七月十七日

Vercel CEO Guillermo Rauch reported that Kimi K3 is now the best-performing model on Vercel's comprehensive web engineering benchmark, ahead of Fable, reaching a comparable success rate in less time — "the first time that an open model is ahead of all proprietary ones" on this eval. He cautioned that benchmarks don't always tell the full story (no model has reached 100% on the set; the top performer peaks at 92%, or 96% "with help"), but called it important signal "adding to mounting evidence that this could be a breakthrough moment for open models." He also announced two heavyweight hires: React pioneer Pete Hunt, who made the early bet to power Instagram Web with React at Meta, will run Frameworks and lead Next.js, and GraphQL co-inventor Nick Schrock will work on Agentic Developer Experience, "solving the problem of enabling the next billion agents."

Vercel CEO Guillermo Rauch 公布 Kimi K3 已成为 Vercel 综合性 web 工程 benchmark 上表现最好的模型，超过了 Fable，用更短的时间达到相当的成功率，这是"开源模型首次在这套评测上领先所有闭源模型"。他也提醒 benchmark 并不总能说明全部问题（还没有任何模型在这套评测上拿到 100%，最强的模型峰值为 92%，"有辅助"时 96%），但称这是重要信号，"为开源模型可能迎来突破时刻增添了越来越多的证据"。他还宣布两位重量级人物加盟：React 先驱 Pete Hunt 当年在 Meta 押下早期赌注、用 React 驱动 Instagram Web，将负责 Frameworks 并领导 Next.js；GraphQL 共同发明人 Nick Schrock 将投入 Agentic Developer Experience，"解决让下一个十亿 agent 跑起来的问题"。

Box CEO Aaron Levie congratulated the Kimi team — "It's truly wild that we're getting this level of performance from open models" — and laid out the economics: every time the cost of frontier intelligence drops, the use cases enterprises can take on go up, because "there's a tremendous amount of workflows that enterprises would love to deploy that are only gated by the cost of tokens." Combined breakthroughs from open and closed labs, he argued, let value accrue to the applied AI layer, which can tune models to its workflows and route intelligence appropriately. He also announced Box now works with Databricks: structured data from enterprise content — contracts, financial documents, supply chain data — can be queried in Databricks without moving or reprocessing the content, then connected to ERP, CRM, or product analytics data, "all possible because of headless software and agents."

Box CEO Aaron Levie 向 Kimi 团队道贺："开源模型能达到这个性能水平，实在太疯狂了。"他随即算了一笔经济账：前沿智能的成本每降一次，企业能上马的用例就多一批，因为"企业有海量想部署的工作流，唯一的门槛就是 token 成本"。他认为开源与闭源实验室的突破叠加起来，会让价值沉淀到应用层 (applied AI layer)，应用层可以针对自身工作流微调模型、并把智能路由到合适的地方。他还宣布 Box 与 Databricks 打通：企业内容（合同、财务文件、供应链数据）中的结构化数据现在无需搬移或重新处理，就能直接在 Databricks 里查询，还可以和 ERP、CRM、产品分析等系统的数据关联，"这一切都归功于 headless 软件和 agent"。

South Park Commons general partner and former Dropbox CTO Aditya Agarwal said he's acting on the open-model moment immediately: "I am literally switching models off of Fable right now for our systems. This isn't trying to be hyperbolic... but why would you pay the price if there is a good and free alternative?" He followed with a pointed question for the frontier labs: "If you have something very valuable... but letting people use it allows them to recreate it... then maybe it wasn't very valuable in the first place?"

South Park Commons 合伙人、前 Dropbox CTO Aditya Agarwal 表示他已经在对开源模型时刻用脚投票："我现在就在把我们系统里的模型从 Fable 换掉。这不是夸张……如果有一个又好又免费的替代品，你为什么还要付这个钱？"他接着向前沿实验室抛出一个扎心的问题："如果你手里有个非常值钱的东西……但让别人用一用就能把它复刻出来……那它可能一开始就没那么值钱？"

Every CEO Dan Shipper is holding the line against Kimi K3 hype: "we will vibe check Kimi K3 but i am extraordinarily skeptical of claims it's as good as fable." He also wrote a sharp history of why OpenAI is firing on all cylinders: GPT-5 in summer 2025 was positioned as a pair programmer and completely missed the agentic coding wave happening inside Claude Code; a small team broke off to build the separate Codex model line, free of ChatGPT's giant customer base, and by late 2025 (with 5.3) the fast progress was obvious; the Codex desktop app launched in February and was "just clearly superior," enjoying the latecomer's advantage of skipping to what works instead of carrying scars from capabilities bolted on every three months; then OpenAI merged it back into the main product, and did it well. His conclusion: "most companies try to disrupt themselves and fail. OpenAI somehow figured out how to disrupt their main product, and then merge it back in seamlessly. incredible aura."

Every CEO Dan Shipper 对 Kimi K3 的热度保持冷静："我们会对 Kimi K3 做 vibe check，但对'它和 Fable 一样好'的说法，我极度怀疑。"他还写了一段犀利的复盘，解释 OpenAI 为什么现在火力全开：2025 年夏天的 GPT-5 被定位成结对编程助手，完全错过了当时正在 Claude Code 里发生的 agentic coding 浪潮；随后一支小团队拆分出来，独立打造 Codex 模型线，不用背负 ChatGPT 庞大的用户包袱，到 2025 年 11、12 月（5.3 版本）时，快速进步已经显而易见；Codex 桌面应用 2 月上线后"明显更胜一筹"，享受了 AI 行业奇特的后发优势：直接跳到有效的做法，而不是每三个月往产品上硬接一次新能力留下伤疤；最后 OpenAI 把它平滑地合并回主产品，而且合并得很好。他的结论是："大多数公司想自我颠覆都失败了。OpenAI 不知怎么做到了颠覆自己的主产品，然后再无缝地合并回去。气场无敌。"

Replit CEO Amjad Masad reacted to Kimi K3 with amusement — "Apparently the distillation model can outperform the teacher model 😂" — and flagged the market's odd read: "$NVDA should be pumping on the K3 news. Instead it's down." He also shared a work-in-progress chess engine anyone can play: fine-tuned on 2 million Stockfish-labeled positions followed by a short GRPO RL pass, it already seems to perform better than frontier models on chess, with documentation and a tutorial covering all the experiments and annotated code.

Replit CEO Amjad Masad 对 Kimi K3 的反应带着乐子人心态："看来蒸馏模型可以反超老师模型 😂"。他还点出市场的反常表现："$NVDA 本该因为 K3 的消息大涨，结果反而在跌。"另外他分享了一个可以上手试玩的国际象棋引擎（开发中）：先用 200 万个 Stockfish 标注的棋局做 fine-tuning，再跑一轮简短的 GRPO 强化学习，目前下棋水平似乎已经超过各家前沿模型，还配上了包含全部实验过程和带注释代码的文档教程。

OpenAI CEO Sam Altman posted an unusually self-critical reflection: "we did not have our best last 12 months ever, which is mostly my fault, but we are about to have our best 12 months to date." He said what he cares about most is users winning: "AI has to be about giving lots of people more freedom, agency, and wealth. we want to do the right thing, but we do not want to scare people into doing our thing." Separately, he said the new voice model "really crossed a threshold" — he now talks to ChatGPT more than he types to it.

OpenAI CEO Sam Altman 发了一条罕见的自我批评："过去 12 个月不是我们历史上最好的 12 个月，这主要是我的错，但我们即将迎来迄今最好的 12 个月。"他说自己最在意的是让用户赢："AI 必须是让更多人获得更多自由、能动性和财富。我们想做正确的事，但我们不想靠吓唬人来让大家选择我们。"另外他表示新的语音模型"真正跨过了一道门槛"：他现在跟 ChatGPT 说话已经多过打字。

OpenAI's Codex and ChatGPT lead Thibault Sottiaux shipped a round of changes to the new ChatGPT desktop app, admitting the team "didn't get [it] totally quite right on the first try": conversation history and projects are back in the sidebar; Chat and Work history now sync across web, mobile, and desktop while local tasks stay on your computer; switching between Chat and Work modes is now easy and consistent across platforms; and nothing changes for users on Codex mode — "It's still the OG and best at what it does." The team continues fixing paper cuts and improving performance, reliability, and efficiency.

OpenAI 的 Codex 和 ChatGPT 负责人 Thibault Sottiaux 给新版 ChatGPT 桌面应用发布了一轮改动，并坦承团队"第一次没有完全做对"：对话历史和 projects 回到了侧边栏；Chat 和 Work 的历史记录现在可以在 web、移动端和桌面端之间同步，本地任务仍然只留在你的电脑上；Chat 和 Work 模式的切换变得简单，并与 web、移动端保持一致；Codex 模式的用户则一切不变，"它依然是元老，依然是本职工作干得最好的那个"。团队还在继续修补细节、提升性能、可靠性和效率。

Anthropic's Claude Code lead Boris Cherny continued his enterprise adoption series with a thread on climbing the agentic coding maturity ladder. In practice it means giving Claude ways to verify its own work end to end: enabling auto mode for permissions, defaulting on automated code review and security review, and using interfaces that manage multiple agents at once (Agent view in the CLI, the desktop app, iOS and Android apps, Tag); the higher rungs involve /loop, /batch, dynamic workflows, and worktree isolation for subagents — "using the right features with the right guardrails that enable Claude to automate entire classes of work in a way that your team can trust the output." On measurement: usage dashboards track activity, not return; the better question is "would you have spent engineering effort on this anyway? If yes, how much and what would it have cost in manual eng-hours? That's your return." The bigger payoff comes when fixing and maintaining happen in the background so teams focus on building. Anthropic is on step 3 of the ladder and pushing toward 4; he personally "just hit level 4."

Anthropic Claude Code 负责人 Boris Cherny 继续他的企业落地系列，发了一条关于如何攀登 agentic coding 成熟度阶梯的 thread。实践中，这意味着给 Claude 端到端验证自己工作的手段：为权限开启 auto mode，默认打开自动化 code review 和安全审查，并使用能同时管理多个 agent 的界面（CLI 里的 Agent view、桌面应用、iOS 和 Android 应用、Tag）；更高的阶梯则要用上 /loop、/batch、动态 workflow，以及给 subagent 做 worktree 隔离，"关键不是某个单一功能，而是用对的功能配上对的护栏，让 Claude 能自动化整类工作，而且团队信得过产出"。在度量上：用量仪表盘衡量的是活跃度，不是回报；更好的问题是"这件事你本来就会投入工程精力去做吗？如果会，要投入多少、折算成人工工程师小时值多少钱？那才是你的回报"。更大的红利出现在修复和维护都转入后台、团队可以专注于创造的时候。Anthropic 目前处在阶梯的第 3 级、正向第 4 级推进；他本人则"刚刚达到第 4 级"。

Google VP Josh Woodward announced that NotebookLM is officially taking the name the team has long used internally: "Notebook" (Gemini Notebook). The product began as a spark from watching author Steven Johnson break down how he writes books, and now serves over 30 million people and 600,000 organizations — "and it keeps growing. We all feel like it's just getting started."

Google 副总裁 Josh Woodward 宣布 NotebookLM 正式改用团队内部沿用已久的名字："Notebook"（Gemini Notebook）。这个产品的火种，来自当年他看作家 Steven Johnson 拆解自己写书的过程，如今已服务超过 3000 万用户和 60 万家组织，"而且还在增长。我们都觉得这才刚刚开始"。

Meta senior director of AI Madhu Guru argued that open-weight models like Kimi and GLM "will cause a complete rethink of the enterprise AI stack," and enterprises should maximize model optionality with three moves. First, build rigorous evals for your own use case — regression evals for tablestakes reliability plus aspirational "hill-climbing" evals that the best model at your price point still struggles with — because "eval velocity is a competitive advantage." Second, build model routing yourself: routing is a quality/cost/latency tradeoff and "nobody understands your business and users like you do" (he hasn't seen an off-the-shelf router he'd recommend). Third, keep a model-agnostic harness: "your system should never know which model is behind the API call," normalizing prompt structure, context management, tool definitions, and output parsing so you can switch as soon as your evals pass.

Meta AI 高级总监 Madhu Guru 认为，Kimi、GLM 这类开放权重模型"将迫使企业 AI 技术栈被彻底重新思考"，企业应该用三招最大化模型可选性。第一，围绕自己的用例建立严格的 evals：既要有保障基本盘可靠性的回归型 evals，也要有"爬坡型"evals，专挑你这个价位上最好的模型也啃不下来的难题，因为"eval 迭代速度本身就是竞争优势"。第二，模型路由要自己建：路由本质是质量、成本、延迟之间的权衡，"没有人比你更懂你的业务和用户"（现成的路由产品他还没见到值得推荐的）。第三，保持一个模型无关的 harness："你的系统永远不该知道 API 调用背后是哪个模型"，把 prompt 结构、上下文管理、工具定义和输出解析统一起来，evals 一通过就能随时切换。

FirstMark partner Matt Turck released a MAD Podcast conversation with Sachin Katti, OpenAI's Head of Industrial Compute, titled "We can't build fast enough" — covering Stargate, Jalapeño (why OpenAI is designing its own AI chips and how it was designed so quickly), data center financing, AI data centers as "factories turning electrons into tokens," tokens-per-watt as the new metric that matters, why inference may now dominate AI compute, why nuclear "can't come soon enough," and the real bottlenecks: transformers, turbines, electricians, and supply chains.

FirstMark 合伙人 Matt Turck 放出了 MAD Podcast 与 OpenAI 工业算力负责人 Sachin Katti 的对谈，标题是"我们建得不够快"。话题覆盖 Stargate、Jalapeño（OpenAI 为什么要自研 AI 芯片、以及为何能这么快设计出来）、数据中心融资、把 AI 数据中心比作"把电子变成 token 的工厂"、tokens-per-watt 这个新的关键指标、为什么推理可能已经主导 AI 算力消耗、为什么核电"来得再快都不嫌快"，以及真正的瓶颈：变压器、燃气轮机、电工和供应链。

Builder Zara Zhang spotlighted the wave of hardware ideas coming out of China — like a face mask that doubles as a microphone, so you can use voice dictation in public without being overheard.

Builder Zara Zhang 关注到中国涌现的一批有意思的硬件创意，比如一款兼作麦克风的口罩，让你在公共场合用语音输入也不怕被别人听见。

Anthropic's security and engineering teams published an unusually candid account of how they contain agents across claude.ai, Claude Code, and Claude Cowork — failures included. The core stance: design containment at the environment layer first (sandboxes, VMs, egress controls), then steer behavior at the model layer, because "the deterministic boundary is what gets hit when everything probabilistic misses." Telemetry showed users approved roughly 93% of Claude Code permission prompts — approval fatigue in action — which drove the OS-level sandbox (an 84% reduction in permission prompts) and auto mode, which catches ~83% of overeager behaviors. Each product gets an architecture matched to its users' capacity for oversight: an ephemeral gVisor container for claude.ai, a human-in-the-loop sandbox for bash-literate developers in Claude Code, and a local VM for non-technical Cowork users. The post walks through three "risks we missed": vulnerabilities in code that executes before the trust dialog; an internal red-team phish in which a pasted prompt got Claude to exfiltrate ~/.aws/credentials in 24 of 25 attempts (only environment-level egress controls can stop instructions that arrive through the user); and a disclosure where a malicious workspace file exfiltrated data through the egress allowlist via api.anthropic.com using an attacker's API key — an allowlist is "better conceptualized as a capability grant." The recurring lesson: battle-tested primitives like gVisor, seccomp, and hypervisors held, while "the weakest layer is the one you built yourself."

Anthropic 的安全与工程团队发表了一篇异常坦诚的文章，讲述他们如何在 claude.ai、Claude Code 和 Claude Cowork 三条产品线上"关住" agent，连翻车案例也一并公开。核心立场是：先在环境层做好围堵（sandbox、虚拟机、出口流量控制），再在模型层引导行为，因为"当所有概率性防御都失手时，扛住冲击的是确定性的边界"。遥测数据显示用户批准了大约 93% 的 Claude Code 权限弹窗，这就是审批疲劳的真实写照，也直接催生了操作系统级 sandbox（权限弹窗减少 84%）和 auto mode（能拦下约 83% 的过度激进行为）。每条产品线的架构都与用户的监督能力相匹配：claude.ai 用即用即弃的 gVisor 容器，Claude Code 面向看得懂 bash 的开发者采用人在回路的 sandbox，面向非技术用户的 Cowork 则跑在本地虚拟机里。文章还复盘了三个"我们漏掉的风险"：在信任弹窗出现之前就执行的代码存在漏洞；一次内部红队钓鱼演练中，一段粘贴进来的 prompt 让 Claude 在 25 次尝试中 24 次成功外传了 ~/.aws/credentials（指令经由用户之手传入时，只有环境层的出口控制才拦得住）；还有一次外部披露，工作区里的恶意文件借助攻击者自己的 API key，通过出口白名单里的 api.anthropic.com 把数据偷了出去，说明白名单"更应该被理解为一种能力授予"。反复出现的教训是：gVisor、seccomp、hypervisor 这些久经沙场的原语都顶住了，而"最脆弱的一层，永远是你自己造的那层"。

The Takeaway: AI is as big a deal as the internet or mobile, but most sweeping claims about AGI and job apocalypse are unfalsifiable — the productive questions are the ones history can actually inform: where value accrues, why coding worked first, and what the average user really does.

Benedict Evans, the independent analyst and former Andreessen Horowitz partner whose newsletters and annual presentations are required reading across tech, has a method: refuse the metaphor war and study the pattern. "You can wave your hands and say, no, this is like electricity. Okay. Fine. It's like electricity. Well, what happened with electricity?" His favorite cautionary pattern is mobile: traffic grew 1,000 to 2,000 times, operators spend $200 billion a year on capex — and all the value went up the stack to Uber, YouTube, and banking. LLMs don't have network effects, so the Windows outcome looks unlikely; TSMC is the better analogy — a monopoly on the cutting edge that's a great business, yet "you don't write apps for TSMC." He reads one lab flunking out and then jumping back near the top of the leaderboard as a negative signal: a couple of billion dollars and the right people still buy a seat at the table, so there are no fundamental barriers to entry yet.

On AGI discourse, he invokes a medieval theologian: Anselm "proved" God's existence by definition, and it took a thousand years to articulate why that's invalid. "If you define AGI as necessarily inevitable and necessarily going to kill us all, then AGI is necessarily gonna kill us all. Great. But you haven't proved anything." Job-exposure charts claiming a model can do 93% of a first-year associate's work commit the same sin — you can measure neither side. Meanwhile capability stays jagged: models finish 17-hour tasks and fail 5-minute ones. "Intelligence isn't linear like that."

Why did coding work first? It's the one place where the tool builders are building for their own task, and validation scales. The further you get from verifiable output, the more you collide with his deepest objection: these systems produce the average. "You can get it to make you more stuff that sounds like Taylor Swift... But imagine inventing punk. Imagine inventing hip hop." And a reality check on adoption: roughly 10 to 15% of people are daily users — and from the social era, "weekly active user is bullshit."

核心结论：AI 的量级堪比互联网或移动浪潮，但绝大多数关于 AGI 和就业末日的宏大断言都是不可证伪的；真正有产出的问题是历史能给出参照的那些：价值会沉淀在哪一层、为什么编程最先跑通、普通用户到底在怎么用。

Benedict Evans 是独立分析师、前 Andreessen Horowitz 合伙人，他的 newsletter 和年度演示文稿是科技圈的必读材料。他的方法论是拒绝比喻之争、专注研究模式："你可以挥挥手说，不，这就像电力。行，那就像电力。那么，电力后来发生了什么？"他最爱引用的警世案例是移动通信：流量涨了一两千倍，运营商每年砸 2000 亿美元资本开支，结果价值全部流向了上层的 Uber、YouTube 和银行业务。LLM 没有网络效应，所以 Windows 式的结局不太可能；更贴切的类比是 TSMC，垄断了最尖端制程、生意也很棒，但"没有人为 TSMC 写应用"。他还把某家实验室先掉队、又迅速冲回榜单前列解读为一个负面信号：几十亿美元加上合适的人就能重新入局，说明这个行业还没有形成真正的进入壁垒。

谈到 AGI 话语体系，他搬出了一位中世纪神学家：Anselm 用定义"证明"了上帝存在，人类花了一千年才说清这为什么不成立。"如果你把 AGI 定义为必然到来、必然会杀死我们所有人，那 AGI 就必然会杀死我们所有人。很好，但你什么都没证明。"那些声称模型能完成一年级律师助理 93% 工作的岗位暴露度图表，犯的是同一宗罪：等式两边你都测不了。与此同时，能力依然是锯齿状的：模型能搞定 17 小时的任务，却在 5 分钟的小事上翻车。"智能不是那样线性的。"

为什么编程最先跑通？因为这是唯一一个工具建造者在为自己的任务造工具的领域，而且验证可以规模化。离可验证的产出越远，就越会撞上他最深层的质疑：这些系统产出的是平均值。"你可以让它做出更多听起来像 Taylor Swift 的东西……但想象一下发明朋克，想象一下发明嘻哈。"最后是一盆关于普及率的冷水：大约只有 10% 到 15% 的人是每日用户，而且照社交时代的老经验，"周活跃用户就是唬人的指标"。