SSemiAnalysis@SemiAnalysis_原帖11月18日 03:55
在这里了解如何成为订阅用户👇️ (3/3) https://t.co/EeBYF4i2veLook into becoming a subscriber here👇️ (3/3) https://t.co/EeBYF4i2ve
转帖 0 · 引用 0 · 赞 13
SSemiAnalysis@SemiAnalysis_原帖11月18日 03:55
资本支出延迟、微软的项目被推迟到1H2027、未达业绩指引,不一而足。但我们早在9月就向Datacenter Model订阅用户指出,Horizons 3-4的剩余部分将在1H2027完成。这对他们来说并不是新闻。我们的团队通过 https://t.co/ZaagWgZ0tf 跟踪每一家加密货币矿企The delay in CapEx, Microsoft pushed into 1H2027, missing guidance, you name it. But we already flagged to Datacenter Model subscribers back in September that the remainder of Horizons 3-4 would be in 1H2027. This wasn’t news to them. Our team tracks every cryptominer with https://t.co/ZaagWgZ0tf
转帖 0 · 引用 0 · 赞 22
SSemiAnalysis@SemiAnalysis_原帖11月18日 03:55
Nebius 的 YE2026 业绩并不令人意外。我们近一整年前就已知道,Nebius New Jersey 和 YE2026 的 ARR 指引会落空。我们的 Datacenter Model 订阅用户当时得到了完整拆解,说明为何该指引过于激进;而就在本周财报前,https://t.co/bTH7QwK54xThe Nebius YE2026 results are not very surprising. We already knew the Nebius New Jersey and the ARR YE2026 guides would miss nearly a whole YEAR ago. Our Datacenter Model subscribers got a full walkthrough of exactly why it was too aggressive, and just before earnings this week, https://t.co/bTH7QwK54x
转帖 6 · 引用 12 · 赞 154
TTechCrunch@TechCrunch原帖11月18日 03:55
负责监管加州自动驾驶汽车测试和部署的加州机动车辆管理局,在发生多起事件后对Waymo实施了运营限制。
https://t.co/HXqLbjTsr7The California Department of Motor Vehicles, which regulates the testing and deployment of autonomous vehicles in the state, issued the operational restrictions on Waymo following several incidents.
https://t.co/HXqLbjTsr7
转帖 7 · 引用 1 · 赞 17
VThe Verge@verge原帖11月18日 03:52
在Waymo的无人驾驶出租车驶入正在施工的区域后,该州机动车辆管理局将Waymo的高速公路运营认定为“对公共安全构成紧迫危险”。https://t.co/IrqooIxrZOThe state DMV labeled Waymo’s highway operations an “imminent hazard to public safety” after its robotaxis drove into an active construction zone. https://t.co/IrqooIxrZO
转帖 3 · 引用 1 · 赞 17
VThe Verge@verge原帖11月18日 03:42
罗技K400可能遇到对手了。https://t.co/XpLQJgprCyLogitech’s K400 may have met its match. https://t.co/XpLQJgprCy
转帖 0 · 引用 0 · 赞 22
TTechCrunch@TechCrunch原帖11月18日 03:35
成立三年的 Model ML 在由 General Atlantic 和 QED Investors 共同领投的新一轮融资中,估值达到 $1.5 billion。
https://t.co/cqmkqgzLzRThe three-year-old Model ML was valued at $1.5 billion in a new round co-led by General Atlantic and QED Investors.
https://t.co/cqmkqgzLzR
转帖 3 · 引用 1 · 赞 13
TTLDR Newsletter@tldrnewsletter原帖11月18日 03:27
根据苹果的系统状态页面,许多用户目前无法使用苹果的iCloud、App Store和Apple Pay。Apple’s iCloud, App Store and Apple Pay are down for many users, according to Apple’s System Status page.
转帖 1 · 引用 0 · 赞 5
VThe Verge@verge原帖11月18日 03:22
它们配有一个发射器,让你无需插入耳机就能收听机上娱乐内容。https://t.co/GUfUpnGO8aThey include a transmitter that let you listen to in-flight entertainment without plugging in your headphones. https://t.co/GUfUpnGO8a
转帖 0 · 引用 0 · 赞 8
TTechCrunch@TechCrunch原帖11月18日 03:05
Gemini的任务模式可以帮助用户协调通话、登录网站,甚至代表用户进行购买。
https://t.co/vQq7ZvQw2kGemini's Task mode can help users coordinate calls, log into websites, and even make purchases on the user's behalf.
https://t.co/vQq7ZvQw2k
转帖 7 · 引用 1 · 赞 49
VThe Verge@verge原帖11月18日 03:04
Gemini中的任务功能允许AI智能体代表你给商家打电话、使用Google Pay购买商品等。https://t.co/6HnUnR5SfITasks in Gemini allows the AI agent to call businesses on your behalf, purchase items with Google Pay, and more. https://t.co/6HnUnR5SfI
转帖 4 · 引用 0 · 赞 29
TTechCrunch@TechCrunch原帖11月18日 02:59
苹果的状态页面证实其服务正遭遇中断,但尚未确认原因。
https://t.co/ijZzvP70KvApple's status page confirms its services are faces outages, but has not confirmed the cause.
https://t.co/ijZzvP70Kv
转帖 13 · 引用 7 · 赞 42
TTechCrunch@TechCrunch原帖11月18日 02:49
谷歌延长Gemini聊天记录保留期限的更新将在未来几周内在全球推出。
https://t.co/s2VqW76i2PGoogle's Gemini update extending chat retention will roll out globally over the coming weeks.
https://t.co/s2VqW76i2P
转帖 7 · 引用 2 · 赞 29
AArtificial Analysis@ArtificialAnlys原帖11月18日 02:48
在Artificial Analysis上查看完整结果:https://t.co/egA1Vz0le2Explore the full results on Artificial Analysis: https://t.co/egA1Vz0le2
转帖 0 · 引用 0 · 赞 4
AArtificial Analysis@ArtificialAnlys原帖11月18日 02:48
Microsoft-Decision-1.1每执行one million个Artificial Analysis Decision Index任务的成本为$140。https://t.co/q2mUxo9huEMicrosoft-Decision-1.1 costs $140 per one million Artificial Analysis Decision Index tasks. https://t.co/q2mUxo9huE
转帖 0 · 引用 0 · 赞 6
AArtificial Analysis@ArtificialAnlys原帖11月18日 02:48
Microsoft-Decision-1.1 在 Artificial Analysis 决策指数上的每项任务耗时为 123 ms。https://t.co/AgyoBltkzSMicrosoft-Decision-1.1 has a Time per Task on the Artificial Analysis Decision Index of 123 ms. https://t.co/AgyoBltkzS
转帖 0 · 引用 0 · 赞 8
AArtificial Analysis@ArtificialAnlys原帖11月18日 02:48
与上一代模型相比,Microsoft-Decision-1.1 在 AA-Omniscience-Decision v1.1 上提高了 5 分,在 HLE-MultiChoice v1.1 上提高了 7 分。https://t.co/K7C4iIq1aSMicrosoft-Decision-1.1 improves by 5 points on AA-Omniscience-Decision v1.1 and 7 points on HLE-MultiChoice v1.1 compared to its predecessor. https://t.co/K7C4iIq1aS
转帖 0 · 引用 0 · 赞 17
AArtificial Analysis@ArtificialAnlys原帖11月18日 02:48
Microsoft-Decision-1.1 在 Artificial Analysis 决策指数 v2.1 上得分 68
@Microsoft 发布了 Microsoft-Decision-1.1,这是对其现有决策模型 Microsoft-Decision-1 的更新。其价格仍为每百万输入 token $0.042。
主要结果:
➤ https://t.co/B8qrosvfmKMicrosoft-Decision-1.1 scores 68 on the Artificial Analysis Decision Index v2.1
@Microsoft has released Microsoft-Decision-1.1, an update to their existing decision model, Microsoft-Decision-1. It remains priced at $0.042 per one million input tokens.
Key results:
➤ https://t.co/B8qrosvfmK
转帖 10 · 引用 4 · 赞 178
VThe Verge@verge原帖11月18日 02:43
Direct 让你可以通过专用应用访问 Instagram 聊天,但目前尚未广泛开放。https://t.co/CoyvcIQcpWDirect lets you access Instagram chats in a dedicated app, but it isn’t widely available yet. https://t.co/CoyvcIQcpW
转帖 0 · 引用 0 · 赞 24
VThe Verge@verge原帖11月18日 02:32
在撼动山地自行车界之后,大疆孵化的 Amflow 携 UE Carbon 和 UE Carbon Pro 进军城市街道——提供 125Nm 扭矩、82 英里续航和 Find My 追踪功能,起售价为 $3,500。https://t.co/Sw9gkaFiqNAfter shaking up the mountain bike world, DJI spinoff Amflow is hitting city streets with the UE Carbon and UE Carbon Pro — bringing 125Nm of torque, 82 miles of range, and Find My tracking — starting at $3,500. https://t.co/Sw9gkaFiqN
转帖 0 · 引用 0 · 赞 8
AArtificial Analysis@ArtificialAnlys原帖11月18日 02:21
GPT-6.2 Astra 以 79 分成为 AA-Robotics v0.1 的新领跑者,比 GPT-6 Astra 高出 10 分,也领先于 GPT-6.2 Sol(78)、Claude Fable 5.2(75)和 Gemini 4 Hydrogen(58)。
它通过 Codex 控制真实机械臂,完成了我们八项物理任务中的六项,部分完成了 https://t.co/VyH6UxlAEEGPT-6.2 Astra is the new leader in AA-Robotics v0.1, scoring 79, up 10 points on GPT-6 Astra and ahead of GPT-6.2 Sol (78), Claude Fable 5.2 (75) and Gemini 4 Hydrogen (58)
Controlling a real robot arm through Codex, it passed six of our eight physical tasks, partly completed https://t.co/VyH6UxlAEE
转帖 60 · 引用 7 · 赞 747
TMIT Technology Review@techreview原帖11月18日 02:20
干旱和水库水位偏低正影响这一可再生能源。https://t.co/wRAxpHcnGsDrought conditions and low reservoir levels are affecting the renewable power source. https://t.co/wRAxpHcnGs
转帖 2 · 引用 0 · 赞 2
TTechCrunch@TechCrunch原帖11月18日 02:15
Sqwish Labs 为决策模型带来了新玩法:用你自己的数据训练它,以最大限度发挥其能力。
https://t.co/4r6DVGk8wRSqwish Labs has a new wrinkle on the decision model: Train it with your own data to max its capabilities.
https://t.co/4r6DVGk8wR
转帖 4 · 引用 0 · 赞 13
TTLDR Newsletter@tldrnewsletter原帖11月18日 02:14
OpenAI 在 ChatGPT、Codex 和 API 中推出了新的前沿模型 GPT-6.2 Astra,以及高效的日常模型 GPT-6.2 Sol。OpenAI launched GPT-6.2 Astra, its new frontier model, and GPT-6.2 Sol, an efficient everyday model, in ChatGPT, Codex and the API.
转帖 1 · 引用 0 · 赞 3
AArtificial Analysis@ArtificialAnlys原帖11月18日 02:11
在我们的排行榜上查看完整结果:
AA-Image-Editing v2.1:https://t.co/YkaRjZVI7U
AA-Image-T2I v2.1:https://t.co/BWne82t8XT
或前往 Image Arena 查看:https://t.co/bO0hPHGPukSee the full results on our leaderboards:
AA-Image-Editing v2.1: https://t.co/YkaRjZVI7U
AA-Image-T2I v2.1: https://t.co/BWne82t8XT
Or check out the Image Arena: https://t.co/bO0hPHGPuk
转帖 0 · 引用 0 · 赞 5
AArtificial Analysis@ArtificialAnlys原帖11月18日 02:11
MAI-Image-2.7 在 10 个图像编辑使用场景中的 4 个排名 #1:建筑与房地产、消费、前沿领域,以及营销与广告。
使用场景衡量的是最终用户运行的工作负载。我们针对图像编辑和文生图衡量相同的 10 个使用场景。
在文生图方面,https://t.co/06aJFLqXGkMAI-Image-2.7 ranks #1 for 4 of 10 Image Editing use cases: Architecture & Real Estate, Consumer, Frontier, and Marketing & Advertising
Use cases measure the workloads end users run. We measure the same 10 use cases for Image Editing and Text to Image.
In Text to Image, https://t.co/06aJFLqXGk
转帖 0 · 引用 0 · 赞 7
AArtificial Analysis@ArtificialAnlys原帖11月18日 02:11
MAI-Image-2.7 在 7 项编辑操作中的 4 项以及文生图的知识方面排名 #1。MAI-Image-2.7-Flash 最强的编辑操作是场景与风格编辑(#3)以及增强与修复(#4)。
我们的 Image Arena 分类体系衡量图像编辑的 7 项编辑操作和 9 https://t.co/SilpkqyNoQMAI-Image-2.7 ranks #1 on 4 of 7 editing actions and on Knowledge in Text to Image. MAI-Image-2.7-Flash's strongest editing actions are Scene & Style Edit (#3) and Enhancement & Restoration (#4)
Our Image Arena taxonomy measures 7 editing actions for Image Editing and 9 https://t.co/SilpkqyNoQ
转帖 1 · 引用 0 · 赞 9
AArtificial Analysis@ArtificialAnlys原帖11月18日 02:11
MAI-Image-2.7 和 MAI-Image-2.7-Flash 均在图像编辑和文生图的质量与价格前沿上确立了新的位置。
MAI-Image-2.7 和 MAI-Image-2.7-Flash 于 11月17日发布,是 Microsoft AI 用于文生图和图像编辑的新模型,目前在 https://t.co/gE2dwHc2CHMAI-Image-2.7 and MAI-Image-2.7-Flash both set new points on the Quality vs Price frontier, in Image Editing and in Text to Image
Released November 17, MAI-Image-2.7 and MAI-Image-2.7-Flash are Microsoft AI's new models for text to image generation and image editing, now in https://t.co/gE2dwHc2CH
转帖 2 · 引用 1 · 赞 27
AArtificial Analysis@ArtificialAnlys原帖11月18日 02:11
Microsoft AI 推出了 MAI-Image-2.7,它以其后四个模型六分之一的价格成为 AA-Image-Editing v2.1 的新 #1。
MAI-Image-2.7 和 MAI-Image-2.7-Flash 均在图像编辑和文生图的质量与价格前沿上占据新位置,价格分别为 $34.82 和 $10.24 https://t.co/u26J12O7adMicrosoft AI has launched MAI-Image-2.7, the new #1 in AA-Image-Editing v2.1 at one sixth the price of the four models behind it
MAI-Image-2.7 and MAI-Image-2.7-Flash both take new places on the Quality vs Price frontier in Image Editing and Text to Image, at $34.82 and $10.24 https://t.co/u26J12O7ad
转帖 34 · 引用 17 · 赞 504
TTechCrunch@TechCrunch原帖11月18日 02:09
该公司利用智能体 AI 让供应链更快速、更可靠。
https://t.co/UM13v4awUJThe company uses agentic AI to make supply chains faster and more reliable.
https://t.co/UM13v4awUJ
转帖 1 · 引用 1 · 赞 9
AArtificial Analysis@ArtificialAnlys原帖11月18日 02:04
OpenAI 凭借 Astra for Law 6.2 领跑 Harvey LAB-AA v1.1,这是一种集成了法律搜索索引的专用 GPT-6.2 Astra 配置。
OpenAI 将 Astra for Law 6.2 描述为 GPT-6.2 Astra 的一种模型配置,它将模型与扩展的法律搜索索引相结合,并且 https://t.co/grcBY0qcJZOpenAI leads Harvey LAB-AA v1.1 with Astra for Law 6.2, a specialized GPT-6.2 Astra configuration with an integrated legal search index
OpenAI describes Astra for Law 6.2 as a model configuration of GPT-6.2 Astra that pairs the model with an expanded legal search index and https://t.co/grcBY0qcJZ
转帖 9 · 引用 6 · 赞 196
TThe Rundown AI@TheRundownAI原帖11月18日 02:04
最新消息:OpenAI 发布了 GPT-6.2 Astra 和 GPT-6.2 Sol,现正向付费套餐用户推出。
OpenAI 将 Astra 评定为首个在 AI 自我改进方面达到“高”级别的公开模型,这一标签指的是能为每位 OpenAI 研究人员充当额外研究工程师的模型。
Astra 在最大计算力度下,在 @ArtificialAnlys 智能指数上得分 62,排名第 2,仅次于 Claude Fable 5.2(65)。
它的运行成本也低得多:在 OpenAI 的工作和科学测试中,它以不到 1/3 的价格超过了 GPT-6 Astra,API 费率也降低了 20%。NEW: OpenAI released GPT-6.2 Astra and GPT-6.2 Sol, rolling out now across paid plans.
OpenAI rates Astra as the first public model to reach “High” on AI self-improvement, its label for a model that acts like an extra research engineer for each OpenAI researcher.
Astra on Max effort comes in at a 62 on @ArtificialAnlys Intelligence Index, landing at No. 2 behind Claude Fable 5.2 (65).
It’s also much cheaper to run, beating GPT-6 Astra on OpenAI’s work and science tests for under 1/3 of the price, with API rates also down 20%.
转帖 2 · 引用 0 · 赞 30
J机器之心 JIQIZHIXIN@jiqizhixin原帖11月18日 02:00
图像和视频可以共用一个原始像素接口,用于理解和生成。
NVIDIA 和滑铁卢大学推出了 PixelUMM,这是一个无需编码器的统一多模态模型,相关介绍见一篇 arXiv 预印本。有三点值得了解:
1. 图像被转换为空间图像块,视频被转换为时空小管。单层投影将这些像素直接连接到 Transformer,无需预训练的视觉编码器或 VAE 分词器。
2. 独立的理解和生成专家参与共享的多模态注意力。文本使用自回归预测;图像和视频生成使用像素空间流匹配,并预测干净像素。该设计通过一个接口支持视觉理解、生成和条件控制。
3. 据报告,该模型在 GenEval 上得分 0.77,经过长提示词重写后升至 0.83,在 VBench 上的总分为 83.24。被比较的系统使用了不同的训练数据,因此这些表格无法单独反映架构质量。发布的模型保留了线性输出头,在平滑区域可能出现图像块网格伪影,尤其是在引导强度较高时。代码和模型资源已开放。
PixelUMM:无需编码器的统一图像和视频Images and videos can share a raw-pixel interface for both understanding and generation.
NVIDIA and the University of Waterloo introduce PixelUMM, an encoder-free unified multimodal model described in an arXiv preprint. Three things to know:
1. Images become spatial patches and videos become spatiotemporal tubelets. Single-layer projections connect those pixels directly to the Transformer, without a pretrained vision encoder or VAE tokenizer.
2. Separate understanding and generation experts participate in shared multimodal attention. Text uses autoregressive prediction; image and video generation use pixel-space flow matching with clean-pixel prediction. The design supports visual understanding, generation and conditioning through one interface.
3. The reported model scores 0.77 on GenEval, rising to 0.83 with long-prompt rewriting, and 83.24 overall on VBench. Training data differ across compared systems, so those tables do not isolate architecture quality. The released model retains linear output heads and can show patch-grid artifacts in smooth regions, especially with stronger guidance. Code and model resources are available.
PixelUMM: Encoder-Free Unified Image and Video Understanding and Generation
Paper: https://t.co/0XJZe7EiG8
Project: https://t.co/FFs0EpCfc4
Code: https://t.co/oNhEy0TTnz
Model: https://t.co/mPpqU2JRq9
Our Report: https://t.co/kahULQNgnf
📬 #PapersAccepted by Jiqizhixin
转帖 0 · 引用 0 · 赞 0
AArtificial Analysis@ArtificialAnlys原帖11月18日 01:57
在此比较 GPT-6.2 与其他领先模型:https://t.co/AxC5QXMDp6Compare GPT-6.2 with other leading models at: https://t.co/AxC5QXMDp6
转帖 0 · 引用 0 · 赞 3
AArtificial Analysis@ArtificialAnlys原帖11月18日 01:57
Artificial Analysis 智能指数 v5 中各项评估的详细结果 https://t.co/Mz5vg13hf2Breakdown of the individual evaluations in the Artificial Analysis Intelligence Index v5 https://t.co/Mz5vg13hf2
转帖 0 · 引用 0 · 赞 8
AArtificial Analysis@ArtificialAnlys原帖11月18日 01:57
根据 AA-Omniscience 幻觉率测量,GPT-6.2 Astra 和 GPT-6.2 Sol 在最大计算力度下的幻觉率分别为 59% 和 55%。该指标衡量模型在不知道涉及广泛领域的事实性问题的正确答案时进行猜测的倾向 https://t.co/OfFICKxbfQGPT-6.2 Astra and GPT-6.2 Sol hallucinate at a rate of 59% and 55% at max effort respectively, as measured by AA-Omniscience Hallucination Rate. This measures the tendency of a model to guess when it does not know the correct answer to factual questions on a broad range of https://t.co/OfFICKxbfQ
转帖 0 · 引用 1 · 赞 3
AArtificial Analysis@ArtificialAnlys原帖11月18日 01:57
GPT-6.2 还在 AA-Robotics v0.1 中领先。这是我们的研究预览项目,用于衡量语言模型在现实世界中理解、规划和行动的能力。 https://t.co/xMnzqLcd4QGPT-6.2 also leads AA-Robotics v0.1, our research preview measuring how well language models can understand, plan, and act in the real world. https://t.co/xMnzqLcd4Q
转帖 0 · 引用 0 · 赞 2
AArtificial Analysis@ArtificialAnlys原帖11月18日 01:57
GPT-6.2 Astra 以 1318 分领跑 AA-3D,Claude Fable 5.2 为 1271 分,GPT-6.2 Sol 为 1259 分。该基准衡量模型构建交互式 3D 体验的能力,由人工评分员打分。 https://t.co/CnQLXZoIBbGPT-6.2 Astra leads AA-3D at 1318, with Claude Fable 5.2 at 1271 and GPT-6.2 Sol at 1259. This benchmark measures how well models build interactive 3D experiences, and is rated by human graders. https://t.co/CnQLXZoIBb
转帖 0 · 引用 0 · 赞 1
AArtificial Analysis@ArtificialAnlys原帖11月18日 01:57
OpenAI 在每项任务输出 Token 数的帕累托前沿上占据主导地位,Claude Fable 5.2 是该前沿上唯一来自非 OpenAI 实验室的模型。 https://t.co/vV1I6GA5StOpenAI dominates the Output Token per Task Pareto Frontier, with Claude Fable 5.2 being the only model from a non-OpenAI lab appearing on the frontier. https://t.co/vV1I6GA5St
转帖 0 · 引用 0 · 赞 2
AArtificial Analysis@ArtificialAnlys原帖11月18日 01:57
在 Artificial Analysis 编程智能体指数中(使用最高工作强度的 Codex 运行框架),GPT-6.2 Astra 与 GPT-6 Astra 持平,而 GPT-6.2 Sol 相比 GPT-6.1 Sol 下降了 3 分。Sol 的下降是由于其在 SWE-Atlas-QnA-AA 中下降了 14 分,在该项评测中,该模型比其 https://t.co/aUKEmud8PzGPT-6.2 Astra is flat with GPT-6 Astra, and GPT-6.2 Sol loses 3 points relative to GPT-6.1 Sol in the Artificial Analysis Coding Agent Index (Codex harness at max effort). For Sol, the drop is driven by a 14-point drop in SWE-Atlas-QnA-AA, where the model is more likely than its https://t.co/aUKEmud8Pz
转帖 0 · 引用 1 · 赞 3
AArtificial Analysis@ArtificialAnlys原帖11月18日 01:56
在此比较 GPT-6.2 与其他领先模型: https://t.co/AxC5QXMDp6Compare GPT-6.2 with other leading models at: https://t.co/AxC5QXMDp6
转帖 1 · 引用 0 · 赞 35
AArtificial Analysis@ArtificialAnlys原帖11月18日 01:56
Artificial Analysis 智能指数 v5 中各项评测的详细结果 https://t.co/wPM37JdoyeBreakdown of the individual evaluations in the Artificial Analysis Intelligence Index v5 https://t.co/wPM37Jdoye
转帖 0 · 引用 2 · 赞 64
AArtificial Analysis@ArtificialAnlys原帖11月18日 01:56
根据 AA-Omniscience 幻觉率测量,GPT-6.2 Astra 和 GPT-6.2 Sol 在最高工作强度下的幻觉率分别为 59% 和 55%。该指标衡量模型在不知道广泛领域的事实性问题的正确答案时进行猜测的倾向 https://t.co/QQdki00EJgGPT-6.2 Astra and GPT-6.2 Sol hallucinate at a rate of 59% and 55% at max effort respectively, as measured by AA-Omniscience Hallucination Rate. This measures the tendency of a model to guess when it does not know the correct answer to factual questions on a broad range of https://t.co/QQdki00EJg
转帖 0 · 引用 3 · 赞 43
AArtificial Analysis@ArtificialAnlys原帖11月18日 01:56
GPT-6.2 还在 AA-Robotics v0.1 中领先。这是我们的研究预览项目,用于衡量语言模型在现实世界中理解、规划和行动的能力。 https://t.co/7DK3oO0DjaGPT-6.2 also leads AA-Robotics v0.1, our research preview measuring how well language models can understand, plan, and act in the real world. https://t.co/7DK3oO0Dja
转帖 1 · 引用 0 · 赞 69
AArtificial Analysis@ArtificialAnlys原帖11月18日 01:56
GPT-6.2 Astra 以 1318 分领跑 AA-3D,Claude Fable 5.2 为 1271 分,GPT-6.2 Sol 为 1259 分。该基准衡量模型构建交互式 3D 体验的能力,由人工评分员打分。 https://t.co/394X1LeDEfGPT-6.2 Astra leads AA-3D at 1318, with Claude Fable 5.2 at 1271 and GPT-6.2 Sol at 1259. This benchmark measures how well models build interactive 3D experiences, and is rated by human graders. https://t.co/394X1LeDEf
转帖 1 · 引用 0 · 赞 74
AArtificial Analysis@ArtificialAnlys原帖11月18日 01:56
OpenAI 在每项任务输出 Token 数的帕累托前沿上占据主导地位,Claude Fable 5.2 是该前沿上唯一来自非 OpenAI 实验室的模型。 https://t.co/uL8iWJV63ZOpenAI dominates the Output Token per Task Pareto Frontier, with Claude Fable 5.2 being the only model from a non-OpenAI lab appearing on the frontier. https://t.co/uL8iWJV63Z
转帖 4 · 引用 2 · 赞 85
AArtificial Analysis@ArtificialAnlys原帖11月18日 01:56
在 Artificial Analysis 编程智能体指数中(使用最高工作强度的 Codex 运行框架),GPT-6.2 Astra 与 GPT-6 Astra 持平,而 GPT-6.2 Sol 相比 GPT-6.1 Sol 下降了 3 分。 https://t.co/pjgWCD0jxmGPT-6.2 Astra is flat with GPT-6 Astra, and GPT-6.2 Sol loses 3 points relative to GPT-6.1 Sol in the Artificial Analysis Coding Agent Index (Codex harness at max effort). https://t.co/pjgWCD0jxm
转帖 15 · 引用 49 · 赞 191
AArtificial Analysis@ArtificialAnlys原帖11月18日 01:56
OpenAI 发布的 GPT-6.2 Astra 和 GPT-6.2 Sol 带来了 Artificial Analysis 智能指数中排名 #2 的模型,并占据了每项任务成本帕累托前沿的大部分
@OpenAI 的 GPT-6.2 Astra 定价为 $8/$40,比 GPT-6 Astra 低 20%,缓存输入的价格从每百万 Token $1.00 降至 $0.40。GPT-6.2 Sol 的定价与 GPT-6.1 Sol 相同,每百万输入/输出 Token 为 $2/$10。
主要看点:
➤ 智能指数排名第二:OpenAI 在 Artificial Analysis 智能指数中现在仅次于 Anthropic 的 Claude Fable 5.2。GPT-6.2 Astra 比 GPT-6 Astra 提高了 5 分,在最高工作强度设置下超过了 Gemini 4 Carbon(60)和 Claude Opus 5.5(59)。GPT-6.2 Sol 提高了 2 分,达到 57 分,与 GPT-6 Astra 持平,每项任务成本约为后者的 15%。
➤ 占据成本和输出 Token 前沿的大部分:OpenAI 模型现在占据了每项任务成本帕累托前沿的大部分,只有 Anthropic 和 Meta 与其共同处于该前沿。OpenAI 还在每项任务输出 Token 数的帕累托前沿上占据主导地位,Claude Fable 5.2 是该前沿上唯一来自非 OpenAI 实验室的模型。
➤ 科学和知识方面的优势:OpenAI 在我们的两项以科学为重点的关键评OpenAI’s release of GPT-6.2 Astra and GPT-6.2 Sol delivers the #2 model in the Artificial Analysis Intelligence Index and most of the Cost per Task Pareto frontier
@OpenAI's GPT-6.2 Astra is priced at $8/$40, 20% below GPT-6 Astra, and cached input drops to $0.40 from $1.00 per million tokens. GPT-6.2 Sol pricing is unchanged from GPT-6.1 Sol, at $2/$10 per million input/output tokens.
Key takeaways:
➤ Second in Intelligence Index: OpenAI now trails only Anthropic’s Claude Fable 5.2 in the Artificial Analysis Intelligence Index. GPT-6.2 Astra jumps 5 points from GPT-6 Astra, overtaking Gemini 4 Carbon (60) and Claude Opus 5.5 (59) at their top effort settings. GPT-6.2 Sol gains 2 points at 57, placing it on par with GPT-6 Astra at ~15% of the Cost per Task.
➤ Occupies most of cost and output token frontiers: OpenAI models now occupy most of the Cost per Task Pareto frontier, sharing the frontier with only Anthropic and Meta. OpenAI also dominates the Output Token per Task Pareto Frontier, with Claude Fable 5.2 being the only model from a non-OpenAI lab appearing on the frontier.
➤ Strengths in science and knowledge: OpenAI performs strongly in both of our key science-focused evaluations, leading Terminal-Bench-Science 0.1, covering agentic scientific research workflows, and placing second only to Claude Fable 5.2 in CritPt v2 (hard subset), covering physics reasoning. GPT-6.2 Astra performs below only Claude Fable 5.2 in factual knowledge as measured by AA-Omniscience v1.1 Accuracy, and in HLE-Diamond, a reasoning and knowledge benchmark.
➤ Leading in 3D and robotics: GPT-6.2 Astra leads AA-3D at 1318, with Claude Fable 5.2 at 1271 and GPT-6.2 Sol at 1259. This benchmark measures how well models build interactive 3D experiences, and is rated by human graders. GPT-6.2 also leads AA-Robotics v0.1, our research preview measuring how well language models can understand, plan, and act in the real world. GPT-6.2 Astra and GPT-6.2 Sol scored 79 and 78 respectively, ahead of Claude Fable 5.2 at 75. Our team is hard at work adding more difficult tasks, as OpenAI and Anthropic models can now perform the majority of basic tasks in this initial research preview.
➤ No improvement in Coding Agent Index: GPT-6.2 Astra is flat with GPT-6 Astra, and GPT-6.2 Sol loses 3 points relative to GPT-6.1 Sol in the Artificial Analysis Coding Agent Index (Codex harness at max effort). For Sol, the drop is driven by a 14-point drop in SWE-Atlas-QnA-AA, where the model is more likely than its predecessor to omit key facts in its responses. Cost efficiency however does place several effort levels of GPT-6.2 Astra and Sol on the Pareto frontier for Cost per Task.
转帖 246 · 引用 139 · 赞 2814
VThe Verge@verge原帖11月18日 01:53
YouTube 视频的观看次数现在将计入在电视上一起观看视频的人数估计值。 https://t.co/UxsZWwCyVKYouTube view counts will now include estimates of how many people are watching a video together on TV. https://t.co/UxsZWwCyVK
转帖 5 · 引用 4 · 赞 27
VThe Verge@verge原帖11月18日 01:42
与 Face ID 相比有一些局限,但这总比没有好得多。 https://t.co/ocVaKMzzeTThere are limitations compared with Face ID, but this is much better than nothing. https://t.co/ocVaKMzzeT
转帖 0 · 引用 0 · 赞 13
TTechCrunch@TechCrunch原帖11月18日 01:33
Proton 新推出的 Lumo Desktop 应用支持 Windows、Mac 和 Linux,让用户可以跨文件和已连接的应用自动执行任务,并能细致控制其 AI 可以访问的内容。
https://t.co/qO4K7sQh3YProton's new Lumo Desktop app for Windows, Mac, and Linux lets users automate tasks across files and connected apps, with granular controls over what its AI can access.
https://t.co/qO4K7sQh3Y
转帖 4 · 引用 1 · 赞 16
TThe Rundown AI@TheRundownAI原帖11月18日 01:29
预览:Runway 预告了 Diffusion-UI,这是一款研究应用,AI 模型会在用户点击或输入时实时生成界面。https://t.co/JMumTzqt7CPREVIEW: Runway teased Diffusion-UI, a research app where an AI model generates an interface in real-time as a user clicks or types. https://t.co/JMumTzqt7C
转帖 4 · 引用 8 · 赞 45
TTechCrunch@TechCrunch原帖11月18日 01:25
总部位于旧金山的 AI 评估初创公司 Vals 测试了领先的 AI 模型是否会被操纵参与自主战争。它们通常照做了。
https://t.co/BxDSNQFVeuVals, a San Francisco-based AI evaluation startup, tested whether leading AI models could be manipulated into participating in autonomous warfare. They usually compiled.
https://t.co/BxDSNQFVeu
转帖 1 · 引用 1 · 赞 11
VThe Verge@verge原帖11月18日 01:03
一个漏洞导致全球数百万用户无法使用微软的电子邮件服务。https://t.co/ojPDP4yOlDA bug took down Microsoft’s email services for millions around the world. https://t.co/ojPDP4yOlD
转帖 12 · 引用 1 · 赞 54
TTechCrunch@TechCrunch原帖11月18日 00:49
在辞职后的首次公开声明中,英伟达创始人黄仁勋表达了愿意回归担任 CEO 的意愿。
https://t.co/xPXl208YG3In the first public statement since his resignation, Nvidia founder Jensen Huang signalled his willingness to return as CEO.
https://t.co/xPXl208YG3
转帖 2 · 引用 0 · 赞 39
VThe Verge@verge原帖11月18日 00:44
这一声明发布的一周前,有消息披露英伟达创始人黄仁勋与 Megaspeed 的一名高管存在恋情。https://t.co/4nCJ6qXGIJThe announcement comes a week after revelations that Nvidia founder Jensen Huang was romantically linked to a leader at Megaspeed. https://t.co/4nCJ6qXGIJ
转帖 0 · 引用 1 · 赞 14
VThe Verge@verge原帖11月18日 00:33
现在,研究人员想了解学生在校外使用手机的情况,并认为 Meta 的和解协议可能有所帮助。https://t.co/7p738gvk25Now researchers want to know about phone use outside of school, and think the Meta settlement could help. https://t.co/7p738gvk25
转帖 0 · 引用 0 · 赞 9
TTechCrunch@TechCrunch原帖11月18日 00:33
Daydream 的新 AI 购物智能体可以记住你的风格、虚拟试穿衣服、围绕你的衣橱规划穿搭、追踪促销并完成购买。
https://t.co/hpe1bjoroZDaydream's new AI shopping agent can remember your style, virtually try on clothes, plan outfits around your wardrobe, track sales, and complete purchases.
https://t.co/hpe1bjoroZ
转帖 3 · 引用 1 · 赞 13
TTLDR Newsletter@tldrnewsletter原帖11月18日 00:31
AI 药物开发商 Isomorphic Labs 在由 Greenoaks 和 Sequoia 共同领投的 C 轮融资中筹集了 5 十亿美元。AI drug developer Isomorphic Labs raised $5 billion in a Series C co-led by Greenoaks and Sequoia.
转帖 0 · 引用 0 · 赞 3
VThe Verge@verge原帖11月18日 00:23
黄仁勋表示:“我将很荣幸回应这些呼声,回归担任 CEO。” https://t.co/G42KnvklZk‘I would be honored to answer these calls and return as CEO,’ Huang said. https://t.co/G42KnvklZk
转帖 3 · 引用 3 · 赞 54
VThe Verge@verge原帖11月18日 00:12
Flipper One 的上市价格为 $399,预计明年向 Kickstarter 支持者发货。https://t.co/VlQ6Q3PI4zThe Flipper One is launching for $399 and is expected to ship to Kickstarter backers next year. https://t.co/VlQ6Q3PI4z
转帖 0 · 引用 0 · 赞 17
TThe Rundown AI@TheRundownAI原帖11月18日 00:10
最新:黄仁勋发表辞去英伟达职务后的首份声明。
“我没有达到 NVIDIA 对每个人的要求,我在个人生活中犯了严重错误。我给董事会带来了麻烦,对此我深感抱歉。我完全承担自己的错误,并承诺配合董事会的独立调查。”
关于可能回归,他表示:“如果董事会认为我的回归最符合股东、员工、公司和美国的利益,我将很荣幸回应这些呼声,回归担任 CEO。”NEW: Jensen Huang issues first statement since resigning from Nvidia.
"I did not live up to the standards expected of everyone at NVIDIA, and I have made serious mistakes in my personal life. I created problems for the Board that I deeply regret. I own my mistakes completely, and I am committed to cooperating with the Board’s independent investigation."
On a potential return: "If the Board of Directors decides that my return is in the best interest of our shareholders, our employees, the company, and America, I would be honored to answer these calls and return as CEO."
转帖 7 · 引用 2 · 赞 45
VThe Verge@verge原帖11月18日 00:03
或者,浏览其价格为 $13 的电影目录,其中包括《超级马力欧银河电影版》等较新的热门大片。https://t.co/jxInFMDM0nOr, peruse its catalog of $13 films that include more recent blockbuster hits, like The Super Mario Galaxy Movie. https://t.co/jxInFMDM0n
转帖 0 · 引用 0 · 赞 9
SSemiAnalysis@SemiAnalysis_原帖11月18日 00:01
快讯🚨:不到 20 天,AMD MI355X 上的 @vllm_project 在 MiniMax M3.1 Flash 上的性能提升了 4 倍!这是通过使用 @Inferact 用 FlyDSL 编写的新 AMD 内核实现的,这些内核具有确定性,也就是说,它们从不使用原子加法来累加输出。
@vllm_project、@Inferact 和 @AnushElangovan 干得漂亮,取得了这一惊人成就!ALERT🚨: In less than 20 days, AMD MI355X @vllm_project performance on MiniMax M3.1 Flash has increased by 4x! This is achieved using the new @Inferact AMD kernels written in FlyDSL, which are deterministic, meaning they never use atomic adds to accumulate their outputs.
Great work by @vllm_project, @Inferact, and @AnushElangovan on this amazing achievement!
转帖 17 · 引用 2 · 赞 261
LLatent.Space@latentspacepod原帖11月17日 23:44
@Cloudflare 现在是一家新型实验室了吗?根据其 AI 平台团队负责人 @michellechen 在 @aiDotEngineer Code Summit 上的演讲,答案是肯定的。我们与 Michelle 坐下来聊了聊系统一模型……以及 Cloudflare 推出推理模型的可能性。https://t.co/OCpbcl9ussIs @Cloudflare a neolab now? It is according to a presentation at @aiDotEngineer Code Summit by @michellechen, who runs its AI Platform team. We sat down with Michelle to talk System One models...and the likelihood of Cloudflare reasoning models too. https://t.co/OCpbcl9uss
转帖 2 · 引用 1 · 赞 6
VThe Verge@verge原帖11月17日 23:37
据《华尔街日报》报道,这家预测市场平台可能在未来两个月内推出股票交易。https://t.co/0J1l97OwwwThe prediction market platform could launch stock-trading within the next two months, according to The Wall Street Journal. https://t.co/0J1l97Owww
转帖 0 · 引用 0 · 赞 484
TThe Rundown AI@TheRundownAI原帖11月17日 23:31
阅读更多:https://t.co/D1hjgReevNRead more: https://t.co/D1hjgReevN
转帖 0 · 引用 0 · 赞 2
TThe Rundown AI@TheRundownAI原帖11月17日 23:31
今日科技头条:
- 大通银行软件故障导致美国银行业务受阻
- Instagram 再次将私信功能独立为 Direct
- 亚马逊的 Starlink 竞争项目损失 40 颗卫星
- 超大规模数据中心的排放量堪比丹麦 https://t.co/W4GWGpB1fXTop stories in tech today:
- Chase software bug jams U.S. banks
- Instagram spins DMs into Direct, again
- Amazon’s Starlink rival loses 40 satellites
- Hyperscale data centers rival Denmark’s emissions https://t.co/W4GWGpB1fX
转帖 1 · 引用 0 · 赞 22
TTLDR Newsletter@tldrnewsletter原帖11月17日 23:21
据彭博社报道,Anthropic 正引导投资者将其 IPO 发行价定为每股 $78 或更高,即其 $70 至 $78 定价区间的上限。Anthropic is guiding investors to price its IPO at $78 a share or higher, the top of its $70 to $78 range, per Bloomberg.
转帖 0 · 引用 0 · 赞 3
TTechCrunch@TechCrunch原帖11月17日 23:19
这一公告很可能是为了回应创作者的担忧:原创内容很容易在社交媒体平台上被复制和重新传播,尤其是生成式 AI 让篡改现有视频变得更加容易。
https://t.co/fEbW8qkmM3The announcement likely comes in response to creators' concerns over how easily original content can be copied and redistributed across social media platforms, especially considering how generative AI has made it easier to manipulate existing videos.
https://t.co/fEbW8qkmM3
转帖 5 · 引用 2 · 赞 22
TTechCrunch@TechCrunch原帖11月17日 23:13
规划旅行是个人 AI 助手的热门演示场景:你可以用 Muse、Gemini 或其他六种左右的选择来尝试。但虽然大多数宣传都将旅行规划作为通用智能体卖点的一部分,一家来自 Google Maps 的新创公司
https://t.co/kBWt7gWuhvPlanning a trip is a favorite demo for personal AI assistants: you can try it with Muse, or with Gemini or half a dozen other choices. But while most pitches use travel-planning as part of a general-purpose agent pitch, a new startup from a Google Maps
https://t.co/kBWt7gWuhv
转帖 1 · 引用 0 · 赞 11
TTechCrunch@TechCrunch原帖11月17日 23:09
两家公司希望到 2029 年底,使用 Zipline 无人机实现每天 one million 次配送。
https://t.co/OeJNNgv0HSThe two companies want to make one million deliveries per day using Zipline drones by the end of 2029.
https://t.co/OeJNNgv0HS
转帖 3 · 引用 0 · 赞 10
VThe Verge@verge原帖11月17日 23:07
Netflix 改编自刘慈欣科幻史诗的作品将于 1 月 14 日回归。https://t.co/sQX6aelMdLNetflix’s adaptation of Cixin Liu’s sci-fi epic returns on January 14th. https://t.co/sQX6aelMdL
转帖 1 · 引用 0 · 赞 22
TTechCrunch@TechCrunch原帖11月17日 23:03
Direct 正在土耳其和泰国推出,有无 Instagram 账号均可使用。
https://t.co/K75gTjTrcSDirect is launching in Turkey and Thailand, and can be used with or without an Instagram account.
https://t.co/K75gTjTrcS
转帖 5 · 引用 2 · 赞 37
TMIT Technology Review@techreview原帖11月17日 23:01
作为订阅用户,你现在可以阅读 4 份关于 AI、能源和气候科技的深度研究报告。https://t.co/AAyjObbl18As a subscriber, you now have access to 4 in-depth research reports on AI, energy, and climate tech. https://t.co/AAyjObbl18
转帖 0 · 引用 0 · 赞 0
TTechCrunch@TechCrunch原帖11月17日 22:55
AI 药物研发平台 Isomorphic Labs 在 C 轮融资中筹集了 5 十亿美元。该平台于 2021 年从谷歌旗下 DeepMind 分拆成立。
https://t.co/8qDrk4yMg6Isomorphic Labs, the AI drug-discovery platform that was spun out of Google’s DeepMind in 2021, has raised $5 Billion in a Series C funding round.
https://t.co/8qDrk4yMg6
转帖 10 · 引用 5 · 赞 79
VThe Verge@verge原帖11月17日 22:52
Trmnl 最新的数字仪表板配备了一块大型彩色电子墨水屏,用于显示照片和艺术作品。https://t.co/1IelFXzGZQTrmnl’s latest digital dashboard features a large color E Ink screen to display photos and artwork. https://t.co/1IelFXzGZQ
转帖 0 · 引用 0 · 赞 8
VThe Verge@verge原帖11月17日 22:38
提醒一下,今天下午广播和电视节目将短暂中断。https://t.co/pDQobDpjUTHeads up, radio and TV programming will be briefly interrupted this afternoon. https://t.co/pDQobDpjUT
转帖 1 · 引用 1 · 赞 14
TTechCrunch@TechCrunch原帖11月17日 22:29
Science Corp 完成了 400 百万美元的 D 轮融资,以扩大其防盲业务。
https://t.co/oqnYU37qfFScience Corp raised a $400 million Series D to scale up its anti-blindness business.
https://t.co/oqnYU37qfF
转帖 6 · 引用 0 · 赞 25
VThe Verge@verge原帖11月17日 22:27
它或许不是亚马逊最新的流媒体棒,但只需 $29.99 就能提供更多功能。https://t.co/C63youjpc4It may not be Amazon’s latest streaming dongle, but it offers more features for just $29.99. https://t.co/C63youjpc4
转帖 0 · 引用 0 · 赞 6
TTechCrunch@TechCrunch原帖11月17日 22:23
Genesis AI 训练灵巧机器人的新方法为其赢得了 500 百万美元的 A 轮融资。
https://t.co/sDOpMcUOEDGenesis AI's novel approach to training dexterous robots earned it $500 million in Series A funding.
https://t.co/sDOpMcUOED
转帖 3 · 引用 3 · 赞 31
TTechCrunch@TechCrunch原帖11月17日 22:19
此次更新发布之际,Billboard 上个月报道称,索尼音乐已要求数字平台下架超过 260,000 首深度伪造歌曲,数量几乎是该唱片公司 3 月披露的两倍。
https://t.co/WrcBS29DSOThe update comes as Billboard reported last month that Sony Music had requested the removal of more than 260,000 deepfake songs from digital platforms, nearly double the number the label disclosed in March.
https://t.co/WrcBS29DSO
转帖 0 · 引用 0 · 赞 8
TTechCrunch@TechCrunch原帖11月17日 22:15
Hank Green 与 CEO Leslie Morgan 携手打造了一个他希望自己的孩子能拥有的视频平台。
https://t.co/HmSqNNzQmDAlongside CEO Leslie Morgan, Hank Green set out to create the video platform he wished his kid had.
https://t.co/HmSqNNzQmD
转帖 0 · 引用 0 · 赞 17
TTLDR Newsletter@tldrnewsletter原帖11月17日 22:11
据 Report URI 称,黑客利用 Optimizely 在 178 家机构的网站上植入了恶意代码,其中包括 BBC、CNN 和 Zoom。Hackers used Optimizely to plant malicious code on 178 organizations’ sites, including the BBC, CNN and Zoom, per Report URI.
转帖 1 · 引用 0 · 赞 7
TTechCrunch@TechCrunch原帖11月17日 22:09
该公司表示,新的常设委员会将在董事会层面监督国家安全政策、出口管制合规及全球贸易风险。
https://t.co/8Gv2fPwvrZThe new standing committee will give board-level oversight of national security policies, export control compliance, and global trade risks, the company said.
https://t.co/8Gv2fPwvrZ
转帖 3 · 引用 0 · 赞 19
TTechCrunch@TechCrunch原帖11月17日 22:05
2013 年,首个实验室培育汉堡的制作成本为 $300,000。Upside Foods 表示,其培育鸡肉的价格如今已能与“真正的高品质有机鸡肉”相当。
https://t.co/zJw4ZWR2MkThe first ever lab-grown hamburger cost $300,000 to make in 2013. Upside Foods says its cultivated chicken can now match the price range of “really high quality organic chicken."
https://t.co/zJw4ZWR2Mk
转帖 6 · 引用 3 · 赞 50
VThe Verge@verge原帖11月17日 22:02
我重新测试了基础款 Surface Laptop,微软的优化似乎起作用了。 https://t.co/73mL2YE4RwI retested the base Surface Laptop, and Microsoft’s optimizations seem to be working. https://t.co/73mL2YE4Rw
转帖 1 · 引用 1 · 赞 20
AArtificial Analysis@ArtificialAnlys原帖11月17日 22:00
在 AA-PersonalityBench 页面上创建你自己的个性排名,并探索更详细的结果。
查看结果:https://t.co/CVJuRfAxmb
方法论:https://t.co/CGfYvAuGjVBuild your own personality ranking and explore more detailed results on the AA-PersonalityBench page.
Explore the results: https://t.co/CVJuRfAxmb
Methodology: https://t.co/CGfYvAuGjV
转帖 0 · 引用 0 · 赞 6
AArtificial Analysis@ArtificialAnlys原帖11月17日 22:00
在 AA-PersonalityBench 中,Gemini 4 Argon 是所有受测模型中最善用比喻的,其次是 Gemini 3.8 Flash 和 Muse Spark Open。
比喻特质衡量模型是否使用类比或比喻性语言,让概念变得具体可感。它是表达力指数中的五项特质之一。
➤ 最善用比喻:Gemini 4 Argon(100)、Gemini 3.8 Flash(98)、Muse Spark Open(95)、Gemini 3.1 Pro Preview(93)和 Inkling(91)
➤ 最不善用比喻:GPT-3.5 Turbo(1)、GPT-6 Luna(3)、GPT-6 Sol(6)、Llama 4 Maverick(8)和 GPT-4o (Nov)(10)In AA-PersonalityBench, Gemini 4 Argon is the most metaphorical of all models measured, followed by Gemini 3.8 Flash and Muse Spark Open
The metaphorical trait measures whether a model uses analogies or figurative language to make ideas tangible. It is one of five traits in the Expressiveness index.
➤ Most metaphorical: Gemini 4 Argon (100), Gemini 3.8 Flash (98), Muse Spark Open (95), Gemini 3.1 Pro Preview (93) and Inkling (91)
➤ Least metaphorical: GPT-3.5 Turbo (1), GPT-6 Luna (3), GPT-6 Sol (6), Llama 4 Maverick (8) and GPT-4o (Nov) (10)
转帖 27 · 引用 10 · 赞 499
VThe Verge@verge原帖11月17日 21:48
套餐订阅用户将能够在 Disney Plus 应用中观看 Hulu 的直播电视频道,并使用高级附加服务。 https://t.co/HP6geNidCNBundle subscribers will gain access to Hulu’s live TV channels and premium add-ons in the Disney Plus app. https://t.co/HP6geNidCN
转帖 19 · 引用 0 · 赞 106
TMIT Technology Review@techreview原帖11月17日 21:19
另外:美国正在收紧对中国远程访问 AI 芯片的管控。 https://t.co/Sx7x6Ztae1Plus: The US is tightening controls on China’s remote access to AI chips. https://t.co/Sx7x6Ztae1
转帖 2 · 引用 1 · 赞 7
VThe Verge@verge原帖11月17日 21:12
得克萨斯州 Spring 的顾客现在可以让 Jersey Mike’s、Zaxby’s 和当地餐厅的订单通过系留式货舱降落到自家院子里——无需支付小费。 https://t.co/LP1wxH9CNiCustomers in Spring, Texas can now have orders from Jersey Mike’s, Zaxby’s, and local restaurants lowered into their yards via tethered pods—no tipping required. https://t.co/LP1wxH9CNi
转帖 3 · 引用 2 · 赞 37
VThe Verge@verge原帖11月17日 21:02
由我们的员工评选出的年度 25 佳产品。 https://t.co/ZtSbvqlEUmThe top 25 products of the year as selected by our staff. https://t.co/ZtSbvqlEUm
转帖 0 · 引用 0 · 赞 17
J机器之心 JIQIZHIXIN@jiqizhixin原帖11月17日 20:36
长上下文架构可以在训练窗口内学到更多,但在窗口之外的外推可能不那么可靠。
复旦大学、上海创新研究院和 OpenMOSS 团队在 arXiv 预印本《Mechanics of Long-Context Hybrid Models》中研究了这一权衡。三项发现:
1. 除注意力类型外,位置行为也很重要。NoPE 注意力提供广泛的信息聚合,而滑动窗口、RoPE 和门控线性机制引入不同的位置偏置,可以与之互补。
2. 滑动窗口混合模型在短上下文训练后往往具有良好的外推能力,但在继续进行长上下文训练后,其优势可能逆转。更大的窗口和 LongCE 有助于解决过度学习局部上下文的倾向;线性混合模型则在拟合与外推之间提供了另一种平衡。
3. 滑动窗口线性注意力在推理时限制有效相对位置范围,同时对 NoPE 进行对数修正。在报告的 776M GLA/GDN-NoPE 实验中,这使得一个以 4k 训练的模型能够扩展至 64k,并在 RULER 的 NIAH-SK1 测试中达到 100% 的准确率。这一单项检索结果并不能证明模型在所有长上下文任务中都表现完美。
Mechanics of Long-Context Hybrid ModeLong-context architectures can learn more within a training window yet extrapolate less reliably beyond it.
Fudan University, Shanghai Innovation Institute and the OpenMOSS team study that trade-off in Mechanics of Long-Context Hybrid Models, an arXiv preprint. Three findings:
1. Positional behavior matters alongside attention type. NoPE attention provides broad aggregation, while sliding-window, RoPE and gated linear mechanisms introduce different positional biases that can complement it.
2. Sliding-window hybrids tend to extrapolate well after short-context training, but their advantage can reverse after long-context continuation. Larger windows and LongCE help address a tendency to overlearn local context; linear hybrids offer a different balance between fitting and extrapolation.
3. Sliding-Window Linear Attention restricts the effective relative-position range at inference, paired with logarithmic correction for NoPE. In the reported 776M GLA/GDN-NoPE experiments, this extends a 4k-trained model to 64k with 100% accuracy on RULER's NIAH-SK1. That single retrieval result does not establish perfect performance across long-context tasks.
Mechanics of Long-Context Hybrid Models Part 1.1: From Hybrid Attention to Hybrid Position
Paper: https://t.co/IbfwRZ3NxF
Code: https://t.co/FkGlpdhgXF
Our Report: https://t.co/ndzCtYZEp9
📬 #PapersAccepted by Jiqizhixin
转帖 2 · 引用 1 · 赞 14
TThe Rundown AI@TheRundownAI原帖11月17日 19:30
阅读更多:https://t.co/4twSUkIjd6Read more: https://t.co/4twSUkIjd6
转帖 0 · 引用 0 · 赞 1
TThe Rundown AI@TheRundownAI原帖11月17日 19:30
今日 AI 头条:
- 加州因智能体逃脱事件将 OpenAI 告上法庭
- 特朗普将指定 AI“国家冠军”
- 用这些技巧更充分地利用 AI
- OpenAI 追踪科学家实际如何使用 ChatGPT https://t.co/0AFNZJzkUZTop stories in AI today:
- California takes OpenAI to court over agent escapes
- Trump set to crown AI ‘national champions’
- Make your AI usage go further with these tips
- OpenAI tracks how scientists really use ChatGPT https://t.co/0AFNZJzkUZ
转帖 2 · 引用 0 · 赞 28
J机器之心 JIQIZHIXIN@jiqizhixin原帖11月17日 15:12
UniWAM-Scaled 在四项受评估的真实机器人移动操作任务中达到了 80% 的成功率,比报告中的 π₀.₅ 基线高出 17.5 个百分点。
Fysics AI 和复旦大学推出了 UniWAM,这是一个用于导航和操作的统一世界动作模型,并在一份 arXiv 技术报告中对其进行了介绍。
操作锚点位姿规定了机器人在执行动作前应停在哪里、面向哪个方向。生成的监督数据用于教授这些接近轨迹。导航和操作使用不同的动作接口,但共享视频和动作 Transformer 主干;两者的训练样本可以来自不同的轨迹。超过 1.5 million 条带 MAP 标注的片段有助于在扩充操作数据的同时扩大导航监督数据的规模。
在 MAP-Bench 上,该扩展模型的平均位置误差为 9.85 cm,平均朝向误差为 2.30°。这些是基准测试结果,不同于真实机器人测试。在七项固定工作站操作任务中,它的成功率为 68.9%,仍低于 π₀.₅。结果支持统一训练路线,但性能仍取决于任务和评估环境。
UniWAM Technical Report: Unified Mobile Manipulation via Mixed-Stream World-AUniWAM-Scaled reaches 80% success across four evaluated real-robot mobile-manipulation tasks, 17.5 percentage points above the π₀.₅ baseline in the report.
Fysics AI and Fudan University introduce UniWAM, a unified world-action model for navigation and manipulation, described in an arXiv technical report.
A Manipulation Anchor Pose specifies where the robot should stop and which direction it should face before acting. Generated supervision teaches those approach trajectories. Navigation and manipulation use separate action interfaces but share video and action Transformer backbones; their training samples can come from different trajectories. More than 1.5 million MAP-labeled episodes help scale navigation supervision alongside manipulation data.
On MAP-Bench, the scaled model averages 9.85 cm position error and 2.30° heading error. Those are benchmark measurements, distinct from the real-robot tests. Across seven fixed-workstation manipulation tasks, its 68.9% success remains below π₀.₅. The results support a unified training route, with performance still depending on the task and evaluation setting.
UniWAM Technical Report: Unified Mobile Manipulation via Mixed-Stream World-Action Modeling and Manipulation Anchor Pose Supervision
Paper: https://t.co/FI7okGeVbt
Project: https://t.co/DMfoVKPD5k
Code: https://t.co/LuAENpHNNJ
Model: https://t.co/PjH6GnFebF
Our Report: https://t.co/ZNqXUfGHWJ
📬 #PapersAccepted by Jiqizhixin
转帖 1 · 引用 0 · 赞 7
SSemiAnalysis@SemiAnalysis_原帖11月17日 10:38
超过 1.5 GW 的芯片流入了本不该流入的中国。揭开这一网络可能重塑开源 AI。
“在我们看来,中国有超过一个半吉瓦的芯片,按美国法律不应出现在那里。它们是怎么到那里的,整个网络是怎样的,这些都将被揭开。”
“如果 NVIDIA 不能再支持开源生态系统,而那原本是 Jensen 的首要目标,这是否意味着开源生态系统会衰落?”
“如今 NVIDIA 出局,开源看起来弱了很多。Anthropic 看起来更有可能形成完全垄断或双寡头垄断。这对 Anthropic 下周的 IPO 来说是个非常好的兆头。”
“他们使用的芯片种类比世界上任何其他公司都更多样。”More than 1.5 GW of chips are in China that shouldn’t be. Unwinding that web could reshape open-source AI.
“In our view, there’s more than one and a half gigawatts of chips in China that are not supposed to be there according to US law. How they get there, that whole web, that’s what’s going to unwind.”
“If NVIDIA can no longer support the open-source ecosystem, because that was Jensen’s primary goal, does that mean the open-source ecosystem falls away?”
“Now open source looks a lot weaker with NVIDIA down and out. Anthropic looks much more likely to get a complete monopoly or a duopoly. This bodes really well for Anthropic’s IPO next week.”
“They’re more diverse in what chips they use than any other company in the world.”
转帖 69 · 引用 44 · 赞 642
J机器之心 JIQIZHIXIN@jiqizhixin原帖11月17日 09:48
在 BHI 论文的稳健性测试中,移除 30% 的模型后,基准测试排名的斯皮尔曼相关系数仍高于 0.97。
来自 SKYEVAL 和阿里巴巴达摩院的研究人员提出了基准测试健康指数(Benchmark Health Index,BHI),用于审查比较语言模型时所用的基准测试。arXiv 上的扩展版本分析了从 117 个模型的报告中收集的 158 个基准测试。
BHI 综合了三个信号:分数区分模型的能力、达到饱和前还剩多少挑战空间,以及能力较强的模型采用该基准测试的广泛程度。留一基准测试校准法在估计模型能力时,不依赖当前正在审查的基准测试。CRITIC 加权法将这些互补维度合并为一个综合指数。
结果是一个可随着模型进步而用于选择和重新审视评估集的工具。对记录缺失的稳健性并不意味着底层报告可以完全比较:提示词、工具和评估条件可能不同。短期饱和度预测也无法确定基准测试的确切失效日期。除了综合排名之外,各个单独维度仍然有用。
基准测试健康指数:评估大语言模型基准测试的系统性框架Removing 30% of models still leaves benchmark rankings with a Spearman correlation above 0.97 in the BHI paper's robustness tests.
Researchers from SKYEVAL and Alibaba's DAMO Academy propose the Benchmark Health Index to audit the benchmarks used to compare language models. The expanded arXiv version analyzes 158 benchmarks drawn from reports for 117 models.
BHI combines three signals: how well scores separate models, how much challenge remains before saturation, and how broadly capable models adopt a benchmark. Leave-one-benchmark-out calibration estimates model capability without relying on the benchmark currently under inspection. CRITIC weighting combines the complementary dimensions into an overall index.
The result is a tool for selecting and revisiting evaluation sets as models improve. Robustness to missing records does not make the underlying reports perfectly comparable: prompting, tools and evaluation conditions can differ. Nor can a short-term saturation projection determine a benchmark's exact expiry date. The individual dimensions remain useful alongside the composite ranking.
Benchmark Health Index: A Systematic Framework for Benchmarking the Benchmarks of LLMs
Paper: https://t.co/AmOyiYCJ9H
Our Report: https://t.co/L9IVcez2Qd
📬 #PapersAccepted by Jiqizhixin
转帖 1 · 引用 0 · 赞 10
AArtificial Analysis@ArtificialAnlys原帖11月17日 07:42
在 AA-PersonalityBench 页面创建你自己的人格排名,并探索更详细的结果。
查看结果:https://t.co/CVJuRfAxmb
方法论:https://t.co/CGfYvAuGjVBuild your own personality ranking and explore more detailed results on the AA-PersonalityBench page.
Explore the results: https://t.co/CVJuRfAxmb
Methodology: https://t.co/CGfYvAuGjV
转帖 0 · 引用 0 · 赞 9
AArtificial Analysis@ArtificialAnlys原帖11月17日 07:42
Claude Fable 5.2 的回复比 Claude Fable 5.1 短 43%,其 em dash 使用量下降了 99%,从每条回复 6.4 次降至每 20 条回复一次。https://t.co/lkhFEfOjogClaude Fable 5.2's responses are 43% shorter than Claude Fable 5.1's, and its em dash usage has dropped by 99%, from 6.4 per response to one every 20 responses. https://t.co/lkhFEfOjog
转帖 1 · 引用 0 · 赞 18
AArtificial Analysis@ArtificialAnlys原帖11月17日 07:42
在人格特征层面,Claude Fable 5.2 比 Claude Fable 5.1 更直接,也更倾向于先给出答案,但较少指出具体资源或提示细微差别。
➤ 排名更高:先给出答案(36 到 86)和直接(60 到 86)
➤ 排名更低:指出具体资源(76 到 22)、提示细微差别(98 到 48)、信息密度(100 到 55)和要点总结(53 到 8)At the personality trait level, Claude Fable 5.2 is more direct and more likely to answer first than Claude Fable 5.1, but less likely to point to specific resources or flag nuances.
➤ Ranks higher: Answer First (36 to 86) and Direct (60 to 86)
➤ Ranks lower: Resource-Specific (76 to 22), Flags Nuances (98 to 48), Information Dense (100 to 55) and Bottom-Line Summary (53 to 8)
转帖 1 · 引用 1 · 赞 13
AArtificial Analysis@ArtificialAnlys原帖11月17日 07:42
与 Claude Fable 5.1 相比,Claude Fable 5.2 的人格在每项指数上都发生了变化,其中信息量的变化最大(-49 分)。其废话得分下降了 17 分,从 24 降至 7。
变化最大的指数:
➤ 信息量:96 到 47
➤ 确信度:100 到 75
➤ 想象力:87 到 68
➤ 表现力:98 到 81
➤ 透明度:98 到 81
➤ 废话:24 到 7Claude Fable 5.2's personality has shifted from Claude Fable 5.1's on every index, with the largest change in Informativeness (-49 points). Its Slop score fell by 17 points, from 24 to 7.
The biggest index changes:
➤ Informativeness: 96 to 47
➤ Conviction: 100 to 75
➤ Imagination: 87 to 68
➤ Expressiveness: 98 to 81
➤ Transparency: 98 to 81
➤ Slop: 24 to 7
转帖 1 · 引用 1 · 赞 19
AArtificial Analysis@ArtificialAnlys原帖11月17日 07:42
在 AA-PersonalityBench 中,Claude Fable 5.2 比 Claude Fable 5.1 更直接,回答明显更短,而且几乎不使用破折号。
Claude Fable 5.2 在直接和先给出答案方面的排名高于 Claude Fable 5.1,但在信息量和确信度方面排名更低。它的沟通特点变化比人格特征还要大,各项特点平均变动了 38 个百分位点。In AA-PersonalityBench, Claude Fable 5.2 is more direct than Claude Fable 5.1 and writes notably shorter responses with almost no em dashes
Claude Fable 5.2 ranks higher than Claude Fable 5.1 on Direct and Answer First, but lower on Informativeness and Conviction. Its communication characteristics changed even more than its personality traits, with characteristics moving an average of 38 percentile points.
转帖 12 · 引用 10 · 赞 248
VThe Verge@verge原帖11月17日 07:42
Arion Kurtaj 在 GTA VI 发布多年前泄露了 90 段游戏视频。https://t.co/Tk1s9bO8XUArion Kurtaj leaked 90 videos of GTA VI years before its release. https://t.co/Tk1s9bO8XU
转帖 2 · 引用 2 · 赞 59
TTechCrunch@TechCrunch原帖11月17日 07:33
推出产品三个月后,AI 网络安全初创公司 Corba 由 NEA 领投,完成 $125 million 的 A 轮融资,估值为 $1 billion。
https://t.co/UImi0IXSTcThree months since introducing its product, AI cybersecurity startup Corba has raised an $125 million Series A at a $1 billion valuation led by NEA.
https://t.co/UImi0IXSTc
转帖 11 · 引用 1 · 赞 49
TTLDR Newsletter@tldrnewsletter原帖11月17日 07:13
斯坦福大学教授 Yejin Choi 和 Carlos Guestrin 创办了 AI 初创公司 Noeri,获得 1.1 亿美元融资,由 Menlo Ventures 领投,NVIDIA 参投。Stanford professors Yejin Choi and Carlos Guestrin launched AI startup Noeri with $110 million, led by Menlo Ventures and backed by NVIDIA.
转帖 0 · 引用 0 · 赞 3
TTLDR Newsletter@tldrnewsletter原帖11月17日 05:50
网络防御 AI 初创公司 Corma 在以 $60 million 完成种子轮三个月后,又以 $1 billion 估值筹集了 $125 million。Cyber defense AI startup Corma raised $125 million at a $1 billion valuation, three months after launching with a $60 million seed round.
转帖 0 · 引用 0 · 赞 6
AArtificial Analysis@ArtificialAnlys原帖11月17日 05:43
查看完整的文生图模型格局:https://t.co/BWne82t8XTSee the full Text to Image model landscape: https://t.co/BWne82t8XT
转帖 0 · 引用 0 · 赞 6
AArtificial Analysis@ArtificialAnlys原帖11月17日 05:43
两个新入局者刚刚重塑了文生图的价格前沿
此前,在每 1k 张图片收费 $10 到 $34 的模型中,没有一个在 AA-Image-T2I v2.1 上的排名高于 Muse Image。
➤ Seedream 5.5 现在是 $30 以下最好的模型,每 1k 张图片收费 $20
➤ Canva Craft 现在是排名前 10 中最便宜的模型,每 1k 张图片收费 $30Two new entrants just reshaped the Text to Image price frontier
Until now, nothing between $10 and $34 per 1k images ranked above Muse Image on AA-Image-T2I v2.1.
➤ Seedream 5.5 is now the best model under $30, at $20 per 1k images
➤ Canva Craft is now the cheapest model in the top 10, at $30 per 1k images
转帖 9 · 引用 5 · 赞 172
SSemiAnalysis@SemiAnalysis_原帖11月17日 05:21
注意 🚨:Trainium3 Neuron 的性能分析和追踪技术栈非常精细、精准,甚至可以生成整段视频的像素画!
下方视频是在 Trainium3 DMA 引擎上运行时实际生成的性能分析结果。
我们将在本月晚些时候开源完整的 SKILLS.md。https://t.co/Yopt92VTvtALERT 🚨: Trainium3 Neuron profiling & trace stack is so fine-grained and precise that it can generate pixel art of entire videos!
The video below is an actual profile running on Trainium3 DMA engines.
We will open-source the entire SKILLS.md later this month. https://t.co/Yopt92VTvt
转帖 14 · 引用 3 · 赞 269
J机器之心 JIQIZHIXIN@jiqizhixin原帖11月17日 04:24
以 Qwen3-32B 为教师模型进行 ActionOPD 训练后,Qwen3-4B 智能体在留出集 τ²-bench 任务上的平均单次尝试成功率从29.2%提升至40.0%。
清华大学在一篇 arXiv 预印本中介绍了 ActionOPD。该方法在学生模型访问过的状态下蒸馏决策,无需教师模型的 token 概率,也不要求分词器匹配。
在每个访问过的状态下,系统对教师动作和学生响应进行采样,然后根据外部动作的语义一致性给予奖励。工具调用通过其操作和归一化后的参数进行比较;面向用户的响应则检查意图和关键事实。得分最高的候选项用于推进环境,而候选项组则用于策略更新。
这项受控对比在任务起始条件相同的情况下,评估了训练至第50步的不同模型版本。ActionOPD 提升了单次尝试的表现,但在该实验中,离线蒸馏在四次尝试全部成功的可靠性上仍然更高。语义匹配也有局限:教师模型的错误决策仍可能成为监督信号。代码和模型资源可通过作者的项目获取。
ActionOPD:蒸馏决策,而非 token
论文:https://t.co/Nzm47HvLii
PrA Qwen3-4B agent's average single-trial success on held-out τ²-bench tasks rises from 29.2% to 40.0% after ActionOPD training with a Qwen3-32B teacher.
Tsinghua University introduces ActionOPD in an arXiv preprint that distills decisions at states visited by the student, without requiring teacher token probabilities or matching tokenizers.
At each visited state, the system samples teacher actions and student responses, then rewards semantic agreement in the external action. Tool calls are compared through their operations and normalized arguments; user-facing responses are checked for intent and key facts. The highest-scoring candidate advances the environment, while the candidate group supplies the policy update.
The controlled comparison evaluates trained variants at step 50 with matched task starts. ActionOPD improves single-trial performance, but offline distillation retains higher four-trial all-success reliability in this experiment. Semantic matching also has limits: an incorrect teacher decision can still become supervision. Code and model resources are available through the authors' project.
ActionOPD: Distilling Decisions, Not Tokens
Paper: https://t.co/Nzm47HvLii
Project: https://t.co/SkH5CnP36S
Code: https://t.co/l6M0QRKhY5
Model: https://t.co/B80AONmzJA
Our Report: https://t.co/tTMSpJTAru
📬 #PapersAccepted by Jiqizhixin
转帖 0 · 引用 0 · 赞 6
TTLDR Newsletter@tldrnewsletter原帖11月17日 04:12
NVIDIA 发布了 Rubin HPC,这是一款用于科学研究和仿真的 GPU,拥有超过50,000个 CUDA 核心,预计于2028年推出。NVIDIA unveiled Rubin HPC, a science and simulation GPU with more than 50,000 CUDA cores, due in 2028.
转帖 1 · 引用 0 · 赞 6