AI 资讯 · 2026-10-08

模型开始按任务动态分配推理，超 10 万 Token 后，输入输出单价均涨 5 倍。    作者丨郑佳美    编辑丨岑   峰                                                                                                       刚刚，Anthropic 发布了 Claude Haiku 5.5。这次升级的幅度不小，尤其是在计算机操作方面，OSWorld 2.1 评测成绩从上一代的 15.7% 提升到了 72.4%。知识工作和复杂推理能力也有明显改善，而按照 Anthropic 公布的数据，新模型的平均运行成本还下降了约 75%。除此之外，Haiku 5.5 还加入了自适应推理机制，并支持 100 万 Token 上下文。过去，Haiku 主要用来处理摘要、分类、信息提取这些高频任务，优势一直是速度快、成本低。这次更新后，它能处理的工作明显复杂了不少，部分评测成绩甚至已经接近 Sonnet 5.5。不过，换到复杂编程任务，两者的差距依然明显。再看这次更新的推理机制和价格规则，Haik

 云栖大会，解开市场对阿里大模型的三个疑惑。    作者丨胡敏    编辑丨刘伟                                                                                                       今年云栖大会，变化格外明显。参加过多年的人，打开主论坛议程就能感受到，和往年很不一样。在吴泳铭发言之后，第一个登台分享的是千问大语言模型负责人刘大一恒；紧随其后的是ATH技术副总裁郑波，他负责视听等多模态模型。接在两场模型演讲之后，是平头哥高慧，与阿里云CTO李飞飞。云栖大会，芯片和云当然重要，阿里对云的投入不会减少，芯片也依然是它绕不开的底牌。但重要的不止是芯片和云，现在站在市场聚光灯下的是AI，大家关心的自然也包括阿里的模型。阿里的模型往哪走、做多大、能不能打，是眼下真正牵动市场神经的问题。那么，在云栖大会上，这些大模型负责人们究竟回答了哪些市场最为关注的问题？01Qwen下一代模型，规模有多大？这是市场最直接的悬念。此前一年多，业内对Scaling的质疑没停过。但到2026年年中，业界风向已明显调转

据雷峰网多方获悉，腾讯混元原视觉大模型算法中心负责人胡瀚创业项目已取得新进展：方向仍聚焦多模态，计划融资数千万美元，目标估值达数亿美金，远识资本担任本轮融资 FA。多位了解该项目的投资人告诉雷峰网，过去一个月，已有多家投资机构与胡瀚接触。商业化是机构与胡瀚沟通时反复追问的问题。大模型创业的融资窗口已不如前两年，仅凭技术团队和模型能力，越来越难打动投资人。在未来商业化上，胡瀚给出的答案也与多数创业团队不同，据一些投资人转述，他认为其模型未来可以出售给大型科技公司，例如DeepSeek。胡瀚此前负责腾讯混元视觉大模型算法中心。今年 3-4 月，其负责的业务线从薄列峰体系划入姚顺雨体系。今年 7 月，市场曾传出胡瀚计划离职创业的消息。

Google is launching a "universal" Gemini AI agent that can work across apps and devices in the background. The tool, announced as part of the Gemini at Work event on Thursday, will be available within the Gemini Enterprise app, allowing users to chat with the Gemini agent and assign it tasks from a single interface. […]

过去一年，“物理AI”几乎成了自动驾驶、车企共同争夺的下一张船票。世界模型、具身大模型、机器人数据……众多公司从不同切入口布局物理AI，试图成为新风口之下的主角，相比之下，如祺出行选择了一条不那么显眼的路径，不自己训练大模型，而是尝试成为模型背后的数据基础设施。实际上，如祺出行在2023年就开始思考，除了提供包括Robotaxi等出行服务，一家出行平台还能在AI产业里扮演什么角色。随后两年，如祺出行逐步完成数据闭环，一端连接驾驶员、Robotaxi和城市道路，另一端服务自动驾驶模型的训练与迭代。而具身智能的兴起，让这套能力有了新的用武之地。在如祺出行COO韩锋看来，自动驾驶与智能机器人需要的数据并不完全相同，因为前者主要在二维道路空间中完成感知、决策和运动，机器人则要进入三维空间，处理抓取、搬运和精细操作。但两者背后的数据处理流程高度相似，从采集、预处理到标注、质检和导出，底层工具链可以复用。然而，由于真实数据采集成本高昂，且周期漫长，越来越多智驾和具身智能公司借助世界模型和仿真平台生成合成数据，补充真实数据不足，以提升模型训练效率。但在韩锋看来，世界模型能力越强，真实数据的重要性并不

大模型并没有“吃掉”机器人学。    作者丨马晓宁    编辑丨岑   峰                                                                                                       9月27日，IROS 2026 将在匹兹堡开幕。重隔31年后，IROS 重回这座钢铁之城。作为全球规模最大的机器人学术会议之一，今年进入正式议程的论文达到 1933 篇。把这近两千篇论文铺开来看，它更像是一张 2026 年机器人研究的横截面：从强化学习、模仿学习，到路径规划、视觉感知、运动控制、灵巧操作，再到人形机器人、触觉和大模型，过去几年机器人领域几乎所有重要技术路线，都同时出现在这张图谱里。但真正值得注意的，并不是 Humanoid、VLA 或 World Model 这些更容易获得产业和媒体关注的关键词。对 1933 篇论文进行多标签梳理后可以看到，Robot Learning / Embodied AI 相关论文约有 809 篇，Navigation / Planning 564 篇，Percepti

Google has released an experimental note-taking app that can transcribe meetings and audio files entirely offline, as reported earlier by TechCrunch. The app, called Google AI Edge Foresight, is free to use and runs on macOS using the company's on-device EmbeddingGemma 2 model. Similar to AI note-taking apps like Granola and Wispr Flow, Foresight summarizes […]

Three fired OpenAI safety researchers dispute allegations of mishandling sensitive information, warning in an open letter that their dismissals are creating a chilling effect on the company’s AI safety culture.

With GPT-5.6, GPT-6 Astra, and GPT‑Image‑2.5, Pollo AI helps creators turn bold ideas into detailed images and cinematic video ads.

Across recruiting, engineering, and operations, Oracle turns specialist knowledge into fast, repeatable workflows with ChatGPT Work and Codex.

Anthropic's updated usage policy explicitly prohibits users from repeatedly abusing Claude in extreme cases, though ordinary frustration and criticism are still allowed. The new rules also address election interference, deceptive campaigns, weapons software, and surveillance.

Google is turning Gemini into an AI agent that can plan, execute tasks, and work across business apps and systems. The agent can delegate work to subagents, use multiple AI models, and even gets its own workplace identity, complete with an email address.

OpenAI disrupted two AI-enabled influence operations that used false-front journalists and a think tank to spread geopolitical messaging.

AI agents can now write CUDA kernels that outperform PyTorch—but proving those speedups are real is the harder problem. I put them to the test on an NVIDIA DGX Spark and found that benchmark design matters just as much as the code. The post AI Agents Beat PyTorch: Writing Faster CUDA Kernels appeared first on Towards Data Science.

Goodfire just launched what it says is a cheaper way to keep AI agents in check: Instead of paying a second AI to read everything an agent does, its monitors peek inside the model while it works and only call in backup when something looks fishy.

Zach Yadegari, the teen co-founder of popular Cal AI calorie tracking app, has launched a new personal AI agent startup that competes with Instinct, Muse, and Bee.

Natura’s $99 Interface smart ring lets you summon AI agents with the press of a finger to complete tasks, capture thoughts, and control devices — while doubling as a health tracker.

Ben Affleck is going viral for his deep knowledge of AI, from neural networks and transformers to open weights. The actor, who sold his AI filmmaking startup to Netflix earlier this year, is proving he's more than just a Hollywood star.