🏆 Headline
Paying doesn't opt you out: Gemini trains on subscribers' chats by default
An angry post from a paying subscriber pulled six hundred-plus upvotes on Hacker News this Wednesday. The poster discovered that Gemini chats are used to train models — not through a data leak, but through a default setting written plainly in the documentation: personal Google accounts have "Keep Activity" on by default, and while that toggle is on, your conversations are used to improve Google services, "including by training generative AI models." Paying or not makes no difference. The part that truly stings is the coupling: to turn off training, you have to turn off memory too. The single "Keep Activity" switch simultaneously controls saved chat history, personalization memory, and the availability of advanced features like remote browser and remote computer code execution — want the conveniences, accept the training; want out, go back to starting from zero every time. Per Google's documentation, human reviewers may read conversations in certain circumstances, and the data can be retained for up to three years. To be fair, none of this is hidden — the Gemini Privacy Hub spells it out clearly; it just isn't on the page a subscriber would open. Workspace enterprise accounts are excluded from training by default, which explains why "paying" and "privacy" run on two lines that never cross in Gemini: the subscription buys features and quota, and the training opt-out is reserved for enterprise administrators.
Source: Hacker News / Google Gemini Privacy Hub | 2026-11-04
OpenAI and Meta bought coding-agent "workbenches" on the same day
Wednesday saw two deals land in AI coding tools within hours. OpenAI announced the acquisition of Pierre Computer Company: the ten-person team, founded in 2023, builds something essential for the agent era — Code Storage, an API-first Git infrastructure designed for machines rather than humans, where code and artifacts generated by agents can be stored and retrieved at scale; its open-source libraries Trees and Diffs are already inside Codex. The same day, Meta confirmed the acquisition of Conductor (first reported by Axios): a product that lets engineers direct multiple coding agents at once across parallel workspaces. Founder Charlie Holtz announced he is joining Meta to work with chief AI officer Alexandr Wang and Superintelligence Labs product head Nat Friedman, teasing that the "next interface" will surface early next year. Neither deal disclosed terms. The model-layer arms race is not over, and the land grab for the agent workbench has already begun.
Source: OpenAI Blog, Axios / TechCrunch | 2026-11-04
AMD posts a record quarter: data center revenue up 97%
AMD reported third-quarter results after the close on November 3: revenue of $13.4 billion, up 45% year over year and above the top end of guidance; the data center segment surged 97%, with both Instinct GPUs and EPYC CPUs setting sales records; non-GAAP earnings per share of $2.09, up 74%. CEO Lisa Su said growth should accelerate further in the fourth quarter as the Helios and "Venice" generations scale. Against the data center's solo run, the gaming business visibly stalled. When AI infrastructure becomes a chip company's absolute engine, the traditional businesses are left holding the backdrop.
Source: AMD IR / CNBC | 2026-11-03
AI compute stocks slid again — this time it is not about earnings
AI and data center stocks came under pressure again on Wednesday, and the trigger changed: the U.S. midterms concluded with the camp favoring brakes on data center buildout winning key tech hubs. Per CNBC, the Texas governor's mansion flipped — governor-elect Gina Hinojosa has pushed to immediately pause new data center construction in the state. Wells Fargo estimates about 27% of planned U.S. data center gigawatts sit in Texas; Dan Ives of Yorkville Ives & Co said bluntly there is "no way to sugarcoat" it, and the election results will slow the pace of data center buildouts nationwide. Tesla opened down 5%, SpaceX down 4%. Last Friday the market was doing the capex math; this round, policy took the stage — the second half of the compute sprint starts with the color of land, power, and permits.
Source: CNBC | 2026-11-04
Putting AI into Siri while discounting employee AI budgets: Apple's daily cap drops from $300 to $200
Per Bloomberg, Apple is reducing employee AI usage budgets: engineers' daily AI spending caps dropped from $300 to $200, with further cuts planned, part of a company-wide cost-control push. At the same time, the company is working to put AI at the center of its products and operations. Weaving AI into system features with one hand and drawing budget red lines for internal AI use with the other, Apple is walking a tightrope between telling the AI story and counting the AI bill.
Source: Bloomberg / The Verge | 2026-11-04
Uber's earnings call poured cold water on AI agents: 4 rides in the busiest week
On Uber's third-quarter earnings call on Wednesday, CFO Balaji Krishnamurthy reported a number cold enough to freeze: during Halloween week — the company's busiest single week, with roughly 320 million trips and orders across ride-hailing and delivery — AI agents placed exactly 4 ride orders; delivery did better, at about 500 orders. CEO Dara Khosrowshahi was candid: "It is very, very early and very limited traction." The agents that can do everything in demos are nearly invisible in production data — the platforms keep building agent interfaces; users' thumbs just haven't stepped aside.
Source: TechCrunch / Uber Q3 earnings call | 2026-11-04
Nothing's first "AI-native" product ships in December: no screen, and it couldn't be an app
Nothing CEO Carl Pei announced on X this Wednesday that the company's first "AI-native product" ships in December. His description is restrained and full of suspense: "It's a tool, but one built with the personality and care that have come to define a Nothing product. It has no new screen, and it couldn't have been an app." He calls it a productivity tool and is deliberately keeping his distance from AI-wearable hype — the company has teased AI glasses and AI-linked earbuds before, but he declines to file this one in that drawer. "The whole industry is hunting for its iPhone moment: the pendant, the pin, the glasses, the companion, some magical consumer device that makes AI part of everyone's life." Then he turns and pours cold water on the trend.
Source: TechCrunch / Carl Pei on X | 2026-11-04
37 skills, 17 agents, 8x the cost: fancy coding pipelines lost to thinking harder
The arms race in coding-agent harnesses just met a set of sobering controlled experiments. Surge AI tested 4 popular coding-agent setups on 100 real software engineering tasks (all driven by Opus 5.5 in Claude Code), and the conclusion stings: the plainest setup — leaving the harness untouched and simply raising the model's thinking effort — beat every fancy add-on, gaining 11.3 points over medium effort, more than any single harness delivered. The most elaborate setup carried 37 skills, spawned 17 agents on one task, cost 8 times as much, and still scored worse than vanilla. The most eye-catching failure: asked to add a Zoom link and passcode to event confirmation emails, the 17-agent setup worked and worked and none of them noticed the email went out at registration, before payment — unpaid registrants would receive the passcode.
Source: Surge AI / Hacker News | 2026-11-03
🏆 今日头条
付费也逃不过：Gemini 默认拿订阅用户的聊天记录训练模型
一个付费订阅用户的愤怒发帖，本周三在 Hacker News 上拿到了六百多个赞。发帖人发现自己的 Gemini 聊天记录正在被拿去训练模型——这不是数据泄露，而是白纸黑字写在文档里的默认设置：个人 Google 账号默认开启「活动保留」，只要这个开关开着，你的对话就会被用于改进 Google 的服务，「包括训练生成式 AI 模型」。付不付费都一样。 真正扎心的是耦合设计：想关掉训练，就得连记忆一起关。「活动保留」这一个开关，同时控制着聊天历史的保存、个性化记忆，以及远程浏览器和远程电脑执行代码等高级功能的可用性——想要方便，就要接受训练；想要退出，就得回到每次都从零开始的状态。按 Google 文档的说法，人类审核员可能在特定情形下阅读对话内容，相关数据最长保留三年。 公平地说，这一切没有藏在暗处——Gemini 隐私中心的文档写得明明白白，只是不在用户会打开的订阅页上。企业版 Workspace 账号默认不参与训练，这也解释了为什么「付费」和「隐私」在 Gemini 这里是两条不相交的线：订阅费买的是功能和额度，退出训练的选项只对企业管理员开放。 > 💬 免费产品用数据换服务，逻辑自洽；付费产品还默认拿数据训练，逻辑就断了——更妙的是把「退出训练」和「关掉记忆」焊在同一颗开关上，用户连选择的资格都没有。默认值就是商业模式，这一课值得所有 AI 厂商抄进笔记。
来源：Hacker News / Google Gemini 隐私中心 | 2026-11-04
OpenAI 和 Meta 同一天出手，抢的全是编码智能体的「工作台」
周三的 AI 编码工具赛道一天两响。OpenAI 官宣收购 Pierre Computer Company：这家 2023 年创立的十人小团队做着一件对智能体时代至关重要的事——Code Storage，一套 API 优先、为机器而非人类设计的 Git 基础设施，智能体生成的代码和产物可以在这里规模化存取；其开源库 Trees 和 Diffs 已经进入 Codex。同一天，Meta 确认收下 Conductor（Axios 首报）：一个让工程师同时指挥多个编码智能体、在并行工作区里协同干活的产品，创始人 Charlie Holtz 宣布加入 Meta，将与其首席 AI 官 Alexandr Wang 和超级智能实验室产品负责人 Nat Friedman 共事，并预告「下一个交互界面」明年初见分晓。两笔交易金额均未披露。模型层的军备竞赛还没打完，智能体工作台的卡位战已经开场。 > 💬 谁买到智能体干活的动线，谁就握住下一个入口——这轮并购买的基本不是收入，是位置。
来源：OpenAI 官方博客、Axios / TechCrunch | 2026-11-04
AMD 交出创纪录一季报：数据中心营收暴涨 97%
AMD 在 11 月 3 日盘后发布三季度财报：营收 134 亿美元，同比增长 45%，超出指引上限；数据中心板块同比暴涨 97%，Instinct GPU 和 EPYC CPU 双双创下销售纪录；非公认会计准则口径每股收益 2.09 美元，同比增长 74%。CEO 苏姿丰表示，随着 Helios 和 Venice 两代产品放量，第四季度增长还会进一步加速。与数据中心一枝独秀相对的，是游戏业务的明显失速。当 AI 基建成为一家芯片公司的绝对引擎，传统业务的版图就只剩下衬托。 > 💬 一份财报两种温度：机架烧到发红，客厅坐冷板凳——AI 基建浪潮的资产负债表版本，就长这样。
来源：AMD 投资者关系 / CNBC | 2026-11-03
AI 算力股又跌了，这次不是因为财报
AI 与数据中心概念股周三再度承压，导火索换成了落幕的美国中期选举：主张给数据中心扩建踩刹车的阵营在多个科技重镇胜出，据 CNBC 报道，得州州长之位易主，当选的 Gina Hinojosa 主张立即暂停该州数据中心新建。富国银行的统计显示，全美规划中的数据中心装机容量约 27% 位于得州；Yorkville Ives & Co 的 Dan Ives 直言「没法粉饰」，选举结果将拖慢全美数据中心的建设节奏。特斯拉开盘跌 5%，SpaceX 跌 4%。上周五市场算的是资本开支的账，这一轮轮到政策登场：算力狂奔的下半场，先要看土地、电力和审批的脸色。 > 💬 需求端的报表再漂亮，也架不住供给端多出一行「政策风险」——算力这门生意的定价模型，刚被塞进一个新变量。
来源：CNBC | 2026-11-04
一边把 AI 放进 Siri，一边给员工的 AI 预算打折：苹果每日上限从 300 美元降到 200 美元
据 Bloomberg 报道，苹果正在下调员工的 AI 使用预算：工程师的每日 AI 开销上限从 300 美元降至 200 美元，后续还有进一步压缩计划，此举是全公司控制成本动作的一部分。与此同时，这家公司正努力把 AI 放到自家产品和运营的舞台中央。左手把 AI 织进系统功能，右手给内部用 AI 划预算红线，苹果在「讲故事」和「算账单」之间走着钢丝。 > 💬 对外讲 AI 故事，对内算 AI 账单——大公司的 AI 战略进入「既要又要」的精算阶段，连最有钱的公司都在数 token。
来源：Bloomberg / The Verge | 2026-11-04
Uber 财报电话会泼了盆冷水：最繁忙的一周，AI 智能体只叫了 4 单车
Uber 周三的三季度财报电话会上，CFO Balaji Krishnamurthy 报了一个冷到结冰的数字：在万圣节那一周——Uber 全年最繁忙的单周，打车与外卖合计约 3.2 亿次出行和订单里——通过 AI 智能体下的打车订单只有 4 单；外卖侧稍好，也只有约 500 单。CEO Dara Khosrowshahi 的总结很坦诚：「现在还非常、非常早期，能看到的使用非常有限。」发布会上无所不能的智能体，在真实生产数据里几乎隐形——平台照旧在搭智能体接口，只是用户的拇指还没让位。 > 💬 演示视频里智能体什么都能替你干，生产数据里一周干 4 单——用户教育这一关，比接口对接难多了。
来源：TechCrunch / Uber 财报电话会 | 2026-11-04
Nothing 首款「AI 原生」产品 12 月出货：没有屏幕，也不可能是应用
Nothing CEO Carl Pei 周三在 X 上官宣：公司首款「AI 原生产品」将于 12 月出货。他的描述克制又充满悬念：「它是一个工具，但带着 Nothing 产品一贯的个性和用心。它没有新屏幕，而且它不可能是一个应用。」他称其为生产力工具，并刻意与 AI 可穿戴的炒作保持距离——此前公司预告过 AI 眼镜和带 AI 的耳机，但这次的新品他不愿归入那一类。「整个行业都在寻找自己的 iPhone 时刻：吊坠、别针、眼镜、伴侣，某个能把 AI 塞进所有人生活的神奇消费设备。」话锋一转，他对这股风潮泼了冷水。 > 💬 手机圈最懂营销的公司用「反类目」话术吊足了胃口——产品是什么不重要，重要的是你已经记住它不是什么。
来源：TechCrunch / Carl Pei 的 X 帖子 | 2026-11-04
37 个技能、17 个智能体、8 倍成本：豪华编码流水线输给「多想一会儿」
编码智能体框架的军备竞赛，遇到了一组冷静的对照实验。Surge AI 用 100 道真实软件工程任务测试了 4 套流行的编码智能体流水线（模型统一用 Opus 5.5 驱动的 Claude Code），结论有点伤人：最朴素的那套——不改框架、直接调高思考档位——跑赢了所有花哨外挂，比中等档位多得 11.3 分，比任何单一外挂的提升幅度都大。最豪华的一套装了 37 个技能、一次任务派出 17 个智能体、成本是 8 倍，成绩反而不如裸跑。最刺眼的失败案例：让智能体给活动确认邮件加 Zoom 链接和密码，17 个智能体忙活半天，没有一个发现邮件在用户付款前就发出去了——没付钱的人也能拿到入场密码。 > 💬 预算花在「加人手」不如花在「让它想」——至少目前，复杂度本身是负资产。
来源：Surge AI / Hacker News | 2026-11-03