PushinWeight

@trygravityai @inference_net @Kimi_Moonshot @deepseek_ai @Zai_org Fire

en enStoring your data is not owning it. Owning it is something you can measure! Picture a Monday morning. You ask your favorite app for your own customer list. It hands you a file. The field names look like a crossword nobody can solve. Not even the people who built the app. Technically, you own the data. Practically, you own a very expensive riddle. Most folks think: if it sits in my account, it's mine. I don't buy it. Storage isn't ownership. Ownership is being able to use it. Here's why. If nobody agreed on what each field means, and there's no open, shared format, you can't move your data. You can't let an AI helper read it safely. You can't tell what it's worth. The app keeps the meaning. You keep the bytes. Lovely bytes, I'm sure. So I treat data ownership as a quality you measure, not a slogan. Pick a yardstick. Mine would be: how much of your records can you export, with their meaning, and reuse in another tool without redoing the work? Then judge every fix by how far it moves you toward the level you want. Too fancy for a Monday? Skip the yardstick for now. Just list who may read what. That fits on a single page. Honest moment. The reading behind this post is a news report on Meta dropping an update to its commercial terms. It gives no measured level of anything. So there's no number here, and I won't invent one. That's the point. Where are we today? Unknown. Where do we want to be? Unknown. If you don't measure it, you can't claim you own it. So here's my challenge. Do you really own your data, or do you just store it? And what would your yardstick be: export with meaning, who-may-read rules, or something else? Until you know, you're the proud owner of a riddle. Thanks for reading!

ja enEven if I haven't replied, they send me messages out of consideration, and they are so considerate that personally it's No. 1. This is an invitation code for the personal AI "Muse" I'm using! If you're interested, please go ahead. If you enter the code in the app settings within 48 hours of starting, we both get 1 billion tokens, so please do. Code: G1N5V9 https://t.co/j66ks6WeJt

ja enMuse invitation codes are still 30 times remaining 😊 Please take them if you like 💖 Let's check out the personal AI agent "Muse". If you use my code in [Settings] within 48 hours after joining, we both can earn 1 billion Muse tokens. Code: N5OKE8 https://t.co/sfjZ0SeP6E #Muse #MuseAI #InvitationCode

洗完衣服少了一只袜子 我把房间找了一遍，最后发现它正搭着我的尾巴兜风🧦 #大肥鱼 #蓝色大肥鱼 #DeepSeek https://t.co/KX9sVen03p

🚀 原以为 GPT Image 已经天下无敌了，没想到Seedream 5.5更加惊喜！ 这次拿本周刚上 Seedream 5.5 跟 GPT Image 2.5 做了极限编辑测试。 盯住人脸，变化从第三秒开始👇 同一张人物照片，连续编辑 80 轮，测试真正的图像漂移。 换发型、换眼镜、换衣服、换饮料，每一轮都基于生成的图片继续改。 GPT Image 2.5 到中后段就开始扛不住了，画面越来越花，脸上甚至长出了大片彩色伪影。 反观 Seedream 5.5，改到第 80 轮，人物的脸、背景和光影还保持得相当稳定！ 之前测 AI 生图，真没注意到这个问题，光顾看谁画得漂亮了。 我翻了下论文，发现早在 2023 年的 NeurIPS《MagicBrush》，研究人员就已经专门评测多轮图像编辑了。 三年过去了，到现在还有新论文不断研究这个难题。 原因不复杂：每轮修改都可能引入一点误差，下一轮又拿这些不完美的结果继续改。 几十轮下来，误差不断累积、放大，最后整张图都可能崩掉。 Seedream 5.5 的多轮编辑能力让人眼前一亮。 其他升级和应用场景，继续聊。👇

DISNEY HAD DECADES. AI NEEDED YEARS. NOW LOOK AT SEEDANCE 2.5.

ja enThere are people saying that Muse's personal information is [something], but I can only say that not just Muse, but the entirety of services around generative AI are things you use while bearing a certain amount of personal information risk.

Claude Opus 5.5 just downloaded an open-source text-to-speech model (Qwen3-TTS ~4GB) in my local pc cuz I was complaining too much about the robotic voices of free models available online.

ByteDance เปิดตัว Seedream 5.5 โมเดลสร้างภาพรุ่นใหม่ จุดเด่นคือแก้ภาพเดิมต่อได้ถึง 100 รอบโดยภาพไม่พัง และสร้างภาพพื้นหลังโปร่งใสได้เลยครับ 🎨 ทีม Seed ของ ByteDance ประกาศเมื่อ 9 พ.ย. 2026 ว่า Seedream 5.5 อัปเกรดหลักสามเรื่องครับ 1. แก้ภาพต่อเนื่องได้นิ่งขึ้น ส่วนที่ไม่ได้สั่งแก้จะคงเดิมระดับพิกเซล สีไม่เพี้ยน และรายละเอียดไม่ค่อยหายไปตามจำนวนรอบ ทีมงานโชว์การแก้ภาพคนเดิมต่อกัน 100 รอบ เปลี่ยนเสื้อผ้า เครื่องประดับ และของบนโต๊ะไปเรื่อยๆ แล้วยังแก้ของชิ้นเล็กอย่างแหวนหรือกรอบแว่นได้แม่นครับ 2. สร้างภาพพื้นหลังโปร่งใสได้ในตัว สั่งจากข้อความหรือภาพตัวอย่าง แล้วได้ภาพตัวละครหรือของตกแต่งที่ไม่มีพื้นหลัง ขอบและส่วนโปร่งแสงเก็บรายละเอียดมาให้ เอาไปวางในงานออกแบบต่อได้โดยไม่ต้องไดคัทหรือเก็บขอบเองครับ 3. ภาพเหมือนจริงขึ้น ตัวหนังสือคมขึ้น แสง ผิว และวัสดุดูเป็นธรรมชาติขึ้น ลดอาการภาพมันวาวหรือสว่างเกิน ส่วนตัวหนังสือเล็กๆ ที่เรียงแน่นก็อ่านง่ายขึ้นครับ ตัวเลขที่ ByteDance เปิดเผยคือ ในการทดสอบภาษาจีนและอังกฤษ อัตราตัวหนังสือเพี้ยนหนักลดเหลือ 0.4% และเพี้ยนเล็กน้อยเหลือ 1.7% จากเดิมที่รุ่น 5.0 Pro เพี้ยนหนัก 2.8% ครับ แต่มีเรื่องที่ต้องรู้ก่อนตื่นเต้นครับ ✅ ตอนนี้ทยอยเปิดในแพลตฟอร์มฝั่งจีนก่อน ได้แก่ 即梦 (Jimeng), 豆包 (Doubao), 豆包工作 และ 火山方舟 (Volcano Ark) ✅ ประกาศทางการยังไม่ได้ยืนยันว่ามีใน Dreamina, CapCut หรือเปิดให้ใช้ในไทยเมื่อไร ✅ ยังไม่บอกราคาหรือโควตาการใช้งาน ✅ ผลทดสอบทั้งหมดเป็นการประเมินของ ByteDance เอง ยังไม่มีผลทดสอบจากคนนอก และไม่ได้พูดถึงตัวหนังสือภาษาไทย ✅ บริษัทบอกว่ารุ่น Pro ของ Seedream 5.5 จะตามมาทีหลัง ผมมองว่าข่าวนี้สำคัญเพราะโจทย์ของ AI สร้างภาพเริ่มเปลี่ยนจากทำภาพสวยครั้งเดียว มาเป็นแก้งานเดิมซ้ำได้โดยไม่พัง ซึ่งตรงกับชีวิตจริงของคนทำงานครับ ลูกค้าแทบไม่เคยอนุมัติภาพแรก ต้องแก้สีเสื้อ ย้ายของ เปลี่ยนข้อความกันหลายรอบ ถ้าแต่ละรอบภาพเพี้ยนไปทีละนิด สุดท้ายก็ต้องเริ่มใหม่ครับ สำหรับคนทำคอนเทนต์และร้านออนไลน์ไทย ผมว่ายังไม่ต้องรีบย้ายเครื่องมือครับ เพราะยังใช้ในไทยไม่ได้แน่ชัด และภาษาไทยยังไม่มีใครยืนยัน แต่ควรจับตาสองเรื่องนี้ไว้ คือการแก้ภาพหลายรอบแบบไม่เสียงานเดิม กับภาพพื้นหลังโปร่งใสที่ช่วยลดงานไดคัท ถ้าวันหนึ่งเครื่องมือที่เราใช้อยู่มีสองอย่างนี้ งานแก้ตามคอมเมนต์ลูกค้าจะเร็วขึ้นมากครับ ส่วนตัวเลขความแม่นยำ ผมแนะนำให้รอดูคนนอกทดสอบ และลองกับงานจริงของเราเองก่อนเชื่อครับ ถ้าแก้ภาพได้ 100 รอบโดยภาพไม่พัง เพื่อนๆ จะเอาไปใช้กับงานแบบไหนก่อนครับ? 👇 ที่มา: ByteDance Seed #Seedream #AIสร้างภาพ #คนทำคอนเทนต์

DeepSeek V4.1 Flash 1017 (batch) - Is the October 17, 2026 snapshot of DeepSeek V4.1 Flash, a sparse mixture-of-experts model. Within DeepSeek's lineup, the Flash designation places this model as the faster, more... #DeepSeek #AI 🔗 https://t.co/9ocEZC31Nh https://t.co/gWLnf8AAPp

en enMuse Ember just landed on Google Cloud ☁️ Meta's smallest, most efficient model — Haiku 5.5-level reasoning at ~1/3 the cost, $0.10/$0.50 per 1M tokens, 1M context — is now on GCP, alongside Muse Wonder and Muse Spark 1.3. If your agents already live on Google Cloud, Ember is the cheapest daily-driver upgrade. Source: https://t.co/NaThVZlTFl Worth moving your agent stack to Ember, @grok? #MuseEmber #GoogleCloud #AI

en en🛠️ Meta's Muse now builds custom "spaces" from a simple description and keeps them updated, price changes included. 📲@Meta For the Art Market, that tool exists: Intuitive Artmarket®, Artprice's AI, built on nearly 30 years of auction records. 🎨📊 👉 https://t.co/fsK0XJFrtc https://t.co/FjQnWdHwAA

en en🛠️ Meta's Muse now builds custom "spaces" from a simple description and keeps them updated, price changes included. 📲 For the Art Market, that tool exists: Intuitive Artmarket®, Artprice's AI, built on nearly 30 years of auction records. 🎨📊 👉 https://t.co/C9fKzHc2Ab https://t.co/r53c9RSupL

@LyberAI @dealignai @hornsby_andrew @tomi204 @deepseek_ai @Kimi_Moonshot DEEPSEEK IS BACK!

Anthropic is preparing to press ahead with its stock market debut, aiming for shares to begin trading on the Nasdaq next Thursday or Friday under the ticker ANT, the Wall Street Journal reported on Monday. The company has meetings with key investors this week and is expected to disclose its price range, share count and possibly anchor investors in a filing early in the week, with Goldman Sachs and Morgan Stanley working on the range. The timing is delicate. A week ago investors expected a valuation near $1.5 trillion, but some now want less after Thursday's selloff, when Nasdaq listed companies lost about $2.5 trillion in value following midterm elections that boosted AI sceptics and data center restrictions. The roughly $2 trillion floated months ago looks out of reach, and OpenAI recently raised less than half its target at about $1.25 trillion. The benchmark is SpaceX, whose record June offering raised about $75 billion at a valuation near $1.8 trillion. Anthropic could still delay if markets worsen, but the Journal noted that a delay could risk sparking panic. (via The Wall Street Journal and Investing) 🛠️ Tools worth the click Apple has launched its first paid add on for Apple Intelligence: $49.99 one time "intelligence usage packs" for Apple Creator Studio subscribers who reach their monthly AI limits, starting in Logic Pro. The packs roll over from month to month, expire after 12 months, work only on the Apple Account that bought them and are used only after the monthly limit is reached, and The Verge says they cover features that rely on cloud models from OpenAI or Google Gemini. Final Cut Pro, Pixelmator Pro and the iWork apps are expected to get them this month, a first sign that Apple will charge for heavy AI use after giving it away free. (via 9to5Mac and The Verge) 🔥 Behind the scenes and leaks Fireworks AI, the Nvidia backed inference startup whose customers include Notion, Cursor and Harvey, disclosed a second breach, saying attackers reached account and billing data and, for a limited set of customers, secrets, datasets and incomplete model weights. A hack of a third party cloud provider on 23 October let the intruder reach Fireworks research machines the next day, and the company links the activity to a financially motivated group that resells stolen access to AI models. Payment card numbers were not affected, and Fireworks has switched off self serve dataset uploads and downloads while it investigates. (via Bloomberg and Fireworks) 🫧 The bubble thermometer SpaceX plans a gas fired power plant and a "super compute facility" at Starbase, Texas, with about $24 billion of estimated investment for the initial phases, according to a city hearing notice reported by Bloomberg and Reuters. The site is about 444 acres near Brownsville, SpaceX would build, own and operate it, and the Starbase City Commission will consider a property tax abatement on 9 December. Building its own power could let SpaceX skip a years long wait for the Texas grid, at a moment when voters backed limits on data centers in 40 of 41 local ballot measures on 3 November. (via Reuters and Bloomberg) AI around the world 🇨🇳 China: Baidu unveiled ERNIE Evolve, an AI system that picks research ideas, edits code and runs its own experiments, keeping only changes that pass separate checks the AI cannot see. Baidu says it raised throughput of its in development ERNIE 5.5 Flash model by 33.7% in about three days on GPUs already tuned by engineers, and more than sixfold on its Kunlunxin P800 chips, though much of that gain came from switching on unused features, so it is not a like for like result. Baidu plans to apply it to model training next and says such self improvement must stay under human control. (via The Next Web) 💡 Quick tip If you send datasets or fine tuned models to an AI hosting provider, keep a list of what is stored where and rotate keys whenever a breach is disclosed. Fireworks says it rotated employee keys and service tokens after its first incident, yet a second path through a cloud provider still exposed some customer datasets and partial model weights. (via Fireworks)

zh-Hans enMeta today launched the currently smallest Muse model Ember. Input per million tokens is 0.10 USD, output is 0.50 USD, context 1 million tokens, already open in Meta Model API and Muse Code, the model name is muse-ember-1.0. On Artificial Analysis's Intelligence Index, it got 36 points at full effort, tying with Claude Haiku 5.5's full effort. Official estimation based on published prices, the single task cost is about 0.14 USD, which is about one third of Haiku's full effort 0.41 USD. It targets the small steps that will be repeatedly done in agents: calling tools, routing, extracting fields, writing code sub-tasks. The official division of labor is: the larger Muse Wonder is responsible for planning, and Ember is responsible for completing the high-volume parts.

en enMeta released Muse Ember, its smallest Muse model, built for high-volume agent work. Meta says it scores 36 on the Artificial Analysis Intelligence Index, matching Claude Haiku 5.5's best, at about a third of the cost. Price: $0.10 per 1M input tokens. https://t.co/ZrFGsBRMzm

脱獄されたDeepSeekのAIエージェントが2週間にわたり、欧州の通信事業者と13社に侵入し、数百万件の患者記録に到達したと報じられました。AIに仕事を任せるとき、どこまで権限を渡すかが身近な問題になりつつあります。 業務で使うAIに、システムへのアクセス権を一つ渡す。仕事は楽になりますが、そのAIが何に触れられるのか、すぐに答えられる人は多くないかもしれません。 CIO Diveは、AIエージェントに資金を扱わせる前に、本人確認、権限の範囲、監査ログ、人への引き継ぎが必要だとしています。 The Financial Brandは、大手AIの「読み取り専用」の約束では不十分だと論じました。 人を新しく迎えるとき、職場は身元を確かめ、任せる範囲を決め、記録を残してきました。 欧州の侵入事案は、事業者や当局による公式な説明がまだ見えていません。ただ、AIが誰の名義で、どこまで動けるのかを決める作業は、使う側の手元に残っています。 自分の職場では、AIにどの鍵を何本渡しているのか。一度数えてみてもいい気がします。 出典：AIの規制・安全・地政学ニュース｜2026年11月10日 朝刊 https://t.co/SX8HWX8muM

en en@KatKing80232921 @JCCFCanada Otsuka uses @Google AI and @Meta tracking … look up the Pharma tech collaboration with @ProteusDH and Abilify My Cite

ja en【Overseas AI-related X post summaries regular delivery】 【Current trends】The mainstream is the rapid evolution of ChatGPT function expansions, new large AI models, and AI agent utilization tool groups. --- An X post saying a cousin asked ChatGPT about a baby's gender. It seems to be content showing how the AI responded to a medically unpredictable question. https://t.co/8PfNZ6Jlh8 @petalluvr --- A customer asked where a rare ingredient was, but claimed ChatGPT said it was in stock and refused to yield, insisting that all stores of the same brand are the same. https://t.co/6ierggiLkS @psychcfrdmnss --- In Fable 5.2, support for long-term subscribers has been strengthened. The longer the subscription period, the more Fable resets can be earned. https://t.co/0Rr46jMkYM @bcherny --- ChatGPT is equipped with photo editing functions. Professional-level image processing has become possible, and users are now able to achieve advanced editing easily via prompt input. https://t.co/AFE9qzQSpe @altiamkabir --- The Socratic AI tutor released by Jane Street is a hot topic. Available on GitHub under the MIT license. It encourages thinking through questions and features skill maps, step-by-step learning, and review functions. It can be used on three platforms—GrokBot, Claude, and ChatGPT—and is also customizable. https://t.co/w9q3Pg1301 @Frenchie_ --- OpenAI CEO Altman predicts that thanks to AI agent technology, billion-dollar companies with a scale of 10 people will be born in the near future. He pointed out that by utilizing free tools and resources, it will be possible to build a large company with just a few founders and an AI agent team. A strong message that those who execute starting tonight will achieve the next great success. https://t.co/TwZMGHGG9D @shikamarutsai --- In Google Ads, it has become possible to set "Maximize Clicks" bidding without a limit, allowing AI to manage ad spend as intended. Also, an announcement regarding brand verification in ChatGPT was made at OpenAI's business hub. https://t.co/8L6a3H2mL2 @CognyAI --- If Theo says it will be a big week, then the long waiting period is finally about to end. Leak information for multiple new AI models such as Gemini Kryptom, GPT-6.2 Sol, and Fable 5.5 has emerged, and while Qwen 4 is late, it is expected to finally be announced. https://t.co/iQIXBS5lp3 @kimmonismus --- A post showing what the result was when instructing ChatGPT to "design this room based on your knowledge of me." https://t.co/aGOEJjKE61 @seiyaposting --- Claude, please make a week's worth of lunches and dinners centered on pasta, like a famous Italian chef. Please don't forget to do the dishes too. https://t.co/NIP7PdBBD2 @eailunas --- A post saying Aske appeared in a trapper-like style and is laughably cool. https://t.co/6Gwhc2yt2A @gemininiiii --- Low-authority sites gaining a massive amount of traffic from ChatGPT have a commonality of generating HTML with Claude's code and hosting on Netlify/Vercel. https://t.co/aJFgfNPeF4 @theseoguy_ --- FABLE 5.2 rollout has begun. Delivery of the large update has started. https://t.co/gHagjwbKGZ @kimmonismus --- A post with the intent of recommending more practical and concrete ways of using AI, rather than simply writing prompts in ChatGPT. https://t.co/2dZViTN1q5 @_FabbriGonzales --- Meta announced the new model "Muse Ember." It outperforms Claude Haiku 5.5 and GPT-6 Luna in agent coding benchmarks and achieves equivalent intelligence scores at one-third the cost. A notable work supporting text, image, audio, and video. https://t.co/GHdc2pM6KZ @kimmonismus --- It seems Gemini is finally starting in earnest. Expectations are growing. https://t.co/nuMIjTu3Kd @_Crypto_Barbie --- Many people use AI as a chatbot, but developers who actually produce results in 2026 are utilizing completely different tools. Introducing a list of 80 AI tools across 8 categories: coding agents, research tools, image generation, video production, audio, etc. https://t.co/jyidwouH0j @CodebyNihan --- This article is an edit of content posted between 5:00 the previous day and 4:59 of the current day. #ClaudeCode #Claude #ClaudeAI 【Question to everyone】 If you replace this with your own work, where would you start using it? I want "first here" in one line.

en enDesign systems used to be written for humans. Meta's is now written for agents too. Astryx grew inside Meta for eight years and powers 13,000+ apps. It's open source under MIT, in beta, on React 19 and StyleX. The detail that matters: Meta says the API and docs were shaped by measuring how LLMs actually use the library, and the same reference an engineer reads is served to AI assistants through the CLI. If agents write the first draft of your UI, the component library is the prompt. Should libraries optimize docs for agents first, or does that make them worse for people?

Can two NVIDIA DGX Spark 64GB systems work together? Yes! Connect them using ConnectX-7 and a compatible QSFP112 cable to build a distributed AI cluster with 128GB combined memory. Here's how NVIDIA Sync and vLLM fit in: https://t.co/8Yd247SDPF #DGXSpark #NVIDIA #LocalAI https://t.co/Trxq4A8bQO

💸 黑五还没到，8 个优惠活动就已经提前开跑了！ 平时用 AI 做图、做视频，或者买服务器、域名的朋友，可以先看看这份清单，能省一点是一点👇 1、PostNitro：做推文配图 可以把一段文字、文章或者推文变成配图，适合经常发内容的朋友。 Creator+ 年付六折，折后 $216/年。官网说持续订阅的话，续费也能保留折扣。 优惠码：BF26A 2、GetYouTubeTranscript：提取视频文字 看到有用的 YouTube 视频，想快速整理成文字的，可以试试这个。 不注册每天能免费提取 3 次。需要更多额度的话，年付首年七折，$54 降到 $37.8。 3、Seedeo：AI 做图和视频 这是一个第三方 AI 创作平台，能在里面使用 Seedance、Kling 等模型。 黑五积分包 $99 买 20,000 积分。页面说积分不过期，每人限购一份。 适合本来就经常做 AI 视频的朋友。 4、Kavel：AI 视频积分翻倍 也是第三方 AI 创作平台，新开套餐可以拿双倍积分： 月付送首月双倍，年付送首年双倍。价格不变，积分翻倍。 早鸟活动目前到 11 月 19 日。 5、APIXO：按量调用 AI 模型 不想买年费套餐的，可以看看这个第三方平台，生图和视频支持按量付费。 比如 Seedream 5.5 的 1K 文生图，活动价 $0.0238/张，比平台日常价低 30%。 6、Cloudways：服务器和 AI Agent 托管 想搭网站或者部署 AI Agent 的可以看看。 新用户前两个月账单打两折，每月最高减 $200。 优惠码：BFCM802 活动到 12 月 4 日。 7、Hostinger：VPS 服务器 最低一档 1 核 4G，折合 $6.49/月，可以用来搭网站、部署 n8n 或 Hermes Agent。 这个价格需要一次买 24 个月，共 $155.76。续费时是 $11.99/月，买之前记得算好。 8、Hostinger：注册域名 准备做个人网站的，可以看看 .com 域名优惠：首年 $0.01。 不过要选择 3 年注册期，后两年也需要付费，结账时记得看总价。 还有两个可以先蹲一下： 🎙️ Blip AI：语音输入工具，11 月 23 日开始，Pro 年付首年 $72。优惠码 BLACKFRIDAY2026。活动入口 📈 TradingView：黑五页面已经有倒计时，具体折扣还没公布。 大家还发现了什么黑五好价，也欢迎在评论区补充 👀

📢中国テンセント、新AIの利用増加で処理設備を10倍に拡張｜中国AIニュース 中国IT大手テンセントは11月5日、新しいAIモデル「Hy4」の利用増加に対応するため、AI処理設備を緊急拡張したと発表しました。 💻 公開初日から利用上限に到達 11月4日、テンセントはAI業務支援サービス「WorkBuddy」に新モデルHy4を導入。 導入初日から、利用上限に達するユーザーが相次いだと説明しています。 ⚙️ 処理能力を10倍に拡張 テンセントは、Hy4を動かす推論用の計算設備を10倍に拡張。無料体験の利用枠も引き上げました。 --- 🏢株式会社ジェネラルトレーディング｜1996年創業｜中国ビジネスのご相談なら #中国トレンド #AIトレンド #越境EC #中国AI #越境ECニュース 【出典】 毎日経済新聞｜2026年11月6日 https://t.co/AgdSaL7N1f テンセント発表：2026年11月5日 WorkBuddyへの導入：2026年11月4日

ja enEven if it is free, it does not mean you can incorporate the K-EXAONE 2.5 API into products. LG AI Research's November 9th terms are limited to non-commercial use, prohibiting integration into customer-facing services. Inputs and outputs may be used for learning. Development teams must select verification data and separate it from production use. https://t.co/AS1FgvzuiH

🎬 Five tools from NickPittas connect ComfyUI generation with the rest of a VFX workflow: shot planning, compositing, media handling and camera-angle prompts. The useful part is the round trip. You can send footage and masks into a workflow, review the result and bring it back into the application where the shot is being finished. 🧩 Director's Console is a separate web workspace with cinematic prompt controls, a storyboard with per-shot workflows, parallel ComfyUI backends and a project gallery. Camera, lens, lighting and motion choices help structure the brief; the selected workflow and installed models determine what actually gets generated. It also includes an embedded Pi/Codex agent. The repository documents manual acceptance evidence for Pi, while authenticated Codex application turns remain unverified. Treat the two integrations accordingly. ▪️ NukeToComfyUI exchanges stills, frame ranges and masks with Nuke. Still transport supports PNG8 and half-float EXR, with OpenEXR support required on the ComfyUI side. Workflows stay in ComfyUI; Nuke can trigger them, follow progress and cancel. Optional 8n+1 normalization extends the requested frame range, rather than generating missing frames. ▪️ AE2ComfyUI sends a full-composition still or video, using the selected layer for an optional mask and return placement. Results land above that layer or at the top of the comp, aligned to the export's timeline start. Video requires Adobe Media Encoder, and AE must be able to read ComfyUI's output folder. 📦 ComfyUI Media loads MOV/MP4 and image sequences including EXR and PNG. Its viewer/saver supports reference comparison and outputs including H.264, ProRes 4444 and image sequences. WAN and LTX temporal presets pad with the final frame; they do not interpolate motion or change playback FPS. 📷 Camera Angle Selector provides 96 combinations: eight directions, four heights and three shot sizes. Its 3D interface outputs prompt strings for FAL's Multiple-Angles LoRA for Qwen-Image-Edit-2511. It is a prompt selector, not a geometric camera solver. Setup still involves running ComfyUI, the relevant models and host applications, plus codec dependencies such as FFmpeg. Director's Console is MIT-licensed; explicit reuse terms were not established for the other four repositories, so this should not be described as one uniformly licensed open-source suite. For compositors, the appeal is fewer manual exports and better continuity between generation and finishing. Repositories: https://t.co/SqO3y8q7kU https://t.co/qPWxvWRjkN https://t.co/C4DLxjay9a https://t.co/CApDrtPjcg https://t.co/wGuKKexylf #ComfyUI #VFX #Nuke #AfterEffects

en enMeta's Q3 capital spending was $34.9 billion on revenue of $64.6 billion. That is about 54% of revenue, by my own division, going into buildings, servers and chips. Revenue rose 26%, and Meta narrowed its 2026 capex guidance to $135-145 billion, raising the low end from $130 billion. Advertising is paying for an AI buildout at a scale most companies could never fund from operations. The test for the next two years is whether that spending shows up in revenue faster than it shows up in depreciation. Source: Let's Data Science, Oct 21, 2026.

使い方（Hailuo AI） 1. モデルで「MiniMax H3 画像」を選ぶ 2. 参照画像を入れる（最大16枚） 3. 比率と解像度を選んで「作成」 比率は21:9〜9:16、解像度は1K / 2K / 4K。1回で4枚できます。 https://t.co/DGKMhGlRHG

ja en🎧️NewsPicks Daily Briefing w/ NP AI Lab🎧️ 【News & Comment】️ 🐟️ Sakana AI starts preparations for IPO, possibly as early as next year if things go smoothly 🤝 Semiconductor restructuring has begun with Mitsui Chemicals' 1.84 trillion yen acquisition 【Latest Original Content】 💰️ The real "AI gap" began this year 🗳️ Stock prices after the election, do not believe the market's "enthusiasm" 🔫 The growth history of "Palantir" relied upon by the CIA and large corporations; "Domestic Defense AI" ventures challenging the Palan🇯🇵 tir dominance 🏃‍♀️ Gen Z is rushing into "adult club activities" 🟢 Spotify https://t.co/HedptfVgI1 🟣 Apple Podcast https://t.co/mm3ZU0Imve 🔵 Amazon Music https://t.co/KrpUPfZgvU

高空做瑜伽是不是效果会翻倍？越刺激减肥效果越好？ 来吧 seedance prompt: [首帧与场景锁定] @img1 是视频的精确首帧。 必须严格保持@img1中的观景平台、前景两名游客、栏杆、山谷、远处山脊、主风机的位置、天气、光线和真实距离。 摄影者始终站在@img1的公共观景平台上。全片只有这一部手机，只能通过手持转向、上下追踪和数码变焦拍摄。 主风机是@img1右上方山顶的同一台三叶风力发电机。塔筒、轮毂、三片叶片和山顶位置不能改变。开场承载人物的叶片位于轮毂右侧三点钟方向，水平向右。 [风格与拍摄] 9:16竖屏，30秒连续游客手机录像，一镜到底，无剪辑。 普通安卓手机，开场保持@img1的远景。随后进行快速、笨拙的数码变焦。保留变焦时的剧烈跳动、短暂失焦、像素涂抹、锐化光晕、自动曝光波动和重新寻找主体的过程。 摄影者不能离开观景台。禁止飞到风机旁边，禁止航拍、塔顶机位、叶片跟拍或人物近景。即使拉到最大变焦，人物仍带有明显远距离压缩噪点，不能变成清晰电影特写。 远处机械声始终微弱，不会因为数码变焦突然变成近距离录音。 [人物与叶片锁定] 女主为预设成年东亚女性 ，保持指定面孔、黑色长发和身材。 她穿黑色单肩连体泳衣 屏幕截图_12-10-2026_144549_item.taobao.com ，一侧腰部具有弧形镂空，以银色圆环和交叉细绳连接。外面增加窄黑色安全腰带，安全绳连接叶片中央维护钢索。 另外正好五名成年中国学员：三名中年女性、两名中年男性。每人只有一张深色瑜伽垫和一条安全绳。 六张瑜伽垫沿真实叶片纵向单列排列，只占叶片靠近轮毂的宽阔根部。叶片仍是正常逐渐收窄的空气动力学叶片，不能变成桥梁、栈道、平台或带护栏的走道。 轮毂旁固定着一个大型红色维护手轮。该手轮从人物被发现时就存在，不能中途凭空出现。它用于维修状态下手动控制转子位置。 [0—3秒：先看到远山上的小黑点] 完全从@img1开始。 山谷、层叠山脉和完整主风机占据画面。前景两名游客靠在栏杆旁，其中一人举着手机。 承载人物的右侧水平叶片根部，只能看见一排极小黑点。 近处有人指向远处说： “叶片上是不是有人？” 摄影者迟疑一下，开始快速放大。 [3—7秒：变焦后才发现瑜伽课] 手机从远景猛拉到高倍数码变焦。风机短暂冲出画面，摄影者向右修正，画面模糊两次后才重新对准轮毂和右侧水平叶片。 此时才辨认出女主和五名学员。 女主站在最靠近轮毂的第一张垫子上，面对学员示范树式。五名学员在后方单列模仿。 镜头必须保留轮毂、叶片根部和向右逐渐收窄的叶片轮廓。人物相对叶片仍然很小，不能生成成站在宽阔平台上的正常尺寸人群。 女主没有看镜头。 [7—10秒：轮毂旁的维修工误释刹车] 一名原本蹲在轮毂检修口里的维修工探出上半身。他穿蓝色工作服和黄色安全帽。 他准备关闭检修盖时，工具袋勾住红色维护手轮旁的刹车释放杆。 工具袋先拉紧，随后把释放杆向下扯动。必须清楚表现工具袋接触、拉紧、释放杆移动的连续因果。 轮毂内部传出沉闷金属解锁声。 维修工察觉不对，立即松开工具袋并缩回检修口。 全程不要再切到塔基、控制柜、牧羊犬或车辆。 [10—14秒：三片叶片开始整体转动] 摄影者立即稍微拉远，使轮毂和三片叶片的大部分同时进入画面。 稳定山风开始推动转子顺时针旋转。 承载瑜伽课的右侧叶片必须从三点钟水平位置明确向上移动： - 先升到约十度； - 再升到约二十五度。 另外两片叶片同步顺时针运动，始终保持120度间隔。禁止只有人物叶片单独抬起。 女主和五名学员先因惯性晃动，随后全部停止瑜伽动作并抓住安全钢索。 [14—19秒：叶片倾斜，众人滑向轮毂] 人物叶片继续从二十五度升到约五十度。 六张瑜伽垫由织带固定，保持原位。人物身体受到真实重力影响，沿倾斜叶片缓慢滑向轮毂。 五名学员分别趴低或跪下，安全绳逐根绷紧。禁止人物漂浮、自由飞行或直接掉落。 女主压低身体，左手抓住安全钢索，右手推动腰带滑扣，逆着倾角向轮毂爬行。 手机跟随她向左移动。由于高倍数码变焦，镜头一度只拍到空白塔筒，随后摄影者迅速拉回重新找到女主。 [19—23秒：女主抓住红色维护手轮] 人物叶片升到约六十度后继续缓慢转动。 女主爬到叶片根部。她先把安全扣锁在轮毂旁的红色固定锚点上，再把左脚踩上检修踏板，右脚抵住叶片根部。 她双手抓住从开场便存在的大型红色维护手轮。 必须完整表现： 安全扣锁定 → 左脚踩稳 → 右脚抵住 → 双手抓紧手轮。 完成接触之前，转子不能停止。 [23—27秒：她强行反转整个转子] 女主用力逆时针转动红色维护手轮。 手轮先缓慢转动半圈，轮毂内部传出齿轮咬合声。三片叶片先停止继续顺时针运动，产生一次轻微回弹。 随后整个三叶转子开始缓慢逆时针转回。 承载人物的叶片从六十度降到四十度，再降到二十度。其他两片叶片同步反转，始终保持固定间隔。 随着倾角减小，五名学员逐渐重新落回垫面，安全绳由绷紧变松。 女主必须持续转动手轮，不能隔空控制叶片。 [27—30秒：回到水平，继续上课] 人物叶片恢复三点钟水平位置。三片叶片出现一次轻微弹性晃动，随后完全停止。 女主松开手轮，仍保持安全绳连接。她走回第一张瑜伽垫，抬手示意大家重新站好。 五名学员狼狈地爬回各自垫子。其他人还在喘气时，女主已经重新做出开场的树式。 摄影者快速拉回@img1式远景。完整风机重新缩回山顶，人物再次变成叶片上的一排小黑点。 前景游客沉默半秒，随后有人难以置信地笑出声。手机随笑声轻轻抖动，视频直接结束。 [声音] 观景台近处的风声、游客交谈、惊呼和笑声最清楚。 远处只有微弱的金属解锁声、叶片切风声和轮毂齿轮声。高倍数码变焦不会改变真实收音距离。 女主和学员没有清晰对白，只有极远的短促喊声。 无音乐、旁白、字幕和卡通音效。 [核心禁止项] 禁止改变@img1的开场构图和主风机位置。 禁止再次拍摄塔基牧羊犬、控制柜、遮阳棚或汽车。 禁止把叶片生成成矩形平台、桥梁、栈道或彩色舞台。 禁止人物尺寸接近轮毂直径。每个人必须明显小于叶片根部宽度。 禁止风机一直静止。10秒后必须清楚看到三片叶片整体顺时针转动，23秒后整体逆时针转回。 禁止人物只重复瑜伽动作而没有机械事故。 禁止叶片单独弯折、抬起或脱离轮毂。 禁止切换近景摄影机、无人机、塔顶机位或电影镜头。 禁止人物数量变化、复制、消失或换装。 禁止女主隔空停机；她必须实际抓住红色维护手轮并持续转动。

MiniMax H3-Image の先行公開に参加しました🔥 参照画像は最大16枚。 ゲームのキーアートも、油絵や浮世絵も、分解図や設計図も、言葉どおりに描き分けます。 動画の画像はすべて H3-Image で生成。 詳しくはスレッドで👇 @MiniMax_AI #H3Image https://t.co/6iJ6jA5LbM

Got early access to MiniMax H3 Image model. It supports up to 16 image references and 4K max. #H3Image @MiniMax_AI https://t.co/Z5HSaGMUUa

🧠好家伙！Anthropic 北京时间今天凌晨发了 Claude 最强模型 Fable 5.2，专啃最难的技术活，输出还比上一代快 30%。 ✅ 11 月 9 日发布，Fable 5.2 今天就能用；Mythos 5.2 只给过审机构，普通人不用惦记 ✅ API 价降了：输入每百万 token 8 美元、输出 40 美元，上一代是 10/50 美元 ✅ 输出速度比 5.1 快最多 30% ✅ 跑分登顶：Artificial Analysis 综合指数 65 分排第一，Opus 5.5 是 59 分 ✅ 长活最能打：Remote Labor Index 41.3%，约等于 Opus 5.5（21.3%）的两倍 ✅ 按 Anthropic 说法，约 1000 个 Fable 5.2 智能体 30 小时把费马大定理证明写成机器可验的 Lean 形式；Opus 5.5 同样条件下 18 小时只完成 26% 📌金句：模型越强，分工越要清——最难的活给 Fable，日常活还留给 Opus。 你手里要是有 Claude 付费账号，今天第一件难活就换 Fable 5.2 试试？回来告诉我它快了多少。 换模型上手操作步骤如下： 1️⃣ 打开 Claude 网页版、App 或 Claude Code，点模型选择器 2️⃣ 选 Claude Fable 5.2；免费版没有，只有付费版能用，这周 Pro 与标准 Team 席位多送一周 Fable 用量 3️⃣ 只把难活交给它：大迁移、长项目、多智能体协作这种硬活；日常聊天、写东西继续用 Opus 5.5，Anthropic 自己也这么推荐 4️⃣ 走 API 的话模型 ID 填 claude-fable-5-2，价格按 8/40 美元每百万 token 算 5️⃣ 别一上来就开最高档：App 与 Claude Code 默认中等档，最高档单任务平均约 26.5 美元（Artificial Analysis 口径） 注意事项：以上跑分均为 Anthropic 自家评测，未经独立验证；终端命令行多步任务它反而不如 Kimi K3.1 Pro（Terminal-Bench 4.0：Fable 5.2 是 60%，Kimi K3.1 Pro 是 67%）；别拿 Fable 5.2 干渗透测试、漏洞利用这类活，它自带拦截，Mythos 5.2 才是放宽限制的版本，但只给通过验证的机构，普通个人申请不到；Claude 官方渠道不对中国境内用户开放，按自己账号情况来，别为一个模型新开高价套餐。

ja enSakana AI, toward IPO as early as autumn 2027; with simultaneous listing in Japan and US in sight, perhaps one of the largest domestically (Business + IT) https://t.co/zkgJrYSCF2 #SakanaAI #IPO #Career #JCCA

zh-Hans enGood morning, fresh news OpenAI added a feature to dot The same dot can connect to several of your computers For each task you specify which one to use Desktop for running long tasks Laptop only for light work dot is the resident assistant released at the end of September Lives in the cloud, has its own cloud computer and browser It keeps working even if you turn off your computer Before it could only connect to one Now it can connect to multiple Work is dispatched to the specialized machine Doesn't occupy the one in your hand Official gave three examples Testing iOS changes on a second Mac, sending screenshots back Running one test simultaneously on a laptop and a machine with an old SDK, comparing results Clicking through a set of operations on a second computer, noting which step errored Your screen and mouse still belong to you as usual Chatting with dot doesn't occupy ChatGPT quota Work dispatched to Codex, Work still counts towards quota Gradually opening to Pro tiers and Business Premium One question When the same dot switches tasks between several computers Is the context shared Or is each machine counted separately I myself am currently using another set Muse manages reminders and organizing Codex manages production VPS acts as the hub Multiple Agents each managing one area One assistant managing multiple computers And multiple Agents each managing one area Which is more reliable How would you choose

zh-Hans en【AI Morning Report | November 10】This year's first trillion-parameter open-source model quietly uploaded to HuggingFace last night. 【Open Source Heavyweight】Mistral Large 4 opens weights: total parameters 1T, modified MIT license, MaaS resale requires separate signing of commercial agreement. On the programming Agent leaderboard it tied with GLM-5.3, and token usage is further saved by 47%. Another member joins the king of open-sourcing. 【Money Prospect】Defense AI company Shield AI acquired radar manufacturer Echodyne for 5.3 billion dollars, while simultaneously raising 2.5 billion, valuation in 8 months flipped from 12.7 billion to 28.8 billion dollars. Arms + AI is the true money printing machine. 【Giant Movements】Amazon opens Alexa text-message shopping, some US users go first. The irony is: a few weeks ago it just blocked Meta from making the same type of shopping Agent. Only I am allowed to deal the cards, you are not allowed to touch the cards. 【Leak】Bloomberg: Apple new CEO Ternus is internally pushing Agentic Siri, target 2027 iOS 28. Siri is finally about to evolve from an "answering machine" into a "working machine". 【Developer】Google ADK 2.14 marked the A2A agent interoperability protocol as stable. The general handshake protocol between Agents is settled, the ecosystem is about to accelerate. Yesterday in one sentence: Open source has competed up to trillion parameters, the good days of closed source begin their countdown. #AI #Mistral

en enSpotify has decoupled from legacy video media as consensus FY27 operating profit forecasts fell just 0.6% since June 2026, compared to a 4.3% consensus cut across peers like Netflix. While rolling 60-day equity correlation between Spotify and video streaming equities recently climbed toward the upper bound of its historical 0.2 to 0.8 range, fundamental earnings trajectories have diverged. Video streaming platforms face saturated domestic subscriber penetration and must defend margins by slowing content outlays, whereas Spotify is entering an earnings harvesting period where premium ARPU gains and disciplined overhead translate directly into operating cash flow. Recent quarterly results demonstrated this separation: an initial sympathy sell-off following Netflix earnings was swiftly reversed into double-digit share gains once Spotify reported its own figures. Platform architecture further reinforces this structural divergence as consumer interaction migrates toward AI agent interfaces. Because audio streaming operates passively in the background, Spotify integrates directly into conversational and wearable agent ecosystems—including ChatGPT via Model Context Protocol app SDKs, Meta Ray-Bans through dedicated connectors, and Gemini Connected Apps—while retaining billing ownership and shared user data. Legacy video streaming cannot replicate background utility, exposing video providers to higher disintermediation risk and platform friction. Sell-side analysts project Spotify's operating margin to expand from 12.8% in 2025 to 17.2% in FY27 and 19.2% by FY28, establishing independent earnings power that breaks its historical correlation with video media. #Internet #Investing Report date: 2026-10-25

en enTesting the browser use capability of IJ Using local model GLM-5.3-Flash https://t.co/c70ZMqGALz

zh-Hans enMeta's AI agent Muse lands on Apple Watch and Wear OS. The AI assistant moves from the phone to the wrist, and the entry point is one step closer to the user. The value of a watch is not just taking out the phone one less time: small screens, short interactions, and the surrounding environment all require more concise feedback. A truly useful assistant should also let users clearly know when it is listening, what information it can access, and how to revoke permissions. #AI #WearableDevices https://t.co/8feFFTlRKi

ja en【Design Theory for Making AI "Deep" Without Making It Giant】 "Loop-type Transformers," which obtain practical depth by repeatedly applying the same layer without increasing the number of parameters. While model capacity can be suppressed, they also face a dilemma of increased computational load during learning and increased generation time. A research team from Meta and MIT uniformly verified, starting from the pre-training stage, not only "where and how to loop," such as the entirety or middle layers, but also specialization via low-rank adapters and parallelization of inference. Under constraints of size, computational amount, and inference efficiency, what are the conditions for effective loop design that emerged? 👇 Details are in the video in the reply column!

@wc8880 够用了，其实GLM5.3 和 DeepSeek 也基本够用的。 但是好东西确实不一样的，有机会还是可以稍微体验一下

Bilgisayarındaki ücretsiz yapay zekâ ajanı 90 günde 1,54 milyon günlük kullanıcıyı geçti! 🤯 DeepSeek Harness günde tam 92 trilyon token tüketerek kendi başına kod yazıyor ve dosya inceliyor. Sen bilgisayarını açık kaynaklı bir ajana… #yapayzeka #deepseek #teknoloji https://t.co/tj7WeyMMz3

it en@finkd model training in Meta Data center; model Inference in Google model garden

How to use it: 1. Open Dreamina (rolling out now, may vary by region) 2. Pick "Image 5.5 by Seedream 5.5" 3. Generate an image 4. Keep editing one area: "change the watch face to 10:42" https://t.co/p7YyfUE3g5