九月五日

Sam Altman said GPT-6 Astra is available in Work/Codex to Pro, Enterprise, and Business Premium users and in the API, followed shortly by access for all Plus and Business users. OpenAI's Thibault Sottiaux added that Plus, Pro, and Business users would receive a full banked reset, and said Astra arrived ahead of schedule. More consequentially, he said internal productivity rose enough while OpenAI had exclusive access that the team moved some plans forward by six months for DevDay instead of shipping them in the middle of next year. Replit CEO Amjad Masad also confirmed Astra is available on Replit, extending the launch into another builder platform.

Sam Altman 表示，GPT-6 Astra 已在 Work/Codex 中向 Pro、Enterprise 和 Business Premium 用户开放，同时进入 API，随后又覆盖全部 Plus 和 Business 用户。OpenAI 的 Thibault Sottiaux 补充说，Plus、Pro 和 Business 用户将获得一次完整的 banked reset，并表示 Astra 比原计划更早上线。更重要的是，他透露 OpenAI 独占使用 Astra 期间，内部生产力提升到足以把部分计划提前六个月，改在 DevDay 发布，而不是等到明年年中。Replit CEO Amjad Masad 也确认 Astra 已登陆 Replit，把这次发布延伸到另一个 builder platform。

Vercel CEO Guillermo Rauch argued that agents must operate through today's web rather than wait for a replacement, and that WebMCP can make this more efficient by letting pages expose their own agent tools. His concrete example is a Next.js development page giving the agent tab-specific debugging tools and context directly, avoiding a separate MCP server and the loss of detail that comes from searching server logs. He paired that near-term architecture with a much more ambitious prediction: AI software factories will eventually produce bug-free, self-improving software.

Vercel CEO Guillermo Rauch 认为，agents 必须在今天已经存在的 web 上运行，而不是等待一套替代基础设施；WebMCP 可以让页面直接暴露自己的 agent tools，从而提高效率。他给出的具体例子是：Next.js 开发页面可以直接向 agent 提供当前 tab 专属的 debugging tools 和 context，不需要另外寻找和配置 MCP server，也不会因为翻查 server logs 而损失细节。他还把这条近期的架构路线与一个更大胆的判断联系起来：AI software factories 最终会产出无 bug、能够自我改进的软件。

Meta Senior Director of AI Madhu Guru proposed a practical curriculum for aspiring AI product builders: choose a workflow you understand, automate it end to end, and keep expanding the ambition. The exercise forces four product decisions into the open: what a great complete experience looks like, how MCPs and tools should be used, where humans must remain in the loop, and how the result should be evaluated. His point is that one real implementation teaches more than a month of reading because it exposes the gaps between a model demo and a working system.

Meta AI Senior Director Madhu Guru 给想提升 AI 产品能力的人提出了一套实践课程：选择一个自己熟悉的 workflow，用 AI 把它端到端自动化，然后不断提高目标。这个练习会迫使 builder 正面回答四个产品问题：完整而优秀的体验应该是什么样，MCPs 和 tools 应该如何使用，哪些环节必须保留 human in the loop，以及如何评价最终结果。他的核心观点是，一次真实实现比读一个月资料学得更多，因为它会暴露 model demo 与可用系统之间的全部缺口。

Y Combinator CEO Garry Tan said he moved GStack's preferred remote-session browser to Aside after comparing it with other agent tools. He described Aside as more than browser access: it combines credential handling, integrations, a harness, memory, and access-control defaults. In his OpenClaw-to-Slack example, a setup that had taken two hours was completed through Aside in under three minutes. The broader product signal is that agent browsers are becoming governed execution environments, not just ways to render and click web pages.

Y Combinator CEO Garry Tan 表示，在与其他 agent tools 比较后，他已把 GStack 首选的 remote-session browser 改为 Aside。他认为 Aside 不只是提供浏览器访问，还把 credential handling、integrations、harness、memory 和默认 access controls 组合到一起。在他连接 OpenClaw 与 Slack 的例子中，原本花了两小时的配置通过 Aside 在不到三分钟内完成。更广泛的产品信号是，agent browsers 正在成为受治理的执行环境，而不再只是渲染和点击网页的工具。

Claude in Chrome is now generally available on every paid Claude plan, and it can autonomously read, type, click, navigate, and fill forms using a user's existing browser session. Anthropic says a classifier checks each proposed action against the user's request before execution, while probes scan tool results for hidden prompt injections. On its stronger red-team evaluation, attacks that reached the model succeeded 17.6% of the time against Opus 4.5 and 3.8% against Opus 5 before extra safeguards; with probes and the action classifier, none succeeded against Sonnet 5 or Opus 5, while Fable 5 had a 0.3% success rate limited to manually verified low-severity cases. The practical constraint remains important: Anthropic says “prompt injection remains a moving target,” so Enterprise admins can restrict approved domains and users can disable automatic approval.

Claude in Chrome 现已向所有 Claude 付费 plan 全面开放，并可利用用户现有的浏览器 session，自主完成读取、输入、点击、导航和填写表单。Anthropic 表示，每次执行前都会由 classifier 对照用户原始请求检查拟议 action，同时 probes 会扫描 tool results 中隐藏的 prompt injection。在更强的 red-team evaluation 中，额外 safeguards 启用前，成功抵达 model 的攻击对 Opus 4.5 和 Opus 5 的成功率分别为 17.6% 和 3.8%；加入 probes 与 action classifier 后，针对 Sonnet 5 和 Opus 5 的攻击均未成功，Fable 5 的成功率为 0.3%，且人工确认都属于低严重度场景。实际限制仍然不能忽略：Anthropic 明确表示“prompt injection 仍是不断变化的目标”，因此 Enterprise admins 可以限制 approved domains，用户也可以关闭 automatic approval。

Claude Cowork now includes a separate built-in browser in the desktop app, letting Claude handle web tasks without an extension or access to a user's normal tabs, bookmarks, and passwords. Anthropic positions it differently from Claude in Chrome: the built-in browser is for delegated work in an isolated environment, such as gathering research or collecting invoices, while the Chrome extension remains suited to acting on a page and signed-in account the user already has open. Logins can be imported site by site, with banking, email, and single sign-on excluded unless explicitly selected. The browser is rolling out to Pro, Max, and Team on macOS, Windows, and Linux beta, is available to Enterprise admins now, and uses the same action checks as Claude in Chrome, though Anthropic cautions that the protections “can't eliminate” prompt-injection risk.

Claude Cowork 现在在 desktop app 中内置了一个独立浏览器，让 Claude 无需 extension，也无需访问用户日常浏览器里的 tabs、bookmarks 和 passwords，就能处理 web tasks。Anthropic 对两种模式的定位很清楚：built-in browser 适合在隔离环境中委派工作，例如收集研究资料或整理发票；Claude in Chrome 则更适合操作用户已经打开、并已登录 account 的页面。用户可以按 site 导入 login，banking、email 和 single sign-on 默认不会被带入，除非明确选择。该浏览器正在向 macOS、Windows 和 Linux beta 上的 Pro、Max 与 Team 推送，Enterprise admins 已可启用；它采用与 Claude in Chrome 相同的 action checks，但 Anthropic 也提醒，这些保护“无法消除” prompt injection 风险。

The takeaway: Arm CEO Rene Haas sees AI as both a force multiplier for chip engineering and a source of sustained demand for CPUs, memory, fabrication capacity, and power-efficient edge compute.

核心结论：Arm CEO Rene Haas 认为，AI 既是芯片工程的生产力倍增器，也会持续推高对 CPUs、memory、fabrication capacity 和高能效 edge compute 的需求。

Arm still primarily licenses CPU and system IP, but it has moved from components to compute subsystems and now to a physical ARM AGI CPU, initially developed after Meta asked for a general-purpose agentic CPU that existing suppliers did not offer. Haas said 80% to 90% of Arm engineers use AI daily. The greatest gains are in verification, validation, debugging, and documentation, which dominate a chip cycle that can last 24 to 36 months. As he put it, “The largest amount of time is in the verification, the validation, the debug.” RTL generation and best-in-class physical design remain harder because much of the necessary training material is proprietary, although Arm's documented IP and test benches give it an advantage. For straightforward designs, Haas thinks an AI-assisted path from idea to a fab-ready GDS II file could become possible more than five years from now.

Arm 的主要业务仍是授权 CPU 和 system IP，但它已经从单独组件扩展到 compute subsystems，并进一步进入实体 ARM AGI CPU 产品；最初的契机是 Meta 想要一款现有供应商无法提供的通用 agentic CPU。Haas 表示，Arm 有 80% 到 90% 的工程师每天使用 AI。最大的收益来自 verification、validation、debugging 和 documentation，而这些环节占据了长达 24 到 36 个月芯片周期中的大部分时间。正如他所说：“花费时间最多的是 verification、validation 和 debug。” RTL generation 和顶级 physical design 仍然更难，因为所需训练材料大多是 proprietary；不过 Arm 完整记录的 IP 与 test benches 构成了一项优势。对于较直接的设计，Haas 认为五年以后，AI 有可能把从想法到可交付 fab 的 GDS II file 串成一条完整路径。

He expects the supply environment to remain constrained for at least three to five years because transformer workloads are both compute- and memory-intensive. Data-center construction may become one bottleneck, while wafers and memory become others; in his view, the industry is “not even close” to oversupplying underlying demand. That makes supply-chain expertise, strategic partnerships, and access to capital as important to semiconductor startups as inventive architectures. It also explains Arm's move into a physical product, which requires relationships with foundries, memory vendors, substrate suppliers, and new operational capabilities beyond its historically asset-light licensing model.

他预计 supply environment 至少未来三到五年仍会受限，因为 transformer workloads 同时高度消耗 compute 与 memory。Data center 建设可能成为一处瓶颈，wafers 和 memory 则会成为另外几处；在他看来，行业距离满足底层需求还“远远不够”。因此，对 semiconductor startups 而言，supply-chain expertise、strategic partnerships 与 capital access 和创新架构同样重要。这也解释了 Arm 为什么进入实体产品：相较于过去轻资产的授权模式，实体芯片要求公司建立与 foundries、memory vendors、substrate suppliers 的关系，并补齐新的运营能力。

Haas rejected the idea that accelerators make CPUs secondary. Accelerators may generate tokens, but CPUs still orchestrate where those tokens go, and the same system logic applies across data centers, vehicles, robots, phones, and wearables. He expects both humanoid and task-specific robots, with Arm technology spanning real-time sensing at the extremities and the processors already used in many robotic “brains.” Smaller edge devices may be an especially strong market because local AI must fit within power limits that rule out placing a 50-watt GPU on a person's head.

Haas 不认同 accelerators 会让 CPUs 变得次要。Accelerators 可以生成 tokens，但 CPUs 仍负责协调这些 tokens 去往哪里；同样的 system logic 也适用于 data centers、vehicles、robots、phones 和 wearables。他预计 humanoid 与 task-specific robots 会同时发展，Arm 技术既能覆盖末端的 real-time sensing，也已经用于许多机器人“大脑”中的处理器。更小的 edge devices 可能尤其适合 Arm，因为 local AI 必须满足严格的功耗限制，人们不可能把一块 50-watt GPU 戴在头上。