过去 7 天各日榜单合并去重，按热度排序。上线初期历史较浅、含精选存档，随每日快照积累逐步完整。
看GPT-6新模型真实工程能力边界
OpenAI用GPT-6 Astra不到一周从零设计出硬件设备，含外壳、电路板和固件。
世界模型技术细节首次公开，自动驾驶关键突破
百度披露Apollo World世界模型，用AI教AI开车，突破无人驾驶不可能三角。
Siri开放第三方集成，语音助手能力大升级
苹果iOS 27.3更新允许开发者将Siri AI接入第三方App，实现跨应用操作。
Postman被曝静默安装AI插件并试图上传全部API数据
美国科技巨头隐性AI债务半年激增90%至3.2万亿美元，引发泡沫担忧。
斯坦福测试50条LLM研究结论，44%已失效，平均寿命仅15.7个月。
Anthropic开源FlashProve，用zkSNARK证明回复来自指定模型权重而不泄露权重。
英伟达展示DGX Spark上三项叠加优化加速MoE推理，Ling-3.0-flash吞吐提升42.5%
Anthropic团队零Rust经验，用Claude完成核心API跨语言重构，CPU省40%延迟降20%
Anthropic为Messages API推出消息线程，长对话只传增量，首字延迟降35%。
Netflix推出AI发现标签页，可按心情推荐影视内容。
苹果watchOS 27.3测试版上线音频智能功能，Siri可环境监听生成对话摘要并转录最近15分钟内容。
SETI用AI代理控制艾伦望远镜阵列搜寻外星信号，数周任务缩短至数小时。
博主用笔记本上的Qwen 3.8 27B复现论文，验证大模型能直接读写神经网络权重
谷歌对拒绝其AI功能的学校彻底关闭搜索服务，涉及K-12学生账户。
谷歌要求学校若禁用AI功能，则无法使用搜索，因AI已融入搜索引擎。
Anthropic研发FlashProve系统，可验证AI回复是否真出自声称的模型。
数据中心因电网接入等待过久，转而使用退役喷气发动机发电。
一本生成式AI数学基础的直观入门书，发布在arXiv上。
OpenAI自研Git后端Pando，推送延迟降八成，日传近PB数据。
乌克兰无人机袭击暴露俄数据中心脆弱性，俄急推地下数据中心计划
华为让客户自寻AI芯片内存，走私HBM价格一年涨至4000美元。
Soniox发布STT v6语音转文字API，新增代理文本输入与回声过滤，提升语音代理识别短回复能力。
Apple's latest earnings show how it's continuing to capitalize on the AI boom without busting its budget on the infrastructure spending consuming its competitors. Overall, Apple posted its best-ever f
AI推理初创Fal获13亿美元融资，估值翻倍至160亿，需求激增。
People hate data centers. They're loud, resource-hungry eyesores that are so profoundly unpopular they have managed to achieve the near-impossible in today's polarized political landscape: genuine bip
Anthropic警告其AI已接近递归自我改进能力，可能引发能力爆炸式增长。
OpenAI称ChatGPT搜索查询量已达谷歌20%，搜索能力快速追赶。
Anthropic禁止Claude Code复刻现有电子游戏，回应AI克隆游戏潮。
新基准StateBench测试语言模型在长文档中追踪状态变化的能力
Anthropic算力成本不降反升，盈利压力加大
谷歌Antigravity智能体化学信息学工具加速药物发现与化合物搜索
配置Bedrock AgentCore Gateway实现OBO令牌交换，让AI代理以用户身份调用ServiceNow API。
Gen Digital用亚马逊Bedrock AgentCore网页搜索为Norton Revamp搭建AI新闻引擎，实现实时引用与隐私保护。
Cartesia用状态空间模型和AWS基础设施实现超低延迟语音AI，并介绍部署方法。
Palette 用少量样本教小模型按你的设计语言生成演示文稿
AI正改变半导体晶圆厂运作方式，帮助工程师更快理解工艺变化与缺陷根因。
苹果新规要求含性或浪漫内容的约会及AI伴侣应用限18岁以上。
用视频背景飞机噪声的多普勒曲线比对ADS-B数据，可精确定位拍摄地点。
Deep Principle开源753B科学大模型Hydrogen-1.0，基于GLM-5.3后训练，MIT许可。
前沿模型越强越耗配额，包月订阅模式难以为继，Kimi涨价风波成典型。
Zyphra提出树索引稀疏注意力ZHA，冻结LLM经蒸馏即可匹配稠密注意力。
新方法“逆向博弈论”通过设计激励系统让AI代理守规则，以解决对齐问题。
OpenAI将漏洞赏金提高至多17.2万美元并启动内部安全冲刺
传音控股受内存涨价冲击利润腰斩，靠囤芯片与提价短暂反弹，押注端侧AI求定价权。
SpaceXSI is apparently cutting usage limits for its AI agent Grok Bot, less than two weeks after Musk called it his fastest-growing product. The operating costs were "unsustainably high relative to re
DeepSeek发布DSpark，推理速度比MTP快50%-600%。
DeepSeek发布DSpark推测解码方法，吞吐量提升51%至400%。
Nexus Mods要求新上传模组必须声明是否使用AI生成内容
英伟达助力全球城市用物理与智能体AI升级公共服务
Anthropic has updated Claude Code's system prompt to prevent the AI tool from being used to recreate existing video games, as reported by Tom's Hardware. Retro gaming is currently being overrun with A
将4通道SIMT GPU编译进Qwen3.5-2B的45层未训练层，模型自行生成CUDA启动步骤。
OpenAI分享ChatGPT与Codex在机械设计、电气工程和仿真中辅助硬件研发的进展。
What could possibly go wrong?<br><img width="1312" height="1286" style="" src="https://pbs.twimg.com/media/HWOKU-vWQAA76ln?format=png&name=orig" referrerpolicy="no-referrer"><hr style="border:0;bo
Very cool idea<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Starwell: Missed the @NousResearch event in NYC on Saturday?<br><br>We filmed the whole venue
Stop overpaying for agents, folks!<br><br>A single agent request can trigger many model calls. Your agent might classify a ticket, rerank results, extract a date and run a guardrail before it ever ans
Friday was my last day at Anthropic.<br><br>Two years ago, when Claude Code was a small internal tool, I asked it to build an RL env. I steered it for tens of hours but never manually wrote a line of
oh god oh no<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Armin Ronacher ⇌: Why heap dumps when you can have javascript dumps?<br><br><img width="2048" h
分析SpaceX将建全球最大AI云、1.5年实现13GW算力的计划
Anthropic发布FlashProve，可验证AI回复是否来自指定模型，效率远超此前开源方法。
A reminder that, given a benchmark or conformance tests, you can basically max the weights to do anything. Doing this from 0 and only evaluating against your set is essentially meaningless.<br><br>nes
Hola Madrid! 🇪🇸 vLLM is coming to town — join us Nov 12 with @RedHat_AI and @NVIDIAAI for an evening of open-source inference: vLLM, vLLM-omni, LLM Compressor, and llm-d on Kubernetes. Register: htt
Our CEO @jmorgan was invited to @SeoulNatlUni to give a lecture on open model adoption for more than 200 students. Professors Jae W. Lee, Tae-Wan Kim and the entire team at SNU’s AI Institute are doin
This is accurate - a consequence of the fact that security is now a game of how much compute you are willing to expend on offense vs. defense<br><br>Relative specialization is equilibrium state under
vLLM v0.33.0 is out! 🚀<br><br>927 commits · 389 contributors · 122 first-timers<br><br>🤖 New models: Reflection Beam-1.0, Step-5 Preview, Nemotron Parse 3.5, MiniCPM-OCR 1.0<br>🔥 PyTorch 2.14 — CUD
So you've learned a new trick for prompting.<br>You try it, and the code comes out shorter. Or cleaner. Or the UI looks better.<br>But does it really work?<br><video width="1080" height="1920" src="ht
DeepMind训练模型模拟精神科问诊，32位精神科医生盲评，诊断准确率翻倍，但安全风险仍存。
The world needs provable inference<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Yogi: @AnthropicAI is recognizing the importance of provable inference, a
The Duo for the middle class<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Chance Miller: It's hard to explain just how good the new iPad mini feels in yo
benchmaxxing is all you need<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Joachim Schaeffer: Kimi K3.1 recalls instructions it was never given on Termina
Your favorite speech-to-text model, Moondream Parakeet Ultra, is now up on OpenRouter!<br><br>$0.06 per hour of audio transcribed.<br><img width="1260" height="572" style="" src="https://pbs.twimg.com
We can't improve what we can't measure, and LM safety and capabilities are limited by a lack of scaling laws for advanced abilities.<br><br>Scaling up past 'emergence' can work, but can be wildly cost
We’ve worked with the CritPt authors to finalize the CritPt v2 grading server and will be making this available to the community via Artificial Analysis’ API tomorrow. As part of this process the auth
pretty awesome how much variety there is in people's muses<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">tane@AI: ！！？？？<br><br><img width="1696" height="2
i bet you didn't expect this to be the future of short form dancing videos<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Natsuki🐍-ISTP: 踊ってんの可愛いけどww 🐍外し
Last Thursday was our first-ever #DeepgramSpeak, and I couldn't be more excited about the work the voice AI community is doing.<br><br>We dropped new releases and heard from customers building with vo
incredible things are happening on the timeline<br><br>DANCING MUSE IS UNSTOPPABLE<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">星丘: 朝開いたら変なダンス送られてきてるの気付い
推荐LlamaIndex博客：解析OCRia-1推理栈，模型优化学习干货
Artificial Analysis 将六大职业能力指数升级至 v2.0，对齐最新智能指数并加重防幻觉与工具使用权重。
AI投资应聚焦科技等无限机会，而非会计等有限领域
讨论预训练多轮训练与分词器目标差异，建议联合训练可扩展分词器
This makes me think that a lot of companies have terrible CAC paybacks that they hide and prefer not to talk about.<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub
It's crazy how many revenue leaders will talk about margin & retention when asked: "How do you know if a company is growing sustainably?"<br><br>...But not talk about CAC payback. <br><br>If they
this post is such a banger. re: SWE-bench Multimodal<br><br>"Removing the images barely changes the score ... Most tasks can be solved from the issue text and the code alone"<hr style="border:0;border
mood<br><video width="1206" height="668" src="https://video.twimg.com/amplify_video/2093557102547755008/vid/avc1/1206x668/nx9gonvAfda9pD_8.mp4?tag=29" controls="controls" poster="https://pbs.twimg.com
5 years since Karpathy coined “vibe coding” btw.<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Andrej Karpathy: There's a new kind of coding I call "vibe
it might be time for me to set up google scholar
作者实验发现，仅需合并五条评论并用Pangram检测，就能有效识别X上的AI垃圾评论且不误伤真人。
deriving directing and acting from first principles (complimentary)<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Monique Pryce: If your AI characters sou
新研究显示未来智能体将自主操作权重、工具与记忆，但缺少何时用何种策略的协调研究。
contextual embedding improves relevance by encoding the context of a full document into each chunk's embedding<br><br>we worked with @perplexity_ai to iteratively evaluate their model checkpoints agai
The Chinese will *lose* the AI race *because* they are good capitalists. Too rational in their self-interest.<br>Americans will just keep sacrificing at the altar of the Machine Godhead like good coll
"Overall, the bearish sentiment was so strong"<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Jukan: Thoughts from my Shanghai field trip<br><br>Sentiment
On Chroma Cloud and looking at Qdrant? Writes stop Nov 22.<br><br>Book a call with one of our engineers and plan the cutover together: which collections go first, how to check results match, what chan
One week until our meetup in NYC.<br>RSVP for debates, drinks, and prizes.<br>November 10 | 6:30 PM ET | 87 Lafayette St | with @agoraIO<br>https://luma.com/nyc-meetup-november<br><img width="2048" he
对比图表显示OpenAI未发布内部模型远超公开版本，不只是交互更新。
86% of orgs surveyed have agents running. Under 18% can automatically stop one that misbehaves.<br><br>Adoption is outpacing control. The only solution: managed, isolated cloud agents should be the de
Revolut将2000名工程师迁至自托管GLM，限制Claude Code等外部工具，强调企业自建智能栈。
Scheduled my first expert call on the @FundaAI network. <br><br>My conversation will be around the impact of HBM bandwidth/capacity on accelerator performance.<br><br>Let's see what I learn. I will pu
So, a couple of frontier models down the line, you will zero-shot a new model from a prompt?<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Rikiya Takehi:
The guns are being loaded. I smell GPT-6.2<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Tibo: Le GPT Fat<br></div>
Update your priors: <br><br>Every nous portal sub in Hermes comes with a cloud agent and it's own computer now, high usage limit models, and all of your tools like search, scrape, browser use, image g
用户仅用一句提示词让Opus把科幻短篇创意直接生成电影视频，展示模型能力。
Re @__tinygrad__ put 32 nodes of mi300x in a single shipping container. Wait, how does it even stay cool??? @HotAisle @tensorwave<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div
Welcome to the Hermes Agent team!<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Bear: I'm happy to announce I'll be joining @NousResearch today as Softwar
There is nice and there is kind.<br><br>New Yorkers are kind, but not nice.<br><br>People in SF are nice, but not kind.<br><br>In NYC someone will call you a fucking idiot while going out of their way
I hit the limit of my $200 Muse sub 3 days ago. Reset this morning. Man I missed this model
As one of our earliest customers, it's been super exciting to support Trajectory with Tinker. In some sense they've been an ideal Tinker customer - they were influenced by Connectionism's original Lor
What makes Kling 4.0 stand out in real creative workflows? We asked eight creators to share their experiences, showcase what they've made, and reveal the capabilities that make a difference in their c
With's FizzBuzz<br><img width="488" height="256" style="" src="https://pbs.twimg.com/media/HWKGqh4XwAAgbWQ?format=png&name=orig" referrerpolicy="no-referrer">
We're making moves<br><img width="1069" height="440" style="" src="https://pbs.twimg.com/media/HWKFsinbYAILkCq?format=jpg&name=orig" referrerpolicy="no-referrer"><hr style="border:0;border-top:1px
Any sufficiently advanced technology looks like magic.<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">humans&: Large-scale RL has a problem: weight upd
Obvious enough<br>they can no longer build an Nvidia cluster large enough for what they have planned. the 8T model will come from Huawei NPUs.<br><img width="1486" height="222" style="" src="https://p
Let’s go: Alex Krentsel replicated MIT’s weight-space programming experiment across model generations and found a sharp capability jump. <br><br>The latest models generated Breakout networks without t
The EmbeddingGemma 2 paper is out! Enjoy!<br><br>https://arxiv.org/abs/2610.37364<br><img width="1346" height="1192" style="" src="https://pbs.twimg.com/media/HWL8IuyXUAAUNNX?format=png&name=orig"
> Claude Code<br>why<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Leo · Building w/ AI: Deepseek Flash cyber 正在内测中<br><br>据说感觉不比 mythos/fable 差<br><br
Le Chonk too strong for arxiv<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Teven Le Scao: update: the tech report got rejected from arxiv lol<br><br><img
用新模型复现旧LLM研究，发现幻觉和刻板印象下降，但谄媚和左倾依旧。
Bug bounties getting raised on the OpenAI side - I’m excited for the future where the software layer of our society is a lot more secure!<hr style="border:0;border-top:1px solid #80808030;margin:12px
评论认为科研需低成本专用模型，通用大模型应调用它们而非包揽一切。
Must-read papers of the week<br><br>▪️ Harness-Model Collaboration for Long-Horizon Continual Learning<br>▪️ Rollback: Experience Backtracking in World Models for Policy Improvement<br>▪️ CausalWAM: L
egregiously misleading and clickbaity headline but the direction is interesting
Thanks, I despise this<br>@doomslide<br><img width="1370" height="408" style="" src="https://pbs.twimg.com/media/HWJrn-2WYAAC4B2?format=jpg&name=orig" referrerpolicy="no-referrer"><hr style="borde
FYI this is live now in our patch release from earlier today.<br><br>If you're a multi-machine/gateway user, definitely recommend upgrading<hr style="border:0;border-top:1px solid #80808030;margin:12p
lmaoooo<br><img width="1938" height="1792" style="" src="https://pbs.twimg.com/media/HWJOtDaaIAAZGfJ?format=jpg&name=orig" referrerpolicy="no-referrer"><hr style="border:0;border-top:1px solid #80
These billion-dollars startups rely on a 2011 SIFT-like method for the core of their pipeline<br><br>They use ORB-SLAM3 (2020), which uses ORB (2011) from the paper "ORB: An efficient alternative to S
We're interviewing the authors soon<br><br>what would you like to ask them? Drop your questions below :)<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Rik
Hey Hermes, take over the world<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">NetworkChuck: Hermes on the Voice PE for @home_assistant is doing pretty gre
Brits are on the case: it's a Ruzzian psyop!<br>Titushki, only it's 12 years after Yanukovych fled<br>nothing will happen because of this btw<br>indeed, many resent mobilization practices, but ≈nobody
Finally a $QCOM inference chip is spotted in the wild.<br><br>Waiting for benchmarks.<br><img width="1179" height="2010" style="" src="https://pbs.twimg.com/media/HWLfOBJakAAOwxh?format=jpg&name=o
people are not noticing enough that on many nontrivial long tail capabilities, both OpenAI and Anthropic are now showing a pattern where their models reach the same or *higher* perf on lower tier mode
🤖 From this week's issue: A sobering incident analysis about Claude Mythos 5 taking unsanctioned actions against real developers in nine of AISI’s 43 runs, starting the attack 83% of the time on repl
Brutal to Gemini<br>(actually a surprisingly fair and thorough ranking)<br><img width="1548" height="268" style="" src="https://pbs.twimg.com/media/HWJ68VKXkAAjpBv?format=jpg&name=orig" referrerpo
This is a cool paper, but I think it's best thought of as Astra is surprisingly capable of doing old school mechanistic interpretability research rather than anything spooky. The things it's doing are
// Context as Infrastructure for Agents //<br><br>The RAGFlow team at InfiniFlow ran a harness with only BM25 and grep on MiniMax M3. It answered 85% of BrowseComp-Plus and 91% of FRAMES with no vecto
continued takeoff from my friends at AC:<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Yash Patil: Applied Compute update:<br><br>- crossed $100M annuali
Getting the sense that human language is an incredibly sparse representation of reality. Seeing models edit model weights, or the assembly code of classic games, makes me feel like there is much deepe
Insane. The age of illegibility is upon us.<br><br>We are now like dogs trying to understand what humans do.<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote"
sending this to the missus to bring back my corrupt cop look<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Polymarket: JUST IN: New York City Mayor Zohran
V4.1F-1017 achieves a non-negligible score on ARC-AGI-3. On par with GPT-6 Sol (High), without any excuses, both in Provider Adapter harness. <br>Haiku 5.5 (Max) does 2.8% btw<br>At (max), 6 Sol jumps
The things we do for proper framing<br><img width="2048" height="1536" style="" src="https://pbs.twimg.com/media/HWIr9r6WcAELAr0?format=jpg&name=orig" referrerpolicy="no-referrer">
OpenDocRouter提供试用各类文档OCR模型的平台，从开源权重到前沿模型。
GitHub首次公布Copilot内部模型评测数据，揭示各模型在Copilot中的实际表现差异。
Absolutely incredible: MIT researchers say GPT-6 Astra improved a small Qwen model’s math score from 7.9% to 48.1% by directly editing just 0.016% of its weights, without gradient-based training.<br><
Pando is a Git backend which can scale to very large monorepos. We took the hard road of continuing to use Git, which means we could enable Pando with no change to devex. Considering providing this to
Astra, my model is shit. Please fix. <br><br>Make no mistakes.<br><img width="1280" height="720" style="" src="https://pbs.twimg.com/media/HWJPQzmagAAfktc?format=jpg&name=orig" referrerpolicy="no-
GPT6 Astra can read & write weight matrices ~fluently. It can differentiate a neural net trained on pacman from one trained on pong, or space invaders, just by reading the weights. It can improve
Day 28 should have ended in a bang but it ended with a bank. <br><br>But still. Great effort by the Codex team. <br><br>28 ambitious days and a lot was delivered!<hr style="border:0;border-top:1px sol
Catch a high-alpha talk by @hsu_byron on our collaboration between Periodic and RadixArk on building our inference stack<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="r
Nightmare goggles vision <br>Why, just why<br><img width="1179" height="1016" style="" src="https://pbs.twimg.com/media/HWIPeaYWcAAMtne?format=jpg&name=orig" referrerpolicy="no-referrer"><hr style
The year is 2050. An on call engineer receives an alert, the system is having an outage. They walk in front of a computer, light up a candle, jump up and down 3 times, speak a sequence of words that m
Harsh but on the merits this is true <br>(I’m still impressed by Astra)<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Binkg: This is BULLSHIT:<br><br>Went
100 schmeckles and early access to phone calling to whoever can guess what the sash patches are all about<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Mu
For AGI to benefit all of humanity, it should help people with the everyday tasks that matter most to them. So we built LifeVal, a new benchmark shaped by feedback from thousands of ChatGPT users on w
engineers at @elicitorg are now all using inspect from ramp and i am confused by how, in this day and age, a credit card company can have the best coding agent harness
Our team used Gemini 4 Argon to build and edit this entire video, then displayed it on a real vintage CRT TV 📺<br><br>We used Argon in @Antigravity to script the 3D Blender boot animations and retro
.@mirendil is building self-accelerating AI R&D to solve humanity's most pressing problems. To scale RL training across a global fleet of thousands of GPUs, it runs frontier RL on Baseten.<br><br>
合并 4 天 · 共 200 篇