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.
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。
MAI mass-produced 265K training environments from 4.87M PRs. The real shift in AI coding is industrial-scale training infrastructure.
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。
MAI uses three mechanisms to keep RL training stable for thousands of steps. DeepSeek optimizes for speed, GLM for memory, MAI for discipline.
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。
MAI-Thinking-1 reveals top AI labs' R&D taste.
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 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 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。
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.
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。
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.
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。
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.
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。
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.
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 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.
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。
When to pursue alpha and when to settle for beta in life decisions. A framework based on half-life, feedback quality, and comparative advantage.
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。
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.
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。
A career guide for AI beginners: four self-diagnosis starting points, three practical routes with week-one action plans, and the pitfalls to avoid.
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。
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.
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。
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.
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。
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.
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。
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.
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。
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.
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。
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.
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。
DeepSWE reveals how SWE-Bench Pro lost resolving power to contamination and verifier flaws. The same models spread 62 points on a fresh benchmark.
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。
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.
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。
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.
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。
From evaluation criteria to algorithmic evolution to industrial practice: a systematic survey of three decades of digital watermarking technology.
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
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.
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。
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.
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。
For two decades data centers were the golden geese of American local governments. Now a systemic backlash is unfolding from coast to coast.
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。
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 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。
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.
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。
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.
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。
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.
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 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.'
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」。
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 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。
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.
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。
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.
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。
Anthropic made four acquisitions in six months. OpenAI bought Astral. Nearly everything acquired is open source. Why spend hundreds of millions instead of forking?
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？
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.
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。
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.
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'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.
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。
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.
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。
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.
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。
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.
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。
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.
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 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 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。
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.
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。
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.
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。
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.
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。
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.
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。
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.
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。
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.
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 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 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。
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.
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。
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.
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。
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.
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 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 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。
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。
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.
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。
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 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。
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.
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.
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。
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。
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.
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。
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.
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。
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.
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。
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.
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。
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.
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。
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.
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。
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.
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。
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.
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。
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.
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。
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 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。
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.
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。
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.
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。
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 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。
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.
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。
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.
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。
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.
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。
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 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。
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 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 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.
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。
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.
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。
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 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。
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.
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。
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.
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。
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.
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。
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.
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.
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。
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.
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。
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.
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。
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.
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。
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.
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。
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.
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。
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.
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。
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.
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。
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.
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。
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.
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。
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 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.
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。
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.
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。
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.
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。
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.
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。
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.
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。
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.
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。
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.
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。
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.
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。
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.
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。
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 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 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.
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。
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.
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。
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.
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。
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 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 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 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'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.
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。
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.
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。
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.
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。
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.
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。
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.
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。
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".
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。
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.
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。
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 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 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 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 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.
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。
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.
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。
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.
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。
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.
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。
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.
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。
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.
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。
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.
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。
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.
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。
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.
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。
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 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 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.
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。
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.
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。
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.
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。
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.
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 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.
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 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.
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。
Ollama switched to MLX as its inference engine on Apple Silicon. This article analyzes MLX's architectural advantages, M5 Neural Accelerator hardware synergy, performance benchmarks (decode vs prefill), the current inference ecosystem, and existing limitations.
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。
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 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 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.
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。
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 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 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 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 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.
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。
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 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。
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 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.
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。
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。
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 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 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.
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。
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 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 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.
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。
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.
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。
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 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。
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。
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：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。
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。
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.
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。
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 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.
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。
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 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 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.
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 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.
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。
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.
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
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。
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。
Source: Jensen Huang GTC 2026 Keynote (2026-03-16, San Jose), multi-source cross-survey
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 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 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。
Source: https://github.com/HKUDS/CLI-Anything
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
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.
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。
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。
> 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
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。
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
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。
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。
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
> 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)
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。
Survey Date: March 9, 2026 | Methodology: 5 parallel librarian agent groups + cross-verification
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。
.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}}