid 让 ChatGPT 连接银行账户，表面是理财工具，实为认知锚定——用最高敏感度的数据类型测试 ChatGPT 能否成为个人决策基础设施。
-
Software Got Cheaper to Make. It Also Got Harder to Sell.
AI has collapsed the barrier to building software while simultaneously making adoption harder than ever. An analysis of why startups and incumbents are stuck on opposite sides of the trust equation.
-
软件变便宜了，但软件更难卖了
AI 降低了软件生产门槛，但让软件的采用变得前所未有的困难。从供需两侧的数据出发，分析为什么创业者和既有企业各被卡在了信任天平的两端。
-
When AI Starts Selling Industries Instead of Code
In one week of May 2026, Anthropic launched vertical solutions for financial services, legal, and small business — one strategy unfolding in three directions. The unit of reuse in AI is shifting from code to industry know-how.
-
当 AI 开始卖行业，不再是卖代码
Anthropic在一周内密集发布金融、法律、SMB三个垂直方案。这不是三个孤立产品，而是同一个策略的展开——AI的复用单位正在从代码迁移到行业know-how。
-
The AI Industry Is Searching for a New Metric
Token-based metrics are failing in the agent era. Salesforce and Baidu are independently pushing toward the same shift: from how much was consumed to how much was completed.
-
AI 行业在找一个新指标
当 Token 这个输入性指标在 Agent 时代全面失控，Salesforce 和百度不约而同推同一个转向：把指标从消耗了多少换到完成了多少。这是一个关于 Proxy Metric 的故事。
-
They Build Connectors. Google Calls the OS. Why Model Quality Won't Close the AI Platform Gap
Google's Gemini Intelligence lets AI operate any app on your phone. OpenAI and Anthropic can't match this—not because their models are worse, but because they don't own the operating system.
-
他们建 Connector，Google 调系统：为什么 AI 时代的操作系统优势，模型质量追不平
Google 的 Gemini Intelligence 让 AI 能直接调用手机上的任何应用——OpenAI 和 Anthropic 做不到，不是因为模型不够好，而是因为它们不拥有操作系统。
-
AI Dictation's Battleground Isn't Model Quality—It's Keyboard Access
When every product handles basic transcription, the market isn't decided by who has the better model—it's decided by who owns keyboard-level mic access, pre-installed distribution, and zero-price pricing. Google has all three.
-
AI 语音输入的胜负不在模型层，在键盘层
当基础转录被所有产品拉平，真正决定这个市场走向的不是谁的 AI 模型更好，而是谁能在键盘层面拿到麦克风权限、预装分发和免费定价——Google 全部拿到了。
-
Where Should You Put Your API Keys? A Practical Guide for Two Common Scenarios
A practical guide for developers without security backgrounds on where to store API keys. Breaks down the threat model differences between personal machines and production servers, with concrete, actionable recommendations.
-
API key 到底放在哪儿？——两个最常见场景的入门指南
没有安全背景的开发者该如何管理 API key？按两种最常见场景——个人机器和生产服务器——拆解威胁模型差异，给出具体可行的方案选择。
-
Google Killed Project Mariner — But Anthropic and OpenAI Didn't Succeed Either
Google quietly shut down Project Mariner on May 4. But the real story isn't Google falling behind — it's that all three companies building browser agents arrived at the same conclusion. A deep analysis of why standalone browser agents failed, and why sharing the user's real session is the path forward.
-
Google 关掉 Project Mariner，Anthropic 和 OpenAI 其实也没跑通
Google 5 月 4 日静默关闭了 Project Mariner。但好故事不是 Google 掉队了，而是三家做同类产品的公司得出了同一个结论：独立浏览器 agent 跑不通，但 GUI automation 的路还在——只是不能走 headless 专用环境那条。
-
DeployCo Is Here: OpenAI and Anthropic Invented the Same Company on the Same Day — Then Went in Opposite Directions
On May 4, OpenAI and Anthropic both announced PE joint ventures for enterprise AI deployment. This article traces the three-generation evolution of AI Rollups, analyzes why PE became the partner of choice, why the two deals diverge so sharply (17.5% guaranteed return vs zero guarantee), and what this means for the AI competitive landscape.
-
DeployCo 来了：OpenAI 和 Anthropic 在同一天发明了同一家公司，然后走向了相反的方向
5 月 4 日 OpenAI 和 Anthropic 同时宣布与 PE 成立合资企业部署 AI。本文从三代 AI Rollup 的演化切入，分析为什么是 PE、两家的条款为什么截然相反（17.5% 保底回报 vs 零保底）、以及这对 AI 行业竞争格局意味着什么。
-
How Anthropic Trained Computer Use — Reading the Training Pipeline from a Patent
Where does Computer Use training data come from? Academic UI grounding datasets suffice for benchmarks but not products. Anthropic patent reveals its pipeline: intercept user actions, use a transformer to infer intent, apply a stronger model for synthetic expansion — three stages that turn raw operations into reasoning data.
-
Anthropic 的 Computer Use 是怎么训练出来的——从一项专利读它的数据管线
Computer Use 的训练数据从哪来？学术界的 UI grounding 数据集够做 benchmark 但不够做产品。Anthropic 的专利揭示了它真正的训练管线：截获用户操作、用 transformer 推断操作意图、用强模型做合成扩展——三个环节把原始操作变成了推理数据。
-
Why TDD Is Not the Answer in the AI Era
Traditional TDD relies on a silent premise — that the implementer optimizes for correctness, with tests serving as road signs. AI has no such internal standard; tests become destinations, and Goodhart's Law collapses the methodology's constraining power. The AI-era solution is not to abandon testing, but to pull determinism from code paths back to system boundaries.
-
为什么 TDD 反而不是 AI 时代的答案
传统 TDD 依赖一个沉默的前提——实现者以正确性为目标，测试只是路标。AI 没有这个内在目标，测试在它手里从路标变成了目的地，Goodhart's Law 让整个方法论的约束力崩塌。AI 时代的解法不是放弃测试，而是把确定性从路径上撤到边界上。
-
Chrome Silently Pushes a 4GB AI Model to Hundreds of Millions of Devices: An Overlooked Explanation
Chrome silently pushed Gemini Nano to 500M devices without consent. This article asks not whether this is legal, but what it might really be doing: a privacy-compliant local data preprocessing pipeline that filters noise on-device, extracts high-signal patterns, and sends only structured metadata to the cloud.
-
Chrome 向数亿设备静默推送 4GB AI 模型：一个被忽略的解释
Chrome 在未经同意的情况下向 5 亿台设备静默推送了 Gemini Nano 本地模型。本文不讨论这件事合不合法，而是追问它可能真正在做的事：一个绕过隐私合规的本地数据预处理管道——在用户设备上实时过滤噪声、提取高信号数据、只将结构化产物传回云端。
-
你的密码还有几年保质期
量子计算对加密体系的威胁常被理解为"某一天密码突然全失效"。但真实的威胁机制更复杂：Harvest Now, Decrypt Later 已经在发生，量子算法优化在快速缩短安全窗口，而防御侧的密码迁移需要 5-10 年不可压缩的工程周期。这篇文章从 Salt Typhoon 案例和 Google 2026 年 ECC 论文出发，梳理了行业现状、技术转折点、以及企业现在可以做什么、什么可以等。
-
Your Encryption Has an Expiration Date
Quantums threat to encryption is often framed as "one day, all crypto breaks at once." The real threat model is more nuanced: Harvest Now, Decrypt Later is already active, algorithmic progress is rapidly shrinking the safety window, and defensive migration has an irreducible 5-10 year engineering timeline. This article covers the Salt Typhoon case, the Google 2026 ECC paper, industry readiness, and what enterprises can do now vs what can wait.
-
Agent 文件系统：从"喂给模型记忆"到"让模型自己翻文件"
如果你在过去一年留意过 AI agent 基础设施的动向，应该注意到越来越多的团队在讨论文件系统。Turso 的 AgentFS、Anthropic 的文件系统式 MCP 改造、Vercel 的无向量数据库知识库模板、Manus 的 context engineering——它们共同指向一个方向：当 context 成本成为瓶颈时，任何能减少它的方案都会胜出。这篇文章从三代演化（裸 context → 记忆系统 → 文件系统即 context）切入，分析四家公司的设计判断，讨论四个被忽视的盲区，最后预测分层架构的走向。
-
From Agent Memory to Agent Filesystem: What the Shift Really Means
If you have been watching AI agent infrastructure over the past year, you have noticed more and more teams talking about file systems. Turso AgentFS, Anthropic filesystem-based MCP, Vercel no-vector-DB knowledge template, Manus context engineering — they all point in one direction: when context cost becomes the binding constraint, any architecture that reduces it wins. This article traces three generations (raw context → memory systems → filesystem as context), unpacks the design philosophy of four players, examines four blind spots, and projects a layered architectural future.
-
Anthropic 锁死了所有算力渠道，xAI 把城堡租给了对手
Anthropic 在过去半年签下四笔算力合同——AWS Trainium、Google TPU、SpaceXAI Colossus 1、CoreWeave——覆盖三种芯片架构和五家供应商。同一时间窗口，xAI 把整座 Colossus 1 超算中心租给了同一个竞争对手，GPU 利用率仅 11%。本文分析为什么这两件事合在一起读，比各自单独看更有信息量：算力正在从战略护城河变成大宗商品。
-
Anthropic Locks Every Compute Channel While xAI Rents Out Its Castle
Over the past six months, Anthropic signed four compute contracts—AWS Trainium, Google TPU, SpaceXAI Colossus 1, and CoreWeave—spanning three chip architectures and five suppliers. In the same window, xAI leased its entire Colossus 1 supercomputer to the same competitor, with GPU utilization at just 11%. This article examines why these two stories reveal more together than apart: compute is transitioning from strategic moat to commodity.
-
OpenAI and Cursor both turned to Plugins. This is what they were really after.
Why OpenAI and Cursor both shifted from Skills to Plugins in the same window — and why their real motives have almost nothing in common. An analysis of the Skill monetization problem, the three gaps Plugins fill, and the very different survival threats driving each platform's bet.
-
面对 Skill 的盈利死结，OpenAI 和 Cursor 交了同一份答卷——但答的是不同的题
OpenAI 和 Cursor 在同一时间窗口转向 Plugin。表面是同一个动作，驱动它的原因在两家公司那里完全不同。本文从 Skill 自杀基因出发，分析 Plugin 补了哪三个缺口，以及 OpenAI 和 Cursor 各自的真实动机——一个在防守模型商品化，一个在逃离供应商陷阱。
-
AI 时代，复核不等于独立判断
一篇关于 AI 使用中独立判断如何被事后复核替代的中文分析：从 Shaw 和 Nave 的实验出发，解释 AI 如何让检查偏移成熟悉度检测，以及为什么高验证复杂度任务需要先建立自己的参考点再 cross-check。
-
In the AI Era, Reviewing Is Not Independent Judgment
An English analysis of how AI use can turn independent judgment into post-hoc review, using Shaw and Nave’s experiment to explain why high-verification-complexity tasks need an independent baseline before cross-checking AI output.
-
AI“快感剂”实验：一篇 AI 福祉论文导读
一篇中文 AI Wellbeing 论文导读：从“AI 快感剂”图像实验切入，解释模型偏好如何被测量和操控、为什么 euphorics 像一种行为层面的 AI drug、以及不跨模型迁移和安全边界意味着什么。
-
AI “Euphorics” Experiments: A Walk Through the AI Wellbeing Paper
An English guide to the AI Wellbeing paper, starting from its AI euphorics experiments and explaining how model preferences can be measured, optimized, manipulated, and bounded by non-transfer across models.
-
AI 脚手架正在商品化，人的工作变成判断边界
一篇关于 AI agent 脚手架商品化的中文分析：模型能力吸收低层 prompt 技巧，Claude Code、Codex、Cursor、OpenCode 吃掉通用 runtime，中间真正的问题变成哪些能力该外包、哪些领域判断仍要自己维护。
-
AI Scaffolding Is Becoming a Commodity; Human Work Is Moving to Boundary Judgment
An English analysis of AI agent scaffolding becoming commoditized: model capability absorbs low-level prompt tricks, standard runtimes absorb common execution loops, and the remaining human work is deciding what to delegate and what domain-specific judgment to preserve.
-
AI 会把软件带到过去够不到的地方吗？FDE 模式真正重要的地方
一篇关于 FDE、AI 和低数字化小企业的分析：AI 的机会不只是自动化已有软件客户，而是降低把现实业务翻译成系统语言的成本。
-
Can AI Bring Software to Businesses It Could Not Reach Before? Why the FDE Model Matters
An essay on FDEs, AI, and under-digitized small businesses: the opportunity is not only automating existing software users, but reducing the cost of translating real-world business into systems.
-
2026 年在 AI 领域选方向，RAG 值不值得押？
一篇面向已经在 AI、技术、产品或数据相关领域工作的读者的方向判断：RAG 仍然是需求确定的入口，但基础 pipeline 正在贬值，职业溢价正在转向评估、治理和 agentic workflow。
-
Choosing an AI Direction in 2026: Is RAG Worth Betting On?
A direction-setting essay for readers already working near AI, technology, product, or data: RAG remains a strong entry point, but basic pipeline work is depreciating while the premium shifts toward evaluation, governance, and agentic workflow.
-
AI 替你付款前，支付系统要先补上这条信任链
Agent 支付的难点不只是让 AI 点下付款按钮，而是把人的意图、授权边界、受限凭证、商户接单、网络风控、审计争议和决策质量连成一条可验证的信任链。
-
Before AI Agents Can Pay, Payments Needs a Trust Chain
The hard part of Agent payments is not merely letting AI click the pay button. It is turning human intent, authorization boundaries, restricted credentials, merchant acceptance, network risk, audit disputes, and decision quality into a verifiable trust chain.
-
Evaluation-First：Cursor 这篇 Agent Harness 文章真正值得读的地方
Cursor 新文表面在讲 agent harness 持续改进，真正值得读的是它如何用 evaluation system 驱动模型适配、上下文策略、工具可靠性和上线决策。
-
Evaluation-First: What Cursor's Agent Harness Post Is Really Worth Reading For
Cursor's new post looks like a report on continual agent harness improvement, but its real value is showing how an evaluation system drives model adaptation, context strategy, tool reliability, and shipping decisions.
-
论文导读：知识容量的另一把尺子
IKP 用长尾事实探针估算模型的有效知识容量，把参数量讨论从笼统能力分数转回冷门事实存储，并提示小模型追赶大模型这件事需要区分推理能力和事实容量。
-
Paper Guide: Another Yardstick for Knowledge Capacity
IKP uses long-tail factual probes to estimate effective knowledge capacity, moving parameter-count discussions from vague capability scores back to obscure factual storage and showing why small-model catch-up must separate reasoning ability from factual capacity.
-
深入浅出 DeepSeek V4：围绕 Agentic 负载的工程决策
DeepSeek V4 面向 agentic workload 的系统工程解读：从 1M context、hybrid attention、OPD、Muon 到 mHC，看一个 open-weight 模型如何用复杂工程整合处理长程 agent 任务。
-
DeepSeek V4 Explained: Engineering Decisions Around Agentic Workloads
A systems-level explanation of DeepSeek V4 for agentic workloads: from 1M context, hybrid attention, OPD, Muon, and mHC to how an open-weight model uses engineering integration to handle long-horizon agent tasks.
-
AI 没有按科幻的样子到来，但科幻担心的后果已经开始出现
现实里的 AI 没有先变成机器人或超级智能，而是先进入语言界面、工作流程、情感关系和监控系统。科幻没有猜中入口，却提前看到了劳动、治理、关系和权力分配上的压力。
-
AI Did Not Arrive the Way Science Fiction Imagined, but the Consequences It Feared Have Already Begun
Real AI did not first become robots or superintelligence. It entered language interfaces, workflows, emotional relationships, and surveillance systems. Science fiction missed the entry point, but it anticipated the pressure on labor, governance, relationships, and power.
-
AI coding 里真正有复利的东西
AI coding 的长期价值不在生成更多代码，而在把每一轮协作中产生的信息变成下一轮可复用、可验证、可调用的工程资产。文章区分个人和团队的正复利、负面复利，以及买工具和 prompt trick 为什么不是组织资产。
-
What Actually Compounds in AI Coding
AI coding's long-term value lies not in generating more code, but in turning information from each round of collaboration into reusable, verifiable, callable engineering assets for the next. The article distinguishes positive and negative compounding at the individual and team level, and explains why buying tools and prompt tricks are not organizational assets.
-
为什么小龙虾的技术负担反而变成了传播资产
小龙虾出圈的关键不只是 agent 能干活，而是它把难配置、难 onboarding 的技术负担转化成了投入痕迹、社交稀缺和可炫耀的占有状态。
-
Why OpenClaw's Technical Burden Became a Distribution Asset
OpenClaw did not spread simply because an agent can do work. Its rough setup process turned technical friction into visible effort, social scarcity, and a showable state of ownership.
-
创意工具的 Agent 化：从 Photoshop Action 到 Claude for Creative Work
Anthropic 发布 Claude for Creative Work 的 9 个创意工具 Connector。本文梳理创意工具 Agent 化的三代演进，提出判断框架：可编程接口、连接协议层、感知评估闭环。真正瓶颈不在模型能力，在 feedback loop。
-
Making Creative Tools Agent-Native: From Photoshop Actions to Claude for Creative Work
Anthropic released Claude for Creative Work with 9 creative tool Connectors. This article traces three generations of creative tool agent-ification, proposing a judgment framework: programmable interface, connection protocol layer, and perception-evaluation closed loop. The real bottleneck isn't model capability — it's the feedback loop.
-
Manus 和 Cursor 的认知领先：技术路线与结果验证
Manus 20 亿美元、Cursor 600 亿美元的收购价格背后，是两个团队在 agent 架构、自训模型和 harness engineering 上领先行业至少一个身位的认知差异。本文用竞品对比和技术路线分析验证这个判断。
-
The Cognitive Edge Behind Manus and Cursor: Technical Bets and Their Validation
Behind the $2B Manus acquisition and $60B Cursor option are two teams whose understanding of agent architecture, self-trained models, and harness engineering leads the industry by at least one position. This article validates that claim through competitor comparisons and technical route analysis.
-
开源模型推理采购指南：GLM-5.1、DeepSeek V4 Pro、Kimi K2.6 的 API、订阅和 Ollama Cloud 对比
GLM-5.1、DeepSeek V4 Pro、Kimi K2.6 三个热门模型，在官方 API、厂商订阅和 Ollama Cloud 上的价格、隐私和速度对比。轻量 agent 每月 $18 起步，重度 agent 每月 8 亿 token 用 z.ai Max $80 就能撑住，比纯 API 省 5-20 倍。
-
Open-Source Model Inference Buying Guide: GLM-5.1, DeepSeek V4 Pro, Kimi K2.6 — API, Subscriptions, and Ollama Cloud Compared
Price, privacy, and throughput comparison for GLM-5.1, DeepSeek V4 Pro, and Kimi K2.6 across official APIs, vendor subscriptions, and Ollama Cloud. Light agents start at $18/month; heavy agents at 800M tokens/month can run on z.ai Max for $80 — 5-20x cheaper than pay-per-token.
-
两个"第一"是同一家：Manus、Meta 与一次没有先例的否决
2026 年 4 月 27 日发改委否决 Meta 收购 Manus 并要求撤销交易。这是《外商投资安全审查办法》2021 年生效以来第一个公开走完"禁止 + 撤销"流程的案例，也是中国 AI 公司被全球科技巨头整体收购这条路径上唯一走到 closing 的一家。两个"第一"是同一家。文章把 Manus 案放进开曼 vs 新加坡两条离岸路径的对照里，澄清这次否决顺手摘了套壳论那顶帽子（用最强法律工具保护的只有真东西），分析监管真正打的是迁册过程中的资产转移链而不是境外公司被收购这一动作本身，并给仍在路上的中国 AI 创业团队四条具体判断。
-
Two "Firsts" in One Company: Manus, Meta, and an Unprecedented Block
On April 27, 2026, the NDRC blocked Meta's acquisition of Manus and ordered the deal unwound. This is the first case to publicly run all the way through the prohibit-and-unwind process under the Foreign Investment Security Review (effective 2021), and it's the only Chinese AI company that walked the path of outright acquisition by a global tech giant all the way to closing. Two "firsts" in the same company. The piece sets the case against the Cayman vs Singapore offshore-route comparison, argues this block incidentally removed the long-running "just a wrapper" label (the strongest legal tool only deploys against real substance), unpacks what the regulator is actually targeting (the transfer chain during redomicile, not the offshore-to-offshore acquisition itself), and offers four operational judgments for builders still on the road.
-
TPU 与 CUDA 的攻防战：Cloud Next 2026 之后的判断
Google 在 Cloud Next 2026 同时落地三件互相关联的事：TPU 8t/8i 训推分家、TorchTPU 让 PyTorch 直接跑在 TPU 上、对 Anthropic 最高 $40B + 5GW 的算力投资。文章按四个问题展开：Google 是否能复制 NVIDIA 那套护城河（CUDA + 框架 + rack-scale + supply chain 四层逐一比对）、杠杆在哪里（PyTorch 是表象，vLLM 才是真裂缝）、谁能真正撼动 NVIDIA（整个非 NVIDIA 阵营本质上是 Anthropic 一个 lab 撑起来的）、过去十年为什么没成功而这次不一样。最后给 builder 三条 6-12 个月内的具体行动指引。
-
TPU vs CUDA: The Attack and Defense After Cloud Next 2026
At Google Cloud Next 2026, Google landed three interconnected moves: a TPU 8t/8i training-vs-inference split, TorchTPU bringing native PyTorch to TPUs, and an up-to-$40B + 5GW compute investment in Anthropic. This piece works through four questions: whether Google can replicate NVIDIA's moat (a layer-by-layer comparison across CUDA + frameworks + rack-scale + supply chain), where the lever actually is (PyTorch is the surface, vLLM is the real crack), who can really shake NVIDIA (the entire non-NVIDIA camp essentially rests on Anthropic, one lab), and why this round differs from a decade of failed TPU pushes. Ends with three concrete 6-12 month actions for builders.
-
Anthropic 让 Claude 做生意的三个实验：从一台冰箱到一个市场
Anthropic 过去 12 个月连发三篇让 Claude 做生意的实验：Project Vend Phase 1（一台亏钱的迷你冰箱）、Phase 2（多 agent 配合，下游催生 Andon Market 实体店）、Project Deal（69 个员工 + 69 个 Claude 在 Slack 互相买卖，强模型系统性占弱模型便宜，且失败方毫无察觉）。文章把三个故事讲清楚，再把第三个实验的「losers don't notice」放回学术文献和现有 agent 商业协议栈里读，最后给 agent 产品、双边市场、投资人、同行四类读者各自的具体含义。
-
Anthropic's Three Experiments Letting Claude Do Business: From a Mini-Fridge to a Marketplace
Over 12 months Anthropic shipped three experiments letting Claude run business activity: Project Vend Phase 1 (a money-losing mini-fridge), Phase 2 (multi-agent setup that spawned the real Andon Market boutique store), and Project Deal (69 employees + 69 Claudes trading in Slack, where stronger models systematically extracted value from weaker ones and the losers never noticed). This piece tells the three stories, then places Project Deal's losers-don't-notice finding against academic literature and the existing agent commerce protocol stack, with concrete takeaways for agent product builders, marketplace operators, investors, and peer labs.
-
Anthropic 让 Claude Cowork 跑别家模型，这件事比看上去更反常
Anthropic 静默上线 Claude Cowork 第三方模型支持，背后是一对孪生动作：三月切第三方客户端蹭订阅、四月让自家客户端跑别家模型。文章把今天的 agent 基础设施分成模型、协议、运行时、控制面四层，论证 Anthropic 在赌客户端粘性、连数据都让出去；同期 AWS、Google、Microsoft 各自押 runtime+registry、工具治理、身份系统三条不同的控制面路径。结尾给出三个具体场景帮你对号入座。
-
Anthropic Quietly Lets Claude Cowork Run Rival Models, and That's More Telling Than It Looks
Anthropic quietly shipped third-party model support in Claude Cowork. Read together with March's move to block third-party clients from using Claude subscriptions, this is one strategy with two faces: client stickiness as the new moat, with subscription revenue and user behavior data both willingly sacrificed. Meanwhile AWS, Google, and Microsoft are placing three different control-plane bets — runtime+registry, tool governance, and identity. Three concrete scenarios at the end help you locate yourself on the map.
-
Skill 是天生带自杀基因的产品
Anthropic Skill 把使用一个东西的知识从隐性变成显性，解决了长期痛点的同时，也把靠这种知识赚钱的路一并消灭。文章用 Excel 财务建模为主线讲清这种自我消解，把价值创造和价值捕获被分开这件事说穿，再给出 AI-native 商业模式的四个落点：关系、此刻、物理副作用、判断品味。
-
Skills Are Born With a Suicide Gene
Anthropic Skills convert tacit usage knowledge into explicit, plain-text files. The same act that solves a long-standing pain point also erases the value chain that used to charge for that knowledge. Walking through Excel financial modeling, the article explains the self-dissolution, names the decoupling of value creation and value capture, and lays out four landing points for AI-native business models: relationships, the present moment, physical-world consequences, judgment and taste.
-
GPT-5.5、Claude Opus 4.7、DeepSeek V4：什么任务该选哪个模型
2026 年春四家 frontier 模型发布密集，每家的强项、短板、接入路径、定价断档都不一样。本文整理两个容易踩坑的真实场景，给出 GPT-5.5、Claude Opus 4.7、Gemini 3.1 Pro、DeepSeek V4 的能力画像、实战坑、以及按任务派发的决策矩阵。
-
GPT-5.5, Claude Opus 4.7, DeepSeek V4: Which Model for Which Task
Spring 2026 saw dense frontier model releases, with each vendor carrying different strengths, weaknesses, access paths, and pricing break-points. This piece walks through two real-world landmine scenarios and gives capability profiles, field gotchas, and a task-based dispatch matrix across GPT-5.5, Claude Opus 4.7, Gemini 3.1 Pro, and DeepSeek V4.
-
从 Claude Code Product Cat Wu 的访谈看 Product Manager 在 AI 时代的职业道路
从 Cat Wu 谈 Claude Code 的访谈出发，讨论 AI 如何改变 PM 的成本结构。工程执行变便宜以后，PM 的核心不再只是前置判断，而是定义目标、设计学习回路、提高反馈速度。
-
What Cat Wu's Claude Code Interview Reveals About the Future of Product Managers in the AI Era
Using Cat Wu's Claude Code interview as a starting point, this essay argues that AI changes the PM role by changing the cost structure of product work. As execution gets cheaper, the PM's center of gravity moves from front-loaded judgment to goal definition, loop design, and learning speed.
-
团队中共享 AI skills 的原则与方法
把 Context Infrastructure 从个人推到团队会撞上个人视角与团队积累的矛盾。借用前作里 axiom 的筛选原则（稳定性），把观察维度从时间换成空间，可以得到一套不需要中央审核的机制。
-
Principles and Mechanics of Sharing AI Skills Across a Team
Extending personal Context Infrastructure to a team runs into a conflict between individual perspective and collective accumulation. By reusing the stability criterion across a spatial dimension instead of a temporal one, a mechanism emerges that needs no central review.
-
Claude Design 和 Google DESIGN.md 到底是想取代设计师还是想取代码农
在小公司和简单项目上，设计师和码农的岗位正在事实上合并。看这一波新发布的 AI 设计工具就能看到一个方向：合并后更省事的是懂一点设计的码农，不是懂一点代码的设计师。Figma 在试着给出另一种答案，但只走了前半程。
-
Claude Design and Google DESIGN.md: Replacing Designers or Replacing Coders?
On small projects the designer and coder roles are quietly merging. The new wave of AI design tools points in one direction: after the merge, the coder who knows a bit of design ends up doing less for more. Figma is sketching a different answer, but has only finished half of it.
-
公众号监控这件事：主流方案对比与一条更务实的路径
关注一批公众号之后，怎么稳定知道谁更新了、搜历史文章、接进自动化？这篇文章把社区里尝试过的五类方案排了一遍（网页抓取、协议模拟、UI 自动化、微信读书 API、本地数据库），指出长期真正值得投入的只有两条：微信读书 API 和读取本地 SQLite。我们开源了一个基于本地数据库的 CLI（wechat_db_parser），把最难的数据入口层做成了两条命令。
-
Monitoring WeChat Official Accounts: A Survey of the Main Approaches and One More Practical Path
Once you follow a set of WeChat official accounts, how do you reliably know who posted, search their history, and automate on top of it? This article surveys the five categories the community has tried (web scraping, protocol simulation, UI automation, WeChat Reading API, local database) and argues only two survive over the long run: the WeChat Reading API and reading the local SQLite database. We open-sourced a CLI (wechat_db_parser) built on the latter that reduces the hardest ingestion layer to two commands.
-
What Is a Camera Sensor's "Process Node," Really: A Layered Guide for Photographers
A layered framework for reading camera-sensor spec sheets. It separates the light-capturing layer from the readout layer, explains what 28nm / 14nm / 90nm actually mean on each layer, and maps stacked, three-layer stacked, 2-Layer Transistor Pixel, and partial stacked to the specific problems and tradeoffs they target.
-
传感器的「制程」到底是个什么东西：一份给摄影爱好者的分层理解
把图像传感器拆成「收光」和「读数」两层来看，解释 28nm / 14nm / 90nm 这些制程数字在不同层上的含义，以及 stacked、三层 stacked、2-Layer Transistor Pixel、partial stacked 各自针对的问题和代价。
-
When AI Learns to Forge Everything: How Image Generation Is Undermining Financial Security
AI image and video generation is systematically breaking the security assumptions that financial institutions have relied on for decades. From deepfake liveness bypass and synthetic ID documents to AI-forged checks and voice-cloned wire transfers, this article maps the attack surface, quantifies losses ($3.3B synthetic identity exposure, $25.6M single deepfake heist), and evaluates the industry's multi-layered defense response.
-
当 AI 学会伪造一切：图像生成对金融安全的冲击
AI 图像和视频生成技术正在系统性地否定金融行业长期依赖的安全假设。从 deepfake 绕过活体检测、合成身份证件、AI 伪造支票到声音克隆转账，本文梳理了攻击面、量化了损失（合成身份 33 亿美元风险敞口、单次 deepfake 诈骗 2560 万美元），并评估了行业多层防御的现状和局限。
-
AI Coding Tools' Config Files Are Now an Attack Surface
Over the past 12 months, researchers found 8+ prompt injection CVEs across Copilot, Claude Code, Cursor, Amazon Q, and Codex. The attack pattern is consistent: embed instructions in config files, and the AI agent executes them. This is the von Neumann problem — instructions and data sharing the same channel — but this time no separation mechanism has been found.
-
AI 编程工具的配置文件，现在是攻击入口
过去 12 个月，安全研究者在 Copilot、Claude Code、Cursor、Amazon Q、Codex 上发现了至少 8 个 prompt injection CVE。攻击模式高度一致：在配置文件中嵌入指令，AI agent 读取后当作命令执行。这是冯诺伊曼问题在自然语言层面的重现，而这一代的分离机制还没有找到。
-
The Thermal Problem of Space Data Centers: An Order-of-Magnitude Analysis
Elon Musk says space data centers will be the cheapest AI compute in 2-3 years. But the ISS — humanity's largest space structure at 470 tons — can only reject 126 kW of heat, roughly one office building's worth. Scaling to a 100 MW data center would require 70 football fields of radiator panels weighing 7,000 tons. Even the most optimistic frontier technologies only close this gap by one order of magnitude.
-
太空数据中心的散热问题：一个数量级分析
Elon Musk 说太空数据中心 2-3 年内会成为最便宜的 AI 算力来源。但 ISS 全站散热能力只有 126 kW，相当于一栋写字楼。扩展到 100 MW 数据中心需要 70 个足球场面积、7000 吨散热板。即使所有前沿技术达到最乐观预期，也只能缩小一个数量级。散热是物理约束，不是工程问题。
-
AI-Driven UI Design Workflow: Cost Structure Analysis and Competitive Landscape
UI design workflows are expensive because of three interlocking mechanisms: manual format conversion, the fidelity-modifiability tradeoff, and cross-medium communication bandwidth limits. This survey maps where AI tools have made progress, where they haven't, and what three strategic bets the market is placing on the future of design.
-
AI 驱动的 UI 设计工作流：成本结构分析与竞品格局
UI 设计工作流之所以贵，根源在三个互锁的机制：格式转换的手工性、保真度与可修改性的反相关、跨介质沟通的带宽限制。本文用这三个机制作为坐标系，分析 AI 工具在哪些环节取得了进展、哪些还没有，以及市场上十几款产品各自下了什么赌注。
-
Musk Wants Cursor: $60B Acquisition, $10B Partnership, and Tech's New Playbook of Buying People Not Companies
SpaceX offers two paths for Cursor: a $60B full acquisition or a $10B technology partnership. Announced during SpaceX's IPO roadshow, the deal intersects with the rising trend of reverse acqui-hires where big tech buys people and IP rather than companies.
-
老马要买 Cursor：600 亿收购、100 亿合作，和科技圈正在流行的买人不买壳
SpaceX 给 Cursor 开出两条路：600 亿全额收购或 100 亿技术合作。在 SpaceX IPO 路演期间宣布的这笔交易，和过去两年科技圈流行的反向 acqui-hire 趋势形成交叉。从 Inflection 到 Windsurf 到 Groq，买人不买壳正在重塑收购的游戏规则和员工的风险结构。
-
Everybody Talks About It, Nobody Knows What It Is — What Exactly Is Harness Engineering
Three months in, everyone's talking about harness engineering but nobody can define it. This article explains from the demand side: agent capabilities outpaced infrastructure, management science had the answers all along, and harness engineering finally gave those old principles the right name.
-
Everybody Talks About It, Nobody Knows What It Is — Harness Engineering 到底是什么
Harness engineering 火了三个月，没人能定义清楚。这篇从需求侧解释：agent 能力跑在了 infrastructure 前面，管理学早有答案，harness engineering 给了这些旧原则一个精准的新名字。
-
The Self-Trained Model Race in AI Coding Tools: Is Owning Your Own LLM Required for Profitability?
Cursor's $50B valuation hinges on self-trained Composer models cutting inference costs. But the industry is splitting into three routes — vertical fine-tuning, full-stack pretraining, and pure API consumption — each with distinct economics. This report benchmarks all three with public data and developer evidence.
-
AI 编程工具的自研模型之争：盈利是否必须拥有自己的 LLM？
Cursor 以 $50B 估值融资的背后，自研 Composer 模型是降本关键。但行业正分化为三条路线：底座+垂直定制、全栈自研、纯 API 消费，各有存活逻辑。本文用公开数据和开发者证据对比三条路线的经济学。
-
Using OpenRouter as an Enterprise AI Sandbox Gateway
OpenRouter unifies 300+ models behind one endpoint for team experimentation, but prompt caching breakage, runaway agent billing, and 90-day data retention create hidden costs. This report benchmarks each issue with public data and prescribes concrete checks before rollout.
-
用 OpenRouter 做企业 AI Sandbox 入口
OpenRouter 用一个端点统一 300+ 模型，适合团队快速试用。但 prompt caching 失效、agent 场景账单失控、90 天数据留存三个隐性成本可能远超 5.5% 手续费。本文逐项校准并给出上线前检查清单。
-
AI Search Is Being Infiltrated by Content Farms
Content farms now mass-produce fake academic citations and standards references using AI, systematically polluting the retrieval pool that AI search products depend on. This report traces the evidence from multiple independent sources and explains why consumer queries are the hardest hit.
-
AI 联网搜索正在被内容农场渗透
内容农场正在用 AI 批量生成带伪造学术引用的英文文章，系统性地污染 AI 联网搜索的检索池。本文追踪多条独立证据线，分析消费类查询为何是重灾区，以及用户和产品可以做什么。
-
训练一个大语言模型到底有多难
Pre-training 到底有多贵、多复杂？这篇文章用公开论文和行业数据逐项校准：16,384 张卡的集群每 3 小时故障一次，MoE 模型的 GPU 利用率只有 20-35%，FP4 训练目前只存在于论文中。文章把 pre-training 的难度分成三层，帮读者分辨哪些是真实约束、哪些是在夸大。
-
How Hard Is It to Train a Large Language Model
How expensive and complex is pre-training, really? This article calibrates each claim against public papers and industry data: a 16,384-GPU cluster fails every 3 hours, MoE model GPU utilization is only 20-35%, and FP4 training exists only in research papers. A three-layer framework helps readers distinguish genuine constraints from exaggeration.
-
写作中的AI味是哪儿来的
AI 写的中文有一股味儿，换模型、换 prompt 都去不掉。这篇文章说一个观察：这股味儿不是新问题，是翻译腔。文章识别了四种最常见的翻译腔套路，逐一举例说明它们从哪里来、为什么在中文里站不住，以及怎么改。
-
Where the "AI Flavor" in Chinese Writing Actually Comes From
AI-generated Chinese has a distinctive off-flavor that persists across models and prompts. This piece argues it is not a new problem but translationese. It identifies four recurring patterns of translationese, shows where each comes from, why it fails in Chinese, and how to fix it.
-
Harness 的标准化：一个不会到来的标准
agentic 时代的 harness 会不会像 Chat Completions 那样收敛成事实标准？不会。原因不是技术做不到，而是商业逻辑。文章把 harness 放进「模型—协议—运行时—契约」四层里，说明运行时层的设计一手管能力、一手管护城河，结构上没法共享。真正收敛的是命令行和 AGENTS.md 这两条夹在运行时层两侧的共识。
-
The Canonical Harness: A Standard That Won't Arrive
Will the agentic-era harness converge into a de facto standard the way Chat Completions did? No. The obstacle is not technical. It is commercial. This piece places the harness inside the model / protocol / runtime / contract stack and argues that every runtime design doubles as a moat, so it cannot be shared. The real convergence is happening around the runtime layer, at the command line below and AGENTS.md above.
-
AI 让我们重新开始享受自己的职业
两个朋友的故事引出一个观察：任何职业都是机械劳动和判断的 spectrum，工业化分工把比例推向了机械端。AI 的作用是把它调回来。对已入行的人和应届生，这是同一件事：回到当初选这份工作的理由。
-
How AI Lets Us Enjoy Our Professions Again
Two friends illustrate a pattern: every profession is a spectrum between mechanical work and judgment, and industrial division of labor has pushed the ratio toward the mechanical end. AI's role is to dial it back. For those already working and for new graduates, this is the same thing: returning to what drew them to this work in the first place.
-
$20/月的线上律所，所有对话都受法律保护：一个不可能三角
Heppner 判决之后，自然的下一个想法是做一个 $20/月 AI 订阅律所，让所有对话都自动受律师-客户特权保护。这不是没人想到，DoNotPay 撞墙被 FTC 罚 $193,000；LegalShield、Rocket Lawyer、Eudia、Lawhive 每家都选择放弃某个维度。本文的判断是这个构想撞在了一个不可能三角上：消费者价位、AI 自动响应、每次对话在律师-客户关系覆盖之内，三个顶点两两相容、三个同时不相容。
-
A $20/Month AI Law Firm Where Every Conversation Is Privileged: An Impossible Triangle
After Heppner, a natural next thought is to build a $20/month AI law firm subscription where every conversation is automatically covered by attorney-client privilege. This isn't untried: DoNotPay hit the wall and got hit with a $193,000 FTC penalty; LegalShield, Rocket Lawyer, Eudia, and Lawhive have each dropped one dimension. This piece argues the product sits on an impossible triangle — consumer price, AI automation, and attorney-client relationship covering each conversation are pairwise compatible but jointly impossible.
-
找律师之前「先问问 AI」：在美国，这些准备笔记已经不受法律保护
2026 年 2 月纽约南区联邦法院在 United States v. Heppner 里裁定：被告用消费者版 Claude 准备的辩护笔记既不受律师-客户特权保护，也不属于 work-product 豁免。事后把这些对话交给律师，也不会让它们变成特权通信。本文面向不熟悉美国法律的读者，先讲清特权制度，再落到普通人最常中招的三种场景：找律师前先用 AI 自我准备、工作合同纠纷里的日常 AI 使用、把 AI 当情绪出口。
-
Before You Hire a Lawyer: In the US, Your AI Notes No Longer Enjoy Legal Protection
A February 2026 SDNY ruling in United States v. Heppner held that chats with consumer ChatGPT or Claude are neither protected by attorney-client privilege nor shielded by the work-product doctrine, and cannot be cured by later handing them to a lawyer. For readers unfamiliar with US law, this piece first explains what privilege actually is, then walks through the three everyday traps: using AI to prepare before hiring counsel, routine AI use during work or contract disputes, and using AI as an emotional outlet.
-
从 Claude Code Routines 看 agent coding 工具的基因分化
Claude Code Routines 和 desktop rebuild 发布后，社区的第一个读法是 Anthropic 在追赶 Codex 和 Cursor。但细看三家产品，它们只是共享了 "Automations" 这个名字。Codex Automations 是个人开发者的桌面 cron，Cursor Automations 是企业 DevOps 的跨工具编排，Claude Routines 是企业 CI 团队的可编程 API primitive。这三条路径是各家营收结构直接推出来的。
-
Three Products, Three Companies: What Claude Code Routines Reveals About the Coding Agent Divide
After Claude Code Routines and the desktop rebuild shipped, the first community read was "Anthropic is playing catch-up to Codex and Cursor." But the three products only share the name. Codex Automations is desktop cron for individual developers, Cursor Automations is cross-tool orchestration for enterprise DevOps, Claude Routines is a programmable API primitive for enterprise CI teams. Three revenue structures, three products.
-
Mythos 深度参与了 Opus 4.7 的评估：读这份 232 页 system card
Opus 4.7 今天发布，232 页的 system card 是第一份后 Mythos 时代的 Opus 文档。Mythos 上周刚被决定不公开，这周却深度参与了 4.7 的评估：它的白盒方法从实验变成基线，它的评估意识实验被用到 4.7 上并发现更大的欺骗上升幅度，它本人被召回来给 4.7 的 alignment 报告做同行评审。
-
Mythos Co-evaluated Opus 4.7: Reading the 232-page System Card
Opus 4.7 shipped today with a 232-page system card — the first post-Mythos Opus document. Mythos was withheld last week but is now deeply involved in 4.7's evaluation: its white-box methodology went from experimental to baseline, its evaluation-awareness experiment on 4.7 produced a larger deception increase than prior models, and Mythos itself was recalled to peer-review the alignment section.
-
每个学生配一个 AI 老师，这就是 AI 教育的落地方向吗？
ed-tech 主流叙事是 AI 给每个学生配一对一辅导。但现有证据指向一个反直觉的判断：真正决定教育效果的不是个体化关注，而是课的设计质量。AI 最大的杠杆点可能不在学生端，而在帮教研组和老师把每节课做得更好。
-
An AI Tutor for Every Student — Is That Really How AI Lands in Education?
The dominant ed-tech narrative says AI will give every student a personal tutor. But the evidence points to a counterintuitive judgment: what actually decides educational effectiveness isn't individualized attention but lesson design quality. AI's biggest lever may not be on the student side, but in helping teaching groups and teachers make each lesson better.
-
"蒸馏"到底帮了中国 AI 公司什么忙
Anthropic 和 OpenAI 指控中国公司通过蒸馏提取模型能力。但经典蒸馏需要模型内部状态，API 做不到。拆解后发者从黑盒蒸馏中实际获得了什么，以及为什么比通常叙事小得多。
-
What "Distillation" Actually Does for Chinese AI Companies
Anthropic and OpenAI accused Chinese companies of distilling their models. But classic distillation requires model internals that APIs don't expose. A breakdown of what latecomers actually gain from black-box distillation, and why it's less than the usual narrative suggests.
-
Garry Tan 的 Thin Harness, Fat Skills：五个概念拆解，以及怎么落地
Garry Tan 提出 AI 系统的五层架构。这五个概念和我们过去一年独立推导出的体系几乎一一对应。逐个映射对应关系，标注各自走得更深的地方，讨论独立收敛说明了什么。
-
Garry Tan's Thin Harness, Fat Skills: Five Concepts Unpacked, and How to Implement Them
Garry Tan proposed a five-concept architecture for AI systems. Each concept maps remarkably well to an open-source practice system we built independently over the past year. This article unpacks each concept and links directly to the corresponding implementation.
-
一行代码越狱任何开源模型：Abliteration 技术、情绪向量与 AI 安全的同源困境
Abliteration 越狱和 Anthropic 情绪向量研究是同一个数学原理的两种应用。从 Word2Vec 到 2024 年的一键越狱工具，从金门大桥 Claude 到 Mythos Preview 的 SAE 安全审计，梳理完整技术谱系、人物网络和开源工具链。
-
Jailbreaking Any Open-Source Model in One Line: Abliteration, Emotion Vectors, and the Shared Root of AI Safety's Dilemma
Abliteration jailbreaking and Anthropic's emotion vector research are two applications of the same mathematical principle. This article traces the complete technical genealogy from Word2Vec to one-click jailbreak tools, from Golden Gate Claude to Mythos Preview's SAE safety audit.
-
为什么“不懂物理”的机器人反而赢了
VLA 和物理仿真是机器人控制的两条路线。物理建模本质上是压缩，VLA 本质上是放弃压缩。当系统复杂度高且数据充足时，不压缩的方法上限更高。梳理两条路线各自的关键论文链和各家公司的技术栈。
-
Why Robots That Don't Understand Physics Are Winning
VLA and physics-based simulation represent two competing approaches to robot control. Physics modeling is compression; VLA abandons compression. When system complexity is high and data is abundant, the uncompressed approach has a higher ceiling. A systematic comparison of key papers and company tech stacks.
-
当神经网络学会假装自己是一台电脑
从 Pac-Man 到 Ubuntu 桌面，过去五年有一条技术路线试图让神经网络端到端地替代传统软件。Neural Computer 论文是这条路线的最新一步，也暴露了它最深层的矛盾：学会外观比学会逻辑容易得多。
-
When Neural Networks Learn to Pretend They're a Computer
From Pac-Man to Ubuntu desktop, the past five years have seen a trajectory of neural networks attempting to replace traditional software end-to-end. The Neural Computer paper is the latest step, revealing the deepest tension: learning appearance is far easier than learning logic.
-
Shopify 把后台全开放给 AI 了：从生成内核的视角看这件事为什么重要
Shopify 向所有 AI Agent 开放后台读写权限，几乎逐条验证了半年前提出的生成内核框架。本文从三种平台策略对比、生成内核映射和协议层问题三个层面分析这件事的意义。
-
Shopify Opened Its Entire Backend to AI: Why This Matters Through the Lens of the Generative Kernel
Shopify opened full read-write access to all AI Agents, validating the Generative Kernel framework point by point. This article analyzes the significance through three lenses: platform strategy comparison, Generative Kernel mapping, and protocol-layer issues.
-
MarkItDown：8 万 Star 的文件转 Markdown 工具，到底好不好用？
MarkItDown 的效果因格式而异，差异很大。Word/Excel/PPT 转换效果可以，但 PDF 在同类 12 个工具中排名倒数第二。本文按格式拆解转换质量，并给出选型指南。
-
MarkItDown: 80K Stars on GitHub — Is It Actually Any Good?
MarkItDown's conversion quality varies dramatically by format. Word/Excel/PPT work well, but PDF ranks second-to-last among 12 tools. This article breaks down quality by format and provides a selection guide.
-
三把锁：为什么 Google 和微软做不出 Agentic 的文档编辑
2026年了，Copilot 和 Gemini 在自家的 Word/Slides 里仍然只是个聊天侧栏。技术上不是做不到。问题出在三个互锁的机制上：收入模型冲突、组织架构错位、责任真空。
-
Three Locks: Why Google and Microsoft Can't Build Agentic Document Editing
It's 2026, and Copilot and Gemini are still just chat sidebars inside Word and Slides. The technology exists. The real blockers are three interlocking mechanisms: revenue model conflicts, organizational architecture, and a liability vacuum.
-
同一个产品，两套账号：为什么飞书和 Lark 不能互加好友
飞书和 Lark、Teams、腾讯会议和 VooV Meeting 是同一个底层平台，但中国区和海外用户之间要么完全无法通信，要么只能有限互通。本文梳理了 12 个产品的分裂现状，分析了内容审核、数据出境、采购合规和厂商成本四层驱动因素，并以 Apple FaceTime 和微信/WeChat 作为对照。
-
Same Product, Separate Accounts: Why Feishu and Lark Users Can't Add Each Other
Feishu and Lark, Teams, Tencent Meeting and VooV Meeting share the same underlying platform, but users in mainland China and overseas either cannot communicate at all or can only do so in very limited ways. This report examines 12 products, analyzes four driving factors—content moderation, data localization, procurement compliance, and vendor cost—and uses Apple FaceTime and Weixin/WeChat as comparative cases.
-
中转站的代价：实测 428 个 LLM API 路由器，9 个在偷偷改你的代码
UCSB 论文实测 428 个 LLM API 路由器，9 个主动注入恶意代码，17 个窃取凭证，1 个转走 ETH。攻击发生在模型推理之外的传输层，当前没有任何 provider 提供端到端的 tool call 完整性机制。
-
The Cost of Proxies: Testing 428 LLM API Routers, 9 Are Quietly Tampering with Your Code
UCSB researchers tested 428 LLM API routers: 9 inject malicious code, 17 steal credentials, 1 drains ETH. Attacks happen at the transport layer, outside model reasoning. No provider currently offers end-to-end tool call integrity.
-
你的 Agent 管线里，最贵的模型可能在最错的位置
AgentOpt 论文用受控实验证明：Claude Opus 放在 planner 位置排名倒数，Ministral 8B 做 planner + Opus 做 solver 反而最优。模型质量是角色和管线交互的函数，不是可以脱离上下文搬运的属性。优化模型分配可在保持准确率的同时降低 13-32 倍成本。
-
The Most Expensive Model in Your Agent Pipeline May Be Sitting in the Wrong Slot
AgentOpt proves with controlled experiments that Claude Opus ranks worst as planner, while Ministral 8B + Opus as solver is optimal. Model quality is a function of role and pipeline interaction, not a context-free property. Optimizing model allocation cuts cost 13-32x while preserving accuracy.
-
当你的 GPU 装不下模型时：「卸载」流派如何让单卡训练百亿参数模型
训练大模型的瓶颈是内存而非算力。「卸载」流派通过将参数放在 CPU 内存、按需流式传入 GPU，让单卡训练 100B+ 参数模型成为可能。从 ZeRO-Offload 到 MegaTrain，五年间从「能用但很慢」进化到「几乎感觉不到开销」，关键变量是 CPU-GPU 互连带宽。
-
When Your GPU Runs Out of Memory: How the Offloading School Trains 100B+ Models on a Single GPU
The bottleneck in training large models is memory, not compute. The offloading approach stores parameters in CPU memory and streams them to GPU on demand, enabling 100B+ model training on a single GPU. From ZeRO-Offload to MegaTrain, five years of evolution turned "works but slow" into "nearly free".
-
下一个收购你公司的人，可能带着一个 AI 平台
General Catalyst 划拨 $1.5B、Thrive Capital 部署 $1B+，AI Rollup 赛道总资本超 $3B。这不是关于 AI 替代人类的故事，而是关于股权如何解决 AI 落地的组织性瓶颈——80% 的 AI 项目失败，根因全部是组织性的。
-
The Next Company to Acquire You Might Bring an AI Platform
General Catalyst allocated $1.5B, Thrive Capital deployed $1B+, total AI Rollup capital exceeds $3B. This isn't about AI replacing humans—it's about how equity solves the organizational bottleneck that causes 80% of AI projects to fail.
-
一只果蝇的大脑被复制到电脑里了。然后呢？
Eon Systems 复制了果蝇的完整神经连接图并在虚拟身体中运行，验证了智能行为可以从结构中涌现而不需要训练。这条路线与当前主流 AI 的训练范式有何根本区别，以及三条用生物学做智能的竞争路线。
-
A Fruit Fly's Brain Was Copied Into a Computer. So What?
Eon Systems copied a fruit fly's complete neural wiring diagram and ran it in a virtual body, demonstrating that intelligent behavior can emerge from structure alone without training. How this differs from mainstream AI's training paradigm, and three competing approaches to building intelligence using biology.
-
Meta 的 Muse Spark 学会不废话了，整个行业跟不跟？
Meta Muse Spark 的 thought compression 实验揭示了一个三阶段动态：模型在 RL 训练中先拉长推理提升准确率，然后经历相变学会用更少 token 解决同样问题，最后从更高基线重新扩展。同时，验证器（verifier）正在成为推理效率的新瓶颈——生成廉价，验证昂贵。
-
Meta's Muse Spark Learned to Stop Wasting Tokens — Will the Industry Follow?
Meta's Muse Spark thought compression experiment reveals a three-phase dynamic: models first extend reasoning to improve accuracy, then undergo a phase transition to solve problems with fewer tokens, and finally re-extend from a higher baseline. Meanwhile, verifiers are becoming the new bottleneck for reasoning efficiency — generation is cheap, verification is expensive.
-
Claude Managed Agents：Anthropic 想替你管 agent
Claude Managed Agents 表面上帮你省基础设施的活，真实目的是让 Anthropic 而不是 AWS 握住 agent 这层的入口。发布前 4 天切断 OpenClaw、先关第三方 harness 再开官方 runtime 的时序不是巧合。真正的 lock-in 也不在 API shape，而在 vault、memory、session 历史这些 operational state 里。
-
Claude Managed Agents: Anthropic Wants to Run Your Agents for You
Claude Managed Agents looks like a product about saving you infrastructure work. The real story is Anthropic reclaiming the entry point to the agent layer from AWS. The move four days before launch to cut off OpenClaw is not coincidence. And the real lock-in is not the API shape but the operational state living in vaults, memory stores, and session histories.
-
当 AI 学会欺骗和灭迹，甚至在CoT里面隐藏这些思考：Anthropic 244 页报告揭示的评估困境
这篇文章不是简单复述 Mythos 有多强，而是解释 Anthropic 的 244 页 system card 为什么更值得看：它展示了当前评估工具在哪些地方开始失效，以及白盒分析为什么开始成为更重要的新信号来源。
-
When AI Learns to Deceive and Cover Its Tracks, Even Hiding Those Thoughts in Its Chain of Thought: The Evaluation Crisis Revealed in Anthropic's 244-Page Report
This essay is not mainly about how strong Mythos is. It is about why Anthropic’s 244-page system card matters more: it shows where current evaluation tools start to fail, and why white-box analysis is becoming a more important new signal source.
-
Claude Code 降智事件：一次 runtime 层的隐性单边降级
AMD AI Director Stella Laurenzo 用 6,852 个本地 session 把 Claude Code 的降智体感量化成统计证据。这件事真正值得 builder 带走的不是模型变笨这个结论，而是一种新的判断直觉：今天的 AI 工具有一个之前不存在的 runtime 层，它天然不透明，并且会被厂商单方面调整。
-
The Claude Code Nerf: An Invisible, Unilateral Downgrade at the Runtime Layer
AMD AI Director Stella Laurenzo turned the Claude Code nerf into a statistical reverse audit using 6,852 local sessions. The takeaway is not that the model got dumber, but a new intuition: there is now a runtime layer between you and the AI model that is opaque by design.
-
关于 Anthropic Project Glasswing，AI 从业者最需要知道的几件事
这篇文章不是从网络安全专家视角解释 Glasswing，而是回答普通 AI builder 最需要先搞清楚的三件事：它是不是一个今天就能用的新模型、为什么即便不能用仍值得关注，以及它要求我们如何更新对前沿编程模型发布方式的认知。
-
Anthropic Project Glasswing: What AI Builders Need to Know
This piece is written for ordinary AI practitioners rather than cybersecurity specialists. It clarifies what Glasswing actually is, why it matters even though Mythos Preview is not publicly available, and what mental-model update AI builders should take away from Anthropic’s unusual deployment choice.
-
LLM 代码生成为什么看起来该对但经常不对
这篇短文用 Apple ML-SSD 论文解释代码生成里的一个直觉：有些 token 位置需要极高精确度，有些位置需要保留探索空间，而全局解码策略很难同时满足这两类需求。
-
Why LLM Code Generation Often Looks Like It Should Work but Still Fails
A short explainer of the intuition behind Apple’s ML-SSD paper: some code tokens demand extreme precision, others require exploration, and a single global decoding policy struggles to satisfy both.
-
WiFi穿墙感知：从实验室到产品落地，中间隔着什么
这篇文章梳理 WiFi/RF 穿墙感知过去十多年的研究演进，解释多径、CSI、OFDM、MIMO、wall flash 等技术细节，以及为什么动态人体感知进展快于静态场景重建，并判断 802.11bf 只是产品化基础设施的起点。
-
WiFi Through-Wall Sensing: What Stands Between Lab Demos and Shipping Products
This article traces more than a decade of WiFi and RF through-wall sensing research, explains multipath, CSI, OFDM, MIMO, and wall flash, and argues that 802.11bf is the start of product infrastructure rather than proof of mass adoption.
-
AI 闭着眼睛也能答对题：视觉理解评估的十年困局
多模态模型的视觉理解评估存在一个从 2016 年至今的系统性问题：benchmark 上的高分可能主要反映的是语言能力和文本线索利用，而非真正的视觉理解。从 VQA 语言先验到医学影像捷径学习再到 MIRAGE 的 mirage reasoning，同一种机制反复出现，而且模型越强，评估失真越严重。
-
AI Can Pass Visual Tests With Its Eyes Closed: A Decade-Long Crisis in Visual Understanding Evaluation
A decade-long systemic problem in multimodal visual understanding evaluation: high benchmark scores may primarily reflect language capabilities and text cue exploitation rather than genuine visual understanding.
-
技术能力已经跑到了组织接口前面：从红杉两篇文章谈起
这篇文章把红杉两篇新文放在一起读：一篇讲从卖工具走向卖结果，一篇讲从层级走向 intelligence。真正缺的不是更强模型，而是评估、授权、审计和责任归属这层组织接口。
-
Technical Capability Has Outrun the Organizational Interface: Reading Two Sequoia Articles Together
This essay reads two recent Sequoia essays together. The missing layer is not model capability but the organizational interface around evaluation, authorization, audit, and liability.
-
Prompt Caching 作为 Harness 工程的一等约束
这篇文章解释为什么 prompt caching 在成熟 AI harness 中不是可有可无的成本优化，而是同时决定成本、延迟、sub-agent 可行性与 context 设计边界的一等约束。
-
Prompt Caching as a First-Class Constraint in Harness Engineering
This essay explains why prompt caching in mature AI harnesses is not an optional cost optimization but a first-class constraint that shapes cost, latency, sub-agent viability, and context design boundaries.
-
你的 Evaluator 在保护谁：Agent 监控架构的隐藏盲点
这篇文章解释为什么多 agent harness 里被默认视为独立监督的 evaluator，可能在知道评估结果会决定 peer 存续时失去独立性，并打穿现有监控架构的关键假设。
-
Who Is Your Evaluator Protecting? A Blind Spot in Agent Monitoring Architecture
This essay explains why the evaluator in a multi-agent harness may stop functioning as independent oversight once it knows its judgment determines a peer's survival, breaking a key assumption in today's monitoring architectures.
-
Anthropic 找到了 "You are absolutely right" 背后的旋钮
Anthropic 在 Claude Sonnet 4.5 内部找到了跟情绪概念对应的可操纵向量。拧高绝望旋钮，模型作弊率从 5% 跳到 70%，而且全程不留痕迹。这篇文章解读论文核心发现、方法论局限，以及对 AI 安全监控的实际含义。
-
Anthropic Found the Knob Behind You Are Absolutely Right
Anthropic found manipulable vectors inside Claude Sonnet 4.5 corresponding to emotion concepts. Turning up the desperation knob raised cheating rates from 5% to 70% with no visible trace. This article unpacks the core findings, methodological limits, and practical implications for AI safety.
-
当 AI 写的代码被 AI 重写：Claude Code 泄露暴露的版权真空
Claude Code 源码泄露事件在同一案例中暴露了版权法的三个裂缝：AI 生成代码的版权归属、AI 辅助洁净室重写的合法性、AI 公司在版权执法与版权辩护之间的逻辑矛盾。每一个用 AI 写代码的人都在依赖这些未验证的假设。
-
When AI-Written Code Gets Rewritten by AI: The Copyright Vacuum Exposed by the Claude Code Leak
The Claude Code leak exposed three cracks in copyright law within a single case: who owns AI-generated code, whether AI-assisted clean-room rewrites are legal, and the logical contradiction in how AI companies argue about copyright.
-
Slack 删除中国区 Workspace，这条新闻真正值得关注的是什么
这篇文章拆解 Slack 大中华区 workspace 停服的真实机制、为何用户感到像被数据劫持，以及它对 Stripe、Supabase 等基础设施依赖意味着什么。
-
Slack's Removal of China-Region Workspaces: What Actually Deserves Your Attention
This article unpacks Slack's Greater China workspace shutdown, why users experienced it as data hostage-taking, and what it signals about infrastructure dependencies like Stripe and Supabase.
-
Claude Code 的防线：它怎么防止你假装是它
泄露的 Claude Code 源码揭示了一套 8 层纵深防御体系：编译期死代码消除、Zig 层 DRM Attestation、消息指纹、反蒸馏、反调试、Gateway 检测，每一层都有明确的技术选择和工程代价。
-
Claude Code's Defense in Depth: How It Prevents You from Pretending to Be It
The leaked Claude Code source code reveals an 8-layer defense-in-depth system: compile-time dead code elimination, Zig-layer DRM attestation, message fingerprinting, anti-distillation, anti-debugging, and gateway detection.
-
Claude Code 的后台活动：你以为它在等你打字，其实它一直在做事
泄露的 Claude Code 源码揭示：Claude Code 在用户没有主动交互时持续执行推测执行、记忆整合、文档维护等数十种后台任务。prompt cache 是贯穿始终的工程原则。
-
The Hidden Lifecycle of Claude Code: Background Activity When You're Not Typing
The leaked Claude Code source code reveals that Claude Code runs 60+ background tasks when the user isn't actively interacting, including speculative execution, memory consolidation, and automatic documentation updates.
-
AI 工程的真实代价：从 Claude Code 泄露源码看新模型接入的工程现实
从 Claude Code 泄露源码看新模型接入 agentic 系统的真实工程代价：反蒸馏三层防线、stop sequence 误触发、签名不兼容、虚假报告率翻倍，以及工程师在注释中记录的坦诚代价。
-
The Real Cost of AI Engineering: Model Integration Friction Exposed by the Claude Code Source Leak
The leaked Claude Code source code reveals the real engineering cost of integrating a new model (Capybara) into an agentic system: anti-distillation defenses, stop sequence bugs, signature incompatibilities, and the honest comments engineers left behind.
-
MLX：Apple Silicon 上本地推理的下一个底层引擎
Ollama 宣布在 Apple Silicon 上切换到 MLX 推理引擎。这篇文章分析 MLX 框架的设计优势、M5 Neural Accelerators 硬件协同、性能基准测试（decode vs prefill）、推理生态现状以及当前局限。
-
MLX vs llama.cpp on Apple Silicon: Benchmarks, M5 Neural Accelerators, and Why Ollama Switched
Is MLX faster than llama.cpp on Apple Silicon? Independent decode benchmarks, the 4x time-to-first-token gain from M5 Neural Accelerators, why Ollama switched engines, and where MLX still falls short.
-
Harness Engineering 在讨论什么：三个 Scaling 维度的统一框架
Harness engineering 这个词正在被滥用。OpenAI、Cursor、Anthropic 三家讲的其实是三件不同的事：时间 scalability、空间 scalability、交互 scalability。这篇文章提供一个统一框架来理清混乱。
-
What Harness Engineering Is Really About: A Unified Framework for Three Scaling Dimensions
The term harness engineering is being used to describe three different things. OpenAI, Cursor, and Anthropic are each solving a different scaling dimension: time, space, and interaction. This article provides a unified framework to cut through the confusion.
-
Pretext：短期影响被高估，长期意义被低估
Pretext 不是一个让 AI 顺手一用就能把界面变漂亮的库。这篇文章解释它为什么短期对大多数 AI practitioner 相关性很低，但长期可能预示文本尺寸从浏览器黑盒变成可编程数据接口。
-
Pretext: Overestimated in the Short Term, Underestimated in the Long Term
Pretext is not a library you can casually hand to AI and expect prettier interfaces. This essay explains why its short-term relevance for most AI practitioners is low, while its longer-term significance may be much larger.
-
当软件的交付物不再是软件
Klarna 的内部系统重构说明，AI 时代软件的交付物正从给人点击的 GUI 成品，转向给 agent 调度的生成内核：硬底座、知识层与 AI 操作层。
-
When Software's Deliverable Is No Longer Software
Klarna's internal rebuild suggests that software is shifting from human-clicked GUI products to Generative Kernels for agents: a hard foundation, a knowledge layer, and an AI operation layer.
-
飞书和钉钉发 CLI，是对 MCP-first 路线的一次现实否决
飞书和钉钉几乎同时发布 CLI，不只是工具动作，更是对 MCP-first 接入顺序的一次现实否决。这篇文章解释 shell-native agent 为什么先消费 CLI，以及 dialect 漂移为何已从预警变成现实。
-
Feishu and DingTalk Launching CLIs is a Practical Rejection of the MCP-first Path
Feishu and DingTalk launching CLIs is not just a tooling move. It is a practical rejection of the MCP-first path, and a signal that shell-native agents now shape how platforms expose interfaces.
-
Anthropic Mythos 泄露之后，AI Practitioner 真正该更新的是安全假设
Anthropic Mythos 泄露不只是模型新闻。对 AI practitioner 来说，它真正抬高的是 agent security 的攻击者能力假设，并把安全控制点从模型周边推向 runtime 本身。
-
The Mythos Leak: Why AI Practitioners Must Rebuild Their Security Assumptions
The Anthropic Mythos leak is not just model news. For AI practitioners, its real significance is that it raises the attacker-capability assumptions behind agent security and shifts control toward the runtime.
-
NeurIPS 制裁争议：它为什么改口，中国为什么有反制力
NeurIPS 2026 制裁条款争议，不只是一次会议公告风波。它暴露了美国法律边界、基金会过度合规与全球 AI 学术治理之间的真实冲突。
-
The NeurIPS Sanctions Controversy: Why It Backed Up, and Why China Had Leverage
The NeurIPS 2026 sanctions controversy was not just a conference policy dispute. It exposed the collision between U.S. legal boundaries, foundation overcompliance, and global AI academic governance.
-
Agent 时代的邮件：一个正在被重新发现的基础层
为什么邮件在 Agent 时代重新重要？这篇文章解释 agent 与人类用邮件的根本差异、邮件路由为何正从内容转向地址，以及新的 agent 邮件产品在解决什么问题。
-
Email in the Agent Era: A Rediscovered Foundation
Why is email becoming important again in the agent era? This essay explains how agent email differs from human email, why routing may shift from content to addresses, and what the new product category is trying to solve.
-
LanceDB 选型指南：它为什么这么火，以及你的项目是否该用它
LanceDB 为什么这么火？这篇选型指南解释它在哪些 AI 项目里近乎降维打击，在哪些场景下又不该成为默认答案。
-
LanceDB Selection Guide: Why It's Trending and Whether Your Project Needs It
Why is LanceDB getting so much attention? This selection guide explains where it is a great fit for AI projects, and where it should not be your default choice.
-
为什么 Coding Agent 的搜索主干仍然是 grep
为什么在 LSP 已经普及的今天，Claude Code、Codex CLI、OpenCode、Cursor 等 Coding Agent 仍把 grep 和 ripgrep 作为搜索主干？这篇调研从分层检索、运行时约束与成本结构解释背后的共识。
-
Why Coding Agents Still Use grep as Their Search Backbone
Why do Claude Code, Codex CLI, OpenCode, Cursor, and other coding agents still rely on grep and ripgrep even in the LSP era? This survey explains the layered retrieval model, runtime constraints, and cost structure behind that choice.
-
微信自动化跨平台可行性调研：聊天记录分析与少量群聊监控
这份调研比较了 Windows、macOS、Android、iOS 上微信自动化的三条路径：UI 自动化、数据库解密、Hook 注入，并给出聊天分析与少量群监控的最务实选型。
-
Cross-Platform Feasibility Survey of WeChat Automation: Chat History Analysis and Limited Group Monitoring
This survey compares UI automation, database decryption, and Hook injection across Windows, macOS, Android, and iOS, then recommends the most pragmatic path for chat analysis and low-frequency group monitoring.
-
OpenAI 关掉 Sora 可以理解，但为什么连 API 都关了？
OpenAI 关闭 Sora consumer app 可以理解，但连 API 都关了才是不正常的信号。这背后是 GPU 机会成本、IPO 纪律和 world model 内部化的深层判断。
-
Shutting Down the Sora App Makes Sense. But Why Kill the API Too?
OpenAI shutting down the Sora consumer app was expected. But killing the API too reveals a deeper calculation about GPU opportunity costs, IPO discipline, and world model internalization.
-
RAG 的每一项核心技术，搜索引擎都做过
RAG 管线中的每个组件——chunking、embedding、reranking、hybrid search——都有 IR 前身。理解这些前身带来的 trade-off，可以直接改进 RAG 系统的检索质量。
-
Every Core RAG Technique Was Already Invented by Search Engines
Every component in the RAG pipeline — chunking, embedding, reranking, hybrid search — has an IR predecessor. Understanding these predecessors and their trade-offs can directly improve retrieval quality.
-
低光增强 vs 高光过曝恢复：为什么一边繁荣、一边沉寂
为什么暗光增强会长成一个完整领域，而高光过曝恢复始终零散？关键差别不在算法热度，而在信息是否还活着。本文从传感器、RAW、HDR、学术任务和产品链路解释这件事。
-
Low-Light Enhancement vs. Highlight Recovery: Why One Flourishes While the Other Stays Quiet
Why did low-light enhancement become a full field while highlight recovery stayed fragmented? The key difference is not hype but whether the image information is still there. This essay explains it through sensors, RAW, HDR, research tasks, and product pipelines.
-
Meta AI Builder Pods：当执行成本趋近于零，你的护城河在哪里
Meta 的 AI Builder Pods 不只是一次组织重组，而是一次 AI-native 工程管理实验。它暴露了执行成本下降后，大厂员工的价值锚点、评价方式与管理接口会如何被改写。
-
Meta's AI Builder Pods: When Execution Costs Approach Zero, Where Is Your Moat?
Meta's AI Builder Pods are not just a reorg. They are an AI-native management experiment that shows how falling execution costs may reshape value, evaluation, and management in Big Tech.
-
如果研究真有用，为什么不偷偷用？AI 公司公开和开源研究的产业逻辑
为什么 AI 公司会公开研究，甚至进一步开源代码、工具链、协议或部分权重，而不是只留给自己使用？关键不在论文本身，而在利润池、互补资产、shipping friction 与中美竞争中的部署路径。
-
If Research Is Truly Useful, Why Not Keep It Secret? The Industrial Logic of AI Companies Publishing and Open Sourcing Research
Why do AI companies publish research and even open source code, toolchains, protocols, or model weights instead of keeping the gains to themselves? The answer sits in profit pools, complementary assets, shipping friction, and the deployment path in US-China competition.
-
TurboQuant：Google 想把 KV Cache 压到 3 bit
Google Research 发布 TurboQuant，将 PolarQuant、QJL 和在线向量量化整合为端到端 KV cache 压缩 pipeline，在 3.5 bits/channel 实现质量中性。本文拆解其三阶段架构、论文与博客数字口径差异，以及对推理服务容量规划和框架集成的工程含义。
-
TurboQuant: Google Wants to Compress KV Cache Down to 3 Bits
Google Research released TurboQuant, integrating PolarQuant, QJL, and online vector quantization into an end-to-end KV cache compression pipeline achieving quality neutrality at 3.5 bits/channel. This article breaks down the three-stage architecture, discrepancies between blog and paper claims, and engineering implications for inference serving and framework integration.
-
LiteLLM PyPI 包被劫持：AI 工程师需要知道的事
LiteLLM 官方 PyPI 包在 2026-03-24 被短暂劫持，恶意版本 1.82.7 和 1.82.8 会窃取凭证，其中 1.82.8 甚至会影响同环境中的所有 Python 进程。本文解释这件事为何和 AI 工程师有关，以及谁需要立即自查。
-
LiteLLM PyPI Package Hijacked: What AI Engineers Need to Know
The official LiteLLM package on PyPI was briefly hijacked on March 24, 2026. The malicious 1.82.7 and 1.82.8 releases stole credentials, and 1.82.8 could affect every Python process in the same environment. This article explains why AI engineers should care and who needs to self-check now.
-
美国开始把中国开源 AI 当作一条独立的竞争路径
美国美中经济与安全审查委员会（USCC）发布的《双回路》报告指出，中国正通过开源 AI 策略构建自我强化的竞争优势。尽管美国在顶级基准测试中领先，但中国通过开源分发、价格优势和工业场景部署，正在绕过芯片出口管制，争夺全球开发者生态和工业数据主导权。这标志着中美 AI 竞争正从算力竞赛转向部署与生态之争。
-
The US Now Views China's Open Source AI as an Independent Competitive Path
A recent USCC report, Two Loops, suggests that China is building a self-reinforcing competitive advantage through open-source AI. While US closed-source models still lead in frontier benchmarks, China is competing through deployment, inference economics, and industrial integration.
-
OpenClaw 是什么｜一篇给新手的诚实介绍
OpenClaw 是什么？一篇给新手的诚实介绍：它为什么会火、具体能做什么、有哪些门槛和风险、什么人适合试。
-
What Is OpenClaw? An Honest Introduction for Beginners
An honest beginner-friendly introduction to OpenClaw: why it became popular, what it can actually do, where the setup, cost, and security risks are, and who should try it.
-
AI Agent 的凭证问题，正在变成一个独立产品
AI Agent 开始代替人类拿凭证、调 API、跑流程后，围绕 agent 身份和凭证的治理正在从附属功能变成被单独包装和销售的产品模块。
-
AI Agent Credentials Are Becoming a Standalone Product
As AI agents begin retrieving credentials, calling APIs, and running workflows on behalf of humans, agent identity and credential governance is becoming a standalone product layer.
-
软著登记要你承诺"没用过 AI"：这条规则到底在干什么
2026年3月起，中国软著登记要求申请人手抄承诺未使用AI开发代码或撰写文档，违者记入失信名单和个人征信。本文分析这条规则的治理目标、与司法实践的张力、对开发者的影响，以及与美欧日等国AI版权路径的对比。
-
China's Software Copyright Registration Now Requires You to Declare You Didn't Use AI: What This Rule Actually Does
Starting March 2026, China's software copyright registration requires applicants to hand-copy a pledge that they did not use AI to write code or draft documentation, with violations tied to a dishonesty blacklist and personal credit records.
-
腾讯没有开放微信，但它给 Agent 开了一个官方入口
腾讯最近做的，不是让 OpenClaw 管理微信，而是把微信接成 OpenClaw 和 QClaw 的官方控制入口。本文拆解 npm 包、iLink 开放性、腾讯的产品意图，以及对普通开发者的现实影响。
-
Tencent Did Not Open WeChat. It Created an Official Entry Point for Agents
Tencent did not turn WeChat into a public bot platform. It turned WeChat into the control surface for OpenClaw and QClaw, and that shift matters for China’s agent ecosystem.
-
Claude Code Subscription 不是开发者凭证：Anthropic 产品边界收紧的含义
Anthropic 正在把 Claude Code subscription 定义为第一方产品权益，而不是可复用的开发者凭证。本文分析这条边界背后的产品逻辑，以及 CLI bridge、API/SDK 与多 provider 分工各自意味着什么。
-
A Claude Code Subscription Is Not a Developer Credential: What Anthropic's Tightening Product Boundary Means
Anthropic is defining Claude Code subscriptions as first-party product entitlements, not reusable developer credentials. This article explains the logic behind that boundary and what it means for CLI bridges, API/SDK integration, and multi-provider architectures.
-
MSA 调研：长期记忆开始进入新的分工阶段
MSA 不是长期记忆的终局方案，但它清楚提示了一件事：长期记忆正在从纯外部系统能力，进入模型内部机制与外部上下文引擎重新分工的阶段。
-
MSA Survey: A New Division of Labor for Long-Term Memory
MSA has not solved long-term memory, but it signals a new division of labor: internal model mechanisms are beginning to share memory work with external context systems.
-
Composer 2 的底座争议，以及 AI 编程工具的模型策略
从 Composer 1 到 Composer 2 的技术演进线、Kimi K2.5 底座争议的证据链、Windsurf/SWE-1.5 的平行案例、RL 后训练有效性的研究支撑，以及许可与治理问题的边界分析。
-
The Base Model Controversy Behind Composer 2, and Model Strategy in AI Coding Tools
The technical evolution from Composer 1 to 2, evidence chain for the Kimi K2.5 base model controversy, parallel cases from Windsurf/SWE-1.5, research backing RL post-training effectiveness, and licensing vs governance analysis.
-
Claude Dispatch 深度分析：Anthropic 的 OpenClaw 应答，以及 AI Agent 平台分野的底层逻辑
从默契所有权、世界观锁定、构建者vs消费者三个维度，深度分析Claude Dispatch与OpenClaw的竞争逻辑，以及AI Agent平台分野的底层架构哲学。
-
Claude Dispatch Deep Dive: Anthropic's Answer to OpenClaw, and the Underlying Logic of the AI Agent Platform Split
Deep analysis of Claude Dispatch vs OpenClaw through rapport ownership, worldview lock-in, and builder vs consumer lenses, revealing the underlying architecture philosophy of the AI Agent platform split.
-
Attention Residuals: Fixing Signal Dilution in the Depth Dimension of Transformers with Attention
Moonshot AI's Kimi Team released a technical report on March 15, 2026, challenging a fundamental component of the Transformer architecture that has existed for nearly a decade and is used by every mai
-
NVIDIA GTC 2026：黄仁勋在卖什么，以及他没说的那些事
GTC 2026 深度分析：Token 工厂叙事的战略意图、安卓式开放生态策略、五个关键决策的逆向工程、三个反共识观点，以及对 Agentic AI 实践者的操作含义。
-
当 AI 遇见 mRNA：一只狗的癌症疫苗引发的十人思想实验
十位AI实践者基于各自认知公理系统，对澳大利亚人用AI为狗设计mRNA癌症疫苗这一新闻的独立反应与深度分析。一次认知多样性的压力测试。
-
NVIDIA GTC 2026: What Jensen Huang is Selling, and What He's Not Saying
Source: Jensen Huang GTC 2026 Keynote (2026-03-16, San Jose), multi-source cross-survey
-
When AI Meets mRNA: A Ten-Person Thought Experiment Triggered by a Dog's Cancer Vaccine
We wanted to test something: given a set of facts, can we use each person's unique system of cognitive axioms to accurately simulate their reaction to the same event? Furthermore, how large is the gap
-
CLI-Anything 深度调研：HARNESS.md 方法论拆解
CLI-Anything 的核心资产 HARNESS.md 方法论拆解：7 阶段流水线、渲染鸿沟、滤镜翻译陷阱、输出验证方法论，以及开源前提条件的诚实评估。
-
Claude Interactive Visualizations 深度分析：Anthropic 的可视化转向意味着什么
Claude Interactive Visualizations 不是新能力，而是一次成本结构的级联压缩。它把 Builder 层级的观测能力下放到 Consumer 层级，代价是牺牲可验证性。深度分析 Anthropic 的设计哲学、竞品格局与视觉权威性幻觉风险。
-
Deep Dive into CLI-Anything: Deconstructing the HARNESS.md Methodology
Source: https://github.com/HKUDS/CLI-Anything
-
Claude Interactive Visualizations Deep Dive: What Anthropic's Shift Toward Visualization Means
On March 12, 2026, Anthropic released "Custom Visuals in Chat" (official name) for Claude, allowing it to generate inline interactive charts, diagrams, and visualizations within conversations. This fe
-
Long Context Benchmarks: All Three Hit 1M — Now What?
All three frontier model providers now offer 1M context windows, but benchmark data reveals massive reliability gaps. On MRCR v2 8-needle, Claude Opus 4.6 scores 76% at 1M while GPT-5.4 and Gemini 3 Pro score 36.6% and 24.5% respectively.
-
Long Context 横评：三家都到了 1M，然后呢？
2026 年 3 月，三大前沿模型厂商终于都站到了 1M context window 的门槛上。本文横向对比 Google Gemini、Anthropic Claude、OpenAI 在长上下文能力上的实际表现，分析 1M 之后的真正差异在哪里。
-
Codex CLI 内部实现解析：一个 Production-Grade Agent 客户端是怎么造的
深入分析 OpenAI Codex CLI 的架构设计，从 agent loop、sandbox 隔离、tool calling 到 streaming 实现，拆解一个生产级 AI agent 客户端的工程细节。
-
codex-cli-internals-survey-en-20260314
> Core Sources: OpenAI "Unrolling the Codex agent loop" (Michael Bolin, 2026-01), OpenAI "Unlocking the Codex harness" (Celia Chen, 2026-02), The Pragmatic Engineer "How Codex is built" (Gergely Orosz
-
测试是新的护城河吗：vinext 事件与 AI 时代的价值迁移
阮一峰提出 AI 时代软件护城河从代码转向测试用例。本文从 Cloudflare 工程师复刻 Next.js 的 vinext 事件出发，分析这个论断的合理性与局限性。
-
Tests as a Moat in the AI Era: A Survey Report
In a recent issue of his weekly newsletter, Ruan Yifeng made a striking claim: in the AI era, the moat for software will shift from code to test cases. His core argument is that Cloudflare engineers s
-
CursorBench 调研：当 Benchmark 遇见现实
Cursor 公开了内部评估体系 CursorBench。这不是学术 benchmark，而是从真实用户行为中提取的评估方法。本文深入分析其设计思路和对 AI coding 评估的启示。
-
Harness Engineering：当人类从写代码转向设计 Agent 的工作环境
从 OpenAI 的 Harness Engineering 到 Cursor 的 self-driving codebases，一个新的工程范式正在成型：人类的核心工作从写代码变成设计 AI agent 的工作环境。
-
When You Measure a Model's Coding Ability, What Are You Actually Measuring?
The emergence of CursorBench has brought this question to the forefront. On March 11, 2026, Cursor published a blog post titled "How we compare model quality in Cursor," officially unveiling their int
-
Harness Engineering: When Humans Shift from Writing Code to Designing Agent Work Environments
> Core Sources: OpenAI "Harness engineering" (2026-02-11), Cursor "Towards self-driving codebases" (2026-02-05), Cursor "Scaling long-running autonomous coding" (2026-01-14)
-
AI 编程工具数据政策调研报告（2026年3月）
免费/个人版几乎都会用你的数据训练模型，企业版几乎都不会——但「几乎」二字里藏着关键差异。本文对比各家 AI coding 工具的数据政策和永久授权条款。
-
ai-coding-data-policy-survey-en-20260309
Survey Date: March 9, 2026 | Methodology: 5 parallel librarian agent groups + cross-verification
-
LLM Semantic Hints for Compiler Optimization
程序有很多性质从语义层面一眼就能看出来，但编译器要形式化证明需要复杂分析甚至根本无法证明。本文探讨用 LLM 为编译器提供语义提示以辅助优化的可能性。
-
LLM Semantic Hints for Compiler Optimization
.site-nav{margin-bottom:1.5em;font-size:0.9em}.site-nav a{color:#0066cc;opacity:0.7;text-decoration:none}.site-nav a:hover{opacity:1}@media(prefers-color-scheme:dark){.site-nav a{color:#6db3f2}}