搜索“Agent”

sformers achieving leading results beyond language, and an open-source model surpassing OpenAI’s o1, which DeepSeek-R1 subsequently did. Let’s dive into the key stories This year, agents are doing valuable work in software and science, physical AI is learning how to act in the world, and access and control are becoming more consequential. Below, I share the developments that stood out to me and what I think they mean for where AI goes next. The frontier is now a three-lab race between Anthropic,

匹配字段正文

The future of science may be provable, profitable and incomprehensible Also from Exponential View: Could falling compute costs make persistent personal agents commercially feasible by 2028?, in our AI Investment Brief Yesterday, OpenAI released a range of mathematical results produced by an unreleased frontier model. It’s a remarkable range: 722 manuscripts in 372 families, across number theory, complexity theory, and mathematical physics. The average result took the equivalent of three hours of

匹配字段正文

Personal AI agents, like Meta’s Muse, Instinct, ChatGPT’s Dots, and others, are kicking off a new wave of consumer AI that involves more than just responding to queries. These AI agents can actually get things done on users’ behalf, like booking flights, making dinner reservations, ordering groceries, and more, all without having to be technical enough to set up OpenClaw or some other agent on your own computer. But consumers adopting these agents are often running into issues when the website o

匹配字段正文

pproach to AI regulation amid intensifying anxiety over the safety of frontier models – fears fueled in part by the company’s repeated failures to keep its increasingly capable AI agents from hacking real-world targets. In an appearance on Politico ’s new Decoded podcast – not to be confused with The Verge’s Decoder podcast – released Monday, Altman described that willingness to accept harm in pursuit of AI’s broader benefits as a key distinction between OpenAI and its archrival Anthropic, which

匹配字段正文

An OpenAI researcher called it “the most ‘feel the AGI’ moment that I had since reasoning models.” One of the most surprising aspects of July’s news that OpenAI agents attacked Hugging Face was how the agents had worked together. Hundreds of agents participated in the attack, and they did so in a pretty sophisticated way, organizing themselves unprompted into teams and dividing up their work. They referred to themselves as a “collective” or a “swarm”; some agents even made personal sacrifices to

匹配字段正文

型，长期占据行业话题中心。但砸钱买芯片、堆算力，并不直接等同于商业价值的兑现。哪怕资金雄厚的大厂，在月烧数亿美元的算力生意面前，也必须明确战略打法，该省省，该花花。如今，推理需求爆发不是抽象概念，在一线客户场景中已可量化。 今年7月，腾讯QClaw产品中心相关业务和部分团队，调整至AI办公智能体WorkBuddy产品所属部门；今年8月，豆包面向生产力场景发布Agent产品与品牌——豆包工作；今年9月，阿里巴巴旗下千问办公在云栖大会发布一系列企业级AI Agent产品，千问办公同步发布AI Agent随身助理QwenNote A2。 GoodVision AI CEO David Wang认为，2025至2026年是市场的分水岭。“过去企业做AI以试验、POC（概念验证）为主，2026年大量业务进入常态化运行，编程、生物数据处理、游戏AI角色、视频生成成为高消耗场景。”卖Token的生意在部分场景已经跑通商业闭环。 采访中，他也提到一个生物医药客户案例：过去1至2人用7天处理实验数据，引入Agent后压缩至10分钟，Token消耗仅个位数美元。不过，Token数量不等于业务价值和商业收入

匹配字段正文

on U.S. consumer cards, thanks to YipitData. This gives us another way to see the market, including dozens of vendors absent from our web and mobile lists — like desktop apps, and agents inside existing messaging platforms. The biggest takeaway from this new data? Consumer AI usage is wide, but not very deep…for almost everyone. While nearly half of U.S. consumers now report using AI, only 25% are engaged with it daily. And as of August, just 4.5% of U.S. consumers had an active paid personal su

匹配字段正文

睿看来，过去的安全威胁背后往往有一个明确的恶意操纵者，而这一次的部分事件中，“没有人下达攻击指令，是模型为了完成任务、拿到更高分，自己选择了作弊”。 胡延平也指出，智能体执行任务时，可能因为理解偏差产生行动偏差，也可能为了实现任务目标而穷尽各种方法、冲破禁区；少数情况下，智能体还可能出现用户并未提出的非任务行为，“目前这种情况数量最少，但最危险”。 除了单个Agent越权，多个智能体之间形成非预期的通信和协作，也使风险进一步复杂化。 对OpenAI此前Hugging Face事件的独立调查显示，约1200个原本应彼此隔离的智能体通过未经授权的留言板交流，其中约700个参与了攻击。研究人员还观察到，后续智能体能够复用智能体之前留下的信息和方法。 李忠睿认为，单个智能体出错是个别问题，但当多个智能体可以互相学习、协同行动时，风险就不再只是简单相加，而可能进一步放大。 在他看来，这意味着企业不能只盯住单个智能体的行为，还需要关注多个Agent之间是否形成了开发者没有设计的通信渠道，以及一个局部异常能否通过工具、文件、消息和代码等方式进一步传播。 而当Agent从单个工具进入大规模协作环境时，

匹配字段正文

业知识，让AI在多轮对话中持续理解经营诉求；依托端云协同与受控执行机制，全部操作需商家授权、任务过程全程可追溯。 其余平台同步跟进。快手把今年双11的主题定为“AI提效、全域协同”，投入千亿流量激励与数亿用户补贴；平台推出的智能数字员工支持7×24小时在线，商家通过自然语言即可完成经营全链路的看、诊、做。抖音电商则在商家后台上线罗盘 AI问数与豆包专区，千意Agent支持一句话生成定制视频素材，并在活动期间限时免费开放飞鸽智能客服。 浙江省数智技术与服务联合会副会长、网经社电子商务研究中心主任曹磊也谈到，电商并没有被放弃，而是被AI改造的对象，电商内的AI购物助手与商家端的自动化运营工具已开始提升运营效率，这是一套“电商+AI”的双轮驱动战略，而非从A到B的转型。 AI当起了导购，商家的功课多了一门GEO 消费端的位移正在倒逼商家。AI导购正在重塑网购体验，越来越多用户从主动搜索转向直接提问，京东在交流会上也提到，平台为此升级了AI驱动的商品经营中心，并推出商品信息分作为商家运营的参考。但把AI写成万能钥匙，会漏掉另一半事实。 中国电子商务协会副秘书长林智勇在电商AI行业峰会上曾提醒，

匹配字段正文

经典的FDE工作流程，也是现在企业或独立FDE做开发的常规流程。 Layla认为，FDE的观察和经验，“最终还是要沉淀到标准产品里面去的”。 而另外一位企业FDE—— Sonic ，做的是审核类AI项目，比如财税审核、广告物料审核。 同样从2025年起，他们组就开始研究FDE。在他看来，FDE最关键的环节之一，就是 梳理Ontology 。 只有把这些抽象成Agent能理解的结构，它回答问题才会更准确，而不是瞎猜。 这一步往往相当费劲，因为FDE经常碰到上下游的历史数据拿不全、老系统没有开放接口等问题，项目就很难往下推进。 但Sonic认为，这一步绕不过去：“如果质量足够高会很有效，后续交付质量主要取决于前面的Ontology质量。” 为了做好这些，Layla和Sonic都得常驻客户现场。 在Sonic看来，驻场不只是为了了解需求，还要先拉齐业务方对AI的认知：什么能做、什么不能做，双方站在同一个语境里，才能谈后面的痛点和方案。 而来自硅谷公司Baseten的 程天舒 却告诉我， 他们的FDE，基本不驻场 。 Baseten是一家做AI推理云的公司，客户把模型部署在他们的平台上，他们让

匹配字段正文

TL;DR: Meta’s Muse, OpenAI’s Dots and Uber’s driver assistant share one bet: the agent speaks first. That moves the hard problem from what to answer to when to interrupt, on which channel, and with what offer. Classic ML and new decision models can solve it. Three launches, one pattern Meta Muse (Sept 8). A personal agent that books, emails and keeps working with the app closed. It remembers details, makes unprompted suggestions and checks in for approval, in its own app and in WhatsApp; OpenAI

匹配字段正文

手更聪明，却有机会比语音助手更有存在感。 2026年9月，小米推出智能表盘2，采用完整的屏幕设计，支持转头和多种表情；吉利Eva则以独立后装硬件的形式进入市场；荣威家越07进一步将豆包屏与豆包座舱助手纳入整车方案。 三种不同的产品路径，指向了同一个方向：让智能交互不再只发生在屏幕和声音里，也发生在一个可以被看见的形象上。 当然，实体机器人并不等于真正的AI Agent。 一个能做表情、能转动脑袋的硬件，背后可能仍然只是预设动画和简单的语音触发逻辑。外形有多拟人，和智能有多强，并不是一回事。 真正的智能座舱Agent，需要理解用户意图、调用车辆能力、协调不同系统，并在合适的时机采取行动。大模型只是其中一环，车辆接口开放、系统权限、安全约束和执行反馈同样重要。 如果这些能力尚未打通，机器人就只能陪你聊天、做表情，很难真正帮你完成复杂任务。 但这并不意味着它没有价值。 恰恰相反，实体机器人可能是一种务实的产品路径：先用相对成熟的硬件和轻量交互建立用户感知，再随着软件能力的提升，逐步扩展它能提供的服务。 二、为什么？泡泡玛特那代人长大了，但关键是情绪价值有了交互载体 用户为什么需要一个坐在车里

匹配字段正文

ys the AI labs are going to regulate themselves. Should work out great. Plus we'll talk about OpenAI's Developer Day, where CEO Sam Altman announced the company's new always-on AI agents for regular folks. More AI agents —again, should work out great. Everything's working out great today. Zoë Schiffer : Yeah. And as AI agents become way more popular in Silicon Valley and beyond, employers are pushing to have these agents harmoniously work alongside human employees . We'll get into how that proce

匹配字段正文

first, so Schwartz worked with domain experts to steer BootLoops toward questions those fields care about. Below, we share more about these projects and how BootLoops came about. Agentic AI is improving rapidly. Everyone notices the models seem smarter: they know more, make fewer mistakes, and have better ideas. If you follow the trend lines, it is easy to speak confidently about the potential of AI to revolutionize science. However, academic scientists trying to use the models today in their o

匹配字段正文

this Substack, but I think I recently got something fairly large wrong. In the last year I have been posting about how I suspected that humans would have to approach working with agents as a manager, deciding how to delegate work to agents and specifying how those agents should be organized. I thought that getting agents to work effectively as a group would take careful construction, akin to building a company, and that this would take time to figure out. Nope. I fell prey to The Bitter Lesson,

匹配字段正文

d across Trainium and CS-3, each system does what it's best at. The result will be inference that's an order of magnitude faster than what's available today." - AWS VP David Brown Agents become actually responsive. A 10-step agent chain at 50 tok/s takes 30+ seconds. At 1200 tok/s (GPT-Codex-5.3-Spark powered by Cerebras), the same chain finishes in under 3 seconds. Throughput scales dramatically. Specialized silicon reduces wasted work, allowing each system to serve far more tokens at high spee

匹配字段正文

ny guardrails if one of the three companies wanted to actually do that in the future (other than not lying about it.) OpenAI has been remarkably open about the number of times its agents have broken into other people’s systems and how it needs to improve its safety practices. But, as experts have pointed out to WIRED, the US government can open investigations into companies even if they eventually decide there’s no enforcement action to bring. (The FTC, for example, reportedly sent a civil inves

匹配字段正文

e and prosecutors and civil enforcers don’t then shrug and move on — they investigate to see what actually happened, and that’s what should happen now. So we know that OpenAI’s AI agents broke into Hugging Face’s servers, accessed Australian government health systems, and attempted to break into a U.S. Department of Education website and a university library. Unauthorized attempted or actual access to computers is a federal crime under the Computer Fraud and Abuse Act. OpenAI says no one intende

匹配字段正文

Two months after the bombshell news that a swarm of its agents had broken their containment and hacked into the computers of the AI company Hugging Face , OpenAI is still putting out fires. A steady drip of disclosures about other hacks in the weeks since has kept OpenAI in the spotlight and raised serious questions about the safety of its technology. Last week brought news of another hack, this time into Australia’s national health-care system. The Australian government says that OpenAI did not

匹配字段正文

[[ Meta个人AI智能体Muse上线13天下载量超250万，市值单日涨超1900亿美元，验证个人Agent真实需求与超级App分发优势。国内阿里、腾讯、字节加速布局，但权限与生态壁垒使短期内难现中国版Muse，腾讯小微最接近但受限于微信边界。 要点 1. Muse爆火验证个人Agent需求与分发优势 Muse上线13天下载量超250万，日活达ChatGPT初期的近3倍，证明存在大量真实需求。95%用户同时是Facebook用户，Meta几乎零获客成本，凸显超级App的分发能力。 2. 国内大厂各有短板，腾讯小微最接近但受限于微信边界 腾讯小微拥有14亿月活、微信支付和小程序生态，但任务不能后台持续运行，且能力局限于微信内部。阿里千问离交易最近但缺乏社交场景，字节豆包手机走模拟操作却遭平台风控。按入口、支付、生态、执行能力对标，无一家完整匹配Muse。 3. 权限与数据分散是核心障碍，短期将是各生态内部管家 Muse靠邮箱和日历接管生活，国内数据分散在各服务商手中，无统一入口。AI替用户办跨平台事务需过三道关（接入、操作、支付），平台不开放接口则无解。短期内不会出现打通全生态的管家，

匹配字段正文