karminski-牙医@karminski3 · 4天前70

教你如何10秒钟训练一个小模型! 教大家如何从0训练一个(电)小(子)模(鹦)型(鹉)! (不包熟啊.....逃...) 只需要10秒钟! 而且完全不用搭建环境! 全程在网页训练! 首先你需要有个Mac, 我试了下N卡应该也行, 但是貌似有点问题适配的不好(我的3080Ti它适配失败了WarpSize不支持), 所以建议还是使用 Apple Silicon (M1-M5) 的 Mac 训练. 然后使用llmistanbul直接把你的纯文本文档拖进去就行, 尽量不要包含奇怪的格式, 比如markdown或者json啥的, 不然输出会很奇怪. 我这里直接把哈利波特1-7拖了进去 (注意, 你私下炼作为研究没人管, 但不要把别人的著作炼完了发出去, 请尊重原作者). 然后按照我这个图1：

译推文介绍了如何使用 llmistanbul 在网页上10秒内训练一个小模型（电子鹦鹉）。只需将纯文本文档（如哈利波特1-7）拖入即可，建议使用 Apple Silicon Mac（M1-M5），避免 markdown/json 等格式。N 卡（3080Ti）适配不佳。提醒尊重版权，勿公开发布他人作品。

Rohan Paul@rohanpaul_ai · 4天前70

New Anthropic research shows AI agents may look brilliant at code, but in biology they can fail before the science starts. Strong AI agents could give very different answers to the exact same biology data request, even when nothing changed in the prompt. In one Ebola sequence task, Claude Sonnet 4 returned 106 sequences in 1 run, then 15, then 5, while the expected answer was 266. Those missing sequences did not just make the dataset messy, they changed the scientific story built on top of it. One bad retrieval made the outbreak look like it traced back to 1922, instead of the manually curated result pointing to early 2014. The biology databases were too hard to use reliably through current AI tools. The agents often understood what they were being asked, but their answers varied a lot because they had to fight through scattered databases, hidden website rules, and fragile scripts. The key finding is that adding a repeatable retrieval tool made agents far more accurate and much more consistent.

译Anthropic 研究发现，AI 智能体在代码任务表现出色，但在生物数据库检索中容易失败。以埃博拉序列任务为例，Claude Sonnet 4 三次运行分别返回 106、15 和 5 条序列，而预期为 266 条。缺失序列导致科学结论严重偏移：智能体推断疫情回溯至 1922 年，人工筛选结果却指向 2014 年初。问题根源在于生物数据库分散、网站规则隐蔽、脚本脆弱。引入可重复检索工具后，智能体准确性和一致性大幅提升。Anthropic 呼吁建设更友好的基础设施。

Krea@krea_ai · 4天前58

AI in architecture studios – Krea Podcast with Nitsan Bartov full interview below 👇

译AI 在建筑工作室——Krea 播客与 Nitsan Bartov 完整采访见下方 👇

swyx@swyx · 4天前62

It's finally out!!! @METR_Evals found that more than half of SWEBench results is unmergeable slop. FrontierCode represents over 1000+ hours of maintainer validated software engineering work most frontier models cannot yet solve, much less solve with high quality. Cog had IOI Gold medalists and top code maintainers Look At The Data — FrontierCode includes 3000+ rubrics covering code quality and anticheat reward hacking plaguing other benchmarks. FC Diamond is so hard that Opus 4.8 scores 13.8%. Three eras of AI coding : Three eras of benchmarks 2021 • Autocomplete : HumanEval 2023 • Passing Tests: SWEBench, TerminalBench 2026 • Maintainable Code: FrontierCode to me the most beautiful chart when I requested a special historical run into all extant old models, the data was finding that the easiest third of FC tasks (in FC Extended) were rapidlly and suddenly solved over late 2025 - Opus almost doubled from a 41% pass rate to 74% in 4 months. This describes the "WTF happened in Dec 2025" vibe shift that a lot of folks from @dhh to @karpathy have called out: it is the difference between getting 95% success in 2 rerolls vs 6, making it finally feasible to go up the next layer of abstraction in agentic coding, eg @GeoffreyHuntley's ralph loops or @bcherny's /goals or @steipete's "loops that prompt your agents" without fearing too much that things go off the rails. My guess: as AI accelerates from here, each FrontierCode tier will saturate in sequence, hopefully ~annually. I've already asked the team to prepare FrontierCode 2027.... The old mountains will be destroyed. Their rubble becomes regolith. And from that regolith, the next model forest grows. Circle of life.

译Cognition 发布 FrontierCode 编码评估，每任务由顶级开源维护者花费 40+ 小时编写。METR 发现 SWEBench 超一半结果为不可合并的垃圾代码。FrontierCode 含 3000+ 评分标准，首次衡量代码是否可合并。最高难度 FC Diamond 上，Opus 4.8 仅得 13.8%。在 FC Extended 最易任务中，Opus 在 2025 年底 4 个月内从 41% 提升至 74%，标志 AI 编码进入"可维护代码"时代。

宝玉@dotey · 4天前54

帮转，豆包手机团队招设计工程师

译字节跳动豆包手机团队招聘设计工程师，需具备 Android 平台经验。团队调研总结了五种设计工程师画像：AI Design Engineer（转译 AI 能力为交互产品体验，处理 agent workflow、tool call、状态反馈等）、Product UI Craft Engineer（打磨高质量前端原型和交互细节）、Design Systems Engineer（建立设计系统与前端基础设施，连接 Figma 变量和代码组件）、Creative Technologist / Motion & Graphics Engineer（负责动效、实时图形、3D/空间交互）、AI Design Workflow Architect（搭建 AI 辅助设计工作流，使用 Claude Code、Cursor、v0 等工具）。多数设计工程师技能交叉，欢迎感兴趣者联系。

Artificial Analysis@ArtificialAnlys · 4天前59

MiniMax-M3 scores 55 on the Artificial Analysis Intelligence Index. Once the weights are released, it will be the leading open weights model M3 is @MiniMax_AI's first multimodal M-series model, adding image and video input and a 1M token context window over the text-only MiniMax-M2.7 (50). At 55 on the Intelligence Index it sits just ahead of open weights peers Kimi K2.6 (54) and MiMo-V2.5-Pro (54). MiniMax has noted they plan to release the weights within ~10 days. When MiniMax released the weights for M2.7, it was under a commercially restricted license. Key takeaways: ➤ MiniMax-M3 improves on MiniMax-M2.7 across most evaluations. HLE +9 points (28% to 37%), GPQA Diamond +6 (87% to 93%), AA-LCR +5 (69% to 74%), IFBench +7 (76% to 83%), and CritPt +3 (1% to 4%), with a small regression on SciCode (47% to 45%) ➤ M3 scores ~1670 on GDPval-AA, behind Claude Opus 4.8 (max, 1890) and GPT-5.5 (xhigh, 1769), and level with Claude Sonnet 4.6 (max, 1676). GDPval-AA measures real-world tasks across 44 occupations and 9 industries ➤ Native multimodality, scoring ~80% on MMMU-Pro. Level with GPT-5.5 (xhigh, 79.9%) and Kimi K2.6 (79.4%), behind Gemini 3.5 Flash (high, 84.3%). Not all open weights models support native vision input ➤ On AA-Omniscience, heavy abstention drives both low hallucination and low accuracy. M3 attempts only 30.9% of questions, the lowest among current peers, yielding a low hallucination rate (16.1%) and low accuracy (15.0%) ➤ MiniMax-M3's token usage is close to M2.7's, using ~91M output tokens to run the Intelligence Index (~81M reasoning) versus ~87M (~79M reasoning), while scoring 5 points higher Key model details: ➤ Context window: 1M tokens, up from MiniMax-M2.7's 200K ➤ Pricing: $0.30/$1.20 per 1M input/output tokens up to 512K context, rising to $0.60/$2.40 for 512K to 1M context ➤ Weights: Not yet released. MiniMax has stated the weights will follow ➤ Availability: MiniMax first-party API, @SiliconFlowAI, @gmi_cloud, and @novita_labs

译MiniMax推出首个多模态M系列模型M3，支持图像/视频输入及1M token上下文窗口。在Artificial Analysis Intelligence Index上得55分，超越开源权重的Kimi K2.6和MiMo-V2.5-Pro（均54）。相比前代M2.7，HLE提升9点至37%，GPQA Diamond提升6点至93%，多项基准均有进步。原生多模态MMMU-Pro约80%与GPT-5.5持平。定价$0.30/$1.20/1M tokens（512K内），512K-1M翻倍。权重计划约10天内开源。

Chubby♨️@kimmonismus · 4天前66

WWDC 2026: A brief assessment At WWDC26, Tim Cook's last keynote before he hands the CEO role to John Ternus on September. I've been waiting for WWDC 2026 for a long time. And somehow I got almost everything I wanted. But somehow I still expected more. Before I jump to conclusions, though, I should try everything out first. Here's the first caveat: Apple Intelligence won't be rolled out in the EU initially. What a surprise. Not. The same disappointment every time. Apple introduced "Siri AI," a full rebuild of the assistant that does the things the company first demoed in 2024 and then quietly pushed back twice. It reads what's on your screen, pulls context from your messages, mail and photos, and chains actions across apps. There's a standalone Siri app now, with a conversation history that syncs through iCloud, so it finally behaves like the chatbots people have spent three years getting used to. Here's the part Apple said quietly and everyone else said loudly: the brains are Google's. Siri AI runs on Gemini under the multiyear deal the two companies announced in January. Reports put that deal at roughly a billion dollars a year for a custom large model. Apple paired it with its own on-device Foundation Models and wrapped the whole thing in a privacy story, with Craig Federighi insisting that privacy in AI is non-negotiable and that data is only used to execute your request. The rest of Apple Intelligence is the steady stuff. Photos gets Spatial Reframing, which improves a photo's composition after it's been taken. Safari can monitor a page and notify you about restocks or price drops. Messages offers one-tap suggestions to create a reminder or note based on the conversation. Image Playground adds photorealistic generation and a "describe a change" edit mode. None of it makes headlines alone, but together it's Apple catching up to where the industry was a year ago. Everything else was housekeeping, and some of it is genuinely good. Liquid Glass now has a slider that runs from ultra-clear to fully tinted. macOS 27, dubbed Golden Gate, brings back the uniform toolbars and edge-to-edge sidebars Mac users missed. Performance got real attention: apps launch up to 30 percent faster, AirDrop is up to 80 percent faster, and Apple retuned the CPU scheduler so older iPhones feel quicker. Oh, and rebuilt search across Spotlight, Photos and Mail. Oh, and for some reason almost no WatchOS updates other than a few performance improvements. Disappointed (big Apple Watch fan tho) tl;dr: *Apple Intelligence & Siri AI* - "Siri AI," an entirely new Siri across iPhone, iPad, Mac, Apple Watch and Vision Pro, built on a new privacy-focused architecture. - Powered by Google Gemini (multiyear deal announced Jan 2026, reported at ~$1B/year for a custom model) combined with Apple's own on-device Foundation Models. - On-screen awareness, personal-context search across messages/email/photos, systemwide app actions, and live web answers with world knowledge. - A dedicated Siri app to revisit or start conversations, with history synced privately via iCloud. - Adjustable pace, expressivity and accent for the conversational experience. - Visual updates: Siri animation in the Dynamic Island; swipe down from mid-screen to launch Siri AI. - Siri mode in the Camera app and expanded Visual Intelligence. - Apple Intelligence in apps: Spatial Reframing in Photos, Safari "Notify Me" page monitoring, one-tap suggestions in Messages, photorealistic generation and "describe a change" editing in Image Playground, a new Top Hits ranking in Mail. - Privacy framing front and center: data only used to execute the request, verifiable by outside experts. *Availability & the regional catch* - Developer betas today, public beta next month, free update this fall. - AI features require iPhone 16 or later / iPhone 15 Pro, M1+ iPads and Macs, Vision Pro, Apple Watch Series 10+. - Siri AI not in the EU on iOS/iPadOS at launch (Mac, Watch, Vision Pro yes), due to the DMA. - No new Apple Intelligence features in China at launch, pending regulation. - Image generation has daily limits; iCloud+ raises them. *Design & performance* - Liquid Glass personalization slider (ultra-clear to fully tinted), plus sharper app icons. - macOS 27 "Golden Gate": uniform toolbars, edge-to-edge sidebars, colored sidebar icons, tighter corner radius. - Apps up to 30% faster to launch, photos up to 70% faster to appear, AirDrop up to 80% faster, iPad external-drive transfers up to 5x faster; CPU scheduler retuned for older devices. - Rebuilt search across Spotlight, Photos and Mail. - iOS 27 supports iPhone 11 and later, the widest iOS reach yet. *Everything else across platforms* - iCloud Shared Albums now full-resolution and cross-platform (incl. Android and Windows). - Health: perimenopause and menopause support in Cycle Tracking. - Apple Watch: dynamic app grid of five Siri-suggested apps, a Smart Stack widget tap gesture, a consolidated Find My app. - AirPods: custom EQ; AirPods Pro 3 heart-rate sync via GymKit. - Vision Pro: panoramas convertible into spatial Environments; Wi-Fi up to 3x faster. - Apple Maps: enhanced Flyover combining aerial imagery with AI. So far this looks like a solid WWDC but not revolutionary. Looking forward to test updated Siri / Apple Intelligence although, as a european, I will have to wait :/

译WWDC 2026 上苹果推出全新 Siri AI，由 Google Gemini（多年代价约 $1B/年定制模型）与自研端侧 Foundation Models 驱动。支持屏幕感知、跨消息/邮件/照片个人上下文检索、系统级应用连携操作及实时网页回答，新增独立 Siri 应用与 iCloud 同步对话历史。其他 Apple Intelligence 更新包括 Photos Spatial Reframing、Safari 页面监控、Messages 一键建议、Image Playground 照片级生成与编辑。性能方面，应用启动快 30%，AirDrop 快 80%。macOS 27 命名 Golden Gate。Siri AI 首发不在欧盟可用（数字市场法案限制）。

Chubby♨️@kimmonismus · 4天前73

Apple: “Our most powerful on-device AI model runs on iPhone 17 Pro.” Me: Great, I have one. Apple: “Do you live in Europe?” Me: …

译Apple 在 WWDC 2026 推出基于 Google Gemini（多年度协议，据报约每年 10 亿美元定制大模型）与自研端侧 Foundation Models 的全新 Siri AI，支持屏幕感知、个人上下文搜索、跨应用操作及独立 Siri app（iCloud 同步对话历史）。Apple Intelligence 还包括 Photos Spatial Reframing、Safari 页面监控、Messages 一键建议、Image Playground 照片级生成等功能。最强端侧 AI 模型仅运行在 iPhone 17 Pro 上，且初期不在欧盟推出。隐私端强调数据仅用于执行请求，可被外部专家验证。

Anthropic@AnthropicAI · 4天前61

New Science Blog: Why has AI advanced faster in coding than in biology? To agents, bio databases are like cities built before cars—maddening to drive in because they're designed for different traffic. How do we build infrastructure agents can use? https://www.anthropic.com/research/agents-in-biology

译新的科学博客：为什么 AI 在编码方面的进步比在生物学方面更快？ 对智能体而言，生物数据库就像汽车发明前建造的城市——开进去会让人抓狂，因为它们是针对不同的交通流量设计的。 我们如何构建智能体可以使用的基础设施？ https://www.anthropic.com/research/agents-in-biology

Ethan Mollick@emollick · 4天前54

Last time around Apple released a lot of information about how their AI version of Siri worked between local and cloud models, not so much this time It is nice to have a Gemma-like model on device, but it is extremely limited unless it can call a smarter cloud model when needed.

译上次苹果发布了很多关于其AI版Siri在本地和云端模型之间如何工作的信息，这次则没有那么多。 在设备上拥有一个类似Gemma的模型很好，但除非它能根据需要调用更智能的云端模型，否则其能力极其有限。

jason@jxnlco · 4天前40

If you use poke check out my autmation to keep track of codex launch’s. Maybe you’ll see @interaction mentioned in one of these one day. http://poke.com/r/7ZaRDldAw3s

译如果你用 poke，可以看看我的自动化工具，用来跟踪 codex 的发布。 也许有一天你会在其中看到 @interaction 被提及。 http://poke.com/r/7ZaRDldAw3s

-Zho-@ZHO_ZHO_ZHO · 4天前18

这次的 Siri AI 满足了我两年前对 Siri 的想象笑死

-Zho-@ZHO_ZHO_ZHO · 4天前33

Siri AI + 空间智能生成 一顿智能后，这构图和效果还不如原图呢，这是来搞笑的？？？

Chubby♨️@kimmonismus · 4天前39

- System wide awareness - Visual Intelligence in Camera App

译- 系统级感知 - 相机应用中的视觉智能

Chubby♨️@kimmonismus · 4天前48

Apple Intelligence sounds like a very good update. But did I miss all the relevant (nerd) stuff like paramters, evals in benchmarks and what not? Hopefully on their website

译Apple Intelligence 听起来是一个非常好的更新。但我是否错过了所有相关的（极客）内容，比如参数、基准评测等等？希望官网上有。

Chubby♨️@kimmonismus · 4天前32

Interesting: New Apple Intelligence Siri only available on iPhone 17 Pro. Of course not be available in the EU (god damn)

译有趣：新的 Apple Intelligence Siri 仅限 iPhone 17 Pro 使用。 当然在欧盟不可用（该死）

Rohan Paul@rohanpaul_ai · 4天前58

The prompt era is ending. That's too linear, too bottlenecked by humans. We are entering the loop machine of AI agents. The value is in moving judgment upstream, so the human designs the process and the model handles the recurring friction.

译提示词时代正在终结。那太线性了，太受人类瓶颈限制了。 我们正在进入AI智能体的循环机器。 价值在于将判断上移，让人类设计流程，模型处理重复出现的摩擦。

jason@jxnlco · 4天前18

codex and computer use is so powerful

译Codex和Computer Use非常强大

Boris Cherny@bcherny · 4天前65

When we first demoed Claude Code internally, it got two reactions on Slack. A year after GA, @_catwu and I sat down to talk about what's changed: why I use auto mode instead of plan mode, how routines fix bugs before I see them, why I do most of my coding from my phone now, and where the product is going

译Claude Code GA一周年之际，Anthropic工程师Boris Cherny与@_catwu回顾产品演进。此前内部首次演示时在Slack上引发两种截然不同的反应。Cherny分享了他为何偏好auto mode而非plan mode，routines如何在bug出现前自动修复，以及他如今大部分编码都在手机上完成。视频访谈还探讨了Claude Code的未来方向。

Chubby♨️@kimmonismus · 4天前38

Apple Intelligence: -Personal Understand in apps. - Browse tools for web - on screen Awareness - in App usage

译Apple Intelligence: - 在应用中的个人理解。 - 网页浏览工具 - 屏幕感知 - 应用内使用

Yuchen Jin@Yuchenj_UW · 4天前57

On the whole: “You shouldn’t be prompting coding agents anymore. You should be designing loops that prompt your agents.” Loops are the temporary workaround: today’s LLMs have poor judgment. They struggle to know when to keep going, when to stop, or when to call a tool. Loops force agents to work longer. Loops are incredibly powerful for verifiable goals for now, as AutoResearch shows.

译总体来说： “你不应该再提示编码智能体了。你应该设计循环来提示你的智能体。” 循环是临时解决方案：如今的大语言模型判断力很差。它们难以判断何时继续、何时停止或何时调用工具。循环强制智能体更长时间地工作。 对于目前可验证的目标，循环非常强大，正如AutoResearch所示。

ClaudeDevs@ClaudeDevs · 4天前74

Claude Code's first demo got two Slack reactions. One year after GA, @bcherny and @_catwu look back: verification best practices, why we built auto mode, routines and loops, and what's next. https://www.youtube.com/watch?v=Hth_tLaC2j8

译Claude Code 的第一个演示收到了两个 Slack 反应。 GA 一周年之际，@bcherny 和 @_catwu 回顾：验证最佳实践、为何构建自动模式、例程和循环，以及下一步计划。 https://www.youtube.com/watch?v=Hth_tLaC2j8

Chubby♨️@kimmonismus · 4天前33

Apple Intelligence last. Let the fun begin!

译蒂姆·库克仍在主持WWDC开幕。Apple Intelligence最后登场，好戏开始了！

Yuchen Jin@Yuchenj_UW · 4天前57

“You should design loops that prompt your agents.” Loops are the temporary workaround: today’s LLMs have poor judgment. They struggle to know when to keep going, when to stop, or when to call a tool. For verifiable goals, loops are incredibly powerful, as AutoResearch shows.

译“你应该设计循环来提示你的智能体。” 循环是临时方案：今天的LLM判断力很差。它们很难知道何时继续、何时停止、何时调用工具。 对于可验证的目标，循环非常强大，正如AutoResearch所示。

宝玉@dotey · 4天前61

微信格局还是不够，总是想着大家都去他们家一亩三分地耕耘，还幻想着未来微信会继续是超级入口，人人都在用微信，所以只需要让 AI 去操作小程序。 但现实是，未来微信的入口属性会越来越少，以后的年轻人，不会再去打开微信，只会问自己的 Agent：去帮我总结一下我昨天的群聊，去给我妈发条消息说晚上不回家吃饭了。 而这个承担超级入口职责的 Agent，大概率不是微信 AI。

译微信发布《开发者接入微信 AI 生态的指引》，引导小程序开发者接入微信 AI，让 AI 控制小程序。宝玉对此评论称，微信试图通过让 AI 操作小程序来维持自身超级入口地位，但未来年轻人不会主动打开微信，而是直接向自己的 Agent（如"帮我总结群聊"或"给妈妈发消息"）发出指令。承担超级入口职责的很可能不是微信 AI。

jason@jxnlco · 4天前4

14 weddings happened this weeekend and here I am working on my @aiDotEngineer talk

译这周末举办了14场婚礼，而我却在准备我的@aiDotEngineer演讲。

SemiAnalysis@SemiAnalysis_ · 4天前63

China's Unitree Will Dominate Global Robotics The Fastest Iteration Cycle In Next-Gen Robotics Should See Unprecedented Acceleration https://newsletter.semianalysis.com/p/chinas-unitree-will-dominate-global

译中国宇树将主导全球机器人 下一代机器人最快的迭代周期将迎来前所未有的加速

Chubby♨️@kimmonismus · 4天前63

What many misunderstand: Apple doesn't actually need the best model in the world. It's similar to Meta. Their model only needs to be good enough for 99% of everyday use cases. They don't even want to compete with Frontier Labs, but primarily reach the consumer market. And Apple actually has a good chance there. Because a well-adapted Gemini model, based on (3.1/3.5?) and well integrated into the OS, could achieve exactly the use case that many need: AI that simplifies their daily work.

译苹果在WWDC 2026上承认无法独立构建前沿AI，转而与Google合作。新Siri将基于定制1.2T参数Gemini模型（版本或为3.1/3.5），每年花费约10亿美元（Gurman）。Siri作为独立应用，支持iMessage式聊天、动态岛弹窗、扩展系统及邮件/日历/网页查询，运行在Private Cloud Compute上，Google不会用查询数据训练。苹果策略类似Meta——模型只需满足99%日常场景。iOS 27被定位为“雪豹”式清理更新，放弃iPhone 11和SE2支持，并可能允许用户选择AI引擎（Gemini或Claude）。

AYi@AYi_AInotes · 4天前37

梁文锋的DeepSeek的问世成功颠覆了什么？

gabriel@gabriel1 · 4天前48

people thought agi is a blank textbox where we enter our intentions and get what we want but ask your smartest friend to "clean up my inbox", you'll realize that for him to do it perfectly you'd need to write down 5 pages of instructions, and these instructions changes daily

译人们以为 AGI 是一个空白文本框，输入意图就能得到想要的东西。 但让你最聪明的朋友去“清理我的收件箱”，你会发现为了让他完美完成，你需要写下 5 页的指令，而且这些指令每天都在变化。

jason@jxnlco · 4天前17

What artifacts do you create in codex outside code?

译你在Codex中除了代码还创建了哪些工件？

elvis@omarsar0 · 4天前65

Great tips. In practice, this is how it roughly looks to run agents autonomously for hours or days. /goal or /loop to keep it going. Verification is crucial here.

译@bcherny 分享5条技巧：1) 开启自动权限模式，免手动确认；2) 采用动态工作流，让Opus协调数百/数千Agent；3) 使用/goal或/loop指令促使持续执行；4) 在云端运行Claude Code，可关闭笔记本；5) 确保Opus能端到端自验证——通过Chrome扩展验证网页、iOS/Android模拟MCP验证移动端、启动完整后端服务验证后端。Elvis Saravia强调/goal/loop和验证是关键。

Chubby♨️@kimmonismus · 4天前54

WWDC 2026 - Apple rents Google's brain to fix Siri. What we can expect: Apple's keynote today is a software reset built around one admission: it couldn't build frontier AI alone. What to expect: - Gemini-powered Siri, a rebuilt assistant on a custom 1.2T-parameter Google model, ~$1B/year (Gurman). Runs via Private Cloud Compute, no Google training on your queries. - Siri as an app, standalone, iMessage-style chat with synced history, a "Search or Ask" Dynamic Island pop-up, and an Extensions system. Drafts emails, pulls from mail, calendar, contacts and the web. - Six OS betas - iOS 27, iPadOS 27, macOS 27 ("Big Bear"), watchOS 27, tvOS 27, visionOS 27. iOS 27 is a "Snow Leopard" cleanup release. iPhone 11 and SE2 lose support. - Liquid Glass 2.0 - system-wide opacity slider, fixes for the shadow and transparency complaints. - AI health coach - the watered-down "Health+", now fitness and wellness instead of an AI doctor (pretty cool!) - Model choice (rumored), users may pick the engine behind Apple Intelligence, with Gemini and Claude floated. - Hardware mostly later - M5 Macs, new iMac, foldable iPhone (~$2.5K, Sept), OLED touchscreen MacBook Pro, smart-home hub. Sources: TechInsider, Bloomberg, Gamebezz

译苹果在WWDC 2026承认无法独自构建前沿AI，与Google合作，基于1.2T参数的Gemini模型重建Siri，年费约10亿美元。新版Siri通过Private Cloud Compute运行，谷歌不训练用户数据；Siri成为独立应用，支持聊天、同步历史、Dynamic Island弹出和扩展系统，可起草邮件并获取信息。六个OS beta发布，iOS 27为Snow Leopard清理版，iPhone 11/SE2失去支持；Liquid Glass 2.0透明度滑块；Health+改为健身聚焦；用户或可选AI引擎（Gemini或Claude）。硬件稍后推出。

Deedy@deedydas · 4天前64

Meta AI has shockingly grown 2.5x in the last 2mos and is poised to be the #3 AI consumer app in the world behind Gemini and ChatGPT. Sadly, this growth is very likely inorganic given it has by far the worst retention by a mile: only 4.5% users stay in 30 days.

译Meta AI 在过去两个月内惊人地增长了 2.5 倍，有望成为仅次于 Gemini 和 ChatGPT 的全球第三大 AI 消费级应用。遗憾的是，这种增长很可能是非有机的，因为它的留存率迄今最差：只有 4.5% 的用户会在 30 天后继续使用。

OpenRouter@OpenRouter · 4天前72

New server tool: Advisor Let smaller models consult a higher-intelligence "advisor" model. Helps them escape doom loops, and helps you migrate to cheaper models! 🧵

译新服务器工具：Advisor 让较小的模型咨询一个更高智能的“顾问”模型。 帮助它们逃出困境循环，并帮助你迁移到更便宜的模型！🧵

歸藏(guizang.ai)@op7418 · 4天前5

最近可以在各种媒体和账号上，看到我的那个 Skills 推荐。感谢量子位。

译最近可以在各种媒体和账号上，看到我的那个 Skills 推荐。感谢量子位。 [引用 @op7418]：http://x.com/i/article/2053655813877870592

Chubby♨️@kimmonismus · 4天前78

New from Hivemind: continual learning for AI coding agents, available to everyone starting today. It takes the traces from every agent your team runs (Claude Code, Codex, Cursor, Hermes, Pi) and turns them into reusable skills, then pushes those skills across all of them, all on your own cloud! With the new SkillOpt built in, those skills get trained as they accumulate: +19.1 points of accuracy in Claude Code +24.8 in Codex best or tied on all 52 setups tested Agents that learn on the job and share what they learn. Really exciting.

译Hivemind发布面向AI编程智能体的持续学习功能，即日起开放。该工具收集团队运行的每个智能体（Claude Code、Codex、Cursor、Hermes、Pi）的轨迹，转化为可复用技能并推送到所有智能体，数据存储在用户自己的云存储中。内置SkillOpt使技能持续训练：Claude Code准确率提升+19.1分，Codex提升+24.8分，在全部52个测试设置中最佳或持平。开源，一行命令安装。

Rohan Paul@rohanpaul_ai · 4天前63

Coinbase CEO Brian Armstrong thinks AI demand is almost limitless, but he expects 80% of workloads to shift to models that are 99% cheaper within 12-18 months.

译Coinbase CEO Brian Armstrong 预测，对智能的需求近乎无限，但 80% 的工作负载将在 12-18 个月内迁移到便宜 99% 的模型，仅 20% 继续运行在追求最高 IQ 的最新模型上（如科学突破、高级编排型 AI 智能体）。他类比高端 MacBook/游戏 PC 的配置占比，但指出模型价格下降远超摩尔定律。Armstrong 认为未来瓶颈是能源和算力，而非更好的模型。Coinbase 正将用户提示词路由到更便宜的模型，部分情况下 token 用量指数增长，成本基本持平。

SenseTime@SenseTime_AI · 4天前56

🙌 Shoutout to @FahdMirza for demoing 𝗦𝗲𝗻𝘀𝗲𝗡𝗼𝘃𝗮 𝗨𝟭’𝘀 𝘁𝗲𝘅𝘁-𝗶𝗺𝗮𝗴𝗲 𝗶𝗻𝘁𝗲𝗿𝗹𝗲𝗮𝘃𝗲𝗱 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻 — showing the step-by-step process of formulating a custom perfume 🧴✨ It doesn't just see images. It thinks in them — and outputs in vivid visuals 🎥 https://youtu.be/-uedweS3_w0 Explore prompt examples in SenseTime Studio's Gallery and build your own 👇 🎛️ SenseNova Studio: https://unify.light-ai.top/ (Try infographics; also join Discord for text-image interleaved gen) 🤗 https://huggingface.co/collections/sensenova/sensenova-u1 🛠️ https://github.com/OpenSenseNova/SenseNova-U1 👾 Discord: https://discord.com/invite/BuTXPHmQub

译商汤 SenseTime 展示 SenseNova U1 的文图交错生成能力，通过定制香水逐步演示，证明模型不仅能识别图像，还能以图像为思考单元输出生动视觉内容。相关示例、Gallery、HuggingFace 模型、GitHub 代码及 Discord 社区链接已同步开放。

gabriel@gabriel1 · 4天前40

every job will turn into explaining your intentions to ai explaining what you want to ai is surpringly time consuming, coders already spend 80% of their time doing it, and this will be true for everyone

译每份工作都将变成向 AI 解释你的意图 向 AI 解释你想要什么其实相当耗时，程序员已经有 80% 的时间花在这上面，而这对每个人来说都将如此。