AI 英文奏折

2026/01/08 | 共收录 30 篇原创长文

📊 情报汇总：本篇奏折已自动过滤转推、广告及雷同新闻。

👤 (9) X 上的 LudovicCreator：“🎨 KLING 2.6 🎨 Prompt : A battle-worn knight clad in dented silver armor and a tattered crimson cape rides a black warhorse through a burning medieval village; ashes drift in the air as flames consume wooden buildings around him, and villagers flee in panic; the camera https://t.co/Tjv0X8bTvY” 05:15

身披残甲的红披风骑士策马穿越燃烧的村庄，镜头低角度跟拍，火光与暮色交织，渲染出史诗般的战争氛围。

📊 ❤️10 | 📝 100字 | 🔗 深度阅读原文

👤 (9) X 上的 Ksenia_TuringPost @CES：“Must-read AI research of the week: ▪️ mHC: Manifold-Constrained Hyper-Connection ▪️ Nested Learning: The Illusion of Deep Learning Architectures ▪️ Fantastic Reasoning Behaviors and Where to Find Them: Unsupervised Discovery of the Reasoning Process ▪️ Shape of Thought: When https://t.co/2rMqHHaRvT” 08:33

本周AI研究摘要：多篇论文聚焦推理过程、记忆架构、异构加速器优化及无监督学习等前沿方向。

📊 ❤️264 | 📝 103字 | 🔗 深度阅读原文

👤 (9) X 上的 Rohan Paul：“Today’s edition of my newsletter just went out. 🔗 https://t.co/CzYv0DLKGh 🧑‍🔬 AI now predicts 130 diseases from 1 night of sleep 🛌. 🏆 Boston Dynamics unveiled a product version of its fully electric Atlas humanoid at CES 2026 and said manufacturing starts immediately. https://t.co/dBIUgULOMr” 05:32

AI预测130种睡眠疾病；波士顿动力发布电动Atlas；英伟达转向全栈战略；xAI融资200亿美元；Cursor优化编码上下文调用。

📊 ❤️4 | 📝 135字 | 🔗 深度阅读原文

👤 (9) X 上的 @levelsio：“✨ I used Claude Code to one-shot a 💁‍♀️ Karen Bot When we walk or drive around we come across a lot of things here that could be improved in terms of public infrastructure Roads near us have lots of potholes, there’s municipal waste bins here that someone put on fire 6 months https://t.co/9eLjb9pzTM” 05:16

利用Claude Code开发市政问题自动上报工具，通过AI生成正式报告并附坐标照片，提升反馈效率。

📊 ❤️271 | 📝 222字 | 🔗 深度阅读原文

👤 (9) X 上的 MayorkingAI：“1. Lamborghini Aventador { “video_type”: “cinematic supercar transformation”, “duration_seconds”: 8, “reference_image”: “uploaded_image”, “camera”: { “initial_view”: “POV first-person shot from an extended arm”, “movement”: [ “steady POV focus on logo”, https://t.co/lzMREEcdDH” 04:46

日落赛道，兰博基尼徽章经慢动作坠落后，通过精密机械变形为亮黄色Aventador跑车，采用电影级广角镜头与追焦运镜。

📊 ❤️4 | 📝 257字 | 🔗 深度阅读原文

👤 (9) X 上的 Rohan Paul：“This should be absolutely great. OpenAI is launching ChatGPT Health, a separate space in ChatGPT for health and wellness conversations. Users can link medical records and wellness apps so responses use their own context and timelines. Health chats are stored separately, with https://t.co/XJ2h6qBN7p” 05:06

OpenAI推出ChatGPT Health，支持医疗记录接入，提供加密健康咨询，不用于模型训练。

📊 ❤️6 | 📝 114字 | 🔗 深度阅读原文

👤 (9) X 上的 Rohan Paul：“AI clusters are getting so large that the hard part is often safe multi-tenant operation, not just raw GPU speed. New official Nvidia blog Vera Rubin NVL72 and BlueField Astra. NVIDIA Vera Rubin NVL72 is a compute tray design that pairs Rubin GPUs and a Vera CPU with a https://t.co/Pptss8E5Yw” 04:46

英伟达新推Vera Rubin NVL72与BlueField Astra架构，通过DPU独立管控南北向及GPU间东西向网络，提升多租户裸金属GPU集群的安全隔离能力。

📊 ❤️5 | 📝 393字 | 🔗 深度阅读原文

👤 (9) X 上的 Rohan Paul：“China has asked some tech firms to halt new orders for Nvidia’s H200 AI chips while Beijing decides whether, and under what conditions, they can be used. This is to slow stockpiling and steer buyers toward domestic AI chips. Nvidia CEO Jensen Huang said at CES 2026 that demand https://t.co/bVoYiYIW5h” 04:33

中国要求部分科技企业暂停订购英伟达H200芯片，以引导采购国产AI芯片。美方出口政策反复，影响企业长期规划。

📊 ❤️9 | 📝 119字 | 🔗 深度阅读原文

👤 (9) X 上的 Rohan Paul：“🚨 BREAKING: Anthropic is lining up another mega funding round of $10B at a $350B valuation, led by Singapore sovereign wealth fund GIC and Coatue. Anthropic raised $13B in September 2025 at a $183B post money valuation, so a $350B price tag is close to a 2x step up in about 4 https://t.co/v4jZl1sMyC” 04:16

Anthropic拟以3500亿美元估值再融资100亿美元，加速IPO进程。

📊 ❤️10 | 📝 131字 | 🔗 深度阅读原文

👤 (9) X 上的 Aakash Gupta：“The architects of attention capture seem to understand better than anyone what these tools cost.​​​​​​​​​​​​​​​​ I find the irony here fascinating. The guy who runs a platform that just hit 1 BILLION monthly active users is telling you to consume less information. The guy” 03:52

平台创始人呼吁用户减少使用，而其产品设计却旨在最大化用户沉迷，形成讽刺对比。

📊 ❤️3 | 📝 191字 | 🔗 深度阅读原文

👤 (9) X 上的 Rohan Paul：“K-EXAONE Technical Report A new 236B MoE model from LG AI Research is here, running 23B active parameters with a 256K context window and 6-language capability. It combines hybrid attention with Multi-Token Prediction for a 1.5x speed boost, performing on par with open-weight https://t.co/RH13nRKyCM” 13:45

LG AI发布2360亿参数MoE模型K-EXAONE，激活参数230亿，支持256K上下文与6种语言，性能媲美同规模开源模型。

📊 ❤️49 | 📝 187字 | 🔗 深度阅读原文

👤 (9) X 上的 Philipp Schmid：“Using skip dangerous or yolo mode is still too risky. If you have to maintain an allow or deny list, share it below. I’ll go first. Here is what I use in @antigravity. Note: Most agents support an allowlist for which commands to auto-execute and deny list, which follows the same https://t.co/nqTw0XxvNq” 02:53

使用允许/拒绝列表管理命令自动执行，拒绝列表优先。

📊 ❤️125 | 📝 123字 | 🔗 深度阅读原文

👤 (9) X 上的 @levelsio：“💬 Do any of you still remember RealPlayer™? Back in the 1990s, websites were just text, links and if you were very lucky some images, but definitely NOT video or audio Internet speeds were extremely slow (think 2.5kb/s, so about 10,000 times slower than an average connection https://t.co/xn4aIvcHge” 01:51

RealPlayer在90年代首创网页音视频播放，因微软捆绑Windows Media Player而衰落。

📊 ❤️202 | 📝 454字 | 🔗 深度阅读原文

👤 (9) X 上的 Sharyph：“Most newsletter creators track the wrong numbers. They celebrate open rates while their bank account stays empty. Here’s the truth: Subscriber count doesn’t pay your bills. Revenue-driving metrics do. I’ve seen creators with 50K subscribers making less than creators with 5K. https://t.co/XC0vVqQGAk” 18:15

简报创作者应关注营收指标，而非订阅量等虚荣数据。关键指标包括打开率、点击率、免费转付费率及付费用户留存率，这些直接决定盈利水平。

📊 ❤️8 | 📝 374字 | 🔗 深度阅读原文

👤 (9) X 上的 Pierrick Chevallier | IA：“🌀 No surprise here… time to animate. Let’s jump into Seedance 1.5 Pro and bring Santa Neo to life 🎄 12s video, native audio, multishots — the output is super clean. Here’s the prompt I used: A calm but visibly surprised Neo-like hacker figure (neutral posture, subtle tension https://t.co/D7BSaRaO2i” 02:14

使用Seedance 1.5 Pro生成12秒多镜头视频，以赛博朋克风格精准呈现黑客角色动画及原生音频。

📊 ❤️1 | 📝 186字 | 🔗 深度阅读原文

👤 (9) X 上的 Pierrick Chevallier | IA：“🎅 Next step: Neo goes Christmas mode. Still in Nano Banana, let’s generate a fun image of Neo… 👉 Wearing a Santa hat, with a playful festive vibe. Here’s the prompt I used: Create a cinematic portrait of Neo facing directly toward the viewer. Neo wears his iconic look: black https://t.co/qCQIjXkwNj” 02:14

为《黑客帝国》中的尼奥生成圣诞主题形象：身着标志性黑衣，佩戴红色圣诞帽，在矩阵风格的暗黑背景中呈现电影级肖像。

📊 ❤️1 | 📝 137字 | 🔗 深度阅读原文

👤 (9) X 上的 LudovicCreator：“Firefly is the go-to AI-powered creative space that lets you get top AI models, all in one place. Firefly puts your entire creative process into one powerful experience for ideating, generating, and creating images, videos, audio, and designs. Her we will focus on two features : https://t.co/3riGbtMVLi” 01:15

Firefly是集成多AI模型的创意平台，支持视频文本编辑与时间线剪辑，提供端到端创作流程。

📊 ❤️43 | 📝 217字 | 🔗 深度阅读原文

👤 (9) Rohan Paul on X: “This paper claims a training-free attention check can spot bad math proofs before a proof checker fails. An LLM’s attention patterns can reveal when its math reasoning is correct, with no training needed. Checking LLM proofs is hard because proof checkers can fail for boring https://t.co/CbCI3axuxl” 12:40

本文提出一种免训练的注意力检测方法，能在证明检查器失败前识别错误数学证明。通过分析LLM注意力模式的结构特征，无需训练即可有效判断推理有效性，在Lean数据集上准确率达85%-96%。

📊 ❤️87 | 📝 152字 | 🔗 深度阅读原文

👤 (9) X 上的 Andrew Ng：“If you’ve never written code before, this is for you. I’ve just launched a course that shows you, in less than 30 minutes, how to describe an idea for an app and build it with AI. In this course, you’ll build a working web application – a funny interactive birthday message https://t.co/P6L1A8QO2s” 01:39

零基础AI构建网页应用课程：30分钟学会描述创意并实现，无需编程经验，支持多AI工具。

📊 ❤️1915 | 📝 253字 | 🔗 深度阅读原文

👤 (9) X 上的 Jerry Liu：“I built a form-filling agent that anyone can use 💫 This is an extremely simple but useful (I hope) app. Upload a fillable form 📋, some context files, and chat with the agent to fill the form out automatically ✍️ 1️⃣ Yes it is a Claude Code SDK wrapper 2️⃣ It is better and faster https://t.co/SLR76xpk0X” 01:30

基于Claude的智能表单自动填充工具，支持多格式文件解析，开源可用。

📊 ❤️110 | 📝 131字 | 🔗 深度阅读原文

👤 (9) X 上的 Rohan Paul：“Significant research on AI helping cancer diagnosis. New AI tool from MIT and Microsoft designs tiny “bait” molecules so a urine test could spot cancer-linked enzymes early. This is a “pee-on-a-strip” idea for cancer, and AI is being used to design the sensors so they react to https://t.co/vbVxITjNMI” 01:22

MIT与微软开发AI工具，设计尿液检测分子传感器，精准识别癌症相关酶，提升早期诊断准确性。

📊 ❤️24 | 📝 299字 | 🔗 深度阅读原文

👤 (9) X 上的 Rohan Paul：“Solar Open, Upstage’s new 102B-param model just droppoed on Huggingface. Massive Training Scale: Pre-trained on 19.7 trillion tokens, ensuring broad knowledge coverage and robust reasoning capabilities across various domains. Built on a Mixture-of-Experts architecture with 102B https://t.co/nYITYBjb8Y” 01:13

Upstage发布102B参数Solar Open模型，采用专家混合架构，通过动态调度合成数据比例与严格过滤机制优化训练流程。

📊 ❤️9 | 📝 199字 | 🔗 深度阅读原文

👤 (9) X 上的 LlamaIndex 🦙：“Check out the form-filling agent that automates PDF forms using AI by @jerryjliu0 📄🤖 Use any fillable PDF with an agent that fills it out based on your prompts and context files. Our new experiment creates a multi-turn chat experience for form completion. 🔍 Upload fillable https://t.co/DnTuIgTcfm” 01:04

AI代理自动填写PDF表单，支持多轮对话修正，可上传自定义提示和上下文文件。

📊 ❤️37 | 📝 139字 | 🔗 深度阅读原文

👤 (9) X 上的 Rohan Paul：“Tom Stevens created Tombot after his mother was diagnosed with Alzheimer’s in 2011 and could no longer safely keep her dog. Trying and rejecting substitutes showed him a clear gap, He went on to a master’s at Stanford, to understand how to ease loneliness, anxiety, and” 00:42

汤姆·史蒂文斯为缓解阿尔茨海默症患者的孤独焦虑，基于对700多名老年人的研究，开发出具备逼真外观与交互功能的陪伴机器人Tombot。

📊 ❤️4 | 📝 100字 | 🔗 深度阅读原文

👤 (9) X 上的 Jürgen Schmidhuber：“3 of the 7 awardees (Drs. Bengio, LeCun, Hinton) [QE] repeatedly republished important AI techniques whose creators they failed to credit, not even in later surveys. They didn’t invent any of the foundational algorithms of modern AI. The evidence is found in well-known reports https://t.co/yk0DPci2kW” 00:00

三位获奖者多次重发重要AI技术而未引原创，未发明现代AI基础算法。多份专家评审报告证实，其未作辩驳。

📊 ❤️166 | 📝 410字 | 🔗 深度阅读原文

👤 (9) X 上的 Rohan Paul：“Bank of America research predicts annual data generation to hit 527 zettabytes by 2029. A zettabyte is roughly 1Bn terabytes. At that scale, “just keep it on fast disks” stops being a serious plan. You end up needing a layered setup: very fast storage close to the GPUs for https://t.co/85DT3vdmND” 12:15

美银预测2029年全球年数据量将达527泽字节。海量数据需分层存储体系，并构建高效数据调度系统。数据基础设施正成为AI发展的关键瓶颈。

📊 ❤️53 | 📝 180字 | 🔗 深度阅读原文

👤 (9) X 上的 Rohan Paul：“New Tsinghua Univ and other top Chinese University paper proposes CogFlow – It trains an LLM to solve diagram math by making it truly use what it sees. It also comes with a new MATHCOG dataset with 120K+ paired perception and reasoning annotations. Current multimodal LLMs can https://t.co/5g9ZvVUSM9” 10:12

清华大学等提出CogFlow模型，通过感知-事实转换-推理三阶段及强化学习奖励机制，提升图表数学题解题准确率，减少推理偏差。

📊 ❤️98 | 📝 183字 | 🔗 深度阅读原文

👤 (9) X 上的 Rohan Paul：“This paper shows why common LLM reasoning training collapses into 1 template, and how Distributional Creative Reasoning (DCR) prevents it. Today’s post training loops sample many reasoning traces, meaning different step by step attempts, reward the highest scoring ones, and https://t.co/WlzTZpyUrJ” 22:58

本文提出分布创造性推理（DCR）方法，通过保持推理路径多样性，避免传统训练导致的策略单一化，提升模型鲁棒性。

📊 ❤️43 | 📝 185字 | 🔗 深度阅读原文

👤 (9) X 上的 Rohan Paul：“TII launched Falcon H1R-7B, a reasoning-focused 7B model that’s already outperforming others in math and coding. Plus, it packs a 256K context window. Leveraging the recently introduced DeepConf approach, a test time method that lets an LLM think in many parallel ways, then https://t.co/t9bwm3qg7w” 10:39

阿联酋TII发布Falcon H1R-7B推理模型，采用混合架构与DeepConf并行推理技术，在数学编码任务中性能领先且生成文本量减少38%。

📊 ❤️27 | 📝 208字 | 🔗 深度阅读原文

👤 (9) Rohan Paul on X: “SWE-Lego shows supervised fine tuning alone can train an LLM, a text model for code, to fix bugs. Most coding agents, setups that let an LLM run commands and edit code, struggle because fixes take many steps, like finding files, changing code, and rerunning tests that check if https://t.co/lvOncgHjxe” 21:57

SWE-Lego通过监督微调训练大语言模型修复代码错误，在SWE-bench上达到58.8%的准确率，表明精心设计的数据与训练方法可有效提升修复性能。

📊 ❤️16 | 📝 209字 | 🔗 深度阅读原文