AI Builders Digest — 2026年7月1日
📌 X/TWITTER
Claude (Anthropic) Claude in Microsoft Foundry is now generally available, hosted on Azure — customers get Claude Opus 4.8 and Claude Haiku 4.5 with Azure authentication, billing, and commitment retirement. Inference runs on Azure infrastructure operated by Anthropic, with prompt caching and extended thinking supported today. 🔗 https://x.com/claudeai/status/2071653958905467027
Claude 在 Microsoft Foundry 正式上线，托管于 Azure 平台——客户可直接使用 Azure 认证、计费和承诺消费获取 Claude Opus 4.8 和 Claude Haiku 4.5。推理运行在由 Anthropic 运营的 Azure 基础设施上，prompt caching 和 extended thinking 已支持。 🔗 https://x.com/claudeai/status/2071653958905467027
Anthropic Claude Code 工程师 Boris Cherny In the next version of Claude Code, subagents will run in the background by default — you can keep talking to Claude while your subagents work. If you want foreground execution, just ask. 🔗 https://x.com/bcherny/status/2071647677591466098
下一版 Claude Code 中，子 agent 将默认在后台运行——你可以在子 agent 工作的同时继续和 Claude 对话。如果需要前台运行，直接告诉 Claude 就行。 🔗 https://x.com/bcherny/status/2071647677591466098
OpenAI Codex 与 ChatGPT 工程师 Thibault Sottiaux
Codex shipped a replacement to coarse sandbox modes: reusable, inheritable permission profiles binding OS-enforced file read/write/deny rules (even **/*.env) to per-domain network + Unix sockets, plus fail-closed admin allowlists. Least privilege per task. Also: Codex usage limits have been fully reset and an extra reset credited after investigating reports of faster-than-expected consumption — the root cause was a combination of over-proactive auto-review, duplicate background suggestions, and aggressive retry behavior, all now fixed.
🔗 https://x.com/thsottiaux/status/2071636285807059315
🔗 https://x.com/thsottiaux/status/2071740419030053227
Codex 发布了粗粒度 sandbox 模式的替代方案：可复用、可继承的权限配置文件，将操作系统级的文件读写/拒绝规则（包括 **/*.env）绑定到按域网络和 Unix socket，外加 fail-closed 管理员白名单，实现每个任务的最小权限原则。同时，针对用户反馈 Codex 用量消耗过快，团队已全面重置额度并额外赠送一次重置——根因是自动审查过于激进、后台建议重复生成以及重试行为过于频繁，均已修复。
🔗 https://x.com/thsottiaux/status/2071636285807059315
🔗 https://x.com/thsottiaux/status/2071740419030053227
Box CEO Aaron Levie On the central AI debate: if a closed stack remains perpetually at the frontier by a wide margin, vertical integration and gatekeeping work — you control the best tech and demand favors you. But if open weights AI stays a close second, the equation reverses. The vast majority of tokens will flow to an alternative stack — the model plus the underlying hardware — controlled and monetized by someone else. Your regulatory philosophy depends on your belief about how close to the frontier open weights can stay. 🔗 https://x.com/levie/status/2071775583072375214
关于 AI 领域最核心的辩论：如果闭源技术栈始终大幅领先前沿，那么垂直整合和管制策略就是可行的——你掌控着最好的技术，需求自然会向你倾斜。但如果开源权重 AI 能紧追前沿水平，等式就会逆转：绝大多数 token 将流向另一个技术栈——模型加上底层硬件——由其他人控制和变现。你对 AI 管制的立场，取决于你相信开源权重能离前沿多近。 🔗 https://x.com/levie/status/2071775583072375214
Vercel CEO Guillermo Rauch Vercel now supports 20x larger functions. “You can deploy anything and everything to Vercel. More tomorrow.” 🔗 https://x.com/rauchg/status/2071716510389662153
Vercel 现已支持 20 倍更大的函数。「你可以把任何东西部署到 Vercel。明天还有更多。」 🔗 https://x.com/rauchg/status/2071716510389662153
前 Google Gemini/Veo 产品负责人 Madhu Guru The rise of strong open-weight models like GLM will actually strengthen Google’s position: more companies experimenting with fine-tuning open-weight models means value accrues to the infrastructure layer. Enterprises want the flexibility to run anywhere. 🔗 https://x.com/realmadhuguru/status/2071637885154148785
GLM 等强开源权重模型的崛起，实际上会巩固 Google 的地位：更多公司尝试微调开源权重模型，意味着价值将流向基础设施层。企业需要随处运行的灵活性。 🔗 https://x.com/realmadhuguru/status/2071637885154148785
Peter Yang（AI 教程作者） For writing and editing, plain vanilla Claude web is still the best — better than Codex and Claude Code. His theory: the coding agent’s system prompts make them crappier writers. 🔗 https://x.com/petergyang/status/2071731343390851519
在写作和编辑方面，普通版 Claude 网页界面仍然是最好的——优于 Codex 和 Claude Code。他的推测：coding agent 的系统提示词让它们在写作上变差了。 🔗 https://x.com/petergyang/status/2071731343390851519
Zara Zhang（建造者） “The market value of writing has gone way up” per a PM at Anthropic — good writing is required both for building products (steering models effectively) and building an audience. Also built a Chrome extension that turns your “read later” list into dedicated reading time on your calendar. 🔗 https://x.com/zarazhangrui/status/2071670108033073365 🔗 https://x.com/zarazhangrui/status/2071766827345285411
据 Anthropic 的一位 PM 所说，「写作的市场价值大幅上升」——优秀的写作对于构建产品（有效引导模型）和建立受众都必不可少。她还做了一个 Chrome 扩展，把你的「稍后阅读」列表变成日历上的专属阅读时间。 🔗 https://x.com/zarazhangrui/status/2071670108033073365 🔗 https://x.com/zarazhangrui/status/2071766827345285411
FirstMark VC Matt Turck “Being the underdog, time and again, is the greatest advantage.” 🔗 https://x.com/mattturck/status/2071806129001164934
「一次次地做弱者，是最大的优势。」 🔗 https://x.com/mattturck/status/2071806129001164934
YC CEO Garry Tan “Build power and datacenters.” 🔗 https://x.com/garrytan/status/2071600933210100074
「建造电力和数据中心。」 🔗 https://x.com/garrytan/status/2071600933210100074
📝 官方博客
Anthropic Engineering: An update on recent Claude Code quality reports Anthropic investigated user reports that Claude’s responses had degraded. Three separate issues were identified, all affecting Claude Code, the Agent SDK, and Claude Cowork — not the API. On March 4, default reasoning effort was lowered from high to medium to reduce latency, which was the wrong tradeoff (reverted April 7). On March 26, a bug in the idle-session clearing logic caused Claude to seem forgetful and repetitive (fixed April 10). A broader degradation was caused by an earlier model tuning change that penalized long-form responses, making Claude overly concise — fixed by reverting the data mix change and adding long-form quality as an explicit eval target. All issues resolved as of April 20 (v2.1.116). 🔗 https://www.anthropic.com/engineering/april-23-postmortem
Anthropic 调查了用户关于 Claude 回复质量下降的反馈，发现三个独立问题，均影响 Claude Code、Agent SDK 和 Claude Cowork，API 未受影响。3 月 4 日，默认推理力度从 high 降到 medium 以减少延迟——这是错误的取舍（4 月 7 日回退）。3 月 26 日，空闲会话清理逻辑的一个 bug 导致 Claude 显得健忘和重复（4 月 10 日修复）。更广泛的质量下降源于早期模型微调中惩罚长回复的调整，导致 Claude 过于简洁——通过回退数据混合变更并将长文质量设为显式评估目标来修复。所有问题已于 4 月 20 日（v2.1.116）解决。 🔗 https://www.anthropic.com/engineering/april-23-postmortem
Anthropic Engineering: Scaling Managed Agents — Decoupling the brain from the hands Managed Agents is a hosted service in the Claude Platform that runs long-horizon agents through stable abstractions designed to outlast any particular implementation. The key insight: harnesses encode assumptions about what Claude can’t do on its own, but those assumptions go stale as models improve. For example, Claude Sonnet 4.5 exhibited “context anxiety” (wrapping up tasks prematurely), but Claude Opus 4.5 did not — the context-reset harness became dead weight. Managed Agents virtualizes agent execution the way operating systems virtualized hardware: abstractions like process and file outlasted the hardware they ran on. The goal is to build for “programs as yet unthought of.” 🔗 https://www.anthropic.com/engineering/managed-agents
Managed Agents 是 Claude 平台上的托管服务，通过稳定的抽象层运行长周期 agent，这些抽象设计为超越任何具体实现而存在。核心洞见：harness 编码了关于「Claude 自己做不到什么」的假设，但随着模型进步，这些假设会过时。例如，Claude Sonnet 4.5 表现出「上下文焦虑」（过早结束任务），但 Claude Opus 4.5 没有——上下文重置 haraness 变成了死代码。Managed Agents 对 agent 执行进行虚拟化，就像操作系统将硬件虚拟化一样：process 和 file 等抽象比它们运行的硬件更长寿。目标是构建一个能容纳「尚未被构想的程序」的系统。 🔗 https://www.anthropic.com/engineering/managed-agents
Claude Blog: New in Claude Managed Agents — self-hosted sandboxes and MCP tunnels Managed Agents can now operate in sandboxes you control and connect to your private MCP servers. Tool execution moves to your own infrastructure (or managed providers like Cloudflare, Daytona, Modal, Vercel), while the agent orchestration loop stays on Anthropic’s infrastructure. Network policies, audit logging, and security tooling remain within your perimeter. MCP tunnels are in research preview. Self-hosted sandboxes are in public beta. 🔗 https://claude.com/blog/claude-managed-agents-updates
Managed Agents 现在可以在你控制的 sandbox 中运行，并连接到你私有的 MCP 服务器。工具执行移至你自己的基础设施（或 Cloudflare、Daytona、Modal、Vercel 等托管提供商），而 agent 编排循环保持在 Anthropic 的基础设施上。网络策略、审计日志和安全工具均保留在你的边界内。MCP 隧道处于 research preview 阶段，自托管 sandbox 处于公开 beta。 🔗 https://claude.com/blog/claude-managed-agents-updates
🎙️ 播客
No Priors: Re-engineering the Semiconductor Supply Chain with Intel CEO Lip-Bu Tan
The Takeaway: Intel CEO Lip-Bu Tan is betting that the AI era will make CPUs cool again — and that America can and must build its own advanced chip manufacturing.
Lip-Bu Tan took the hardest job in the semiconductor industry at age 66. Fourteen months into his tenure as Intel CEO, the company has returned 6x to shareholders — but he’s aiming for 10x, venture-style. His playbook: crawl, walk, run. First, strengthen the balance sheet (US government as shareholder, Jensen Huang investing $5 billion, SoftBank’s Masayoshi Son lending a hand). Second, focus on products and simplify the line. Third, build toward a full-stack future where Intel provides not just silicon but the entire rack — CPU, GPU, advanced packaging, and software.
On the foundry business: it’s a trust business. “People want to trust you before they give you the wafer to count on you.” He’s focusing on yield, defect density, and cycle time. The goal by 2030-2032 is to be a credible alternative to TSMC. On TeraFab, the collaboration with Elon Musk: “We share the same view that semiconductor infrastructure is not catching up with AI growth.” Musk questions every traditional assumption — “it’s very refreshing.”
On the macro picture: power constraints, helium supply, and memory shortages are the bottlenecks. On AI’s impact: “The company most impacted is the one not embracing AI.” On what winning looks like in 10 years: laser focus on one niche, find the right partners, build a full-stack platform — like NVIDIA did with CUDA.
Most memorable quote: “I always look for 10x. Being a venture at heart, you want to look for 10x.”
🔗 https://www.youtube.com/watch?v=asCgCv2XB4s
核心观点： Intel CEO 陈立武押注 AI 时代将让 CPU 重新变得重要——而且美国可以且必须建立自己的先进芯片制造能力。
陈立武在 66 岁接手了半导体行业最艰难的工作。担任 Intel CEO 十四个月以来，公司已为股东带来 6 倍回报——但他的目标是 10 倍，以风投的思维方式。他的打法：爬、走、跑。第一步，强化资产负债表（美国政府成为股东，黄仁勋投资 50 亿美元，孙正义伸出援手）。第二步，聚焦产品、简化产品线。第三步，构建全栈未来——Intel 提供的不仅是硅片，而是整个机架：CPU、GPU、先进封装和软件。
关于晶圆代工业务：这是个信任生意。「人们必须先信任你，才会把晶圆交给你。」他正聚焦于良率、缺陷密度和周期时间。目标是到 2030-2032 年成为台积电的可信替代选择。关于与 Elon Musk 合作的 TeraFab：「我们都认为半导体基础设施跟不上 AI 的增长。」Musk 质疑每一个传统假设——「这非常令人耳目一新。」
宏观格局上，电力约束、氦气供应和存储芯片短缺是瓶颈。关于 AI 的影响：「受影响最大的公司是那些不拥抱 AI 的公司。」关于十年后的赢家：聚焦一个细分领域、找到合适的合作伙伴、构建全栈平台——就像 NVIDIA 凭借 CUDA 做到的那样。
最令人印象深刻的原话：「我始终追求 10 倍回报。作为风投出身的人，你的目标是 10 倍。」
🔗 https://www.youtube.com/watch?v=asCgCv2XB4s
通过 Follow Builders 生成: https://github.com/zarazhangrui/follow-builders