OpenAI GPT-6 Astra 全面上线OpenAI GPT-6 Astra Rolls Out Widely
OpenAI CEO Sam Altman 宣布 GPT-6 Astra 已向所有 Pro、Enterprise 和 Business Premium 用户在 Work/Codex 中开放,并已进入 API。随后开始向 Plus 和 Business 用户推送。OpenAI Codex & ChatGPT 团队成员 Thibault Sottiaux 表示,Astra 在内部未全面可用时是最大竞争优势,上线后生产力大幅提升,部分计划提前 6 个月,将在 DevDay 而非明年中期发布。团队还同步为 Plus、Pro 和 Business 用户进行了完整额度重置。
OpenAI CEO Sam Altman announced GPT-6 Astra is now available to all Pro, Enterprise, and Business Premium users in Work/Codex and the API, with rollout starting for Plus and Business users. OpenAI Codex & ChatGPT team member Thibault Sottiaux noted Astra was their biggest competitive advantage while limited, and productivity jumped so much that plans shifted 6 months ahead for DevDay shipping instead of mid-next year. The team also completed a full banked reset for Plus, Pro, and Business users.
OpenAI CEO Sam Altman announced GPT-6 Astra is now available to all Pro, Enterprise, and Business Premium users in Work/Codex and the API, with rollout starting for Plus and Business users. OpenAI Codex & ChatGPT team member Thibault Sottiaux noted Astra was their biggest competitive advantage while limited, and productivity jumped so much that plans shifted 6 months ahead for DevDay shipping instead of mid-next year. The team also completed a full banked reset for Plus, Pro, and Business users.
OpenAI CEO Sam Altman 宣布 GPT-6 Astra 已向所有 Pro、Enterprise 和 Business Premium 用户在 Work/Codex 中开放,并已进入 API。随后开始向 Plus 和 Business 用户推送。OpenAI Codex & ChatGPT 团队成员 Thibault Sottiaux 表示,Astra 在内部未全面可用时是最大竞争优势,上线后生产力大幅提升,部分计划提前 6 个月,将在 DevDay 而非明年中期发布。团队还同步为 Plus、Pro 和 Business 用户进行了完整额度重置。
查看原文 →查看原文 →查看原文 →查看原文 →
Arm CEO Rene Haas:芯片与 AI 的下一程Arm CEO Rene Haas on Chip Architecture and AI
The Takeaway: CPU 永远是计算的核心,AI 验证正大幅压缩芯片设计周期,而供应链与数据中心建设才是真正的瓶颈。
Arm CEO 兼 SoftBank Group International 负责人 Rene Haas 指出,微处理器是一切计算问题的心脏,“所有道路都通向它”。芯片设计的实际架构并非最耗时,验证、验证和调试才占大头,AI 在这方面表现卓越,Arm 内部 80-90% 工程师每天使用。若关闭 AI 工具,就像 1990 年代只允许每天两小时上网。他预测更直接的设计在 5 年以上有望从概念到 GDSII 大幅缩短。
Arm 已从纯 IP 授权演进到计算子系统甚至自有 AGI CPU 产品(与 Meta 合作),并得到 Nvidia、Amazon 等客户支持。供应链约束预计持续 3-5 年,数据中心建设与能源将成下一个瓶颈。对机器人,Haas 认为通用与专用形态并存,Arm 将无处不在。他强调美国需要更多本土晶圆厂以保障国家安全与供应链多元化,并认为 AI 不会导致大规模失业,数据中心实际在创造高技能岗位。
The Takeaway: The CPU remains the heart of every computing problem; AI is already slashing verification time in chip design, while supply chain and data-center buildout are the real governors on growth.
Arm CEO and SoftBank Group International lead Rene Haas stresses that the microprocessor is the heart of everything—“all roads lead through it.” Design itself is not the long pole; verification, validation, and debug dominate the 24-36 month cycle, and AI is excellent at those tasks. 80-90% of Arm engineers use it daily; turning it off would be like 1990s internet limited to two hours a day. Straightforward designs could reach GDSII far faster in five-plus years.
Arm evolved from pure IP licensing to compute subsystems and its own AGI CPU product (starting with Meta), with support from Nvidia, Amazon, Microsoft, and Google. Supply constraints will last at least 3-5 years; data-center construction and energy are emerging bottlenecks. On robotics, both humanoid and task-specific form factors will thrive, and Arm will power the sensing and brains. Haas advocates more U.S. fabs for national security and diversification, and rejects the idea that AI destroys jobs—data centers create skilled work such as electricians.
The Takeaway: The CPU remains the heart of every computing problem; AI is already slashing verification time in chip design, while supply chain and data-center buildout are the real governors on growth.
Arm CEO and SoftBank Group International lead Rene Haas stresses that the microprocessor is the heart of everything—“all roads lead through it.” Design itself is not the long pole; verification, validation, and debug dominate the 24-36 month cycle, and AI is excellent at those tasks. 80-90% of Arm engineers use it daily; turning it off would be like 1990s internet limited to two hours a day. Straightforward designs could reach GDSII far faster in five-plus years.
Arm evolved from pure IP licensing to compute subsystems and its own AGI CPU product (starting with Meta), with support from Nvidia, Amazon, Microsoft, and Google. Supply constraints will last at least 3-5 years; data-center construction and energy are emerging bottlenecks. On robotics, both humanoid and task-specific form factors will thrive, and Arm will power the sensing and brains. Haas advocates more U.S. fabs for national security and diversification, and rejects the idea that AI destroys jobs—data centers create skilled work such as electricians.
The Takeaway: CPU 永远是计算的核心,AI 验证正大幅压缩芯片设计周期,而供应链与数据中心建设才是真正的瓶颈。
Arm CEO 兼 SoftBank Group International 负责人 Rene Haas 指出,微处理器是一切计算问题的心脏,“所有道路都通向它”。芯片设计的实际架构并非最耗时,验证、验证和调试才占大头,AI 在这方面表现卓越,Arm 内部 80-90% 工程师每天使用。若关闭 AI 工具,就像 1990 年代只允许每天两小时上网。他预测更直接的设计在 5 年以上有望从概念到 GDSII 大幅缩短。
Arm 已从纯 IP 授权演进到计算子系统甚至自有 AGI CPU 产品(与 Meta 合作),并得到 Nvidia、Amazon 等客户支持。供应链约束预计持续 3-5 年,数据中心建设与能源将成下一个瓶颈。对机器人,Haas 认为通用与专用形态并存,Arm 将无处不在。他强调美国需要更多本土晶圆厂以保障国家安全与供应链多元化,并认为 AI 不会导致大规模失业,数据中心实际在创造高技能岗位。
查看原文 →
Claude in Chrome 正式全面可用Claude in Chrome Now Generally Available
Claude Blog: Claude in Chrome is generally available。Claude in Chrome 现已对所有付费 Claude 计划全面开放。Claude 现在可以在浏览器中自主采取行动,而无需每一步都获批准。安全分类器会在执行前验证每个动作是否安全且符合用户请求。它能访问内部仪表板、遗留系统和供应商门户等无连接工具,使用现有登录查看页面、点击、导航和填写表单。
Anthropic 重点加强了 prompt injection 防护:模型训练对抗不断增长的攻击库、探针扫描工具结果、动作分类器在自动批准前核对原始请求。评估显示,在探针加安全分类器下,针对 Claude Sonnet 5 和 Opus 5 的攻击成功率为 0%,Fable 5 为 0.3%。企业管理员可限制到批准域名。
Claude Blog: Claude in Chrome is generally available. Claude in Chrome is now generally available on every paid Claude plan. Claude can now take actions autonomously in the browser instead of needing approval for every step. A safety classifier validates each action before it is performed. It reaches internal dashboards, legacy systems, and vendor portals that lack connectors, using existing logins to view pages, click, navigate, and fill forms.
Anthropic strengthened prompt-injection defenses: model training against a growing attack library, probes that screen tool results, and action classifiers that check against the original request before auto-approval. Evaluations show 0% success against Claude Sonnet 5 and Opus 5 with probes plus classifiers, and 0.3% against Fable 5. Enterprise admins can limit it to approved domains.
Claude Blog: Claude in Chrome is generally available. Claude in Chrome is now generally available on every paid Claude plan. Claude can now take actions autonomously in the browser instead of needing approval for every step. A safety classifier validates each action before it is performed. It reaches internal dashboards, legacy systems, and vendor portals that lack connectors, using existing logins to view pages, click, navigate, and fill forms.
Anthropic strengthened prompt-injection defenses: model training against a growing attack library, probes that screen tool results, and action classifiers that check against the original request before auto-approval. Evaluations show 0% success against Claude Sonnet 5 and Opus 5 with probes plus classifiers, and 0.3% against Fable 5. Enterprise admins can limit it to approved domains.
Claude Blog: Claude in Chrome is generally available。Claude in Chrome 现已对所有付费 Claude 计划全面开放。Claude 现在可以在浏览器中自主采取行动,而无需每一步都获批准。安全分类器会在执行前验证每个动作是否安全且符合用户请求。它能访问内部仪表板、遗留系统和供应商门户等无连接工具,使用现有登录查看页面、点击、导航和填写表单。
Anthropic 重点加强了 prompt injection 防护:模型训练对抗不断增长的攻击库、探针扫描工具结果、动作分类器在自动批准前核对原始请求。评估显示,在探针加安全分类器下,针对 Claude Sonnet 5 和 Opus 5 的攻击成功率为 0%,Fable 5 为 0.3%。企业管理员可限制到批准域名。
查看原文 →