七月二十六日

OpenAI's Thibault Sottiaux said ChatGPT Work has officially overtaken Codex in active users. He highlighted mobile as a game-changing way to use it and framed the underlying shift simply: people could always speak to computers, but now the computer can do meaningful work in return. Together, the posts point to voice and mobile becoming practical control surfaces for agentic work, not just chat interfaces.

OpenAI 的 Thibault Sottiaux 表示，ChatGPT Work 的活跃用户数已正式超过 Codex。他特别强调移动端的体验堪称 game changer，并把底层变化概括得很直接：人们一直都能对电脑说话，只是过去电脑无法真正完成多少工作。这几条动态共同指向一个趋势，voice 和 mobile 正在成为 agentic work 的实用控制界面，而不只是聊天入口。

Vercel CEO Guillermo Rauch argued that a product is only as good as the agents autonomously maintaining it, so founders should build a factory that can start, maintain, and grow an idea instead of repeatedly giving ad hoc prompts to an agent. His own research setup is intentionally plain: a `research/` folder, an `AGENTS.md` that specifies format and best practices, and agent CLIs that can correlate prior sessions, sync through iCloud or git, and render findings as deployable HTML reports. Linear Head of Product Nan Yu pushed the idea one recursive step further: if you can build a SoftwareFactory, you can build a SoftwareFactoryFactory, and the pattern generalizes beyond software to designed intentions in areas such as public health or law.

Vercel CEO Guillermo Rauch 认为，产品的上限取决于负责自主维护它的 agents，因此创业者不该反复向 agent 临时下 prompt，而应构建一座能够启动、维护并持续发展想法的工厂。他自己的研究环境刻意保持简单：一个 `research/` 文件夹、一份规定输出格式与最佳实践的 `AGENTS.md`，再加上能够关联历史会话的 agent CLI；文件可通过 iCloud 或 git 同步，研究结果还能被渲染成可部署的 HTML 报告。Linear Head of Product Nan Yu 又把这个思路递归推进了一层：如果能造 SoftwareFactory，就能造 SoftwareFactoryFactory，而且这种模式不只适用于软件，也可以泛化到公共卫生、法律等领域中被设计和实现的意图。

Meta Senior Director of AI Madhu Guru said the US AI community's rapid convergence around open-weight models came from repeated contact with reality, not abstract certainty. He named DeepSeek, the Microsoft-OpenAI breakup, GLM, Kimi, Fable, and the OpenAI-Hugging Face episode as public experiments that exposed different incentives, geopolitical pressures, and first- and second-order effects. Box CEO Aaron Levie called Google's support a complete industry endorsement of open-weight AI, while OpenClaw's Peter Steinberger welcomed competition and OpenAI's signature but noted Anthropic's silence.

Meta Senior Director of AI Madhu Guru 表示，美国 AI 社区之所以迅速形成对 open-weight 模型的支持，并非源于抽象推演，而是因为不断与现实接触。他列举了 DeepSeek、Microsoft-OpenAI 分裂、GLM、Kimi、Fable 以及 OpenAI-Hugging Face 事件，认为这些公开实验分别揭示了不同的激励机制、地缘政治压力，以及一阶和二阶影响。Box CEO Aaron Levie 称 Google 的加入意味着行业对 open-weight AI 的完整背书；OpenClaw 的 Peter Steinberger 则欢迎竞争和 OpenAI 的签名，同时指出 Anthropic 仍未表态。

OpenClaw's Peter Steinberger described running Codex all day for massive parallel QA ahead of a release. His instruction called for 12 subagents, multiple development gateways, worktrees, autonomous pull requests, stress testing, a target of 200 bugs, root-cause fixes, and a continuously updated Markdown report. He said Sol had become much better at understanding intent and finding complex behavioral issues, while avoiding the compaction-boundary failures and shortcuts that had broken this workflow before.

OpenClaw 的 Peter Steinberger 描述了自己如何在发布前让 Codex 全天运行大规模并行 QA。他的指令要求调用 12 个 subagents、启动多个开发 gateway、使用 worktree、自动创建 pull request、执行压力测试、以发现 200 个 bug 为目标、坚持修复根因，并持续更新一份 Markdown 测试报告。他表示，Sol 如今更能准确理解意图并发现复杂的行为问题，同时避免了过去会让这套流程在 compaction 边界失效或开始走捷径的问题。

Replit CEO Amjad Masad deployed a chess system approaching an estimated 1200 Elo and set a goal of exceeding 2000. The constraint is the interesting part: one small fine-tuned LLM must produce every move without assistance from a chess engine, ruling out custom pretraining and specialized architectures.

Replit CEO Amjad Masad 部署了一套国际象棋系统，估算 Elo 已接近 1200，目标是突破 2000。真正有意思的是约束条件：只允许一个经过 fine-tuning 的小型 LLM 独立生成每一步棋，不能借助 chess engine，也不能使用定制 pretraining 或特殊架构。

Anthropic traced a month of quality complaints to three product-layer changes, not its API or inference layer. Claude Code's default reasoning effort moved from high to medium on March 4 and was reverted April 7; a March 26 caching optimization mistakenly erased older reasoning on every subsequent turn after an idle session and was fixed April 10; and an April 16 system-prompt rule limiting text between tool calls to 25 words and final answers to 100 words caused a 3% eval drop before being reverted April 20. All fixes were in v2.1.116, and Anthropic reset subscriber usage limits. The company says it will put more staff on the exact public build, broaden per-model evals and prompt ablations, tighten review and audit tooling, and use soak periods and gradual rollouts for changes that might trade away intelligence.

Anthropic 将一个月来的质量投诉追溯到三个产品层变更，而非 API 或 inference layer。Claude Code 在 3 月 4 日把默认 reasoning effort 从 high 调到 medium，随后于 4 月 7 日撤回；3 月 26 日上线的一项缓存优化存在 bug，导致闲置会话恢复后，每一轮都会继续清除更早的 reasoning，直到 4 月 10 日才修复；4 月 16 日加入的 system prompt 规则把 tool call 之间的文字限制为 25 个词、最终回答限制为 100 个词，造成 eval 下滑 3%，并于 4 月 20 日撤回。所有修复均已包含在 v2.1.116 中，Anthropic 也重置了订阅用户的 usage limits。公司表示，未来会让更多员工使用完全一致的公开版本，扩大逐模型 eval 和 prompt ablation，收紧 review 与 audit 工具，并对可能牺牲智能的变更增加 soak period 和渐进式 rollout。

Anthropic's Managed Agents architecture separates Claude and its harness, the "brain," from sandboxes and tools, the "hands," and from an append-only session log. Both harnesses and containers can then be disposable: a failed harness can reload the event log and resume, while a failed sandbox becomes a tool error that Claude can retry against a newly provisioned container. Credentials stay outside generated-code sandboxes, either attached to resources or held in a vault behind an MCP proxy. The session log also preserves recoverable context beyond Claude's active context window. By provisioning sandboxes only when needed, Anthropic reports roughly 60% lower p50 time-to-first-token and more than 90% lower p95.

Anthropic 的 Managed Agents 架构把 Claude 及其 harness，也就是“大脑”，与 sandbox 和工具，也就是“双手”，以及 append-only session log 相互分离。这样一来，harness 和 container 都可以被随时替换：harness 故障后可重新加载 event log 并继续运行；sandbox 故障则会变成 Claude 可重试的 tool error，由新 provision 的 container 接手。凭证不会进入执行生成代码的 sandbox，而是绑定在资源上，或保存在 MCP proxy 背后的 vault 中。session log 还能在 Claude 当前 context window 之外保存可恢复的上下文。由于只在确有需要时才 provision sandbox，Anthropic 称 p50 time-to-first-token 大约降低 60%，p95 降幅超过 90%。

Claude Managed Agents added self-hosted sandboxes in public beta and MCP tunnels in research preview. Enterprises can keep code execution, files, packages, services, runtime sizing, network policy, and audit controls within infrastructure they control, while Anthropic continues to host orchestration, context management, and error recovery. Supported sandbox providers include Cloudflare, Daytona, Modal, and Vercel. MCP tunnels use a lightweight customer-deployed gateway with one outbound connection, no inbound firewall rules or public endpoint, and end-to-end encryption to reach private databases, APIs, knowledge bases, and ticketing systems. The tunnels work with Managed Agents and the Messages API.

Claude Managed Agents 推出了 public beta 阶段的 self-hosted sandbox，以及 research preview 阶段的 MCP tunnel。企业可以把代码执行、文件、package、service、runtime 资源配置、network policy 和 audit control 留在自己掌控的基础设施内，同时由 Anthropic 继续托管 orchestration、context management 和 error recovery。支持的 sandbox provider 包括 Cloudflare、Daytona、Modal 和 Vercel。MCP tunnel 通过客户部署的轻量 gateway 建立单一 outbound connection，无需 inbound firewall rule 或公开 endpoint，并以端到端加密访问私有数据库、API、知识库和 ticketing system；它同时适用于 Managed Agents 与 Messages API。

The Takeaway: Benedict Evans sees AI as a major platform shift, but argues that today's consequential questions are about product-market fit, value capture, and deployment, not unfalsifiable claims that models will inevitably automate everything.

核心结论：Benedict Evans 认为 AI 确实是一次重大的平台迁移，但眼下真正重要的问题是 product-market fit、价值捕获和落地部署，而不是模型终将自动化一切这种无法证伪的论断。

Evans compares foundation models with earlier infrastructure layers whose economics did not guarantee that they captured the value created above them. Mobile networks carried a 1,000-fold or 2,000-fold increase in data traffic, yet applications such as Uber and YouTube captured much of the upside; likewise, he notes that LLMs lack the network effects that made Windows defensible. Coding already has unquestionable product-market fit, but broader usage remains shallow and capabilities remain jagged: models can complete some long tasks while failing at tasks that take people five minutes. His sharpest warning is against treating jobs as measurable bundles of automatable tasks, because technology also changes the job and creates second-order effects outside the original analysis. As he puts it, “You can always see the jobs that will go away, and you don't know the new jobs.” Model companies still need to invent durable products and distribution while facing high marginal costs and a new leaderboard leader every four to six weeks.

Evans 把 foundation model 与过去的基础设施层相比较，指出底层的经济规模并不保证它能捕获上层创造的价值。移动网络承载的数据流量增长了 1,000 到 2,000 倍，但 Uber、YouTube 等应用拿走了大量上行价值；同样，LLM 也缺少让 Windows 建立护城河的 network effect。编程已经拥有无可争议的 product-market fit，但更广泛的使用仍然较浅，模型能力也依旧 jagged：它们可以完成某些耗时很长的任务，却可能在一个人五分钟就能做完的事情上失败。他最尖锐的提醒是，不要把职业当成一组可以精确测量和自动化的任务，因为技术也会改变职业本身，并在原始分析之外制造二阶影响。正如他所说：“你总能看到哪些工作会消失，却无法知道会出现哪些新工作。”与此同时，model company 仍需在高 marginal cost、每四到六周 leaderboard 榜首就可能易主的环境中，发明真正持久的产品与分发方式。