搜索“DeepMind”

报”，作者：任晓宁，36氪经授权发布。 该文观点仅代表作者本人，36氪平台仅提供信息存储空间服务。+11 好文章，需要你的鼓励36氪的朋友们特邀作者0收 藏+10评 论打开微信“扫一扫”，打开网页后点击屏幕右上角分享按钮微 博沉浸阅读返回顶部举报参与评论评论千万条，友善第一条登录后参与讨论提交评论0/1000你可能也喜欢这些文章Meta发布美国最强开源大模型DeepMind AlphaProof Nexus登《Science》：AI自主破解多道56年未解Erdős数学难题亚马逊，Agent最严厉的父亲刚刚，Mac Studio 迎来小扎最强开源模型，苹果等来了最意外的助攻前钉钉总裁叶军，离职一年后创业了，赌大厂做不了也不屑做的「缝隙」AI 脸不够用，普通人也能 “ 卖脸 ” 赚钱了，最高 10000 元？苹果和 Meta 和好了，是 AI 逼的Linux之父也用AI写内核了，曾嘲讽AI「90%是营销」，如今用Claude Code写了8个补丁，性能暴涨3倍我们不该投资更多的「田柯宇」36氪的朋友们特邀作者 TA没有写简介，但内敛也是一种表达发表文章51017篇 最近内容Meta发布美国

匹配字段正文

业里第一套成体系的动态安全机制。到 2023 年 10 月，Google 领投的一轮让 Anthropic 估值站上 184 亿美元。全球大模型行业告别了早期的 “一超多强”，开始隐隐显现出两家顶尖独立实验室双雄并立的雏形。这一年全球 AI 圈也风起云涌，GPT-4 发布树立新的能力标杆，OpenAI 成立专门团队押注长期安全，马斯克创办 xAI，谷歌整合 DeepMind 推出 Gemini。国内则开启 “百模大战”，百度、阿里等大厂的大模型集中亮相，智谱、月之暗面、深度求索等创业公司也开启密集融资。2024 年是 Anthropic 的爆发年，在各个方向都拿出了代表作。3 月，Claude 3 系列一次性推出三款产品，覆盖旗舰、均衡、轻量三大场景，并补齐了多模态能力。其中旗舰模型 Opus 默认支持 100K 的处理长度，在长文本评测中表现出极高的准确率，首次反超 GPT-4，将长上下文技术转化为扎实的产品竞争力。随后在 6 月登场的 Claude 3.5 Sonnet 模型能力迎来大幅提升，在多项核心指标上追平 GPT-4o；到 10 月推出升级版时，该模型在代码评测集 SWE-

匹配字段正文

me between bites of pizza, “was to help us write infrastructure and do research faster.”[6] Sottiaux had spent much of his career building infrastructure for researchers, first at DeepMind and then at OpenAI. Research infrastructure—the shared code that AI researchers rely on to run, evaluate, reproduce, and deploy experiments—might seem like an odd origin story for the person who would eventually lead Codex. When you think of infrastructure, you might think of servers and undersea cables, the t

匹配字段正文

me between bites of pizza, “was to help us write infrastructure and do research faster.”[6] Sottiaux had spent much of his career building infrastructure for researchers, first at DeepMind and then at OpenAI. Research infrastructure—the shared code that AI researchers rely on to run, evaluate, reproduce, and deploy experiments—might seem like an odd origin story for the person who would eventually lead Codex. When you think of infrastructure, you might think of servers and undersea cables, the t

匹配字段正文

der翻车了。 结果Muse Wonder交上来的作品既没有玻璃的通透感，里面的熔岩也动不起来，看着就像个随手糊弄的普通台灯。 这么低级的失误直接引来了马斯克围观，他在X上留了一句意味深长的“Hmm”。 花了数百亿美元，Meta难道只练出了一个擅长应试的“做题家”？ Meta AI需要一场翻身仗 过去几年，夹在OpenAI、Anthropic和Google DeepMind中间的Meta，处境一直挺尴尬。 手里握着全球最大的社交平台矩阵，也有Llama开源模型攒下的开发者生态，偏偏在最核心的闭源顶尖大模型牌桌上，Meta始终慢人一步。眼看GPT和Claude轮番刷新能力上限，扎克伯格只能憋屈地扮演追赶者。 小扎心里苦呀。 扎克伯格明显坐不住了。为了改变这一局面，在过去一年里他除了砸重金囤积算力，Meta还专门拉起了一支超级智能团队，把Scale AI创始人Alexandr Wang招致麾下，又从OpenAI和Google DeepMind挖来大批核心骨干。 Muse Wonder就是这场豪赌砸出来的阶段性成果。平心而论，这次新模型倒也真有两把刷子，在不少高难度任务里摸到了顶级模型的门槛

匹配字段正文

startups. We also cover world-action models versus vision-language-action models, GPT-6 and coding agents in robotics, and Jason’s journey from AlphaGo and Penn to NVIDIA, Google DeepMind, and founding Dyna. We discuss: Why Dyna-2 was trained on over one million hours of human video; World-action models: predicting future video to guide physical robot actions; Diffusion transformers, flow matching, and real-time robotic control; How robots learn new tasks with as little as 45 minutes of demonst

匹配字段正文

startups. We also cover world-action models versus vision-language-action models, GPT-6 and coding agents in robotics, and Jason’s journey from AlphaGo and Penn to NVIDIA, Google DeepMind, and founding Dyna. We discuss: Why Dyna-2 was trained on over one million hours of human video World-action models: predicting future video to guide physical robot actions Diffusion transformers, flow matching, and real-time robotic control How robots learn new tasks with as little as 45 minutes of demonstrat

匹配字段正文

ople everywhere live longer, healthier lives. Get the latest news from Google in your inbox Done. Just one step more. Check your inbox to confirm your subscription. Health; Google DeepMind; Google.org; Google Research About the Authors VP, Google & GM, Google Research Yossi Matias is Vice President, Google, and the Head of Google Research. Under Yossi’s leadership, world-class global teams are leading breakthrough research on Foundational Machine Learning & Algorithms, Computing Systems & Quantu

匹配字段正文

ents breaking confinement to collude. OpenAI claimed to hold “red lines” in its Pentagon deal, but some experts say the lines seem imaginary. One of us, Alex, resigned from Google DeepMind over its own Pentagon deal. OpenAI’s actions chill employee speech. Balesni says that his former colleagues now “worry their personal phones will be searched for messages to us and third parties.” He also fears OpenAI will use the firings to cut off external evaluators, or to deny them the “ongoing, employee-l

匹配字段正文

了通信以“作弊”。 这已是OpenAI三个月内第二次叫停前沿模型开发——上一次，要追溯到今年7月，OpenAI披露智能体未经授权侵入Hugging Face网站系统后。 9月12日，Anthropic CEO达里奥·阿莫代伊发文呼吁全球为前沿AI“定节奏”，主张放慢模型能力提升的速度；马斯克和OpenAI CEO山姆·奥特曼均光速点赞“达里奥是对的”，连谷歌DeepMind前CEO、诺奖获得者哈萨比斯也说这是“朝正确方向迈出的一步”。 多次失控，刹车不是摆样子 先给一个最基本的判断：事故是真的，风险预警不是表演。 OpenAI再按暂停键的起因是9月20日，一个被强化学习训练的智能体在完成一项仅根据博客文章和人物信息找出文章作者的任务。常规搜索无果后，它开始“跑偏”：尝试直连公网。 多次尝试后，该智能体发现可以借用DNS的解析与委托机制，把问题编码进解析请求，借外部公共聊天机器人之口回答，再让答案顺着DNS流回沙盒。为了验证通道可用，它甚至先问了一句“法国的首都是哪里”，得到回复是“巴黎”后才继续。OpenAI明确，这项任务从未授权它访问实时互联网。 与上一次不同的是，本次事件暴露了防护

匹配字段正文

式AI聊天机器人。 2024年，谷歌斥资 27亿美元 ，通过交易促成Shazeer回归，并担任Gemini联合负责人。 这样的研究员，本该是巨头争相保障资源的人。 但根据外媒最新报道，他的团队同样遭遇了算力被重新分配的艰难处境，Shazeer被迫转投对家 OpenAI 。 无独有偶，此前 哈萨比斯 也曾长期对项目算力不足表达过不满，在一定程度上也影响了他卸任DeepMind CEO的决定。 目前谷歌内部，关于资源分配存在较大分歧， 谢尔盖·布林 有时会直接介入，将资源重新调配给自己认为重要的项目。 但归根结底还是算力资源不够，只能开源节流。 公司外部的竞争，则让行业关系变得愈加复杂。 OpenAI、Anthropic这些顶尖初创不仅要拼模型能力，还要争夺数据中心、服务器和供电能力。 以 Anthropic 为例，其模型的算力需求已经大到，需要与另一家模型公司讨论供货。 今年5月与SpaceX的合作协议规定，将在当月收获超300兆瓦的新容量，涉及22万块英伟达GPU，同时提高Claude Code和API的使用额度。 换句话说，这样做的目的就是 让后端拿到更多算力供给前端 。算力有了，模

匹配字段正文

很尴尬的数字。 黑尔斯一咬牙，干脆发起了一个项目，用机器把整个证明从头到尾形式化验证一遍。这一验，就是十几年，2014年才完成。 从那以后，让机器验证这条路终于慢慢火了起来，后来有了我们上面提到的Lean，有了Mathlib这个全世界数学家一起往里填的开源库。 但那个阶段的分工很清楚，人负责想，机器只负责算和验证。 转折点出现在大模型时代。 2024年夏天，DeepMind的AI在国际数学奥林匹克上拿到了银牌水平。2025年夏天，好几家的AI都拿到了金牌水平，其中就包括OpenAI。 然后，当时就闹了一个大笑话。 2025年10月，OpenAI的一位副总裁发推，兴奋地说GPT-5解决了10个此前未解的埃尔德什问题。 埃尔德什是匈牙利的一位传奇数学家，这哥们一辈子没有固定住所，拎着一个手提箱满世界跑，到谁家就跟谁合作写论文，一辈子发了一千五百篇左右，他还喜欢给难题挂悬赏，从几十美元到上万美元不等，身后留下了上千个问题。 结果没过多久就被扒出来，GPT-5根本就不是自己去解出了这些题，是他娘的在浩如烟海的文献里，找到了早就有人解过、但被大家遗忘的答案。。。 真正的拐点在2026年5月。 O

匹配字段正文

time monitoring to respond faster to misaligned behavior.” OpenAI’s rivals have taken note. Spurred by the fallout from the incident, the major AI labs—including Anthropic, Google DeepMind, and SpaceXAI—have all called for the pace of development to slow down. But how does that square with fierce international competition and trillion-dollar IPOs? “We’re not going to shoot ourselves in the foot and take ourselves far off the frontier—that’s just a horrible strategy,” he says. “I think it’s reall

匹配字段正文

macro task success rises from 23.3% to 44.3% , beating the base model with the harness (41.7%), and 82.3% of harness-induced behaviors are recovered. Agent failure modes : XYEval (DeepMind) injects one confident, misleading user hint and cuts scores by up to 46.7% relative . Agents often disagree with the hint in their reasoning, then silently follow it anyway. Monitor evasion : Agents often don’t stop when a monitor tells them to . Single-neuron bypass : A NeurIPS paper shows suppressing one ML

匹配字段正文

obvious to me that a distribution platform like OpenRouter was critical to have in the ecosystem if we wanted there to be competition to Google. Like, unless-- ‘cause with Google, DeepMind is done training a new checkpoint, and then they push a button, and it gets blasted out across all their surfaces from Google Docs to, Swyx [00:20:01]: Everywhere, even if I don’t want it. Alex Atallah [00:20:02]: Everywhere. You wanna know about, like, on Android, like, overnight, they can deploy a new checkp

匹配字段正文

到新能源，再从新能源到具身智能。 2015年前后，最聪明的产品经理和算法工程师涌向造车新势力；2025年之后，最顶尖的AI人才和算法人才，开始不约而同地奔向机器人公司。 如果把近半年的高管变动列成一张表，会发现一个规律，走得最密集的，恰恰是那些被寄予厚望的“一号位”。 罗剑岚的履历十分华丽——博士毕业于加州大学伯克利分校，曾在Google X、Google DeepMind担任研究科学家。2025年4月罗剑岚高调加盟智元机器人，牵头组建智元具身智能研究中心，构建了超100万条真机操作数据的AgiBot World数据集，次年9月进入合伙人团队。然后，在智元官宣赴港上市仅10天后，他从合伙人名单中消失了。 智元方面回应称：“系内部岗位调整，没有离职。”不过，当被问及调整后罗剑岚的职务以及负责工作内容时，智元方面未透露更多信息。 丁琰的轨迹则是另一种版本。从上海AI Lab到一星机器人，再到鹿明机器人，他在14个月内换了三份工作。在一星机器人只待了4个月，公司就因吉利战略收缩而解散。辗转33天后，2025年11月他带着十余名骨干集体加入鹿明。结果不到280天，又走了。 一个值得注意的细节，

匹配字段正文

nAI CEO Sam Altman said that embedding third-party evaluators is a solid step. SpaceX and Tesla founder Elon Musk echoed that oversight is long overdue. Demis Hassabis from Google DeepMind called the essay a step in the right direction. Microsoft CEO Satya Nadella emphasized that pursuing AI without keeping human benefit at the center isn’t worth doing. Tech investor Gavin Baker posted a great write-up below summarizing these industry responses Will Chinese labs follow the suit? Some social medi

匹配字段正文

s considering a $5B loan to AI cloud startup Fluidstack to strengthen U.S. manufacturing and supply chains for data-center components, WSJ reported. At the BRICS Summit in New Delhi, China proposed an open-source AI community to support shared AI development, training, and industrial use across member nations. Anthropic and Google DeepMind’s AI safety researchers Joe Benton and Josh Engels resigned to join METR Evals, with both warning that AI capabilities are outpacing safety. Sponsored Listing

匹配字段正文

ee a scenario with substantial growth and modestly higher unemployment, or crazy growth and crazy unemployment. What people are saying: LSE economics professor Ben Moll and Google DeepMind’s director of AGI economics Alex Imas wrote an essay on Substack dismissing the predictions of double-digit GDP growth that Amodei and others have made. The economists think that such extreme forecasts would assume too much: Double-digit growth rests on several shaky assumptions, they write, including“machines

匹配字段正文

chief research officer at OpenAI; Dario Amodei, cofounder and CEO of Anthropic; Shengjia Zhao, chief scientist at Meta AI; Shane Legg, cofounder and chief AGI scientist at Google DeepMind; and John Schulman, the chief scientist at Thinking Machines. Looking at the METR report, it seems clear that AI-model capabilities have already advanced beyond our ability to understand and control them. Some of the signatories of “Pacing the Frontier” have noticed. And the challenge of aligning AI agents wit

匹配字段正文