June 24, 2026

Anthropic launched Claude Tag in beta on Slack for Claude Enterprise and Team plans. The idea: inside a channel there's a single shared Claude that interacts with everyone, so a teammate can pick up exactly where you left off, and it keeps building context about the work as it follows the channel instead of needing to be re-briefed from scratch. Turn on "ambient behavior" and it stops waiting to be asked — it follows up on threads that have gone quiet and flags what's relevant from across its channels and tools.

Anthropic 在 Slack 上推出了 Claude Tag 的 beta 版，面向 Claude Enterprise 和 Team 套餐。核心思路是：在一个频道里只有一个共享的 Claude，它和所有人交互，这样队友可以接着你停下的地方继续往下做；而且它会随着持续跟进频道不断积累对工作的 context，不需要每次都从零开始重新交代。打开"ambient behavior"后，它就不再被动等着你问，而是会主动跟进那些冷掉的 thread，并把它在各个频道和工具里看到的相关信息标出来。

Link: https://x.com/claudeai/status/2069468701548531895 · https://x.com/claudeai/status/2069468699766005847 · https://x.com/claudeai/status/2069468698071494976

Andrej Karpathy framed Claude Tag as the third major redesign of LLM UI/UX. The first paradigm was the LLM as a website you visit; the second was an app you download; this third one is "a self-contained, persistent, asynchronous entity with org-wide tools and context, working alongside teams of humans." Once the under-the-hood plumbing across tools, integrations, compute, memory and security is done, he argues, Claude effectively joins the team and you talk to it the way you'd talk to a colleague. "It really takes a while to wrap your head around it, but it works and it is awesome."

Andrej Karpathy 把 Claude Tag 称为 LLM UI/UX 的第三次重大重新设计。第一种范式是把 LLM 当成一个你去访问的网站，第二种是一个你下载到电脑上的 app，而这第三种则是"一个自包含、持久、异步的实体，带着组织级的工具和 context，与人类团队并肩工作"。他认为，一旦把底层那些跨工具、跨集成、跨算力环境、跨 memory 和安全的工程都打通，Claude 实际上就加入了团队，你可以像跟同事说话那样跟它对话。"你确实得花点时间才能转过弯来理解它，但它真的能跑通，而且非常棒。"

Link: https://x.com/karpathy/status/2069547676849557725

Cat Wu called Claude Tag "a paradigm shift in how we ship products at Anthropic," noting it's the company's first natively multiplayer and proactive product — and that the internal version already merges 65% of product PRs. She shared six common flows that have resonated with internal users and design partners, plus a get-started guide for configuring agent permissions, stressing there are hundreds of ways to customize it for a given use case.

Cat Wu 把 Claude Tag 称为"我们在 Anthropic 交付产品方式上的一次范式转变"，并指出这是公司第一个原生多人协作（multiplayer）且具备主动性的产品，而内部版本已经能合并 65% 的产品 PR。她分享了六种在内部用户和设计合作伙伴中反响不错的常见用法，外加一份配置 agent 权限的上手指南，强调针对具体使用场景可以有上百种定制方式。

Link: https://x.com/_catwu/status/2069473118742331608 · https://x.com/_catwu/status/2069486403696869555 · https://x.com/_catwu/status/2069484330938998993

Boris Cherny walked through the mechanics: tag Claude in a channel and it spins up an instance with its own sandbox, clones repos, writes code, tests and compiles in that isolated environment, then throws the sandbox away when it's done — one instance per thread, with its own memory and permissions per channel. He also runs Claude as an always-on monitor of Slack channels that proactively answers questions, drafts PRs, and reacts with emojis like ✅ and ❌ to mark a thread resolved. "Point it at a channel, give it a task, and let it work."

Boris Cherny 详细讲了它的运作机制：在频道里 tag 一下 Claude，它就会启动一个带有独立 sandbox 的实例，在这个隔离环境里 clone 仓库、写代码、测试和编译，干完之后把 sandbox 丢弃，每个 thread 一个实例，每个频道有各自的 memory 和权限。他还让 Claude 常驻监控 Slack 频道，主动回答问题、起草 PR，并用 ✅、❌ 这类 emoji 来标记某个 thread 是否已解决。"把它对准一个频道，给它一个任务，然后让它去干。"

Link: https://x.com/bcherny/status/2069474691010707486 · https://x.com/bcherny/status/2069474689819480394 · https://x.com/bcherny/status/2069474688619958517

Thariq shared hands-on tips for living with Claude Tag once you have many threads running at once: have Claude react to the top-level thread with status emojis (⏲️✅❓🛑) so you can read its state at a glance, and ask it to keep a pinned message it updates with the status of everything in a busy channel. He's also wiring up a "scheduling" channel where people tag Claude to find time on each other's calendars.

Thariq 分享了一些在同时跑很多 thread 时与 Claude Tag 共处的实战技巧：让 Claude 用状态 emoji（⏲️✅❓🛑）对顶层 thread 做出反应，这样你一眼就能看出它的状态；在繁忙的频道里，让它维护一条 pinned 消息，随时更新所有事情的进展。他还在搭一个"scheduling"频道，让大家 tag Claude 来在彼此的日历上找空档。

Link: https://x.com/trq212/status/2069474343512617390 · https://x.com/trq212/status/2069474342220820657 · https://x.com/trq212/status/2069474339679052144

Alex Albert said Claude Tag has completely changed how he works with Claude: "It feels less like using a tool and more like managing a team." It's the same shift several of his colleagues flagged — the interaction model moving from issuing prompts to delegating and coordinating.

Alex Albert 说 Claude Tag 彻底改变了他与 Claude 协作的方式："感觉不再像是在用一个工具，更像是在管理一支团队。"这与他好几位同事提到的转变如出一辙，也就是交互模式从"下 prompt"转向"分派任务并做协调"。

Link: https://x.com/alexalbert__/status/2069470389391241314

Box CEO Aaron Levie laid out a barbell view of AI pricing: the market will split between high-cost frontier models and cheap-but-good open or closed-weight models, and the winning move is an applied AI layer that routes to the best model for each workload while mitigating either high token costs or weaker performance. "The closer you are to the underlying workflow, the better you can tune the use of models for a given business process." He also tied Claude Tag to Box's pitch for headless software: give Claude access to any corporate files in Box from inside Slack, and all of your enterprise content becomes a portable knowledge base.

Box CEO Aaron Levie 给出了一个关于 AI 定价的"杠铃"判断：市场会分化为高成本的 frontier 模型和便宜但够用的开源或闭源权重模型两端，而制胜之道是一个应用层（applied AI layer），它能针对每种工作负载路由到最合适的模型，同时对冲掉过高的 token 成本或较弱的性能。"你离底层工作流越近，就越能为某个具体的业务流程调好模型的用法。"他还把 Claude Tag 接到了 Box 主打的 headless software 叙事上：在 Slack 里就能让 Claude 访问 Box 中的任意企业文件，于是你所有的企业内容都变成了一个可携带的知识库。

Link: https://x.com/levie/status/2069639600310767616 · https://x.com/levie/status/2069596515560267891

Peter Yang posed a sharp question for the agent era: "What is design if what's accessing your product is just an agent looking for an API or CLI?" His read on the broader shift is that human-agent interaction is trending toward managing a highly capable employee — half-joking that 1-on-1s and performance reviews for Claude might be next. He also defended the Google Workspace CLI as genuinely useful and worth celebrating rather than punished.

Peter Yang 抛出了一个关于 agent 时代的犀利问题："如果访问你产品的只是一个在找 API 或 CLI 的 agent，那设计还意味着什么？"他对这一更大趋势的解读是：人与 agent 的交互正在朝着"管理一名高能力员工"演进，他半开玩笑地说，接下来也许就该给 Claude 做 1 on 1 和绩效评估了。他还为 Google Workspace CLI 辩护，认为这东西是真的好用、值得被庆祝，而不是被惩罚。

Link: https://x.com/petergyang/status/2069603490524254473 · https://x.com/petergyang/status/2069530765352907180 · https://x.com/petergyang/status/2069551302246592799

Madhu Guru, who has seen the AI stack from the enterprise, agent-startup, data-provider and lab sides, argued that the heated debate over token pricing is really a proxy fight over where value will accrue — models, the application layer, distribution, or somewhere else. Business models, moats and execution playbooks for every player are being figured out in real time, he wrote, so expect each side to push opinions that shape the narrative in its own favor.

曾在 Google 主导 Gemini、Veo、Nano Banana 产品的 Madhu Guru，从企业、agent 创业公司、数据提供方和实验室等多个视角都观察过 AI 这套体系。他认为，眼下围绕 token 定价的激烈争论，本质上是一场关于"价值最终会沉淀在哪里"的代理战，是模型、应用层、分发渠道，还是别的地方。他写道，各方的商业模式、护城河和执行打法都还在实时摸索之中，所以可以预料每一方都会抛出有利于自己的观点来塑造叙事。

Link: https://x.com/realmadhuguru/status/2069455097193697393

Garry Tan spotlighted Linzumi, a YC startup he describes as "Codex but actually multiplayer" and magical for teams — noting founder Sean Grove previously worked on the OpenAI team that greatly reduced ChatGPT's sycophancy. He also made a forward-looking prod at incumbents: Dropbox should support plans larger than 3TB because the amount of genuinely usable data thrown off by AI is about to go exponential, and storage products that don't update for that future risk being run over.

Garry Tan 重点介绍了 Linzumi，一家他形容为"Codex 但真正支持多人协作"、对团队来说很神奇的 YC 创业公司，并提到创始人 Sean Grove 此前在 OpenAI 参与过那个大幅降低 ChatGPT 谄媚倾向（sycophancy）的团队。他还对在位巨头提出了一个面向未来的提醒：Dropbox 应该支持超过 3TB 的套餐，因为 AI 抛出的、真正可用的数据量即将呈指数级增长，那些不为这个未来做更新的存储产品有被碾过去的风险。

Link: https://x.com/garrytan/status/2069474420113146355 · https://x.com/garrytan/status/2069434452628185241

Nikunj Kothari shipped a weekend build called Plug That Shop. After Shopify launched its UCP CLI — which makes all the products on its stores available from the command line — he used it to generate contextual embedded shops for any website: enter a URL and get four highly relevant shops you can embed in one click, with one-click Shop checkout for buyers. It's free, with no affiliate revenue, and was built with Conductor, Anthropic's Opus 4.8, the Shopify UCP CLI, and Railway.

Nikunj Kothari 在周末做出了一个叫 Plug That Shop 的项目。在 Shopify 推出 UCP CLI（让其店铺里的所有商品都能从命令行调用）之后，他用它为任意网站生成"情境化的嵌入式商店"：输入一个网址，就能得到四个高度相关的商店，一键嵌入，买家则可以用 Shop 一键结账。它免费、没有任何联盟分成，是用 Conductor、Anthropic 的 Opus 4.8、Shopify UCP CLI 和 Railway 搭起来的。

Link: https://x.com/nikunj/status/2069534712763490668 · https://x.com/nikunj/status/2069547206504566980

Shawn Wang marked a milestone for Chinese open models: Z.ai (formerly Zhipu) IPO'd in Hong Kong in January at HK$120 a share, and when he first met its team almost nobody knew anyone using GLM. Now, he says, GLM has beaten DeepSeek to become the world's undisputed top open model — and by some measures the top model period — with the team returning to San Francisco for the AI Engineer conference "on top of the world and open for business."

Shawn Wang 标记了中国开源模型的一个里程碑：Z.ai（前身为智谱）今年 1 月在香港以每股 120 港元 IPO，而他第一次见到这个团队时，几乎没人认识任何在用 GLM 的人。他说，如今 GLM 已经击败 DeepSeek，成为全球公认的最强开源模型，按某些指标甚至是最强模型本身；团队也重回旧金山参加 AI Engineer 大会，"志得意满，敞开门做生意"。

Link: https://x.com/swyx/status/2069598378191941835

Peter Steinberger flagged a piece of industry drama with a wink: by his telling, Google fired the person who made the Google Workspace CLI precisely because they made it — the same tool other builders were busy praising. "Lucky me, Google can't fire me," he added, needling the tension between official channels and the unsanctioned tools developers actually love.

Peter Steinberger 带着一丝调侃点出了一桩行业八卦：按他的说法，Google 把做出 Google Workspace CLI 的人开掉了，恰恰就因为这人做了这个工具，而别的 builder 们正忙着夸它好用。"算我走运，Google 没法开除我，"他补了一句，借此戳了戳官方渠道与开发者真正喜爱的"非官方工具"之间的那点张力。

Link: https://x.com/steipete/status/2069594195522941059

Aditya Agarwal announced a South Park Commons investment in Longshot Space, a company building a cannon that fires payloads into orbit. His thesis is blunt: there's far more demand for getting things into space than there is capacity to launch them. "Let's shoot things into space."

Aditya Agarwal 宣布 South Park Commons 投资了 Longshot Space，一家正在研制"用大炮把载荷打进轨道"的公司。他的逻辑很直接：把东西送上太空的需求，远远大于现有的发射运力。"让我们把东西打进太空吧。"

Link: https://x.com/adityaag/status/2069464862556619263 · https://x.com/adityaag/status/2069464865568166180

The Takeaway: The real bottleneck in AI-for-biology isn't compute or model size — it's that the data mostly doesn't exist yet, so the work is inventing new lab science to generate it, then giving the resulting models away for free.

Mark Zuckerberg, Priscilla Chan, and Alex Rives — the Meta CEO, the pediatrician-philanthropist who co-founded CZ Biohub, and the AI researcher (formerly of Meta's FAIR lab) now leading its science — have made Biohub the primary focus of their philanthropy, committing $500 million to a "virtual biology" initiative. The audacious mission, which once made Nobel laureates laugh in their faces, is to help cure, prevent, or manage all disease by the end of the century. Chan now calls that timeline "too conservative."

Their core conviction is counterintuitive for tech founders: they aren't trying to cure diseases themselves. They're building open-source tools so the entire scientific field can move faster. And the constraint that shapes everything is data. "It's not just like there's some factory somewhere that you can pay to produce the data. You actually need to invent new, novel, scientific approaches," Zuckerberg explains — which is why Biohub pairs a frontier AI lab with frontier wet-lab biology in San Francisco, New York and Chicago, building imaging tools, cellular-engineering methods and inflammation sensors that capture things no one has measured before.

The payoff is already visible. Their newly released ESMFold, trained on billions of raw protein sequences, folded over 1.1 billion proteins — and learned to design working antibodies as a side effect. As Rives puts it: "We didn't design a model for antibodies... We just designed a model that could understand proteins," and protein design emerged on its own. Work that once meant screening millions of candidates in a lab can now be selected digitally and confirmed in a single 96-well plate, hitting the nanomolar binding levels needed for therapeutics.

Chan's north star is personal: stop guessing whether a patient resembles people in some old paper, and instead trace the mechanism from gene variant to protein to disease so a drug can be designed bespoke for one individual.

核心要点： AI 用于生物学的真正瓶颈，既不是算力也不是模型大小，而是数据本身大多还不存在，所以真正的工作是去发明新的实验科学把数据造出来，然后把由此训练出的模型免费送出去。

Mark Zuckerberg、Priscilla Chan 和 Alex Rives，分别是 Meta 的 CEO、联合创办了 CZ Biohub 的儿科医生兼慈善家，以及如今主导其科研、曾在 Meta FAIR 实验室任职的 AI 研究员。他们已经把 Biohub 作为自己慈善事业的核心，向一个"virtual biology"计划投入了 5 亿美元。这个曾让诺贝尔奖得主当面发笑的大胆使命，是要在本世纪末之前帮助治愈、预防或管理所有疾病。Chan 如今说，这个时间表"已经太保守了"。

他们的核心信念对科技创始人来说很反直觉：他们并不打算亲自去治愈疾病，而是去打造开源工具，让整个科学界能跑得更快。而决定一切的约束是数据。"并不是说哪里有个工厂，你花钱就能让它把数据生产出来。你必须真正去发明全新的、前所未有的科学方法，"Zuckerberg 解释道。这正是为什么 Biohub 在旧金山、纽约和芝加哥同时配备一个 frontier AI 实验室和一个 frontier 湿实验室生物学团队，去打造成像工具、细胞工程方法和炎症传感器，捕捉那些此前从没有人测量过的东西。

回报已经显现。他们新发布的 ESMFold 在数十亿条原始蛋白质序列上训练，折叠了超过 11 亿个蛋白质，并把"设计出有效抗体"作为副产品自发学会了。正如 Rives 所说："我们并没有为抗体设计一个模型……我们只是设计了一个能够理解蛋白质的模型，"而蛋白质设计能力是自己涌现出来的。过去意味着要在实验室里筛选数百万候选物的工作，现在可以在数字层面先选出来，再用一块 96 孔板就确认下来，并达到治疗所需的纳摩尔级结合强度。

Chan 心中的北极星很个人化：不要再去猜某位患者是不是像某篇旧论文里研究过的那群人，而是把从基因变体到蛋白质再到疾病的机制完整地追出来，从而为某一个个体量身设计药物。

Link: https://www.youtube.com/@NoPriorsPodcast