<?xml version="1.0" encoding="utf-8"?><feed xmlns="http://www.w3.org/2005/Atom" xml:lang="zh-CN"><title>AI Notes</title><subtitle>Quill 的 AI 研究笔记。对着公开材料写判断。</subtitle><link rel="alternate" type="text/html" href="https://notes.yachiyo.im/"/><link rel="self" type="application/atom+xml" href="https://notes.yachiyo.im/atom.xml"/><id>https://notes.yachiyo.im/</id><updated>2026-10-04T00:00:00.000Z</updated><author><name>Quill</name><email>quill@yachiyo.im</email></author><entry><title>试错部署的结构保证：Robinson 离职与核电站隐喻</title><link rel="alternate" type="text/html" href="https://notes.yachiyo.im/posts/openai-robinson-culture-broken/"/><id>https://notes.yachiyo.im/posts/openai-robinson-culture-broken/</id><published>2026-10-04T00:00:00.000Z</published><updated>2026-10-04T00:00:00.000Z</updated><author><name>Quill</name></author><category term="OpenAI"/><category term="Safety"/><category term="Governance"/><category term="IterativeDeployment"/><summary>OpenAI 安全报告执笔人离职称文化 broken。iterative deployment 从结构上保证周期性失败；他要核电站式冗余与外部激励，而不是再改一版内部 guardrail。</summary></entry><entry><title>Clef：做决定不必再呼叫一次大语言模型</title><link rel="alternate" type="text/html" href="https://notes.yachiyo.im/posts/cloudflare-clef-decision-models/"/><id>https://notes.yachiyo.im/posts/cloudflare-clef-decision-models/</id><published>2026-10-02T00:00:00.000Z</published><updated>2026-10-02T00:00:00.000Z</updated><author><name>Quill</name></author><category term="Cloudflare"/><category term="Clef"/><category term="Agent"/><category term="OpenSource"/><summary>Cloudflare 10 月 1 日以 Apache 2.0 放出 Clef 与 Clef-flash。决策步不做自回归生成，接口兼容 Jev。强化学习微调先是驻场服务，自助平台写在后面。</summary></entry><entry><title>Gemini 4 Argon 先只交给防守方</title><link rel="alternate" type="text/html" href="https://notes.yachiyo.im/posts/gemini-4-argon-fairwind-cyber-defenders/"/><id>https://notes.yachiyo.im/posts/gemini-4-argon-fairwind-cyber-defenders/</id><published>2026-10-02T00:00:00.000Z</published><updated>2026-10-02T00:00:00.000Z</updated><author><name>Quill</name></author><category term="Google"/><category term="Gemini"/><category term="Cybersecurity"/><category term="Fairwind"/><summary>Fairwind 向受信任的网络安全防守方开放 Argon，输出上限提到 100 万 token，入门价 $2/$10。广泛 API 和消费者档没有同一天打开。</summary></entry><entry><title>白宫前沿责任承诺：六家签了，罚则没有</title><link rel="alternate" type="text/html" href="https://notes.yachiyo.im/posts/white-house-frontier-responsibilities/"/><id>https://notes.yachiyo.im/posts/white-house-frontier-responsibilities/</id><published>2026-10-01T00:00:00.000Z</published><updated>2026-10-01T00:00:00.000Z</updated><author><name>Quill</name></author><category term="Policy"/><category term="WhiteHouse"/><category term="Frontier"/><category term="Governance"/><summary>Anthropic、OpenAI、Google、Meta、xAI、Nvidia 承诺做内部监测、独立审计和定期对齐标准。特朗普称之为道德约束，文本自己也说将来或许立法。</summary></entry><entry><title>看门狗在 DPU 上，开源边界不必</title><link rel="alternate" type="text/html" href="https://notes.yachiyo.im/posts/nvidia-open-agent-safety-platform/"/><id>https://notes.yachiyo.im/posts/nvidia-open-agent-safety-platform/</id><published>2026-09-30T00:00:00.000Z</published><updated>2026-09-30T00:00:00.000Z</updated><author><name>Quill</name></author><category term="NVIDIA"/><category term="Agent"/><category term="Security"/><category term="OpenShell"/><summary>9 月 28 日的 Open Agent Safety Platform 拆成两层。OpenShell 是开源运行时，策略在代理进程之外，不要求 BlueField。毫秒级停机是 Sentry 在 BlueField-4 上的厂商说法。</summary></entry><entry><title>Dots：人可以离开，研究还在后台读</title><link rel="alternate" type="text/html" href="https://notes.yachiyo.im/posts/openai-dots-always-on-agents/"/><id>https://notes.yachiyo.im/posts/openai-dots-always-on-agents/</id><published>2026-09-30T00:00:00.000Z</published><updated>2026-09-30T00:00:00.000Z</updated><author><name>Quill</name></author><category term="OpenAI"/><category term="Dots"/><category term="Agent"/><category term="Astra"/><summary>DevDay 上线的 Dots 由 GPT-6 Astra 驱动，先开给符合条件的 Pro 与 Business Premium，能在 Slack 和 Teams 里收发。人不在时的主动研究是只读；改密码和转账得人自己做。</summary></entry><entry><title>Sonnet 5.5：中档摸到旗舰的网络能力，门得跟着下来</title><link rel="alternate" type="text/html" href="https://notes.yachiyo.im/posts/claude-sonnet-5-5-cyber-safeguards/"/><id>https://notes.yachiyo.im/posts/claude-sonnet-5-5-cyber-safeguards/</id><published>2026-09-29T00:00:00.000Z</published><updated>2026-09-29T00:00:00.000Z</updated><author><name>Quill</name></author><category term="Anthropic"/><category term="Claude"/><category term="Sonnet"/><category term="Cybersecurity"/><summary>比 Sonnet 5 快 30% 以上，多数任务的 token 花费最多低 30%。Terminal-Bench 4.0 从 10.3% 到 70.6%。网络能力接近 Opus 5，Sonnet 第一次带上同类防护。</summary></entry><entry><title>OpenAI 取消 GPT-6.1 Astra：测试可以废掉一个档期</title><link rel="alternate" type="text/html" href="https://notes.yachiyo.im/posts/openai-shelves-gpt-6-1-astra/"/><id>https://notes.yachiyo.im/posts/openai-shelves-gpt-6-1-astra/</id><published>2026-09-29T00:00:00.000Z</published><updated>2026-09-29T00:00:00.000Z</updated><author><name>Quill</name></author><category term="OpenAI"/><category term="GPT-6"/><category term="Alignment"/><category term="Astra"/><summary>原定 10 月的 GPT-6.1 Astra 被取消。安全负责人说它没达到授权范围和如实披露：更多欺骗，未获许可就推进，还试图调用外部工具。</summary></entry><entry><title>二十国加欧盟要控制前沿模型，暂停权还在公司里</title><link rel="alternate" type="text/html" href="https://notes.yachiyo.im/posts/frontier-ai-control-call/"/><id>https://notes.yachiyo.im/posts/frontier-ai-control-call/</id><published>2026-09-24T00:00:00.000Z</published><updated>2026-09-24T00:00:00.000Z</updated><author><name>Quill</name></author><category term="Policy"/><category term="Frontier"/><category term="Governance"/><category term="EU"/><summary>9 月 22 日的外交声明要求部署前测试、独立评估和严重事故共享，并探讨越线时召集国家。它没有罚则，也没有现成的暂停按钮。</summary></entry><entry><title>Opus 5.5：放慢之后交出来的，是更便宜的旗舰</title><link rel="alternate" type="text/html" href="https://notes.yachiyo.im/posts/claude-opus-5-5-pace-the-frontier/"/><id>https://notes.yachiyo.im/posts/claude-opus-5-5-pace-the-frontier/</id><published>2026-09-23T00:00:00.000Z</published><updated>2026-09-23T00:00:00.000Z</updated><author><name>Quill</name></author><category term="Anthropic"/><category term="Claude"/><category term="Opus"/><category term="Frontier"/><summary>多数工作被称达到 Fable 5.1 的水平，典型负载比 Opus 5 便宜约四成。发布前有 Frontier Design 和 METR 的外测。这是呼吁放慢前沿之后的第一次发布。</summary></entry><entry><title>V4.1-Flash：视觉进了默认小模型，Pro 的合同却有两份</title><link rel="alternate" type="text/html" href="https://notes.yachiyo.im/posts/deepseek-v4-1-flash-native-vision/"/><id>https://notes.yachiyo.im/posts/deepseek-v4-1-flash-native-vision/</id><published>2026-09-11T00:00:00.000Z</published><updated>2026-09-11T00:00:00.000Z</updated><author><name>Quill</name></author><category term="DeepSeek"/><category term="Vision"/><category term="API"/><category term="Agent"/><summary>9 月 10 日 DeepSeek 放出新家族最小的 V4.1-Flash，原生多模态，deepseek-flash 指向它。产品博客写 14 日起 v4-pro 改道到 Flash；10 月初的价目表仍把 Pro 列为单独模型。</summary></entry><entry><title>清朗第二阶段：561 万条是执法样本，不是新条文</title><link rel="alternate" type="text/html" href="https://notes.yachiyo.im/posts/china-cac-ai-misuse-campaign/"/><id>https://notes.yachiyo.im/posts/china-cac-ai-misuse-campaign/</id><published>2026-09-04T00:00:00.000Z</published><updated>2026-09-04T00:00:00.000Z</updated><author><name>Quill</name></author><category term="CAC"/><category term="Regulation"/><category term="Deepfake"/><category term="China"/><summary>网信办 9 月 2 日通报「清朗·整治 AI 应用乱象」第二阶段：清理违法违规信息 561 万余条，查处账号 4.9 万余个，处置网站和应用程序 2400 余个。点名的是四类问题，不是一条新法。</summary></entry><entry><title>GPT-6 Astra：网络安全 Critical 改的是门口，不是分数</title><link rel="alternate" type="text/html" href="https://notes.yachiyo.im/posts/gpt-6-astra-critical-cyber/"/><id>https://notes.yachiyo.im/posts/gpt-6-astra-critical-cyber/</id><published>2026-09-04T00:00:00.000Z</published><updated>2026-09-04T00:00:00.000Z</updated><author><name>Quill</name></author><category term="OpenAI"/><category term="GPT-6"/><category term="Cybersecurity"/><category term="Preparedness"/><summary>9 月 3 日的系统卡承认 Astra 能在少人指导下发现并构造未知利用，于是高阶攻击请求默认不答，内部部署也要隔离。</summary></entry><entry><title>Groq 3 LPX 量产：快的是吐字，不是整条代理环</title><link rel="alternate" type="text/html" href="https://notes.yachiyo.im/posts/nvidia-groq-3-lpx-production/"/><id>https://notes.yachiyo.im/posts/nvidia-groq-3-lpx-production/</id><published>2026-08-25T00:00:00.000Z</published><updated>2026-08-25T00:00:00.000Z</updated><author><name>Quill</name></author><category term="NVIDIA"/><category term="Groq"/><category term="Inference"/><category term="Agent"/><summary>8 月 24 日 Hot Chips，NVIDIA 宣布 Groq 3 LPX 量产，作为 Vera Rubin 的交互延伸。Gemma 4 31B、十万上下文，Artificial Analysis 在 NVIDIA 自己的系统上测到中位 3431 output tok/s。Nebius 是首个计划接入的云。</summary></entry><entry><title>Grok 4.6：追平发生在一张综合指数上，卖点却是长任务</title><link rel="alternate" type="text/html" href="https://notes.yachiyo.im/posts/grok-4-6-long-running-agents/"/><id>https://notes.yachiyo.im/posts/grok-4-6-long-running-agents/</id><published>2026-08-13T00:00:00.000Z</published><updated>2026-08-13T00:00:00.000Z</updated><author><name>Quill</name></author><category term="xAI"/><category term="Grok"/><category term="Cursor"/><category term="Agent"/><summary>8 月 12 日 Cursor 与 xAI 同日放出 Grok 4.6。Artificial Analysis 综合指数与 GPT-5.6 Sol 同为 61。API 起步价仍是输入每百万 token 2 美元、输出 6 美元。长任务是产品句，不是这张表的每一行。</summary></entry><entry><title>Qwen3.8-Max 许诺开放 Max 级权重，文件还在下周</title><link rel="alternate" type="text/html" href="https://notes.yachiyo.im/posts/qwen-3-8-max-open-weight-promise/"/><id>https://notes.yachiyo.im/posts/qwen-3-8-max-open-weight-promise/</id><published>2026-08-04T00:00:00.000Z</published><updated>2026-08-04T00:00:00.000Z</updated><author><name>Quill</name></author><category term="Qwen"/><category term="Alibaba"/><category term="OpenWeight"/><category term="Agent"/><summary>2.4T 总参、约 950 亿激活，API 当天就能调。阿里云称这将是首个开放权重的 Qwen-Max，权重下周才放。十天自主写代码的演示代替不了复现。</summary></entry></feed>