AI Builders
Digest

Anthropic 正式推出 Claude 统一记忆功能。现在 Chat 和 Claude Cowork 共享同一套记忆，你可以一次性告诉 Claude 记住某事，它就能在不同界面中调用这些上下文。记忆以主题列表形式保存在 Settings 中，可随时查看、编辑或删除；敏感话题（健康、宗教信仰等）默认不进入记忆，需手动开启。记忆默认在 Free、Pro、Max 计划开启，并会随对话自动更新，也可主动说“remember this”来保存。Boris Cherny 和 Cat Wu 均确认记忆变得更简单且更强大，跨表面统一后体验显著提升。

Claude official and team posts confirm the launch of unified memory. Claude now maintains one memory across Chat and Cowork, so context from past chats is available when handing tasks to Cowork. Users control what is stored via Settings; sensitive topics stay out unless enabled. Memory is on by default for Free/Pro/Max and auto-updates, with manual “remember this” support.

OpenAI CEO Sam Altman 发布简洁声明：“we made a chip and it is fast”。这条推文迅速引发大量关注，暗示 OpenAI 在自研硬件上取得了实质性进展，可能进一步优化推理效率与成本结构。

OpenAI CEO Sam Altman posted a terse announcement: “we made a chip and it is fast.” The statement drew massive engagement and points to tangible progress on OpenAI’s custom silicon efforts, potentially improving inference efficiency and cost.

The Takeaway：人类点击数据是 bug，Agent 搜索必须依赖 Agent 反馈而非人类反馈；Parallel 正在为 AI Agent 重建搜索基础设施与激励相容的商业模式。

前 Twitter CEO Parag Agrawal 创立 Parallel Web Systems，目标是让 Agent 像人类使用浏览器和搜索引擎一样高效使用网络，而且用量会是人类的千倍。他指出传统搜索为懒惰人类设计（两个拼错的词加自动补全），而 Agent 会发出更长、更精确的查询，需要完全不同的索引、排序与接口。

Parallel 起步时先做搜索 Agent（按需爬取），用实时计算弥补索引不足，先服务保险核保、销售 enrichment、金融数据收集等替代人工的场景，再逐步扩展索引。他们认为“human click data is a bug”，模型擅长压缩信息，可把大模型研究直接用于索引与排序。最近已将延迟从数秒压到 200ms，并与 Google Cloud 合作，成为 GCP 企业 Agent API 的搜索与 grounding 选项之一。

在商业模式上，Agrawal 认为广告支撑的互联网正在因 Agent 替代人类注意力而破裂。内容方需要可扩展的、激励相容的变现方式。Parallel 采用 Shapley 值（博弈论中的贡献归因）来评估每条内容对 Agent 输出的边际价值，实现按内容独特性与工作价值差异化定价。长期愿景是从“拉取”转向“推送”——网络主动通知 Agent 可行动的变化，形成真正的并行 Agent 网络。

The Takeaway: Human click data is a bug—agentic search must rely on agent feedback, not human clicks. Parallel is rebuilding both the infrastructure and the business models for a parallel web designed for AI agents.

Former Twitter CEO Parag Agrawal founded Parallel Web Systems to let agents use the web as fluidly as humans once used browsers and search engines—only a thousand times more. Traditional search was built for lazy humans typing two misspelled words; agents issue longer, better-specified queries and need different indexes, ranking, and interfaces. Parallel began with on-demand search agents that competed with outsourced human research, then grew the index incrementally. “Our view at Parallel is that human click data is a bug, and agent doing work with search should rely on agent feedback, not human feedback.” They have cut latency to 200 ms and partnered with Google Cloud so Gemini and other models on GCP can use Parallel for grounding.

On economics, the ad-supported web breaks when agents replace human eyeballs. Parallel applies Shapley values to attribute how much each source improves an agent’s output, enabling differential pricing by content uniqueness and the value of the work performed. The end state is a shift from pull to push: the web notifies agents when something actionable changes.

OpenAI Codex 与 ChatGPT 负责人 Thibault Sottiaux 宣布新团队计划上线。该计划类似 Pro $100 方案，但专为团队与小公司设计：包含全部 ChatGPT、ChatGPT Work 与 Codex 功能，可连接 Google Workspace、Slack、GitHub、Microsoft 365 等，支持 SAML/SSO/MFA 安全工作区、集中计费与管理、用量分析与支出控制，且无 5 小时限制。他同时表示“Good products take time. At least 34 days.”

Thibault Sottiaux (Codex & ChatGPT at OpenAI) announced a new team plan that mirrors the Pro $100 tier but is built for teams and small companies. It includes all ChatGPT, ChatGPT Work and Codex features, connectors to Google Workspace, Slack, GitHub and Microsoft 365, SAML/SSO/MFA, centralized billing, usage analytics and spend controls, with no 5-hour limits. He also noted “Good products take time. At least 34 days.”

Vercel CEO Guillermo Rauch 宣布两项重要更新。Run SDK 提供安全的轻量级 eval 环境，让 Agent 生成的代码可在 QuickJS 安全上下文中执行，无需完整沙箱，更快更省成本（npm i run）。Vercel Connect 正式 GA，可一键创建到 Notion 等服务的 MCP 客户端，以认证用户身份查询数据，解决 Agent 安全连接服务与数据的核心难题。

Vercel CEO Guillermo Rauch introduced Run SDK for secure, lightweight evaluation of dynamic Code Mode execution. Agents can run code inside a QuickJS secure context instead of a full sandbox—faster and cheaper (npm i run). Vercel Connect is now generally available: run “vercel connect create notion” to obtain an MCP client that queries on behalf of the authenticated user, addressing the hardest problem in agent building—secure connectivity to services and data.

Google Labs 发布新实验 Putty，一款实时协作的 vibe coding 工具，允许多人一起构建工具与网站。目前仅限美国、18 岁以上用户，开放 waitlist。

Google Labs launched Putty, a collaborative vibe-coding experiment that lets multiple people build tools and websites together in real time. Waitlist is open (US only, ages 18+).

Peter Yang 开源了 /fuck-cancer AI 技能，帮助患者与照护者整理信息、准备对话并自我倡导。技能自动生成并更新一份实用简报，包含五部分：患者与医护团队信息、下一步仅限三项具体行动、已确认事实与未明事项、医学术语通俗解释、护理日志。它从用户提供的文件与上下文构建，需要时可调用 National Cancer Institute 与 ClinicalTrials.gov。支持本地 Markdown 或可分享的 Google Doc，可与 ChatGPT/Codex 或 Claude Code 配合使用。

Peter Yang open-sourced /fuck-cancer, an AI skill that helps patients and caregivers navigate diagnosis and treatment. It builds and maintains a practical brief with five sections: patient/care-team info, next three concrete actions, confirmed vs. unclear facts, plain-English medical terms, and a care log. It draws from user documents and trusted sources (NCI, ClinicalTrials.gov) and works with ChatGPT/Codex or Claude Code, saving to local Markdown or a shared Google Doc.

Meta AI 高级总监 Madhu Guru 继续其 Evals 系列，第 9 篇聚焦“Eval Roadmap Problem”。多数 evals 失败是因为团队把它们当作静态产物，而用户期望与行为早已演进。以金融研究 Agent 为例：从总结 5 页财报，到复盘 5 份报告讲增长故事，再到 15 份文件构建投资论点，最终变成监控组合并主动预警。Evals 必须同步演进：短上下文→长上下文、单轮→多轮、简单 QA→复杂综合、反应式聊天→主动 Agent。实用方法：映射演进维度、优先关键用例、挖掘生产轨迹、为下一阶段建 P0 evals，然后 hill-climb。

Madhu Guru (Sr Director, AI at Meta) published part 9 of his eval series on the “Eval Roadmap Problem.” Most evals fail because teams treat them as static while user expectations evolve. A financial-research agent may start summarizing a 5-page earnings report and later need to synthesize 15 filings into an investment thesis or proactively monitor a portfolio. Evals must track the same dimensions: short-to-long context, single-to-multi-turn, simple QA to complex synthesis, reactive to proactive. Map the dimensions, prioritize, mine production traces, build P0 evals for the next stage, and hill-climb.

Swyx（smol.ai / Cognition 等）发出 PSA：当前不要使用 Codex 的 “locked use” 能力。该功能依赖不稳定的 macOS 特性，本周已两次完全锁死他的 macOS 钥匙串。Apple 开发者论坛已承认这是已知 bug。他建议在云端能力成熟前尽量避免。

Swyx (smol.ai, Cognition, etc.) issued a PSA: do not use Codex “locked use” capabilities right now. It relies on unstable macOS features and has completely locked him out of his macOS keychain twice this week. Apple developer forums acknowledge it as a known bug. Prefer cloud alternatives until the local path stabilizes.

Claude Code 团队成员 Thariq 表示，很快会分享更多让 Claude Code 更可 hack 的进展，呼应社区对可扩展性与自定义能力的强烈需求。

Thariq (Claude Code at Anthropic) teased upcoming work to make Claude Code more hackable, responding to strong community demand for extensibility.

Box CEO Aaron Levie 转发并点评一篇关于规模化应用 AI 策略的文章。他指出模型与企业真实工作流之间存在巨大鸿沟，这正是应用 AI 公司的机会。“世界不只想要原始模型和 Agent，它想要问题被解决、结果被达成。”成功需要理解上下文、推动变革管理、路由到合适模型、连接关键业务系统、解决 UX、掌握领域 evals 等。这些价值远超模型智能本身，当前窗口期巨大。

Box CEO Aaron Levie highlighted a post on applied AI strategy at scale. There is a wide gap between raw models/agents and the actual workflows of an enterprise. “The world doesn’t just want raw models and agents; it wants problems resolved and outcomes achieved.” Capturing that premium requires deep context, change management, model routing, system connectors, UX design, domain evals and more—creating a large window for applied AI companies.

SPC 普通合伙人、前 Dropbox CTO Aditya Agarwal 指出，普通民众讨厌数据中心建设并不令人意外，因为当前 AI 主要惠及知识工作者与高收入群体。真正的转折将出现在 AI 能治愈普罗大众的疾病之时。行业此前过多渲染恐惧而非正面愿景，也没有帮到自己。

Aditya Agarwal (GP at SPC, former Dropbox CTO) noted that public hostility to datacenter buildout is unsurprising: AI today mainly helps knowledge workers and high earners. The big shift will come when AI delivers cures for diseases that affect everyone. The industry has also hurt itself with excessive fear-mongering instead of painting a positive future.

Y Combinator 总裁兼 CEO Garry Tan 强调“You need to be cross-harness memory maxxing”，呼应当前 Agent 记忆统一与跨工具上下文共享的趋势。

Y Combinator President & CEO Garry Tan stressed “You need to be cross-harness memory maxxing,” echoing the broader move toward unified, portable agent memory across tools.