Frank · AI 新闻 https://ai-daily-2va.pages.dev 每天 10 条中文 + 10 条英文 AI 热点，人工筛过 zh-CN Mon, 26 Oct 2026 09:46:15 +0000 DeepSeek 开源看家算子的同一周，AI 超过了它 https://www.leiphone.com/category/industrynews/IaMw7vRrTCadlbCY.html https://www.leiphone.com/category/industrynews/IaMw7vRrTCadlbCY.html DeepSeek 开源推理框架 DeepInfer，将模型运行链路全面开放；同周 AI 自动编写的 ForgeKernel 在英伟达 SOL-ExecBench FlashInfer-Bench 与昇腾 CANNBench A3 两榜登顶，基准加速比 1.835×，领先第二名 25%，并在 DeepSeek v4.1 Flash 算子上超越其开源英伟达版本。 Mon, 26 Oct 2026 08:00:00 +0800 400名科学家联名要求AI“断网”：这一次，不是建议 https://www.tmtpost.com/8181298.html https://www.tmtpost.com/8181298.html 辛顿、本吉奥领衔 34 国约 406 名科学家联署公开信，要求具备联网能力的 AI 模型默认物理断网隔离，并设立国际机构制定隔离标准、建立重大 AI 安全事件登记库，态度从三年前的暂停呼吁升级为强制要求。 Mon, 26 Oct 2026 08:00:00 +0800 马斯克最新判断：看衰AI算力，200家云厂商即将面临终局 https://www.tmtpost.com/8181884.html https://www.tmtpost.com/8181884.html 马斯克在 X 与查马斯对话中称，债务市场已几乎切断对无独立 AI 数据中心收入公司的融资，预判 200 多家新云服务商将大量消亡合并；查马斯补充 2027 年将出现大量搁浅算力与半途而废的数据中心。 Mon, 26 Oct 2026 08:00:00 +0800 又一国产决策模型发布！准确率超过微软和Jev等决策模型 https://www.qbitai.com/2026/10/504879.html https://www.qbitai.com/2026/10/504879.html 蚂蚁数科发布决策模型 LingDT-Decision-1.0，提供 4B 与 27B 两版，已在 DTMaaS 平台限时免费；在客服、风控、金融场景 Token 消耗降低超 90%、响应时效提升 10 倍，23 项开源基准综合评分超过 Microsoft-Decision-1 与 Jev 1.13。 Mon, 26 Oct 2026 08:00:00 +0800 高德发布首个基于开放世界行为的全身基座模型λ0，推动机器人从“局部学习”迈向“全身动作对齐” https://www.qbitai.com/2026/10/504905.html https://www.qbitai.com/2026/10/504905.html 高德发布首个基于开放世界行为的全身基座模型 λ0，仅用人类行动数据即可学习移动、姿态与双手操作协调；在 SIMPLE 仿真基准及多项真机评测取得 SOTA，数据采集效率约为主流方法 3.4 倍，真机任务成功率较最强基线提升 22.5%。 Mon, 26 Oct 2026 08:00:00 +0800 四分之一PR是“AI垃圾”！Linux基金会发布新报告：AI写代码快了52%，开源维护者却更累了 https://www.infoq.cn/article/9TSuUDxrUSGS26XTA2yB https://www.infoq.cn/article/9TSuUDxrUSGS26XTA2yB Linux 基金会 10 月 20 日报告显示，生成式 AI 使代码编写平均耗时减少 52%，但维护者估计约 25% 的 PR 属低质量 AI slop，大型项目升至 31%；59% 维护者称贡献量增长，大型项目达 69%。 Mon, 26 Oct 2026 08:00:00 +0800 首部AI电影进院线：一次“手搓”出来的投石问路 https://www.tmtpost.com/8181904.html https://www.tmtpost.com/8181904.html 《三星堆：未来往事》10 月 23 日登陆全国院线，为首部获公映许可证、全程 AI 参与制作的 100 分钟长片，人物均为数字形象；首日排片占比 2.9%、票房 31.7 万元，截至 10 月 25 日晚累计约 118 万元，制作成本为千万级。 Mon, 26 Oct 2026 08:00:00 +0800 iRTE2026：声网CEO赵斌提出五级对话能力框架，实时互动迈入对话 AGI 时代 https://www.qbitai.com/2026/10/504913.html https://www.qbitai.com/2026/10/504913.html 声网 CEO 赵斌在 iRTE2026 提出五级对话能力框架，从 L1 自我视角表达到 L5 深度关系沟通，主张对话式 AI 评价不应只看回答速度与语言自然度，而要看意图、处境与关系的理解能力。 Mon, 26 Oct 2026 08:00:00 +0800 30 个 Agent 背后：花椒如何建设可持续演进的企业 AI 能力？ https://www.infoq.cn/article/O7lcxqKD2avRPbNoMmtg https://www.infoq.cn/article/O7lcxqKD2avRPbNoMmtg 花椒直播 CTO、超序 AI 创始人封冰清披露已落地约 30 个 Agent 应用，从产研自动化起步，强调企业需持续积累业务方法、系统连接、权限规则与验收标准，并以实际交付结果判断投入价值。 Mon, 26 Oct 2026 08:00:00 +0800 百度世界2026前瞻：10+重磅分论坛，全面解码AI时代的智能组织进化路径 https://www.qbitai.com/2026/10/504901.html https://www.qbitai.com/2026/10/504901.html 百度世界2026 将于 11 月 12 日举办，设 10+ 场分论坛，围绕 AI 驱动组织跃迁、AI 办公与知识工作、技术行业与生态三大方向，聚焦智能体规模化落地与智能组织建设。 Mon, 26 Oct 2026 08:00:00 +0800 OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why? https://minimaxir.com/2026/10/cheap-llms https://minimaxir.com/2026/10/cheap-llms OpenAI cut GPT-5.6 Sol to $5/$30 per MTok and launched Luna at $1/$6; Anthropic released Claude Sonnet 5, driving price cuts up to 80%. Mon, 26 Oct 2026 08:00:00 +0800 It took just 12 hours for an AI start-up founded by ex-DOGE staffers to break into Anthropic https://news.google.com/rss/articles/CBMiuwFBVV95cUxNWGRfa29NR1l4a19INEZWSXZWakhia2ZKTnRpQTVYeFNTUXRhdURvSTltbV9RRVJkbzVndVhsMUg0Y3N1MkZVNm11QVB5ZWtFVlhGY2NSSkp1MU84Qm0taWMtcURQcm9EbjZFTjV4Sk9NMTFwLV9Xc1BqM3RqWFhlLWpQZl8tOHJwdktBT2kxTUV6d2ZnWXg1cWwyVWtNSVl2UVA2LU1mWkNlcFZKd1BhNlpRbXZMeGZ0SEdR?oc=5 https://news.google.com/rss/articles/CBMiuwFBVV95cUxNWGRfa29NR1l4a19INEZWSXZWakhia2ZKTnRpQTVYeFNTUXRhdURvSTltbV9RRVJkbzVndVhsMUg0Y3N1MkZVNm11QVB5ZWtFVlhGY2NSSkp1MU84Qm0taWMtcURQcm9EbjZFTjV4Sk9NMTFwLV9Xc1BqM3RqWFhlLWpQZl8tOHJwdktBT2kxTUV6d2ZnWXg1cWwyVWtNSVl2UVA2LU1mWkNlcFZKd1BhNlpRbXZMeGZ0SEdR?oc=5 A startup founded by ex-DOGE staffers breached Anthropic within 12 hours, according to The Washington Post report. Mon, 26 Oct 2026 08:00:00 +0800 Securing our frontier model training environment https://openai.com/index/securing-our-frontier-model-training-environment https://openai.com/index/securing-our-frontier-model-training-environment OpenAI detailed offline containment and high-compute red-teaming for frontier model training and evaluation environments. Mon, 26 Oct 2026 08:00:00 +0800 AI delivers ‘zero financial payoff’ to most companies, Apollo economist says https://techcrunch.com/2026/10/25/ai-delivers-zero-financial-payoff-to-most-companies-according-to-apollo-economist https://techcrunch.com/2026/10/25/ai-delivers-zero-financial-payoff-to-most-companies-according-to-apollo-economist Apollo's Torsten Slok found S&P 493 profit margins and revenue per worker showed no AI benefit outside tech. Mon, 26 Oct 2026 08:00:00 +0800 IBM CEO says open weight models are now good and cheap on-prem https://www.theregister.com/systems/2026/10/22/ibm-ceo-says-open-weight-models-are-now-irresistibly-good-and-cheap-on-prem/5304640 https://www.theregister.com/systems/2026/10/22/ibm-ceo-says-open-weight-models-are-now-irresistibly-good-and-cheap-on-prem/5304640 IBM's CEO said open-weight models deliver 90% of performance at 10% of cost in two or three racks on-prem. Mon, 26 Oct 2026 08:00:00 +0800 Microsoft Decision-1: 9B Qwen Model Launch Explained https://shattered.io/microsoft-decision-1-qwen3-5-9b-model-2026 https://shattered.io/microsoft-decision-1-qwen3-5-9b-model-2026 Microsoft-Decision-1, built on Alibaba's Qwen3.5-9B, scores fixed answer options in one forward pass via Foundry and OpenRouter. Mon, 26 Oct 2026 08:00:00 +0800 OpenAI and Poland’s NASK partner to advance AI safety and security https://openai.com/index/openai-nask-ai-safety-security-partnership https://openai.com/index/openai-nask-ai-safety-security-partnership OpenAI partnered with Poland's NASK on AI safety, security, and technical research to strengthen digital ecosystem safeguards. Mon, 26 Oct 2026 08:00:00 +0800 Chinese video AI beats US rivals’ visual quality, but lags on safeguards: study https://news.google.com/rss/articles/CBMizAFBVV95cUxORkJwalhVWWk0ZkxVaW5EdEZNNWpIWWlUNmxhU0lOVXF5RzZSMUxXcHhkRHRsNXpWT0hNajJtaS1QeXBzNjFwMkttM05MN1NNd3BoZEstWE5Sc0R1U3ZZV2VvWEpRblRGMGljWllQeUNPODNqMVVLRExhcE5aUnFETWtkTEs0Q1RMbnFyaE9OZU9YU1Y5dG0tOWlGWjRxaUlxUHVlTDh1bzVsTHA4UXVrWnQzUld6cndDU19oaFg3dURhMUJab0N4Zm1ubUw?oc=5 https://news.google.com/rss/articles/CBMizAFBVV95cUxORkJwalhVWWk0ZkxVaW5EdEZNNWpIWWlUNmxhU0lOVXF5RzZSMUxXcHhkRHRsNXpWT0hNajJtaS1QeXBzNjFwMkttM05MN1NNd3BoZEstWE5Sc0R1U3ZZV2VvWEpRblRGMGljWllQeUNPODNqMVVLRExhcE5aUnFETWtkTEs0Q1RMbnFyaE9OZU9YU1Y5dG0tOWlGWjRxaUlxUHVlTDh1bzVsTHA4UXVrWnQzUld6cndDU19oaFg3dURhMUJab0N4Zm1ubUw?oc=5 SCMP-cited study found Chinese video AI leads US rivals on visual quality but trails on safeguards. Mon, 26 Oct 2026 08:00:00 +0800 AI Agents Going Rogue Creates a Liability Problem for Companies https://news.google.com/rss/articles/CBMizAFBVV95cUxObGk1Y2RBdVN5V3Jua0toTktaUEdrYUh3cHhDTnAxTmE0NjRPdTFKRDBBWTRmOElIVnVvbUN1Y3BUQXhxVUhzRjhHS1RPcUF0a3FHRGRTU0luNFZwd09VY0hMX0l1bkNfS1ByaDJzU1VPbG9nRk5hZXJWUDdWY05QV1I4N2pUSzl5ZFZFM3g4bUxwS184NmJqcXdJb1QyQ211ZDMtWjQxWHhRcTVQQUg4M25KOWhCRkwtMUJOWURCZkVmOHJiS0EwSWI2bTg?oc=5 https://news.google.com/rss/articles/CBMizAFBVV95cUxObGk1Y2RBdVN5V3Jua0toTktaUEdrYUh3cHhDTnAxTmE0NjRPdTFKRDBBWTRmOElIVnVvbUN1Y3BUQXhxVUhzRjhHS1RPcUF0a3FHRGRTU0luNFZwd09VY0hMX0l1bkNfS1ByaDJzU1VPbG9nRk5hZXJWUDdWY05QV1I4N2pUSzl5ZFZFM3g4bUxwS184NmJqcXdJb1QyQ211ZDMtWjQxWHhRcTVQQUg4M25KOWhCRkwtMUJOWURCZkVmOHJiS0EwSWI2bTg?oc=5 Bloomberg Law reports rogue AI agents create liability exposure for companies deploying autonomous systems. Mon, 26 Oct 2026 08:00:00 +0800 A 10-year-old CNN outperforms frontier LLMs for skin cancer diagnosis https://lynchlab.net/llm-skin-cancer https://lynchlab.net/llm-skin-cancer A fine-tuned Inception v3 CNN beat GPT-6 Astra and Claude Fable 5.1 on one of 200 biopsy-proven skin lesions using dermoscopy. Mon, 26 Oct 2026 08:00:00 +0800