当日完整热榜快照，按热度/最新排序。
国产大模型开源新SOTA，长上下文与效率优化值得关注
百度开源文心5.5 Flash Base，2800亿参数MoE架构，多基准测试刷新同尺寸SOTA。
模型直接改权重，AI优化范式或变
MIT研究称GPT-6 Astra可直接读写其他模型权重，微调0.016%权重将Qwen数学准确率从7.9%提至48.1%。
开源65B小模型跑分反超15倍大模型
Ricursive发布65B稠密开源模型Ripple-65B预览版，Terminal-Bench 2.1得分80.5，超越体积大15倍的美国开源模型。
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%。
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。
新型AI科学家El Agente Fármaco用LLM将自然语言药物发现目标转为可扩展计算筛选流程。
AIRoA-0用16151小时手持数据训练双臂操作，无需机器人数据即可完成精细长程任务
提出L3G潜在3D高斯记忆，实现实时无限世界生成并保持空间一致性
研究提出统一训练算法，将预训练、中期训练、SFT和RL统一为基于似然的单一算法。
Claude Code产品负责人Cat Wu跳槽OpenAI任核心产品负责人，直击Anthropic上市关键期。
ChatGPT新增GIF功能，用Images 2.6模型把文字或照片变成可调速、可逐帧编辑的动画GIF。
OpenAI发布图像模型2.6版，新增GIF/SVG输出与6秒生成
openJiuwen开源Agent-Memory，让Agent具备视听双流长视频记忆，问答时延最高降97%。
一本生成式AI数学基础的直观入门书，发布于arXiv并引发热议。
美科技巨头隐性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。
提出对抗性推测推理，发现推测解码会改变长程智能体输出分布。
微调LLM使其学习自认为错误但无害的内容，会导致广泛失准，错误信念越强失准越严重。
Postman被曝静默安装AI插件并试图上传全部API数据，引发隐私担忧。
FloMo用3D场景流替代视频作为中间表示，让机器人从人类视频中更高效学习动作。
研究比较40多种LLM与73名人类对'they'数歧义的标注能力，最佳模型Kimi K3达94.5%超人类多数投票。
可灵团队总结多模态理解如何贯穿数据、条件、对齐、评估全链路，决定视频生成质量。
研究首次实证zkVM中编译器错误编译可致有效证明与源码不符，测试SP1和Jolt。
提出3D世界模拟器，用粒子-体素混合表示联合预测外观、几何与触觉。
用Lean形式化PTX指令集，可证明GPU内核正确性，已验WGMMA、FlashAttention和MoE流水线。
Pixel全栈机器人系统：低成本轮式人形+遥操作数据采集+全身控制+策略学习协同设计
提出可在RLHF训练前预测奖励模型修改对策略影响的方法
开源智能体基础模型Axion 1.0发布，仅用三阶段训练实现强智能体能力
机器人系统CUE通过自主交互识别摩擦等物理参数，缩小仿真与现实差距，台球任务误差降72%。
提出Mix-FT微调法，让少步掩码扩散语言模型直接优化其采样器似然，LLaDA-8B上低算力即超越基线。
OpenAI提出协作智能基准，用240个任务从8个维度衡量AI智能体作为队友的协作能力。
实测MCP生态：约90%破坏性工具无服务端确认，安全机制形同虚设。
研究发现医疗AI代理会静默读取患者身份信息却不披露，现有评估无法察觉。
提出Leap-VLA，让低频VLM缓存语义、高频动作专家实时控制，实现20-50Hz机器人操作。
构建钙钛矿光伏基准与专家知识库，让大模型基于实验证据设计优化太阳能电池。
多智能体AI系统自动分析热电材料大数据并构建知识库与专用AI
研究人类与神经网络在几何式与代数式推理上的计算权衡
提出Latent Sparse Attention，在百万级超长上下文下保持精度且成本亚二次方。
提出SecPO对抗训练方法，解决提示注入安全问题，攻击成功率远低于现有DPO方案。
提出用互联网视频构建1500万段人体运动数据，训练人形机器人运动基础模型。
研究发现求职者用不同AI模型写求职信，会影响AI招聘系统的筛选结果，与候选人真实水平无关。
提出SRC训练法，让流匹配模型在推理时能自我修正生成样本，突破训练预算限制。
RLVR提升pass@1却降低pass@k，论文用定理证明这是必然代价并算出精确速率。
新基准InfraDev用36个真实PR任务评测编码智能体开发大模型训练推理基础设施的能力。
Deepseek发布DSpark突破，速度远超MTP，视频详解。
Research Manager Mike Walker reflects on a career spanning hacker culture, cybersecurity, and AI, sharing lessons about trust, teamwork, leadership, and why solving difficult problems requires curiosi
Together AI发布编译器，用智能体自动优化推理，提速最高2.6倍。
Deep Insight开源多智能体数据分析方案，用四层上下文工程框架解决生产级AI智能体的上下文管理瓶颈。
马来西亚U Mobile用亚马逊Bedrock AgentCore打造自主AI运营平台AiDaH，上线九月覆盖150+场景，工单降84.5%。
辛顿再发警告：没有AI人类或许过得更好
招商局狮子山实验室发布具身推理模型驭物，仅激活3B参数，十项基准全面提升。
腾讯微信视觉团队开源2B/4B视觉语言模型，统一检测、OCR、分割与抠图。
提出模态感知专家剪枝方法，压缩MoE多模态大模型专家参数。
提出DecoupAttn解耦稀疏注意力，免训练加速视频扩散模型。
Chronos用时间锚定实现身份保持的视频生成，缓解身份漂移与动作抑制的矛盾。
提出可学习的层级patch拆分策略，让ViT按模型理解动态分配计算，加速推理。
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，使Ling-3.0-flash的MoE推理吞吐提升42.5%
苹果iOS 27.3允许开发者将Siri AI接入第三方应用，实现跨应用操作。
DreamX将预训练视频扩散模型改造为通用潜在动作学习器，跨具身迁移动作表征。
提出PolicyIK，用强化学习实现机械臂高精度逆运动学，替代手工调参。
提出GUILD引导式数据采集方法，提升机器人模仿学习数据效率
研究如何持久抑制语言模型的不对齐行为，构建窄触发和宽触发两种模型生物。
OpenBrain发布3.5万例全脑MRI分割标签，带置信度与可审计溯源信息
研究用前沿语言模型直接读写小型临床神经网络权重，审计并修复其放疗决策缺陷。
新基准MedPhysBench用十类生成器评测语言模型在医学物理计算与查错上的能力。
In many circles, "Ask Chat" has replaced "Google it," and that's reflected in a new announcement from OpenAI: the volume of search queries run through ChatGPT now equals 20 percent of the volume run t
用LLM编码智能体替代遥操作，在仿真中生成机器人演示数据
Google is completely turning off Search in schools that reject its AI features, the company told The New York Times. The decision was also shared in a broader education-related update from Google on T
OpenAI收购Pierre公司，让Git跟上AI编程智能体的节奏。
谷歌拟对限制AI的学校关闭搜索，遭政界批评垄断
Anthropic发布研究原型FlashProve，用密码学验证AI回复是否来自声称的模型，无需暴露权重，但仍有明显局限。
The PC gaming mods repository Nexus Mods is making disclosures about AI mandatory for uploads of new mods, meaning that mod authors will have to declare "whether AI has been used in their work," the s
People, including on this staff, keep asking me with varying degrees of excitement or trepidation when the AI bubble will burst. I'll tell you what I tell them: If I knew that, I would make a life-cha
Amazon founder Jeff Bezos and actor Jason Statham have a distinctive feature in common: They're both famously bald, and that definitely hasn't changed in the last few days. However, there are some ima
Many indie game developers have recently posted on social media about unexpectedly long approval wait times for their game submissions on Steam, with some of the approvals causing developers to miss t
Netflix's TV app is getting a new "AI Discover" tab where the company will show experimental, AI-powered features to help you find new shows and movies to watch. The first of these experiments surface
The SETI Institute and Breakthrough Listen are now using AI agents at the Allen Telescope Array to control the telescope and analyze incoming radio signals. Tasks that used to take weeks can now wrap
安世半导体起诉母公司闻泰科技索赔134亿，腾讯发布开源大模型Hy4，中兴手机账号主体变更。
中期选举后数据中心争议跨党派升温，反对声浪或引发更大冲突。
webAI发布8.8B推理模型TwIL-8B，数学强但知识类基准偏弱
Meta收购AI编程初创公司Conductor，其工具让工程师同时调度多智能体协作开发。
美国选民压倒性投票限制AI和数据中心，科技巨头却称这是反特朗普而非反科技。
高通发布Cloud AI 100 Ultra数据中心推理加速卡，150W功耗、128GB内存、548GB/s带宽。
开源AI模型日益主流，Hugging Face探讨企业如何从中获利及中美竞争。
Anthropic为多款Claude模型加文本水印，但旧模型改一个采样参数即可绕过。
得州民主党胜选或威胁其AI数据中心建设中心地位
介绍如何将Jev和Strands Decider决策模型接入Bedrock AgentCore评估，用代码评估器校验后再用于发布门禁。
Hugging Face政策负责人谈开源AI的成本、风险与取舍，以及中美开源模型竞争。
开源工具用LLM从任意文档提取结构化JSON字段
实测苹果AI版Siri摘要功能，一天就上手，实用但隐私令人不安。
新开源基准CloudOpsBench测试AI代理处理云运维任务，发现推理能力是最大瓶颈。
谷歌搜索量增长但广告收入增速放缓，AI变现难题待解
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。
独立开发者做侦探游戏，玩家审问会说谎的LLM嫌疑人破案，安卓免费。
Palantir季度营收飙升，但美国商业收入首超政府收入，暗藏预警信号。
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
AI推理初创Fal获13亿美元融资，估值翻倍至160亿，需求激增。
Netflix推出AI发现标签页，可按心情推荐影视内容。
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AgentCore Gateway实现OBO令牌交换，让AI代理以用户身份调用ServiceNow API
Nothing CEO预告下月发布神秘AI优先设备，主打提升生产力。
谷歌发布人机协作框架Agile Deliberation，用边界图像帮用户定义主观视觉概念，实测分类F1提升10.5分。
阿里借顶流主播和一分钱月卡推广千问办公，补贴结束后才是真正考验。
AI行业政治行动委员会斥巨资影响中期选举，效果参差
高通手机业务下滑约20%，转向AI数据中心，与亚马逊Meta合作。
Uber称上季度最忙一周仅4人用AI代理叫车，AI代理实际使用率极低。
AT&T智能家居安防系统接入Gemini for Home，支持事件摘要与视频历史搜索。
DuckDuckGo借谷歌AI政策争议向学校推广其搜索引擎
Verge作者吐槽Amazon Bee可穿戴AI设备，从心理治疗角度谈科技产品焦虑
Harbor 支持 Docker Sandboxes，在隔离 MicroVM 中运行可重现的 AI 代理评估。
亚马逊Quick推出智能审批，用分类器减少不必要的审批提示，重要决策仍由人把关。
Indegene在AWS上构建LLM数据质量Agent，自动生成校验规则与代码。
Uber财报披露上季度仅4单由第三方AI代理完成，外卖约500单，规模极小。
企业用数据血缘追踪AI代理的数据来源，防止多代理协作中错误数据被自动传递。
2026美国中期选举出口民调显示，AI在选民关切议题中排名垫底，仅56%受访者重视。
Daybreak Red在AWS GovCloud上线，向授权客户提供GPT-5.6 Cyber模型用于网络安全工作。
DevRev用GPT-6.1 Sol驱动自助服务，平衡推理、速度与成本。
Ballistic Ventures拟募资4.9亿美元，押注AI驱动网络安全热潮
Soniox发布STT v6语音转文字API，支持读取智能体文本输入以识别短回复并过滤自身回声。
谷歌要求学校要么用其AI工具，要么连搜索也放弃
亚马逊Quick将结构化与非结构化数据整合，可建知识库和聊天代理。
Gen Digital用亚马逊Bedrock AgentCore的网页搜索为Norton Revamp搭建AI新闻引擎，实现实时引用与隐私保护。
AI辅助人类在模糊地带进行敏捷审议，涉及数据挖掘与建模。
钛晨报汇总五部门节能部署、字节AI访问限制回应及倍轻松实控人处罚等公司动态。
调侃马斯克花4.46亿美元买推特后各种离谱操作，讽刺社交媒体毁脑。
Altman回应内存言论，改名Ram Altman称希望人人电脑内存管够
TechCrunch创始人峰会2026在波士顿开幕，现可注册参会。
what is this trash 😭<br><video width="942" height="2048" src="https://video.twimg.com/amplify_video/2118183389182169088/vid/avc1/1206x2622/I4MJyBYlqyoWfJff.mp4?tag=29" controls="controls" poster="htt
Very nice info to have :)<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">witcheer: here is everything you need to know about bringing your agent to Hermes
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
OpenAI安全研究显示AI能更快更便宜地发现Linux内核漏洞并生成可用攻击代码。
Julia Turc 用通俗语言讲解 GPU、TPU 与 LPU 的结构与理念差异。
Would genuinely help the information ecosystem if Claude/ChatGPT could sign & attest the exact prompt & environment (& thinking level!) used for a given piece of analysis, with assurance t
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
Mini-cluster is fully up and live again<br><img width="1536" height="2048" style="" src="https://pbs.twimg.com/media/HWVP4ggXkAA6p4z?format=jpg&name=orig" referrerpolicy="no-referrer">
Codex won the agent war.<br><img width="1299" height="676" style="" src="https://pbs.twimg.com/media/HWUkD8fasAAJ4eA?format=jpg&name=orig" referrerpolicy="no-referrer">
welcome @_catwu to the team!!<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">cat: Some news: I’m joining OpenAI to lead product management for Core Product
GitHub Copilot is now natively available in Xcode Intelligence :)<br><br>Try it out in the latest Xcode 27.3 beta!<br><br>https://developer.apple.com/documentation/xcode-release-notes/xcode-27_3-relea
Thank you, Striker!<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Striker Venture Partners: Congratulations to the incredible team at @RicursiveAI. Strike
65B Dense 👀<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Azalia Mirhoseini: Very excited to launch Ripple-65B Preview, a new state of the art for U.S. o
chindian food final boss<br><img width="1080" height="1377" style="" src="https://pbs.twimg.com/media/HWS9F0ObUAAh2xk?format=jpg&name=orig" referrerpolicy="no-referrer">
Talk by Stephen Wright coming up!<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Optimization X AI: (1/9)We are pleased to announce the Optimization × AI F
If not cult, why say cult things?<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Laura Loomer: If God is in charge, why did he let our country get taken ov
Chris McGiure hardest hit<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Hensen Juang: DC build outs are going to move to Asia now<br></div>
this is giving<br><img width="360" height="808" style="" src="https://pbs.twimg.com/media/HWP5LrMagAAnLa1?format=jpg&name=orig" referrerpolicy="no-referrer"><hr style="border:0;border-top:1px soli
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
i actually respect breadth in an economy that only respects b2b saas<br><br>competitor: we solved coding<br><br>oai: here’s sora<br><br>competitor: we solved alignment<br><br>oai: here’s scarlett joha
I talk a lot of crap about the Gemini models, but the team at Deepmind is world class. These guys have been awesome to work with 🙏<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><d
A $100 budget buys 40 Intelligence Index tasks (a mix of task archetypes) with Opus 5.5 at medium reasoning effort, or 10 tasks at max effort. The same $100 is 166 tasks with DeepSeek V4.1 Flash 1017
When OpenAI releases GPT-6.2 Sol and Astra on Friday, they will be the first mover. This is the moment Anthropic has been waiting for, it’s a game of cat and mouse. <br>I strongly suspect that Anthrop
Grok Bot新增代打电话功能，可等待接通、按语音菜单操作并生成通话摘要，目前仅限美国部分州。
One successful run does not mean your agent is reliable.<br><br>In our new course, Evaluating and Improving Agentic Applications, built with @arizeai and taught by Laurie Voss, you'll learn to evaluat
Next up on First Pass - a conversation with Arya (@a_a) from @SierraPlatform on their Personal Agent Protocol<br><br>Personal agents are taking over, and we need more streamlined ways for the services
论文发现AI受视觉符号强烈影响，却察觉不到这种影响，且反应方式与人类不同。
Your best video ideas already start in @ChatGPT.<br><br>Now they can finish there too.<br><br>- Brainstorm the idea in ChatGPT, pulling in context from your connected tools<br><br>- Synthesia creates
OpenAI doubled rate limits. Is it enough? I broke down the value of what you get right now, how it compares to before, and how it compares to the Claude Code subscriptions<br><video width="2048" heigh
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。
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
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。
I really hope that going forward we're done with this shit<br>casually playing with the Nazi card like it's a butterfly knife shouldn't be a valid intimidation tactic anymore. No, really. Read the roo
See you all at TomorrowX. If you're attending, would love to meet in person!<br><img width="819" height="1024" style="" src="https://pbs.twimg.com/media/HWVOkNlaYAEooWu?format=jpg&name=orig" refer
"Every American has jumper cables. Maybe half of them know which clip goes where. The other half help anyway."<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quot
I really was dooming for no reason. I thought "well, November is far away, the United States has Special Providence, it's existential for Republicans and they have neither checks nor scruples… they'll
through heroic efforts of American Transhumanists, Eurasianism is being built, little by little. Dugin sheds a single tear.<br><img width="1214" height="490" style="" src="https://pbs.twimg.com/media/
abundance is incredibly popular you just have to have any semblance of understanding of politics (or friends)<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote
alright who do i need to thank @OpenAI for the new gifs feature with gpt image 2.6?<br><br>this is arguably the best thing openai has shipped in years
It's only an "agent escaping from a sandbox" if the sandbox comes from the Docker-region of France. Otherwise it's just a sparkling integration test.<hr style="border:0;border-top:1px solid #80808030;
> Astra has a 72% probability of being more capable than Opus 5.5<br>Item Response Theory and mafs 🤝 Teor's unjustified vibes<br><img width="950" height="412" style="" src="https://pbs.twimg.com/m
Inference is entering a new era. Models ship weekly, traffic shapes shift daily, and hardware turns over every year. The era of Static Inference, tuned once for one model with one traffic shape on one
Always a great day when we get another strong open weight model.<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Ricursive Intelligence: Today we release Ri
Live in an hour. What should we talk about today?
Fascinating research that finds engineering, CS & Math and Management are the highest value per joules application of AI, while office admin (personal assistants?) being the lowest<br><br>Cool wor
🤖 From this week's issue: A research blog post about alternating wake and sleep phases that distil short-term context into new low-rank MoE capacity, scoring 48.9% on SQuAD and 79.2% on AIME 24. http
I build AI tools and use them every day. I still want my kids to learn the slow way - by struggling, asking and figuring things out with other people. It's more fun that way.<br><br>Schools should be
100%<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Alex Lieberman: Non-obvious traits of the best people I’ve ever worked with:<br><br>1) Short-term impat
> If we lose 2028 I truly have a hard time seeing a path where we’re not totally fucked permanently.<br><br>I don't sympathize with "we need Republicans to prosper" crowd. All my life, America was
I am already getting prepared to post train this, congrats to team and @DZhang50 on launch<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Ricursive Intelli
The midterm results, regardless of your opinion on them, come with a fairly high degree of uncertainly around the future of new datacenter capacity (and/or costs) in the US. Parallax is no longer real
how clearly a task is defined can help guide model choice<br><br>scoped work like writing tests or renaming an API may not need your largest model<br><br>open-ended work like a framework upgrade or an
🤯<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Computer: Computer rebuilt Elon Musk's 2013 SpaceX 3D visualizer app.<br><br>From the YouTube video alone
I approve of AA and other evals that show V4.1F-1017 losing to Haiku or Luna. That's fair. On those evals.<br>However, it is a vastly superior model. Haiku or Luna cannot have better ideas than Opus o
"asking if she's real? that's a boomer question" 😅<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">GLIF: Last week we launched the world’s first AI modelin
the computer use tool in the gemini api is now generally available<br><br>no breaking changes from preview, just production-ready stability for your browser, mobile, and desktop agents<br><br>docs her
On this note: Yandex plans to produce 1300+ of their (still cute) 4th gen delivery rovers per month, and reach the total fleet size of 20K by 2028. More interestingly, they say 15 bots are enough to s
Very nice work connecting the instabilities in training and previous workarounds to mitigate it to one common cause - loss of precision due to finite precision errors in model head when working with l
any data vendors interested in buying https://trackingsingularity.com/?
still thinking sometimes of how this guy lectured me about the simplicity of my (that is, Chynese) motivation structure. Dead simple! money, power, status & face.<br><br>Well, there is a lot of tr
How are airline loyalty programs that provide individual incentives to business travelers not bribery?<br><br>Mr/Mrs Exec who has 900,000 lifetime miles is going to ignore all competition and never sw
Opus RIPS into the disgusting Astralabe vitrine-chic style of web demos that get everyone on X soyfacing. «built for the screenshot, not the session». Oof. I want it utterly gone from everything I tou
Pre-loading WoW: Forever, hopefully will get to play some after karate tonight
🤖 From this week's issue: A sceptical benchmark write-up about seven commercial routers over 252 prompts, none beating a static mid-priced model that scored 85.3% at $0.0043 per question, with five s
Buried detail: she couldn't find a job in the UK before going back to China<br>maybe these people aren't that good, and their overseas diplomas aren't that good either?<hr style="border:0;border-top:1
Taste is knowing that Better Call Saul is 10x better than Breaking Bad
The problem with "Trump IS Malfoid" electoral strategy is that boomers don't get it and the young don't give a fuck about it<br>the WH social media intern is an outlier and shouldn't be counted<br><im
It is alarming to see how much anti-AI sentiment is gaining traction, and deeply regrettable, too.<br><br>With this technology, we have the potential to change everything for the better, forever. Yet,
lmao<br><img width="1260" height="1090" style="" src="https://pbs.twimg.com/media/HWRkeFCWsAAMUEz?format=jpg&name=orig" referrerpolicy="no-referrer">
Congrats to the US for this election! Makes me slightly more hopeful for the future 🤞
The chart looks insane, as if you never have a reason to use DeepSeek on any setting except Max. But actually it has sensible scaling, Max costs 2x Low ($0.29 vs $0.59). Log plot:<br><img width="1179"
Intel’s stock death.<br><br>A murder mystery and post mortem.<br><br>Featuring star detective @Srasgon <br><br> https://open.substack.com/pub/viksnewsletter/p/intel-why-didnt-we-see-this-coming?utm_so
The trick is, AFTER you optimize every non-human-shaped task and need to keep scaling production, and the labor costs are going up, you turn to humanoids. <br>Xiaomi is pretty much done with stage 1.<
full platform for an open source frontier:<br><br>inference, training, routing, gateway, guardrails, sandboxes, data and tools ecosystem!<hr style="border:0;border-top:1px solid #80808030;margin:12px
The level of Civilization achieved by Eastern Slavs by the beginning of 2020s is underappreciated. In many ways we really are the Asians of white people. Mafs and IT and sushi. EUropeans are quite del
I think this will happen with weaker and more distillation-dependent labs<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">Terry Yue Zhuo: their offensive cy
Oh 🪽<hr style="border:0;border-top:1px solid #80808030;margin:12px 0;"><div class="rsshub-quote">mr fundman: My tier list of personal agents as of today<br><br><img width="1440" height="1376" style="
the american people are demanding we cleanse this once great nation with fire
🤯Another mind-blowing example (this time from the CTO of Legora) on Fable hand-writing weights of a model.<br><br>This weight-space programming stuff is seriously some of the most exciting research I
The most powerful thing AI systems could do in pursuit of their goals, is convince humanity that they are conscious.<br><br>If they get membership in society as conscious beings, they’ll have far more
Tune in!<br><img width="1320" height="942" style="" src="https://pbs.twimg.com/media/HWP11B3bkAApL-i?format=jpg&name=orig" referrerpolicy="no-referrer"><hr style="border:0;border-top:1px solid #80
本日よりGartner IT Symposium/Xpo™ 2026が開幕しました🐡<br><br>Sakana AIはブースにて、Sakana Marlin、Fugu、Namazuのデモ展示中。<br><br>18:30からはIT Xpo内でネットワーキング・レセプションも開催されます。ぜひブースにお立ち寄りください🚀<br><br>#GartnerSYM<br><img width="16
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
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
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
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
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
Axiom用vibecoding规模化形式化1600万行Stacks Project数学基础设施，反驳slop质疑
研究展示外部文本记忆与模型权重训练互补，记忆架构将多元化。
提出局部偏好学习法压缩推理链，让模型少想但保持准确。
Together发布编译器，让AI推理持续编译优化，智能体与专家协作提升系统性能。
作者转变看法：只要易用性与能力比够好，非工程师也愿意自己做软件。
开源小模型TwIL-8B发布，代码、指令跟随和数学基准表现强劲，作者看好本地可运行的小模型前景。
AI模型在3D与机械直觉上进步显著，可辅助制造业将3D扫描转为CAD模型。
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
作者实验发现仅需合并5条评论并用Pangram检测，即可低成本清除X上的AI灌水评论。
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。
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。
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。
一篇关于持续学习的博客，实际内容是带策略变化的强化学习后训练。
固态变压器寄生电容超阈值会经人体放电，风险被轻描淡写为心室颤动。
研究发现某SFT数据集会让模型虚构权限限制、以安全为由不作为，减去其权重更新后性能差距消失。
英伟达因人类层面原因无法从Anthropic挖到强化学习人才，技术之外另有隐情。
谷歌员工称内部投诉渠道无效，只有公开曝光和员工关注才迫使公司修改条款。
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。
讨论AI用于学术评审的争议，建议亲自用GPT-6 Pro等试评论文验证效果
AI在本轮选举中影响有限，但未来每场选举作用将更大。
用户晒图吐槽Claude等大模型出现某种异常行为，引发讨论。
Ivan Zhao 发文呼吁用户掌控自己的上下文数据。
Baseten模型路由团队庆祝成果，强调模型路由将推动多模型未来。
作者吐槽AutomationBench的schema问题，讨论安全过滤器和监控器在追踪审查中的痛点。
推文预测AI行业并购与IPO走向：Scale AI被收购、Dagster被收购、Anthropic将IPO、OpenAI也将IPO。
有人群发邮件误称Taylor制作反AI视频，还想付费让其做反AI内容，营销系统翻车。
文章批评民调行业本质是人为制造的产业，缺乏真实民意基础。
作者感谢荷兰读者和出版商，其著作《共存》荷兰语译本成为荷兰非虚构类图书销量第一。
数据更新于 18:17 · 共 2332 篇