九月三日

Anthropic says Claude can now operate a computer in the background from Claude Cowork and Claude Code, clicking, typing, and opening desktop apps while the user works on something else. The capability is in beta for Pro and Max plans in the macOS desktop app and can be enabled under Settings, General, Computer use. Claude Code's Boris Cherny calls background computer use underrated, pointing toward a shift from supervised turn-by-turn automation to parallel desktop work that remains active without occupying the user's foreground attention.

Anthropic 表示，Claude 现在可以通过 Claude Cowork 与 Claude Code 在后台操作电脑，在用户处理其他事情时点击、输入并打开 desktop apps。该能力已在 macOS desktop app 中向 Pro 与 Max plans 开放 beta，可在 Settings、General、Computer use 中启用。Claude Code 的 Boris Cherny 认为 background computer use 的价值被低估了，这意味着自动化正从需要逐步监督的操作，转向不占用用户前台注意力、能够并行执行的 desktop work。

Anthropic's Boris Cherny demonstrates Fable 5.1 using Claude Tag to assemble a last-minute leadership deck from a metrics spreadsheet and Slack data. During the task, it notices that a vendor report conflicts with the numbers and flags the discrepancy before continuing. Claude Tag is available in Slack on Team and Enterprise plans, making the more important product signal not deck generation itself but an agent's ability to reconcile evidence across workplace sources before producing an executive-facing artifact.

Anthropic 的 Boris Cherny 演示了 Fable 5.1 如何使用 Claude Tag，根据 metrics spreadsheet 与 Slack 数据制作一份临时需要的 leadership deck。执行过程中，它发现 vendor report 与数字存在冲突，并在继续之前主动标记差异。Claude Tag 已在 Slack 的 Team 与 Enterprise plans 中提供。比生成 deck 更重要的产品信号，是 agent 能在制作面向管理层的 artifact 之前，跨 workplace sources 核对证据。

Google VP Josh Woodward says Gemini 3.8 Flash is rolling out to the Gemini app and positions it around quality at an attractive price. Meta Senior Director of AI Madhu Guru argues that AI products should ultimately hide model selection from users because choosing among models adds cognitive load without helping them complete the job. He says reliable abstraction requires a deep, use-case-specific view of the model frontier, strong workflow evaluations, and an AI-native team that can continually reroute work as models change.

Google VP Josh Woodward 表示，Gemini 3.8 Flash 正在向 Gemini app 推出，并把它定位为兼顾质量与价格的 model。Meta Senior Director of AI Madhu Guru 认为，AI 产品最终应对用户隐藏 model selection，因为选择 model 增加了 cognitive load，却不能直接帮助用户完成任务。他指出，要可靠地实现这种抽象，需要深入理解特定 use case 下的 model frontier、建立强有力的 workflow evaluations，并由 AI-native team 随着 models 变化持续调整任务路由。

Anthropic's Thariq corrects an important rollout detail: changing effort levels without breaking the prompt cache is live in the API, but not yet in Claude Code. He expects Claude Code support within roughly a day. For builders, the distinction matters because the API can already vary reasoning effort by task while retaining cached context, whereas Claude Code users should not yet assume the same cost and latency behavior.

Anthropic 的 Thariq 更正了一项重要的 rollout 细节：切换 effort levels 而不破坏 prompt cache 的能力已经在 API 上线，但 Claude Code 尚未支持；他预计 Claude Code 会在大约一天内跟进。对 builders 来说，这一区别很关键：API 现在已经可以按任务调整 reasoning effort，同时保留 cached context；Claude Code 用户则暂时不应假设自己已经获得相同的成本与延迟表现。

South Park Commons General Partner Aditya Agarwal says it is difficult to trust frontier labs' assurances that customer data will not be used for training, making open-source models attractive for privacy-sensitive work and making strong US open-source options strategically important. Box CEO Aaron Levie separately calls Muse a major model update and says an open-weight release would materially change US competitiveness in open models. Agarwal also warns startups against designing around today's model limitations, arguing that the more valuable product thesis anticipates capabilities and problems one year ahead.

South Park Commons General Partner Aditya Agarwal 表示，用户很难完全相信 frontier labs 关于客户数据不会用于训练的承诺，因此 open-source models 对 privacy-sensitive work 更有吸引力，而美国拥有强大的 open-source 选项也具有战略重要性。Box CEO Aaron Levie 则把 Muse 称为一次重大的 model update，并认为如果它以 open weights 发布，将实质改变美国在 open models 上的竞争格局。Agarwal 还提醒 startups 不要只围绕今天的 model 局限设计产品，更有价值的 product thesis 应预判一年后的 capabilities 与 problems。

Builder Zara Zhang argues that meetings are no longer recorded primarily so people can read AI summaries or replay the conversation. Their emerging purpose is to capture transcripts that can feed agents. This reframes meeting software from a documentation endpoint into an ingestion layer for systems that can retrieve decisions, update context, and act on organizational conversations.

Builder Zara Zhang 认为，记录会议的主要目的已经不再是让人阅读 AI summaries 或回听对话，而是捕获可供 agents 使用的 transcripts。这让 meeting software 从 documentation endpoint 转变为系统的 ingestion layer，使系统能够检索决策、更新 context，并基于组织内部对话采取行动。

AI educator Peter Yang observes a vocabulary shift toward “Member of Product Staff,” “Member of Design Staff,” and “Member of Technical Staff.” He welcomes the possibility that these labels flatten levels, but sees their functional separation as evidence that a single all-purpose “AI builder” role is not replacing specialization, even at leading labs. OpenAI's Thibault Sottiaux adds an operating-context clue, describing the company as a “mega startup” defined by extreme ownership, care, and pace.

AI educator Peter Yang 注意到一种职位命名变化：“Member of Product Staff”、“Member of Design Staff” 与 “Member of Technical Staff” 正在出现。他欢迎这些称呼可能带来的层级扁平化，但认为不同职能依然分开，说明即使在领先 labs，单一的全能 “AI builder” 角色也没有取代专业分工。OpenAI 的 Thibault Sottiaux 还补充了一条组织语境线索，把公司描述为由 extreme ownership、care 与 pace 驱动的 “mega startup”。

The Takeaway: Every staff writer Katie Parrott treats AI writing as a compounding system in which durable context, fresh human evidence, structured review, and remembered feedback increase creative capacity without outsourcing judgment.

核心结论：Every staff writer Katie Parrott 把 AI writing 视为一个持续复利的系统，其中 durable context、鲜活的人类证据、结构化 review 与被长期记住的 feedback 会扩大创作能力，但不会把判断力外包出去。

Parrott's path began with using ChatGPT as a lower-cost career coach after losing an editor and ghostwriter role at a crypto firm. The tool helped her externalize decisions and later supported a freelance workload by handling repetitive production. The real unlock was context engineering: she supplied brand messaging, product details, audience needs, competitive differentiation, style guides, and example work before asking for output. Her sharper point is that model knowledge is commoditized; the human closes the “last mile” with recent research, proprietary data, lived experience, and the unique observations that make a piece worth reading.

Parrott 的实践起点，是在失去 crypto firm 的 editor 与 ghostwriter 工作后，把 ChatGPT 当作成本更低的 career coach。这个工具帮助她把决策过程外化，后来又通过处理重复性生产任务，支撑了高负荷的 freelance 工作。真正的突破来自 context engineering：在要求输出之前，她先提供 brand messaging、product details、audience needs、competitive differentiation、style guides 与 example work。她更尖锐的观点是，model 已有的知识是 commoditized；人类需要用最新 research、proprietary data、真实经历与独特观察补上 “last mile”，让文章真正值得阅读。

Her experimental Compound Writing plugin adapts Kieran Klassen's Compound Engineering workflow to brainstorming, outlining, drafting, substantive editing, line editing, and a final publication pass. “Every piece of feedback that you give to AI should feed back into the system to improve the next output.” The plugin evaluates structure, voice, and evidence, and lets writers invoke codified lenses inspired by storytellers such as Vonnegut and Hitchcock. Parrott describes it as a gym rather than a shortcut: the system gives the writer more informed material to accept or reject, helping taste develop through repeated decisions.

她的实验性 Compound Writing plugin 把 Kieran Klassen 的 Compound Engineering workflow 改写为 brainstorming、outlining、drafting、substantive editing、line editing 与最终 publication pass。“Every piece of feedback that you give to AI should feed back into the system to improve the next output.” 这个 plugin 会评估 structure、voice 与 evidence，也让作者调用受 Vonnegut、Hitchcock 等 storytellers 启发的结构化视角。Parrott 把它形容为 gym，而不是 shortcut：系统提供更多有依据、可接受或拒绝的材料，作者则通过反复做决定来培养 taste。

The same philosophy extends beyond writing. Her Codex-based career coach maintains priorities and a Kanban board using role context, OKRs, content-performance data, and reader feedback, while agents handle “computer errands” that reduce daily friction. She also describes the boundary of borrowed expertise: after AI helped her build an MCP for Tastemaker, a Codex review found five security vulnerabilities. Her closing thesis is therefore social as well as technical. Education and access must spread beyond people with the money, time, and proximity to early AI networks, or the technology's multiplicative value will compound inside a small early group instead of broadening opportunity.

同一套理念也延伸到写作之外。她基于 Codex 的 career coach 利用 role context、OKRs、content-performance data 与 reader feedback 管理 priorities 和 Kanban board，agents 还会处理降低日常摩擦的 “computer errands”。她也说明了借用专家能力的边界：AI 帮她为 Tastemaker 构建 MCP 后，Codex review 发现了五个 security vulnerabilities。因此，她最后的 thesis 同时关乎社会与技术：education 与 access 必须覆盖那些缺少资金、时间或早期 AI 网络连接的人，否则这项技术的 multiplicative value 只会在少数先行者内部复利，而不能真正扩大机会。