🌐 双语
Archive

AI Builders
Digest

2026-11-13 16 builders · 38 tweets · 1 podcasts · 0 blogs

🔥 热点话题

Gemini 4 Hydrogen 发布:更快更准的轻量模型Gemini 4 Hydrogen Launches: Faster, Smarter Lightweight Model

Google Labs VP Josh Woodward 宣布 Gemini 4 Hydrogen 现已在 Gemini App 上线,号称面向所有人的前沿智能,速度快、更聪明,是一次大幅跃升。Box CEO Aaron Levie 分享早期测试结果:在 Box AI Complex Work Eval 中,Hydrogen 准确率比 Gemini 3.8 Flash 高出 5 个百分点,任务完成时间不到一半,token 消耗约少 12%。该模型适合金融、咨询、生命科学等领域的高效复杂任务,很快将进入 Box AI Studio 供客户自定义 agent 使用。

Google VP Josh Woodward announced Gemini 4 Hydrogen is now live in the Gemini App, describing it as cutting-edge intelligence for everyone that is fast, smart, and a big jump. Box CEO Aaron Levie reported early access results on the Box AI Complex Work Eval: Hydrogen is 5 percentage points more accurate than Gemini 3.8 Flash, completes tasks in less than half the time, and uses about 12% fewer tokens on typical workloads. It is positioned as a strong model for efficient yet complex enterprise document tasks across finance, consulting, and life sciences, with rollout to Box AI Studio coming soon.
Google Labs VP Josh Woodward announced Gemini 4 Hydrogen is now live in the Gemini App, describing it as cutting-edge intelligence for everyone that is fast, smart, and a big jump. Box CEO Aaron Levie reported early access results on the Box AI Complex Work Eval: Hydrogen is 5 percentage points more accurate than Gemini 3.8 Flash, completes tasks in less than half the time, and uses about 12% fewer tokens on typical workloads. It is positioned as a strong model for efficient yet complex enterprise document tasks across finance, consulting, and life sciences, with rollout to Box AI Studio coming soon.

Google Labs VP Josh Woodward 宣布 Gemini 4 Hydrogen 现已在 Gemini App 上线,号称面向所有人的前沿智能,速度快、更聪明,是一次大幅跃升。Box CEO Aaron Levie 分享早期测试结果:在 Box AI Complex Work Eval 中,Hydrogen 准确率比 Gemini 3.8 Flash 高出 5 个百分点,任务完成时间不到一半,token 消耗约少 12%。该模型适合金融、咨询、生命科学等领域的高效复杂任务,很快将进入 Box AI Studio 供客户自定义 agent 使用。
查看原文 →查看原文 →查看原文 →

Claude Meetings 上线:AI 直接加入会议并做笔记Claude Meetings Launches: AI Joins Calls as Visible Participant

Anthropic 正式推出 Claude Meetings,目前支持 Google Meet,其他视频会议平台即将跟进。Claude 会作为可见参与者加入会议,团队可以口头或聊天提问,让它实时起草文档,所有人都能看到并纠正。会后笔记、行动项、完整 transcript 以及 Claude 生成的任何内容都会自动出现在会议页面。Anthropic Research 的 Alex Albert 指出,现在用好 Claude 的关键是补全缺失的 context,Claude Meetings 正是把这些剩余切片带进来的有效方式,内部已经证明非常实用。

Anthropic launched Claude Meetings, starting with Google Meet and expanding to other platforms soon. Claude joins as a visible participant: ask it questions out loud or in chat, or have it start a document while the team watches and corrects in real time. After the call, notes, action items, the transcript, and anything Claude built appear on the meeting page. Anthropic researcher Alex Albert noted that getting good work from Claude is largely about finding the missing context and closing that gap; Claude Meetings is one practical way to bring in those remaining slices and has already proved useful internally.
Anthropic launched Claude Meetings, starting with Google Meet and expanding to other platforms soon. Claude joins as a visible participant: ask it questions out loud or in chat, or have it start a document while the team watches and corrects in real time. After the call, notes, action items, the transcript, and anything Claude built appear on the meeting page. Anthropic researcher Alex Albert noted that getting good work from Claude is largely about finding the missing context and closing that gap; Claude Meetings is one practical way to bring in those remaining slices and has already proved useful internally.

Anthropic 正式推出 Claude Meetings,目前支持 Google Meet,其他视频会议平台即将跟进。Claude 会作为可见参与者加入会议,团队可以口头或聊天提问,让它实时起草文档,所有人都能看到并纠正。会后笔记、行动项、完整 transcript 以及 Claude 生成的任何内容都会自动出现在会议页面。Anthropic Research 的 Alex Albert 指出,现在用好 Claude 的关键是补全缺失的 context,Claude Meetings 正是把这些剩余切片带进来的有效方式,内部已经证明非常实用。
查看原文 →查看原文 →查看原文 →查看原文 →

Jensen Huang 相关争议持续发酵,AI 圈反应强烈Jensen Huang Controversy Continues; AI Builders React Strongly

多位 AI 建设者对围绕 Jensen Huang 的最新争议发表看法。FPV Ventures 合伙人 Nikunj Kothari 指出,相关消息已持续 9 小时仍无定论,预计还会进一步恶化,并警示类似情况可能发生在任何组织。FirstMark VC Matt Turck 总结,AI 生态因此失去了“央行”、最活跃投资者以及大胆收购方。SPC 普通合伙人 Aditya Agarwal 直言“没有 Jensen 就没有我们今天的位置”,看到现状感到切肤之痛。Every CEO Dan Shipper 则用长文类比,认为真正的“不忠”是美国政策本身同时要求遏制中国又要求主导中国市场,Jensen 只是被夹在中间的人。

Several AI builders weighed in on the ongoing Jensen Huang controversy. FPV Ventures partner Nikunj Kothari observed that nine hours without resolution is extreme and expects the situation to worsen significantly before improving, warning it could hit every organization. FirstMark VC Matt Turck listed the losses for the AI ecosystem: its central banker, most active investor, and a bold acquirer. SPC General Partner Aditya Agarwal wrote that the industry would not be where it is without Jensen and called the developments viscerally painful. Every CEO Dan Shipper framed the deeper issue as policy itself demanding both containment of China and dominance of the Chinese market, placing Jensen in an impossible position.
Several AI builders weighed in on the ongoing Jensen Huang controversy. FPV Ventures partner Nikunj Kothari observed that nine hours without resolution is extreme and expects the situation to worsen significantly before improving, warning it could hit every organization. FirstMark VC Matt Turck listed the losses for the AI ecosystem: its central banker, most active investor, and a bold acquirer. SPC General Partner Aditya Agarwal wrote that the industry would not be where it is without Jensen and called the developments viscerally painful. Every CEO Dan Shipper framed the deeper issue as policy itself demanding both containment of China and dominance of the Chinese market, placing Jensen in an impossible position.

多位 AI 建设者对围绕 Jensen Huang 的最新争议发表看法。FPV Ventures 合伙人 Nikunj Kothari 指出,相关消息已持续 9 小时仍无定论,预计还会进一步恶化,并警示类似情况可能发生在任何组织。FirstMark VC Matt Turck 总结,AI 生态因此失去了“央行”、最活跃投资者以及大胆收购方。SPC 普通合伙人 Aditya Agarwal 直言“没有 Jensen 就没有我们今天的位置”,看到现状感到切肤之痛。Every CEO Dan Shipper 则用长文类比,认为真正的“不忠”是美国政策本身同时要求遏制中国又要求主导中国市场,Jensen 只是被夹在中间的人。
查看原文 →查看原文 →查看原文 →查看原文 →查看原文 →

🛠️ 开发者工具与技巧

OpenAI 更新:Codex 改进征集、网络安全工具与 DotsOpenAI Updates: Codex Feedback, Cybersecurity Tools, and Dots

OpenAI Codex & ChatGPT 负责人 Thibault Sottiaux 公开征集 Codex 改进建议(明确要求不要只说模型本身),引发大量回复。同时宣布 OpenAI 正在开发自动化补丁和自动化红队等网络安全新工具,并分享了录制回放。Dots 也推出了更新,banked reset 已激活,将陆续到账。Peter Steinberger 分享团队关于在 OpenAI 内部运行 persistent agents 的文章,并询问用户是否会为 dots 开启 yolo 模式。

OpenAI Codex & ChatGPT lead Thibault Sottiaux asked the community how Codex can be improved, specifically requesting ideas that go beyond “the model,” drawing thousands of replies. He also announced new cybersecurity tools in development for automated patching and automated red teaming, with a recording available. Dots received updates and the banked reset has been activated. Peter Steinberger shared a team write-up on running persistent agents inside OpenAI and asked whether users would enable yolo mode for dots despite the risk of unwanted actions.
OpenAI Codex & ChatGPT lead Thibault Sottiaux asked the community how Codex can be improved, specifically requesting ideas that go beyond “the model,” drawing thousands of replies. He also announced new cybersecurity tools in development for automated patching and automated red teaming, with a recording available. Dots received updates and the banked reset has been activated. Peter Steinberger shared a team write-up on running persistent agents inside OpenAI and asked whether users would enable yolo mode for dots despite the risk of unwanted actions.

OpenAI Codex & ChatGPT 负责人 Thibault Sottiaux 公开征集 Codex 改进建议(明确要求不要只说模型本身),引发大量回复。同时宣布 OpenAI 正在开发自动化补丁和自动化红队等网络安全新工具,并分享了录制回放。Dots 也推出了更新,banked reset 已激活,将陆续到账。Peter Steinberger 分享团队关于在 OpenAI 内部运行 persistent agents 的文章,并询问用户是否会为 dots 开启 yolo 模式。
查看原文 →查看原文 →查看原文 →查看原文 →查看原文 →

Vercel UI SDK:为 Agent 时代重新定义组件抽象Vercel UI SDK: Lower-Level Primitives for the Agent Era

Vercel CEO Guillermo Rauch 讨论了 agent 时代框架与组件库的命运。他提出需要“向下钻取”:删除多余部分、优化、加速、自动化。UI SDK 被描述为 agent 构建新 UI 组件的 x86 指令集——提供无障碍、高性能、一致的封装行为,同时不限制可生成的 UI 形态。Rauch 认为 agent 是新的设计工程师,高层刚性抽象反而会成为无限创造机器的阻碍,UI SDK 试图同时保留便利与自由度。

Vercel CEO Guillermo Rauch examined what happens to frameworks and component libraries in the agent era. The answer, he argues, is to drill deeper: delete the unnecessary, optimize, accelerate, and automate. UI SDK is positioned as the x86 instruction set for agents to spontaneously build new UI components—packaged behaviors that deliver accessibility, performance, and cohesion while imposing zero limitations on what UIs can be generated. Agents become the new design engineers; high-level rigid abstractions are kryptonite to an infinite creation machine.
Vercel CEO Guillermo Rauch examined what happens to frameworks and component libraries in the agent era. The answer, he argues, is to drill deeper: delete the unnecessary, optimize, accelerate, and automate. UI SDK is positioned as the x86 instruction set for agents to spontaneously build new UI components—packaged behaviors that deliver accessibility, performance, and cohesion while imposing zero limitations on what UIs can be generated. Agents become the new design engineers; high-level rigid abstractions are kryptonite to an infinite creation machine.

Vercel CEO Guillermo Rauch 讨论了 agent 时代框架与组件库的命运。他提出需要“向下钻取”:删除多余部分、优化、加速、自动化。UI SDK 被描述为 agent 构建新 UI 组件的 x86 指令集——提供无障碍、高性能、一致的封装行为,同时不限制可生成的 UI 形态。Rauch 认为 agent 是新的设计工程师,高层刚性抽象反而会成为无限创造机器的阻碍,UI SDK 试图同时保留便利与自由度。
查看原文 →

Replit Agent 新增商业能力:可直接为业务采购与投放广告Replit Agent Can Now Buy Ads and Business Supplies

Replit CEO Amjad Masad 宣布,用户现在可以直接让 Replit Agent 为业务采购物品,包括购买广告来获取新客户。这一功能把 agent 从代码与应用构建进一步延伸到实际商业执行环节。

Replit CEO Amjad Masad announced that users can now ask Replit Agent to buy things for their business, including purchasing ads to acquire new customers. The update extends agents from building software into executing real commercial actions.
Replit CEO Amjad Masad announced that users can now ask Replit Agent to buy things for their business, including purchasing ads to acquire new customers. The update extends agents from building software into executing real commercial actions.

Replit CEO Amjad Masad 宣布,用户现在可以直接让 Replit Agent 为业务采购物品,包括购买广告来获取新客户。这一功能把 agent 从代码与应用构建进一步延伸到实际商业执行环节。
查看原文 →

🌍 其他动态

Isomorphic Labs 总裁 Max Jaderberg:把药物发现搬进计算机Isomorphic Labs President Max Jaderberg: Moving Drug Discovery Fully In Silico

The Takeaway:药物发现最大的机会在于把经验驱动的试错过程搬到计算机上,用达到实验精度的 AI 预测模型,在进入人体之前就大幅降低临床失败率。

Max Jaderberg 是 Isomorphic Labs 总裁与创始团队成员。该公司由 Alphabet 支持,目标是“解决所有疾病”。他此前在 DeepMind 工作多年,参与过 AlphaGo、AlphaStar 以及早期生成模型和强化学习,也是将公司卖给 DeepMind 的计算机视觉创业者。AlphaFold 的成功让团队确信,深度模型已经可以在有意义的生物学空间中泛化,因此值得单独成立公司全力推进。

传统制药高度依赖经验试错:假设疾病机制、设计分子、合成、测试,再迭代。低垂果实已被摘完,新药审批数量长期持平,成本却从数亿美元涨到约 15 亿美元,周期从 8-9 年拉长到 15 年。失败主要发生在临床阶段——毒性或脱靶效应只有进入人体后才暴露。Jaderberg 认为,核心瓶颈是“合成真实原子并在真实实验中测试”,因此 Isomorphic 的北极星是把疾病生物学理解与分子设计全部搬到计算机上,目标是让模拟模型达到实验精度。

公司内部同时建设自有管线(聚焦肿瘤、免疫与炎症)并与制药伙伴合作。模型本身是疾病、靶点和模态无关的。他们强调生成专有数据以获得更好的泛化,并探索用体外终点、类器官、器官芯片等新技术直接建模人类临床毒性,而不是继续依赖预测力仅约 30% 的动物模型。Jaderberg 表示:“如果五年后我们对疾病生物学的预测理解没有发生巨大跃迁,我会非常失望。”

The Takeaway: The biggest opportunity in drug discovery is moving the empirical trial-and-error process onto computers so that AI predictive models reaching experimental accuracy can surface toxicity and off-target problems long before human trials.

Max Jaderberg is President and a founding team member of Isomorphic Labs, the Alphabet-backed company whose stated goal is to solve all disease. He spent years at DeepMind working on AlphaGo, AlphaStar, generative models and reinforcement learning, after selling a computer-vision startup to the lab. AlphaFold’s success convinced the team that deep models could finally generalize in meaningful biological spaces, justifying a dedicated company and culture.

Traditional medicine discovery remains highly empirical: form a hypothesis, design a molecule, synthesize it, test it, and iterate. The low-hanging fruit has been picked; approvals have stayed roughly flat while costs climbed to roughly $1.5 billion and timelines stretched to 15 years. Most failures occur in the clinic because toxicity or unintended effects only appear once drugs reach patients. Jaderberg argues the fundamental bottleneck is “synthesizing real atoms and testing them in real assays,” so Isomorphic’s north star is to bring disease-biology understanding and molecular design fully in silico until models match experimental accuracy.

The company runs both an internal pipeline (oncology, immunology and inflammation) and partner programs with pharma. Its models are disease-, target- and modality-agnostic. It prioritizes generating proprietary data for better generalization and is exploring in-vitro endpoints, organoids and organs-on-chips to model human clinical toxicity directly rather than relying on animal models that are only about 30% predictive. “I’d be really disappointed if we fast forward even five years’ time [and] the predictive understanding of disease biology has not step changed in a massive way,” Jaderberg said.
The Takeaway: The biggest opportunity in drug discovery is moving the empirical trial-and-error process onto computers so that AI predictive models reaching experimental accuracy can surface toxicity and off-target problems long before human trials.

Max Jaderberg is President and a founding team member of Isomorphic Labs, the Alphabet-backed company whose stated goal is to solve all disease. He spent years at DeepMind working on AlphaGo, AlphaStar, generative models and reinforcement learning, after selling a computer-vision startup to the lab. AlphaFold’s success convinced the team that deep models could finally generalize in meaningful biological spaces, justifying a dedicated company and culture.

Traditional medicine discovery remains highly empirical: form a hypothesis, design a molecule, synthesize it, test it, and iterate. The low-hanging fruit has been picked; approvals have stayed roughly flat while costs climbed to roughly $1.5 billion and timelines stretched to 15 years. Most failures occur in the clinic because toxicity or unintended effects only appear once drugs reach patients. Jaderberg argues the fundamental bottleneck is “synthesizing real atoms and testing them in real assays,” so Isomorphic’s north star is to bring disease-biology understanding and molecular design fully in silico until models match experimental accuracy.

The company runs both an internal pipeline (oncology, immunology and inflammation) and partner programs with pharma. Its models are disease-, target- and modality-agnostic. It prioritizes generating proprietary data for better generalization and is exploring in-vitro endpoints, organoids and organs-on-chips to model human clinical toxicity directly rather than relying on animal models that are only about 30% predictive. “I’d be really disappointed if we fast forward even five years’ time [and] the predictive understanding of disease biology has not step changed in a massive way,” Jaderberg said.

The Takeaway:药物发现最大的机会在于把经验驱动的试错过程搬到计算机上,用达到实验精度的 AI 预测模型,在进入人体之前就大幅降低临床失败率。

Max Jaderberg 是 Isomorphic Labs 总裁与创始团队成员。该公司由 Alphabet 支持,目标是“解决所有疾病”。他此前在 DeepMind 工作多年,参与过 AlphaGo、AlphaStar 以及早期生成模型和强化学习,也是将公司卖给 DeepMind 的计算机视觉创业者。AlphaFold 的成功让团队确信,深度模型已经可以在有意义的生物学空间中泛化,因此值得单独成立公司全力推进。

传统制药高度依赖经验试错:假设疾病机制、设计分子、合成、测试,再迭代。低垂果实已被摘完,新药审批数量长期持平,成本却从数亿美元涨到约 15 亿美元,周期从 8-9 年拉长到 15 年。失败主要发生在临床阶段——毒性或脱靶效应只有进入人体后才暴露。Jaderberg 认为,核心瓶颈是“合成真实原子并在真实实验中测试”,因此 Isomorphic 的北极星是把疾病生物学理解与分子设计全部搬到计算机上,目标是让模拟模型达到实验精度。

公司内部同时建设自有管线(聚焦肿瘤、免疫与炎症)并与制药伙伴合作。模型本身是疾病、靶点和模态无关的。他们强调生成专有数据以获得更好的泛化,并探索用体外终点、类器官、器官芯片等新技术直接建模人类临床毒性,而不是继续依赖预测力仅约 30% 的动物模型。Jaderberg 表示:“如果五年后我们对疾病生物学的预测理解没有发生巨大跃迁,我会非常失望。”
查看原文 →