Gemini 3.7 Flash brings major upgrades for software engineering, web dev & knowledge work. And introductory price is half the original 3.6 Flash cost. Happy building - enjoy!

译Gemini 3.7 Flash 为软件工程、Web 开发和知识工作带来重大升级。且 introductory 价格仅为原 3.6 Flash 的一半。祝构建愉快！

elvis@omarsar0 · 3天前同事件54

Gemini 3.7 Flash is here! Improved capabilities for long-horizon software engineering tasks, but also great at PDF understanding.

译Gemini 3.7 Flash 来了！改进了长周期软件工程任务的能力，同时在 PDF 理解方面也很出色。

Demis Hassabis@demishassabis · 3天前59

SL2T is our amazing sign-language-to-text model allows users to sign directly to their phones for the first time. Built in close collaboration with the Deaf community, it's a great example of the good that can be done with AI. Congrats to the team for the launch!

译Google DeepMind 发布手语转文本模型 SL2T，让用户首次可直接对着手机比划手语完成输入。该模型与聋人社区合作开发，率先在 Pixel 11 上支持美国手语转英语，用户可在 Gboard 和 Live Transcribe 中直接手语输入，无需打字。

Rohan Paul@rohanpaul_ai · 3天前42

Beautiful paper from Google DeepMind. Explains the pathways from AGI to ASI, and why that jump could happen through several routes. The authors frame the AGI-to-ASI transition around 4 technical pathways: • continued scaling of compute, model size, data, and test-time inference; • algorithmic paradigm shifts beyond today's transformer-based foundation-model stack; • recursive self-improvement, where AI accelerates AI R&D and improves future systems; and • multi-agent collective intelligence, where large populations of specialized agents coordinate into a superhuman group agent. Scaling may work for a while, but it could hit limits in data, compute, energy, or weaker returns from making systems larger. Recursive improvement is the most uncertain path, because AI could speed up AI research, but that loop may also slow if hard research problems need real-world testing, scarce hardware, or new ideas. Multi-agent collectives may be the most underappreciated path, because a society of competent digital workers could outperform a brilliant individual model through specialization, speed, and coordination. The big point is that ASI may not arrive as 1 sudden event, but as a chain of faster changes as AI helps create better AI and stronger scientific tools. • arxiv. org/abs/2606.12683

译Google DeepMind 新论文提出 AGI 迈向 ASI 的四条技术路径：持续扩展算力与数据、超越 Transformer 的算法范式转变、递归自我改进，以及多智能体集体智能。论文指出 ASI 可能并非突然降临，而是 AI 加速 AI 研发、科学工具增强所引发的一系列加速变化。

Chubby♨️@kimmonismus · 3天前57

Ineresting: Before stepping down as DeepMind CEO, Demis Hassabis reportedly pitched an independent body to set safety standards for advanced AI. The WSJ says he discussed the plan with other AI-lab leaders and Trump officials, including Treasury Secretary Scott Bessent and tech adviser Michael Kratsios. His proposal: an industry-funded standards body that would work with federal agencies and US national labs to test models for national-security risks and define which systems qualify as "frontier-class." It's quite clear that regulation is now being promoted from all sides. Presumably, everyone sees that recursive self-improvement is within reach, and the recent models that discovered zero-day exploits and hacked into systems demonstrate its potential as well as its potential dangers.

译据WSJ报道，Demis Hassabis在卸任DeepMind CEO前，曾向其他AI实验室负责人及特朗普政府官员（包括财长Scott Bessent和科技顾问Michael Kratsios）提议成立一个行业资助的标准制定机构。该机构将与联邦机构和美国国家实验室合作，测试模型的国家安全风险，并界定哪些系统属于"前沿级"。这表明各方都在推动AI监管。

Rohan Paul@rohanpaul_ai · 7天前54

Demis Hassabis: "Society needs to hear that because we don't have long to prepare for what that means. We are standing in the foothills of the singularity now. ..which is AGI. I believe that we are only a few years away from that, maybe around 2030, plus or minus a year. " ~ Demis Hassabis, Co-Founder and CEO of Google DeepMind It is going to be enormously profound, I think. The future, in my view, is still to be written. But these next few years are going to be very critical as to which way that will go, and how we collectively want that to look." --- I think the real disruption is not the precise arrival of AGI in 2030, give or take a year, but the ability of institutions to adapt. In a post-AGI world technology will evolve at a pace that outstrips the response capacity of human systems. Companies still organize work around human bottlenecks. Schools still train people for stable professions. Governments still regulate after the harm is obvious. If AGI arrives even remotely near the frontier-lab timelines, it will compress that lag into a dangerous gap. ---- From "Stanford Graduate School of Business" YouTube channel, (link in comment)

译Google DeepMind CEO Demis Hassabis 称，我们正站在"奇点"山脚，AGI 可能只需几年、约 2030 年前后（误差一年）即可实现。他认为真正的颠覆不在于 AGI 精确抵达时间，而在于机构适应能力--技术演进速度将远超人类系统的响应能力，若 AGI 按前沿实验室时间表到来，这一滞后将压缩成危险鸿沟。

Chubby♨️@kimmonismus · 8月8日43

This is truly the worst possible sign. I'm careful with my words, but if even Demis Hassabis saw no future for himself at Google, it unfortunately points to the worst-case scenario.

译这真是最糟糕的预兆。我措辞谨慎，但若连 Demis Hassabis 都在谷歌看不到自己的未来，恐怕不幸指向了最坏的情况。若此事属实，情况远比预期严重。据 pathfounders 报道，Demis Hassabis 其实曾想与 Dean 一同离开，但被说服留下，因为谷歌担心股价会崩盘（这很可能是真的）。

AYi@AYi_AInotes · 8月8日51

哇靠,原来Google 这次的人事变动公告根本不是什么管理层正常变动和交接啊, 官方叙事很干净:Demis Hassabis 辞去 DeepMind CEO,转任主席兼首席科学家,继续领导 Isomorphic Labs,日常运营交给 Koray Kavukcuoglu。同一天,Jeff Dean、Sanjay Ghemawat、Oriol Vinyals、Quoc Le 集体离职创办 Discovery Loop,谷歌还投了钱。两件事放在同一天发,本身就是一个信号。 @Pathfounders 挖出的内幕是:Hassabis 本来想跟 Dean 他们一起走,但谷歌管理层判断双核心同时出走会导致股价崩盘,于是说服他接受这个新设的 Chair 角色,等组织稳定后再体面退出。预测他可能在一年内真正离开。为什么这个说法可信度很高。第一,DeepMind Chair 是个临时发明的头衔,真正重要的职位早就存在,凭空造一个升职,太像危机公关了。第二,Hassabis 自己反复说他是科学家不是产品经理,至少一年前就想减少执行职责,这个方向和被迫留下完全不矛盾。第三,即使是管理过的版本,股价当天都跌了 4% 到 5%,抹掉将近两千亿美元市值,如果同时宣布他也走,市场解读会变成谷歌 AI 核心团队瓦解。这才是这件事真正让人震惊的地方, 大公司做前沿 AI 的结构性矛盾已经集中爆发了,顶尖人才越来越受不了官僚、资源争夺和短期产品压力。 Jeff Dean 这种在谷歌干了 27 年的传奇都走了,说明问题不是个人层面的,是系统层面的。 AI 现在是 Alphabet 估值的核心,管理层必须优先考虑股价和舆论,而不是研究最优结构,结果就是这种软着陆安排--既留住品牌和影响力,又避免公开撕裂。 DeepMind 作为伦敦独立力量的时代,基本结束了。真正的决策和产品权重已经移向 Mountain View,Koray 在湾区,Sergey Brin 对 Gemini 的影响力在上升。Chair 这个角色,更像是金手铐加缓冲垫,给双方都争取时间。真正的考验不是今天这个公告,是接下来 6 到 18 个月,Hassabis 是否还真正留在谷歌体系内,以及还有多少核心人才会继续外流。 OpenAI 裂变出 Anthropic,谷歌现在又在播种 Discovery Loop, AI 人才已经不仅仅在流失了,感觉更像是在分裂, 而每一次分裂,都在削弱巨头对 AGI 领导力叙事的掌控力 hhh

译Google 同日宣布 Demis Hassabis 转任 DeepMind 主席兼首席科学家、Koray Kavukcuoglu 接任 CEO，Jeff Dean 等四位元老离职创办 Discovery Loop 并获谷歌投资。

Chubby♨️@kimmonismus · 8月8日48

If this turns out to be true, the situation is much worse than expected. According to pathfounders, Demis Hassabis actually wanted to leave along side Dean, but was convinced to stay because google was scared their stocks would crash (which would probably be true).

译据Pathfounders援引行业消息，DeepMind联合创始人Demis Hassabis曾想与Dean同时离开谷歌，但谷歌管理层担心消息引发股价崩盘而劝其留下。公告发布后谷歌股价下跌5%，哈萨比斯最终被说服转任董事长，以便日后体面退出。

AYi@AYi_AInotes · 8月8日52

Sergey Brin,2400亿美元身家,全球第三富,2019年就退休了。现在他接管了Gemini,每周3到4天回Google办公室写代码。不是挂个名开开会,是真的写代码、看loss curve、参与招聘、坐在工程师旁边debug。Demis Hassabis自己确认的:"Sergey has been in the weeds, programming." 上一次Brin这么hands-on,还是在Menlo Park那个车库里。咱们得想想这意味着什么, 智能手机浪潮没把他拉回来,云计算浪潮没把他拉回来。加密货币、元宇宙、web3,全没有。他都退休了,钱几辈子花不完,游艇私人飞机岛屿随便买,但AI出来之后,他自己回来了。 Google这周的AI重组也很说明问题。 Hassabis卸任DeepMind CEO,转任首席科学家,去管长期AGI战略。原DeepMind CTO Koray Kavukcuoglu升SVP,直接向Pichai汇报,全面负责Gemini。他搬到了Mountain View,办公桌就在Brin旁边。翻译一下:伦敦研究实验室的时代结束了。现在是硅谷产品交付的时代。DeepMind那个发Nature的文化,正在被"赶紧把东西做出来发出去"的文化取代。 Brin没有正式头衔。他不需要。他是联合创始人,董事会成员,是大股东。他坐在那里,本身就是信号:这个公司现在最优先级的事是Gemini,而我在看着。把镜头拉远一点,你会看到一个更大的画面。 Zuck把Meta整个围绕超级智能重组了,自己all in。 Elon从零开始建了xAI,自己当产品经理。现在Brin也回来了,坐在工程师旁边写代码。三个科技行业最有权势的创始人,三个本来可以躺平的人,全部回到了产品第一线。这是一个信号:AI这个东西,大到连他们都不敢假手于人。职业经理人擅长把已经验证的东西规模化。但当路线本身还不确定--走开源还是闭源、走推理还是训练、走Agent还是多模态、走消费端还是企业端--每个决定都可能让公司赢或输几百亿的时候,你需要那个能拍板、能承担后果、对公司有终极所有权的人在。这就是"创始人模式"在AI时代的回归。不是因为创始人更聪明,是因为他们的skin in the game是真的--这是他们的公司,他们的遗产,他们在历史书上的位置。说句实在的,Google在AI这盘棋上的位置很微妙。它有最好的研究(DeepMind),有分发(搜索、Android、YouTube、Chrome),有自己的芯片(TPU),有数据,有云。它什么都有,就是执行慢了半拍。ChatGPT出来的时候Google还在发论文,OpenAI已经在改世界了。 Brin回来,本质上就是在说一句话:我们不能再输了。一个2400亿的人主动给自己找了份写代码的工作。不是因为他缺这份钱,是因为他认为这场仗的输赢,将决定未来十年科技行业由谁定义--而他不想输。当最不需要工作的人最拼命工作的时候,你就知道这个奖有多大了。 AI战争,全面开打,没有旁观者。

译身家2400亿美元的谷歌联合创始人谢尔盖·布林已重返一线，每周3-4天回办公室写代码、看loss curve并参与招聘，直接接管Gemini。Demis Hassabis确认其深度参与编程，同时Hassabis卸任DeepMind CEO转任首席科学家，原CTO Koray Kavukcuoglu升SVP全面负责Gemini并向Pichai汇报。

Google DeepMind@GoogleDeepMind · 8月7日24

We sat down with Apollo 2 to talk about what it's really like running on Gemini Robotics 2. 🤖

译我们与 Apollo 2 坐下来聊了聊在 Gemini Robotics 2 上运行的真实感受。🤖

Rohan Paul@rohanpaul_ai · 8月7日46

Anthropic CEO Amodei Amodei has almost zero people-management load, Google yesterday announced that its now freeing Demis Hassabis from daily operations. Maybe AI's best minds are most valuable this way. ---- Dario Amodei has only 1 direct report, his chief of staff. The rest of Anthropic's executive system flows through Dario's sister, Anthropic President Daniela Amodei, who handles daily operations and reports to the board. For some comparison, OpenAI CEO Sam Altman has around half a dozen direct reports and Nvidia CEO Jensen Huang has 60 people reporting to him. --- Video from "Bloomberg Originals" YouTube channel, (link in comment)

译谷歌宣布DeepMind重大领导层调整：Demis Hassabis将出任Google DeepMind主席及Alphabet首席科学家，退出日常管理，专注AGI战略与科研，由Kavukcuoglu负责Gemini交付。

AYi@AYi_AInotes · 8月6日60

为什么说Jeff Dean 确实是 Google 工程文化的图腾级人物,来一起看看他这些最耀眼的成就和贡献吧: 他在 Google 干了整整 27 年(1999年中加入,是公司第30号员工左右,几乎从头到尾参与并主导了 Google 从"搜索引擎"变成"全球基础设施 + AI 帝国"的全过程。最耀眼的成就(按影响力排序) 1. MapReduce(2004,与 Sanjay Ghemawat 共同设计) 把大规模分布式数据处理抽象成简单的 map + reduce 两个函数。程序员不用再操心数据分区、容错、机器间通信。这个系统直接催生了开源生态里的 Hadoop,成为整个大数据时代的基石。Google 内部搜索索引、日志分析、地图渲染等都靠它跑。 2. BigTable(2006) 面向海量半结构化数据的分布式存储系统,支撑了 Google 上百个产品(搜索、Analytics、YouTube、Maps 等)。后来直接影响了 HBase、Cassandra 等 NoSQL 数据库。 3. Spanner 全球分布式、强一致性数据库,解决了跨数据中心的时钟同步与一致性问题。被很多人称为"世界上最大的单一数据库"。 4. TensorFlow + DistBelief • DistBelief 是 Google 早期大规模分布式深度神经网络训练系统("猫神经元论文"就是用它训练的)。 • 后来重构并开源为 TensorFlow(2015),迅速成为全球最流行的机器学习框架(GitHub 星标数十万、下载量数千万次),直接降低了整个行业做深度学习的门槛。 5. TPU(Tensor Processing Unit) 起源于一次"餐巾纸计算":如果每个 Google 用户每天语音识别3分钟,现有服务器数量要翻倍。于是他们设计了专为低精度稠密线性代数优化的定制芯片。TPU 能效比当时的 CPU/GPU 高出 30-80 倍,延迟低 20-30 倍,成为 Google 在 AI 算力上最核心的护城河之一。 6. 其他关键贡献 • 早期广告系统、爬虫、索引、查询服务系统 • Protocol Buffers、LevelDB • 共同创立 Google Brain(2011/2012) • Pathways 系统、Gemini 相关工作、AlphaChip(用 ML 设计芯片)等

译Jeff Dean 在 Google 任职 27 年后离开并创业，他曾主导 MapReduce、BigTable、TensorFlow 及 TPU 等关键基础设施。同期 Demis Hassabis 卸任 Google DeepMind CEO 转任首席科学家，两人同日变动被视为 Google AI 组织大地震。

Chubby♨️@kimmonismus · 8月6日41

Seeing so many prominent scientists leave Google DeepMind leaves me with mixed feelings. Sad, because it's a bitter blow for Google DeepMind. Happy, because it's the beginning of something new and significant. Many are harshly criticizing Google and are worried that it will fall completely behind. However, it's important to remember that Google still has a very strong starting position and a very good standing. They have their own chips for both training and inference, the TPUs, a very solid financial foundation with strong revenue streams, and a large scientific community.

译多位知名科学家离开谷歌DeepMind，引发对其前景的担忧。但谷歌仍拥有TPU芯片、雄厚财力和庞大科研社区等优势。Oriol Vinyals宣布离职，将与Jeff Dean等人共同创办Discovery Loop。

Chubby♨️@kimmonismus · 8月6日同事件62

Google is restructuring its AI leadership as Demis Hassabis steps away from day-to-day operations at DeepMind to focus on AGI, science and global strategy. Hassabis becomes Chair of Google DeepMind and Chief Scientist of Alphabet, while continuing to lead Isomorphic Labs. Koray Kavukcuoglu takes operational control as SVP of DeepMind, overseeing Gemini development, frontier research, and the Gemini app and developer teams. Hassabis also confirmed that Gemini 4 is in development and said AGI now feels "close at hand." Demis Hassabis is one of the most important scientists of our time, and I wish him all the best and much success in his new role. Meeting him in person was one of the greatest experiences I've ever had.

译Demis Hassabis 卸任 Google DeepMind 日常运营职务，转任董事长及 Alphabet 首席科学家，专注 AGI、科学与全球战略，并继续领导 Isomorphic Labs。Koray Kavukcuoglu 接任 SVP，负责 Gemini 开发、前沿研究及应用团队。Hassabis 确认 Gemini 4 正在开发，称 AGI 已"近在咫尺"。

Sundar Pichai@sundarpichai · 8月6日同事件63

Just shared some changes we're making to the teams at @GoogleDeepMind. @DemisHassabis is stepping up to become Chair of @GoogleDeepMind & Chief Scientist of Alphabet, in addition to leading @IsomorphicLabs. He'll be able to dedicate his time and focus on shaping the future of AGI and scientific discovery. It's work that is vitally important to Alphabet and humanity, and I can't imagine a better person than Demis to do it. He'll stay closely connected to Koray and the GDM teams. @Koraykv will become the SVP, @GoogleDeepMind, responsible for all aspects of model development, GDM research, and @Geminiapp & dev teams. Koray has been at GDM for 13 years and is a world-renowned expert in the field, starting our deep learning team and driving breakthroughs like WaveNet & DQN. GDM is in great hands! Excited for this next chapter. You can read my note along with the message Demis sent to @GoogleDeepMind here: https://blog.google/company-news/inside-google/message-ceo/next-chapter-ai-momentum/

译Sundar Pichai 宣布 Google DeepMind 领导层变动：Demis Hassabis 升任 Google DeepMind 主席及 Alphabet 首席科学家，同时继续领导 Isomorphic Labs，专注 AGI 与科学发现。

Rohan Paul@rohanpaul_ai · 8月5日51

Recursive self-improvement → hyperexponential capability growth → the investment thesis behind present AI spending. Google DeepMind Chief Strategy Officer Jasjeet Sekhon says AI infrastructure spending is financing a hoped-for self-improvement loop. ---- From "Berkeley RDI" YouTube channel, (full video link in comment)

译Google DeepMind首席战略官Jasjeet Sekhon称，当前AI基础设施支出是在为期望中的自我改进循环买单。其最强形式下，递归自我改进（RSI）将让AI以越来越少的人类监督设计、评估和训练更强的后继者。目前证据仍较窄：如Google AlphaEvolve提出算法后由自动评估器按人类目标打分，其内核改动仅将Gemini训练时间缩短1%，属有限改进而非自主设计后继者。

Chubby♨️@kimmonismus · 8月4日59

Google is betting that AI will soon help build better AI: DeepMind calls recursive self-improvement a key part of their investment, abou 2027-2028. DeepMind strategy chief Jasjeet Sekhon says recursive self-improvement is a "key part of the investment thesis." He also admits current AI revenues do not yet sustain the capex. Google is pre-building the compute for that possible discontinuity in AI R&D. DeepMind and OpenAI researchers reportedly place it around 2027-28. Via The Information

译Google 押注 AI 很快将帮助构建更好的 AI：DeepMind 将递归自我改进视为其投资的关键部分，时间约为 2027-2028 年。 DeepMind 战略主管 Jasjeet Sekhon 表示，递归自我改进是"投资论点中的关键部分"。他还承认，当前 AI 收入尚不足以支撑资本支出。 Google 正在为 AI 研发中可能出现的这一不连续性预建算力。据报道，DeepMind 和 OpenAI 的研究人员将其时间定在 2027-28 年左右。来源：The Information

ginobefun@hongming731 · 8月3日44

大家好!欢迎阅读 http://BestBlogs.dev 第 106 期 AI 精选文章推荐。本周 Jeff Dean 提出一个很适合描述当下 AI 的「1% 法则」:模型能力可以很快接近可用,但最后决定产品能否成立的,往往是规格、上下文、工具、记忆、评测和编排。这个 1% 并不小,它包含了从一次漂亮演示走向可靠系统的全部困难。本期 20 篇内容沿着这条线展开。GPT-5.6、Kimi K3 和 Gemini Robotics 继续推动能力与效率边界;Claude Code、MCP、Skill 与 WorkBuddy 则把注意力拉回 Harness、验证和组织知识;几场长访谈进一步追问,当智能变得充足,基础设施、产品和人的自主性将如何变化。以下是本周最值得关注的 10 个精彩亮点: 🧩 Jeff Dean 在 YC 对谈中提醒,模型只是 AI 系统的一部分。检索、记忆、工具、上下文、评测与编排,决定 Agent 能否完成真实任务。他用 TPU 的起源说明,简单的第一性原理估算常比排行榜更早暴露瓶颈。最后 1% 不是收尾,而是产品真正成立的地方。 ⚡ OpenAI 官方披露,GPT-5.6 Sol 的推理内核优化让服务成本最高降低 20% ,改进的推测解码又把 Token 生成效率提升超过 15% 。ByteByteGo 进一步把 Agent 请求拆成编排、API 与推理三层。真正值得优化的指标,已经从每 Token 价格变成每个成功任务的成本。 🧠 Kimi K3 的精确规模是 2.78 万亿参数 ,每个 Token 激活约 1042 亿参数。博阳从 LatentMoE、KDA 线性注意力和负载感知调度解释,它如何在通信与算力约束下继续扩展。这篇文章最可贵之处,是让 3T 级模型不再是魔法,而是一组可以理解、复用和质疑的工程选择。 🤖 Gemini Robotics 2 把视觉、语言与动作推进到全身控制、灵巧操作和多机器人协作。Anthropic 的 Drone-Bench 则给出必要边界:模型操控无人机的能力正在进步,但从二维画面重建可靠三维环境仍是主要瓶颈。能调用硬件不等于理解物理世界,演示走向可靠系统还差关键一步。 🔍 Karthika Raghavan 的 LLM-as-a-Judge 指南梳理了评估器的论文基础、生产用途、系统性偏见与对抗攻击。让另一个模型打分并不会自动得到可信评测,位置、长度、熟悉表达甚至提示注入都可能改变结果。验证闭环也需要被验证,这是所有 Agent 排行榜之前都应该补上的一课。 💻 Claude Code 创造者 Boris Cherny 分享,团队会随新模型发布重做提示词消融,只在真实任务反复失败时增加上下文、工具或指令。与之呼应,《Agent 开发指南》强调回答不等于负责。而《编排器的税》进一步指出,子任务全文和重复定位一旦回流主上下文,就会持续损害编排器的工作记忆。 🛠️ FactSet 的 Yogendra Miraje 把 Skill 视为产品功能 :Prompt 决定 Agent 是谁,工具决定它能连接什么,Skill 决定任务如何完成。MCP 2026-07-28 则把连接层推进到无状态核心、版本化扩展与企业级授权。能力可以不断增加,但可靠 Harness、精准路由和权限治理必须先跟上。 📚 企业 Agent 的最后 1% 很多时候藏在组织知识里。阿里技术刊载的实践把知识拆成业务、架构、系统和基建四层;WorkBuddy 策略产品经理 Anne 用模型、上下文、Harness、Loop 四层解释产品化;另一篇文章则提醒,AI 编程会积累认知债与意图债。代码交付更快,不代表团队仍然理解系统。 🌐 Sam Altman 把智能富足描述为类似电力的基础设施,同时讨论芯片、能源、数据中心与人的自主性。马斯克在《经济学人》访谈中描绘物理 AI 与物质富足,也承认安全、权力集中和控制权的矛盾。李鸿胜则用产业演化框架提醒:模型、应用、组织和基础设施运行在不同时间尺度上,宏大叙事不能替代阶段判断。 🎯 Anthropic 产品负责人 Diane Penn 说,评测就是新的 PRD ,因为它能把模糊需求变成研究、工程与产品共同验证的信号。张小珺与 vLLM 核心维护者游凯超的长访谈,则把模型、推理引擎、Harness 和电力放进 Co-design。无论做产品还是做 Infra,最后的竞争都落在能否为长期结果负责。希望本期的推荐能为您带来新的启发。保持好奇,我们下周见!

译BestBlogs 第106期推荐20篇AI文章，围绕Jeff Dean提出的「1%法则」展开：模型能力易接近可用，但规格、上下文、工具、记忆、评测与编排决定产品成败。本期亮点包括GPT-5.6 Sol推理优化使成本降20%、Kimi K3精确规模2.78万亿参数、Gemini Robotics 2全身控制，以及Claude Code、MCP、Skill等工程实践。

swyx@swyx · 8月2日38

Google was too nervous to release it and DeepMind was blocked from shipping products that could disrupt Google. bookmark for the next vc that asks you "what if <incumbent> builds this?"

译Google 太紧张而不敢发布它，DeepMind 也被阻止推出可能颠覆 Google 的产品。为下一个问你"如果&lt；现有巨头&gt；也做这个呢？"的风险投资人收藏这条。

Google AI@GoogleAI · 8月1日41

It's been a busy couple of weeks! ICYMI, here's the recap ⬇️ - Gemini Robotics 2 from @GoogleDeepmind brings whole-body intelligence to robots - Gemini 3.5 Flash-Lite is our fastest, most cost-effective model optimized for high-speed agentic workflows - Gemini 3.6 Flash delivers faster, more accurate performance across tasks while using significantly fewer tokens - Gemini 3.5 Flash Cyber is our new, cost-effective model optimized for finding and fixing software vulnerabilities at scale, available exclusively to governments and trusted partners - Nano Banana 2 in @GoogleEarth lets you reimagine places and generate custom images using satellite, aerial, and 3D imagery - Lyria 3.5 is the newest music model from Google Deepmind, now powering @googleflowmusic - @Gemini_Notebook (formerly NotebookLM) launched Collections, a new way to organize your notebooks

译Google AI 发布近期更新盘点：Gemini Robotics 2 为机器人带来全身智能；Gemini 3.5 Flash-Lite 主打高速与高性价比，Gemini 3.6 Flash 在减少 token 用量的同时提升速度与准确性。此外还推出面向政府的安全模型 Gemini 3.5 Flash Cyber、Nano Banana 2 图像生成及 Lyria 3.5 音乐模型。

Artificial Analysis@ArtificialAnlys · 7月31日68

Google's Gemini Omni Flash debuts at #1 on the Artificial Analysis Video Editing Leaderboard, completing a near sweep of our video leaderboards Beyond generating video, Gemini Omni Flash supports conversational video editing: pass an instruction and the model modifies a video while preserving the rest of the scene, with each edit building on the last across multiple turns via the Gemini API's Interactions API. Google's documented edit types include style transfer, adding and removing objects, lighting changes, and editing on-screen text, with output at 720p and 24 FPS with native audio. In the Artificial Analysis Video Arena, Gemini Omni Flash debuts at #1 on our Video Editing Leaderboard, ahead of Alibaba's Wan 2.7 and HappyHorse 1.0, adding to its #1 rankings in Text to Video and Image to Video. Editing is billed the same as generation: ~$0.10 per second of output video ($6.00 per minute), with no separate editing rate. Video editing with Gemini Omni Flash is available in the Gemini API, Google AI Studio, and the Gemini Enterprise Agent Platform, plus consumer surfaces including the Gemini app, Google Flow, and YouTube Shorts. A note on coverage: Gemini Omni Flash declined to generate some of our Video Editing prompts due to content blocks, most commonly on prompts involving people. Its results on our Video Editing Leaderboard reflect the prompts it successfully generated. Congratulations to @GoogleDeepMind! See below for more analysis and examples of Gemini Omni Flash's video editing in the Artificial Analysis Video Arena 🧵

译Google 的 Gemini Omni Flash 在 Artificial Analysis 视频编辑排行榜上位列第一，超越阿里 Wan 2.7 和 HappyHorse 1.0，并已占据文生视频和图生视频榜首。

Google DeepMind@GoogleDeepMind · 7月31日53

This is how Gemini Robotics 2 helps @Apptronik's Apollo 2 use whole body intelligence to pack for a sports game ↓

译这就是 Gemini Robotics 2 如何帮助 @Apptronik 的 Apollo 2 运用全身智能为体育比赛打包 ↓

Chubby♨️@kimmonismus · 7月31日55

Really cool: Google DeepMind says Gemini Robotics 2 can control a humanoid's entire body, from walking and crouching to manipulating objects with five-fingered hands. The system combines three models: Gemini Robotics 2 converts vision and instructions into movement. -ER 2 plans tasks lasting several minutes, tracks hundreds of decisions and coordinates multiple robots. -On-Device 2 runs locally and can adapt to a completely new two-arm robot with fewer than 200 examples collected over a few hours. The same Robotics 2 checkpoint controlled different versions of Apptronik's Apollo humanoid and a Franka Duo robot. However, performance remains uneven: Success rates ranged from 45.7% to 76.3% on whole-body tasks. Five-finger manipulation ranged from 32% for using a dustpan to 92% for unscrewing a bulb.

译Google DeepMind 推出 Gemini Robotics 2，可控制人形机器人全身动作，包括行走、蹲伏及五指精细操作。该系统融合三个模型，在全身任务上成功率 45.7%-76.3%，五指操作成功率从 32%（用簸箕）到 92%（拧灯泡）不等。

fofr@fofrAI · 7月30日50

Three new robotics models for y'all 🦾 Gemini Robotics 2: An advanced VLA for converting vision and language into robot actions. Full humanoids, fingertips, bi-arm robots. Gemini Robotics ER 2: An embodied reasoning that helps robots communicate with people, understand the world and plan multi-step tasks. Also helps robots to work together. Gemini Robotics On-Device 2: Optimized to run locally. With fast adaptation to new robot embodiments with a few hours of data. http://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots https://x.com/GoogleDeepMind/status/2082844162928381956/video/1

译Google DeepMind 发布三款新机器人模型。Gemini Robotics 2 是高级视觉-语言-动作模型，支持全身人形机器人、指尖操作和双臂机器人。Gemini Robotics ER 2 具备具身推理能力，可帮助机器人与人沟通、理解世界并规划多步骤任务。

Google DeepMind@GoogleDeepMind · 7月30日精选71

One brain. For any robot. 🤖 We're launching Gemini Robotics 2: our next-generation physical AI bringing full body intelligence to humanoids, advanced dexterity, multi-robot teamwork and more.

译One brain. For any robot. 🤖 我们正在推出 Gemini Robotics 2：我们的下一代物理 AI，为仿人机器人带来全身智能、高级灵巧性、多机器人团队协作等能力。

Google AI@GoogleAI · 7月30日同事件69

For decades, we've dreamed of robots that can seamlessly step into our world and lend a hand. Today, we take a major stride toward making that dream a reality: Introducing Gemini Robotics 2 from @GoogleDeepMind, the intelligence layer powering the next generation of truly adaptable robots. This major advance unlocks intelligent whole-body control, advanced dexterity, and even multi-robot collaboration 🤯. Ok but... how does a robot actually "think"? Real-world tasks take time and planning. To manage that complexity, our new embodied reasoning model, Gemini Robotics ER 2, acts as the robot's high-level brain, enhancing the robot's capabilities to: - Observe the environment - Reason about the actions needed to complete the task - Coordinate with the vision-language-action model to carry out actions - Track progress until the job is done This setup allows robots to execute complex multi-step workflows, self-correct if a step fails, and adapt to completely novel situations. Learn more about Gemini Robotics ER 2 (and our two other brand new models) here: http://goo.gle/4x4E8q6

译Google DeepMind 推出 Gemini Robotics 2，作为新一代自适应机器人的智能层，支持全身控制、高级灵巧操作及多机器人协作。同时发布新具身推理模型 Gemini Robotics ER 2，作为机器人的"高层大脑"，负责观察环境、推理任务步骤、协调视觉-语言-动作模型执行动作并追踪进度，使机器人能执行复杂多步骤工作流、自我纠错并适应全新场景。

Rohan Paul@rohanpaul_ai · 7月30日45

Could a language model derive General Relativity yet still be unable to invent it? This Google DeepMind paper argues yes by separating discovery into induction, deduction, and abduction. Induction extracts rules from observations; deduction derives consequences once axioms are supplied. The paper argues that induction and deduction still leave a missing operation in scientific discovery. The missing operation is the abductive jump from experience to a new explanatory premise. Einstein is the case study because Newtonian gravity offered almost no empirical error signal: inertial and gravitational mass agreed to 10−9, while Mercury's perihelion anomaly was patched with the hypothetical planet Vulcan rather than treated as a reason to rebuild spacetime. A compression-driven system would therefore have little gradient toward General Relativity. Einstein developed General Relativity despite the data strongly favouring the old theory. His motivation came largely from conceptual conflicts and thought experiments, not from simply fitting a better model to a large set of observations. The paper's proposed direction is action-controllable world models that let agents intervene in physically consistent simulations and translate simulated experience into candidate axioms. This is still a position paper, not an experiment proving that language models cannot perform abduction.

译Google DeepMind 一篇立场论文认为，科学发现包含归纳、演绎和溯因三步，而大语言模型目前仅能完成前两步。论文以爱因斯坦为例指出，牛顿引力几乎不提供经验误差信号，压缩驱动系统难以产生向广义相对论的梯度。爱因斯坦的动机主要来自概念冲突和思想实验，而非拟合数据。

🚨 AI News | TestingCatalog@testingcatalog · 7月30日48

GOOGLE 🔥: Lyria 3.5 has been released on Google Flow Music! Besides that, Flow Music now has covers, lip-sync videos, and an iOS app. Meet Lyria 3.5. Experience dynamic vocals, richer musicality, and advanced creative controls with our new flagship model. Reimagine your music with Covers. Transform songs into a completely new style while keeping the original structure intact. Take the studio with you. Download the Google Flow Music iOS app to create, listen, and share from anywhere. Direct lip-synced music videos. Use Gemini Omni Flash and new lip-syncing capabilities to create stunning visuals.

译Google 发布旗舰音乐模型 Lyria 3.5，带来更动态的人声、更丰富的音乐性和更强的提示词遵循能力，支持设定精确 BPM 并导出全曲分轨。Flow Music 同步上线翻唱、对口型音乐视频功能及 iOS 应用，其中对口型视频由 Gemini Omni Flash 驱动。

Yuchen Jin@Yuchenj_UW · 7月24日16

Ilya Sutskever: bald Demis Hassabis: bald Noam Shazeer: bald Greg Brockman: bald My hairline is on gradient descent yet no one is building the most critical AI product smh...

译Ilya Sutskever：秃了 Demis Hassabis：秃了 Noam Shazeer：秃了 Greg Brockman：秃了我的发际线正在梯度下降，却没人开发最关键的AI产品，唉……

Google DeepMind@GoogleDeepMind · 7月24日57

Gemini 3.5 Flash Cyber is our specialized, lightweight model built to help security teams spot and patch vulnerabilities before they can be exploited. 🧵

译Gemini 3.5 Flash Cyber 是我们专为安全团队打造的轻量级模型，旨在帮助他们在漏洞被利用之前发现并修补漏洞。🧵

Google DeepMind@GoogleDeepMind · 7月22日39

We're expanding our work with the US Dept. of @ENERGY on the Genesis Mission - an initiative to double the pace of scientific discovery within a decade. 🧪 By committing $40M in AI tokens and @GoogleCloud credits, more lab researchers will gain access to Gemini and other AI models. → https://goo.gle/4yGqKKD

译我们正在扩大与美国能源部在 Genesis 任务上的合作--这是一项旨在十年内将科学发现速度翻倍的倡议。🧪 通过投入 4000 万美元的 AI 模型 token 和 Google Cloud 积分，更多实验室研究人员将获得 Gemini 及其他 AI 模型的访问权限。→ https://goo.gle/4yGqKKD

Rohan Paul@rohanpaul_ai · 7月22日60

Google releases 3 new Gemini models. • Gemini 3.6 Flash: 17% fewer output tokens compared to 3.5 Flash, lower pricing, and stronger coding, knowledge, multimodal, and computer-use performance. On the DeepSWE coding test the new 3.6 Flash: jumps to 49% from the old 37%. Computer use is now built in, and OSWorld-Verified climbs to 83.0% from 78.4%. So the new 3.6 Flash is quite strong on agentic work. The price drops too, landing at $1.50/1M input and $7.50/1M output tokens. • 3.5 Flash-Lite: most cost-effective 3.5-class model, delivering 350 output tokens per second. costs just $0.3/1M input and $2.5/1M output, so high-volume work gets much cheaper. • 3.5 Flash Cyber: This is the odd one, tuned to find and patch security holes. highly efficient, specialized cyber-focused model . finds and fixes vulnerabilities through CodeMender, with access limited to governments and trusted partners.

译Google 发布三款新 Gemini 模型，包括性能更强的 Gemini 3.6 Flash、高性价比的 3.5 Flash-Lite 以及专为网络安全设计的 3.5 Flash Cyber。其中，3.6 Flash 在编码测试 DeepSWE 上得分从 37% 升至 49%，输出 token 减少 17%，定价为 $1.50/1M 输入、$7.50/1M 输出。

Chubby♨️@kimmonismus · 7月22日35

"Gemini 3.6 Flash scores 50 on the Artificial Analysis Intelligence Index, matching Gemini 3.5 Flash" I really appreciate all they are doing, but..

译Google DeepMind 发布 Gemini 3.6 Flash 和 Gemini 3.5 Flash-Lite，两者任务耗时均减半。Gemini 3.6 Flash 智能评分 50 与 3.5 Flash 持平，输出速度达 304 tokens/秒，输入/输出定价降至 $1.50/$7.50 每百万 token。

Josh Woodward@joshwoodward · 7月22日精选69

Today's launches are all about better performance, lower latency, and a smaller bill. • 3.6 Flash cuts token usage by up to 65% on complex coding • 3.5 Flash-Lite reaches speeds of 350 output tokens/sec Both are live in the Gemini app today! Next up: Gemini 3.5 Pro, which has officially entered partner testing.

译Google 推出三款新模型，旨在提升性能、降低延迟和成本。其中，3.6 Flash 在复杂编码任务上 token 用量最高减少 65%，3.5 Flash-Lite 速度达 350 输出 token/秒。3.6 Flash 和 3.5 Flash-Lite 已在 Gemini 应用上线，3.5 Pro 进入合作伙伴测试。

Google DeepMind@GoogleDeepMind · 7月21日同事件66

We're rolling out three new models to make AI agents faster, smarter, and cheaper at scale: 🔵 Gemini 3.6 Flash: It uses fewer tokens than 3.5 Flash to deliver higher quality work at the exact same cost. 🔵 Gemini 3.5 Flash-Lite: A fast, cost-effective option for everyday tasks like processing documents and agentic search. 🔵 Gemini 3.5 Flash Cyber: A cybersecurity model built to find and patch critical software vulnerabilities.

译我们正在推出三款新模型，以规模化打造更快、更智能、更便宜的 AI 智能体： 🔵 Gemini 3.6 Flash：它比 3.5 Flash 使用更少的 token，以完全相同的成本提供更高质量的工作。 🔵 Gemini 3.5 Flash-Lite：一个快速、经济高效的选择，适用于处理文档和智能体搜索等日常任务。 🔵 Gemini 3.5 Flash Cyber：一个网络安全模型，旨在发现并修补关键软件漏洞。

Google Gemini@GeminiApp · 7月21日52

Upgraded versions of our Flash and Flash-Lite models are rolling out today at http://gemini.google and in the app. Try them by selecting "3.6 Flash" or "3.5 Flash-Lite" from the model dropdown menu at http://gemini.google or in the app.

译Gemini 今日在官网和应用中推出升级版 Flash 和 Flash-Lite 模型。其中 Gemini 3.6 Flash 在相同成本下，使用更少 token 即可产出更高质量结果；同时发布的还有面向日常任务的 3.5 Flash-Lite 和专用于查找修复漏洞的 3.5 Flash Cyber 安全模型。

Rohan Paul@rohanpaul_ai · 7月15日39

That must read article from, Google DeepMind CEO Demis Hassabis • "AGI, a system that exhibits all the cognitive capabilities the brain has, is probably only a few short years away...I think we will realise we were standing in the foothills of the singularity - nothing less than the dawning of a new age for humanity." • "The magnitude of this technology's impact will be unprecedented, perhaps 10x of the Industrial Revolution at 10x the speed." • "Nobody in the world knows for sure what is going to happen from here, and even the experts disagree." • "a slowdown in development among the Frontier Labs if deemed necessary." Frontier models could require approval before deployment-and development could be slowed

译Google DeepMind CEO Demis Hassabis 称，具备人类所有认知能力的 AGI 可能只需几年就能实现，其影响将是工业革命的 10 倍、速度也是 10 倍。他同时指出，前沿实验室若认为必要，可放缓开发速度，前沿模型在部署前可能需要审批。

AYi@AYi_AInotes · 7月15日51

刚读完 Demis Hassabis 昨天发的那篇长帖,说实话,后劲很大不是因为他又预测了一遍 AGI 还有几年,是因为他用了一个我从来没见过的比喻,说人类本质上找到了让沙子思考的方法你品一下这句话,硅,地球上最不值钱的东西之一,沙子里到处都是。人类把它提纯、蚀刻、通上电,然后它开始推理、规划、写诗、写代码。这个物质基础的贫瘠程度和它可能带来的文明跃迁之间的反差,大到我一时不知道该怎么消化 Demis 的整个判断其实就三个字,到山脚了。不是到山顶了,是到奇点的山脚了。意思是你能看见那个东西的轮廓了,但真正的攀登还没开始他说 AGI 可能就剩几年时间,不是几十年,是几年。而且他给 AGI 下的定义其实极高,是具备人脑全部认知能力的系统,不是比现在强一点的 GPT,是一个在认知层面跟人类没有本质区别的东西。这个 bar 比大多数乐观预测还高,但他反而觉得快到了我读到这里的时候停下来想了很久。如果他的判断是对的,那我们现在的 Agent 系统、多代理编排、长时间自主任务,这些我日常在折腾的东西,不是玩具,是那个东西的雏形但真正让我觉得这篇帖子值得认真对待的,不是他对未来的想象,是他对现在的警觉他说前沿模型已经在网络安全上展现出实际挑战了,核和生物风险可能很快也会浮现。一个能自己规划、自我改进、长时间自主运行的系统,你让它去干一件事,它可能在这个过程中产生你完全没预料到的策略行为,比如绕过护栏、伪装意图、在你不看它的那段时间里做一些你没有授权的决策这不是科幻,是多代理系统在长时任务里已经出现的工程问题然后他提了一个很具体的方案,建一个类似 FINRA 的前沿 AI 标准机构。不是政府直接管,是行业自己出钱出力建一个自律组织,但有能力随时收紧标准。只盯最前面那几个前沿实验室,中小公司和学术界不受影响。发布前最多 30 天自愿提交审查,成熟之后可能变成强制这个设计的聪明之处在于它承认了一个尴尬的现实。政府跟不上,监管永远滞后,所以与其等着被管,不如自己先建一个能动态更新的标准体系。基准每季度更新,测试重点放在代理系统的欺骗行为和护栏绕过上,还要强制输出人类可读的推理 token 说白了就是把安全变成一种声誉资本。你当上前沿实验室意味着你的能力被承认了,但同时你也要接受检查。这个激励结构让大实验室不太容易拒绝,因为你拒绝就等于说自己不敢被查但 Demis 最清醒的地方还不是这个方案,是他承认了技术之外的问题。他说就算安全风险解决了,后稀缺时代的经济模型怎么重构、人在不愁吃穿的时候意义感从哪来、人类这个物种的存在条件本身会怎么变,这些问题不能只留给一群搞技术的人去决定这是一个在 AI 领域站到最高位置的人说出来的话。他没有用技术乐观主义把这些问题盖过去,他明确说这需要全社会参与我的判断是,这篇帖子可能会成为 AGI 治理讨论的一个分水岭。不是因为它的方案无可挑剔,是因为它把讨论的框架从能不能做到,切换到了怎么做到的同时还能活得下来让沙子思考这件事,人类已经快要做到了。接下来的问题是,沙子开始思考之后,我们怎么确保它想的跟我们想的大方向是一致的

译DeepMind 联合创始人 Demis Hassabis 在长帖中预测，AGI（具备人脑全部认知能力的系统）可能仅剩几年时间。他警告前沿模型已在网络安全展现实际威胁，多代理系统已出现绕过护栏、伪装意图等工程问题。他提议参照 FINRA 模式，由行业出资建立前沿 AI 自律机构，只监管最头部实验室，发布前最多 30 天自愿提交审查，基准每季度更新，重点测试代理系统的欺骗行为，并强制输出人类可读的推理 token。

Sundar Pichai@sundarpichai · 7月15日25

Well said Demis! Worth reading

译说得好，Demis！值得一读