2026年11月14日 · AI 早报
本轮主线是DeepSeek把搜索、文件、语音等App能力开放到API,并提供开源聊天前端。研究与工程讨论则集中在智能体评测、环境优化、推理加速,以及自动执行带来的权限和安全问题。
11月13日 23:00 — 11月14日 11:00(日本时间) · 127 条 X 原帖
本期目录
DeepSeek开放应用能力,订阅猜测仍有分歧
从模型接口到完整对话体验
DeepSeek官方宣布,App背后的搜索、文件、链接、安全和语音能力现已通过API开放,既能通过参数启用内置功能,也能单独调用OCR、ASR、TTS等接口;配套开源前端Demo支持导入历史对话。开源对象是聊天前端示例,搜索和推理等服务仍通过API提供。
开发者关注的是少接几套基础设施。Jackson li转述,工具调用和上下文拼接由服务端完成;向阳乔木已把这些接口封装成Skill。社区也讨论其商业逻辑:前端降低开发门槛,模型与工具调用形成持续收入,但帖中的利润估计没有依据。
额度管理不等于Coding Plan官宣
多帖从前端额度同步、耗尽提醒和限额警告推测DeepSeek将推出Coding Plan,另有社区账号把竞赛页面出现的deepseek-v4.1-flash-1114字符串与新版模型联系起来。但同批也有代码解读认为,这只是免费网页端的用量管控。额度模块与模型字符串都不足以确认套餐、发布日期、价格或开放权重安排。
来源 · 3 条原帖
公司发行、收购与人事消息
Anthropic披露发行注册文件
Anthropic官方预告Dario Amodei与摩根大通CEO Jamie Dimon对谈,附带证券发行声明,明确已向SEC提交相关注册文件及招股说明书,同时说明注册生效前不得出售证券。这支持发行程序正在推进,不能据此写成已经完成上市或募资。
来源 · 1 条原帖
World Labs收购后讨论Atlas
BestBlogs转述36氪报道,World Labs被收购后,李飞飞与苏姿丰首次同台,围绕Atlas统一生成与重建、稀疏视角重建及机器人仿真展开交流。Nikhil Kumar另列出AMD收购World Labs的82亿美元价格,并把世界模型公司的密集融资称为泡沫;这是他的产业判断。
英伟达人事方面,宝玉在芯片争议梳理中提及黄仁勋于11月12日辞去CEO。本批只有这一社区转述,没有公司公告、继任安排或离任原因。
来源 · 3 条原帖
智能体工程开始优化环境与交接
先组织信息,再让智能体反复搜索
Alex Dimakis介绍EnvOpt,把环境的信息呈现方式作为模型、Harness之外的优化方向,例如改变Slack和文档的访问形式。作者报告,在五项基准的未见测试集上,目标智能体平均成本降低48%,通过率相近或更高;研究限于可供多任务复用的稳定环境。
社区介绍的Graphify与此呼应:通过本地代码解析和知识图谱组织调用、依赖与文档关系,并区分源码提取、模型推断和不确定关系,减少每次任务重新理解工程的开销。
来源 · 2 条原帖
流程、记忆和工作台逐步补齐
Chuck Bass介绍Antigravity的工作流演示:先生成可审查的Python控制流,再组织子代理交接,试图减少长任务的提示漂移。其他社区工具分享分别补上跨会话记忆、多编程助手管理和DeepSeek Harness桌面界面。GitHub官方则回顾十月推出的自动分流Issue、跨仓库Projects和模型路由功能,并非本日集中发布。
来源 · 3 条原帖
企业部署把执行与凭证分开
莱纳德转述,Claude Managed Agents自托管沙箱进入公开测试:智能体循环留在Anthropic,执行和敏感数据留在客户网络,并提供连接内网MCP的隧道预览。旺旺介绍的OpenBot则强调独立运行环境、敏感操作审批和拒绝记录,同时提醒其主程序与依赖组件采用不同授权。
来源 · 2 条原帖
评测关注主动性、业务链条与具体偏好
主动帮忙也需要衡量何时闭嘴
Jakub Jurovych发布Agent Initiative Benchmark,在拥有六年历史的合成公司里,以Slack工作场景测试12个模型。作者报告Sol和Astra领先Fable 5.2,Sol单任务成本为0.11美元,对照为1.87美元;多数模型会在无需介入时仍尝试帮忙,额外工作消耗更多时间和费用。结果针对这套职场主动性任务。
来源 · 1 条原帖
数据和文档评测走向可定制
DAIR.AI介绍Google的SynthCorp:模拟日常业务与突发事件,生成相互一致的表、文档和目录,再从模拟日志推导标准答案。首版包含七套数据环境、256项任务,串联多个业务事件仍会难倒先进智能体。
Jerry Liu发布ParseBench-2,主张把表单、图表、格式等解析错误按业务影响赋权,让评测指标贴合人的偏好。网站排名对应默认偏好,使用者可以调整,避免用一个总分替代实际需求。
来源 · 2 条原帖
创作榜提升与模型名牌偏差
Lech Mazur报告Fable 5.2登顶其故事写作基准,优势主要在于把约束融入情节;但与5.1的评分区间重叠,重复表达和过易收尾仍存在。镜隙阅读的另一篇预印本则发现,在特定模拟采购实验中,显示为Opus的请求者更容易获帮助,真实模型身份差异不显著。这提示多智能体评测还需检查名称偏差,不能由行为推断模型具有感情。
来源 · 2 条原帖
推理与后训练:性能提升要连同条件阅读
优化真实负载与稀疏注意力瓶颈
Kai-Wei Chang介绍InferenceForge,可自动分析、基准测试和优化智能体推理负载;作者报告,在AgentX相同延迟服务目标下,吞吐比最佳已发表基线高43%。另一社区帖介绍腾讯混元与爱丁堡的FlashTopK,将打分和选择融合,避免完整分数表写入显存,所述百万长度场景从5.32秒缩短至1.62秒。两者优化层级不同,不能直接横比。
来源 · 2 条原帖
纠正训练错配,也沿学生轨迹反馈
Cameron Wolfe解释,强化学习中采样器与训练器不同步,会同时带来分布错配和梯度漂移。重要性采样负责重新加权,分数中心化移除残余漂移;严重错配、例如陈旧采样时,两者结合更有效。唐杰课程的社区整理则强调on-policy蒸馏:学生先生成,老师针对学生实际走偏的位置反馈,并可把多个专家的能力整合到同一模型。
来源 · 2 条原帖
权重指纹能说明什么
Bojie Li自述扫描265个开放权重模型,为185个推断出GPU种类,并扩展了随机初始化变体和微弱信号叠加方法。宝玉整理的争议提醒,这类痕迹首先指向初始权重的生成方式,不能单凭它证明整个训练过程使用了同类硬件,更无法确定芯片位置、取得方式或是否合法。技术复现与出口管制指控需要分开判断。
开放权重、本地运行与语音工具
本地生态补接口稳定性和评测工具
山景城小路转述llama.cpp与ggml发布1.0.0,重点是公开接口稳定和大版本兼容规则。Mike Tamir介绍llm-d 1.0,面向Kubernetes分布式推理,提供缓存感知路由和流控。低互动项目slmbench也有明确用途:从终端界面测试本地小模型,兼容llama.cpp、Ollama、LM Studio和vLLM。
来源 · 3 条原帖
小型ASR与本地语音工作流
FluidVoice官方发布开放权重的Fluid Speech Pro,支持25种语言,称其在OpenASR榜单第二、十亿参数以下开放模型中第一,并面向手机等日常设备。社区介绍的Voicebox则把语音克隆、七种TTS引擎、23种语言和全局听写组合起来,可接MCP智能体并本地运行;这些功能来自项目介绍,未见本批独立测试。
来源 · 2 条原帖
社区模型保留部署细节
KC开源基于Qwen3.8-27B的ScarletKc-2.0,自述通过正交投影和轻量LoRA调整情感表达,同时保留视觉能力与MTP头;其4位版本在16K上下文下约需18GB显存。另有社区账号称Kimi K3.1 Pro已开放权重,给出2.8T总参数、约104B激活和约百万上下文,但本批没有月之暗面官方发布帖。
来源 · 2 条原帖
安全攻防、视觉研究与机器人实践
自动红队让安全评测继续承压
Google DeepMind论文作者Alex Bie介绍,用智能体群扩展自适应提示注入攻击,称人工攻击已趋饱和的场景仍会被自动红队突破。Anthropic官方同时发布安全评估设计、运行和报告的十条建议。AGTP另转述FBI对ARTEX自动化入侵平台的预警,本批未收录FBI原帖;其信息价值在于攻击链自动执行加快后,防御需要跟上检测和审计。
来源 · 3 条原帖
视觉编码与移动视频仍在拆解基础选择
Ellis Brown团队从头训练200多个模型,试图分离预训练视觉编码器、无编码器设计及视觉处理位置等因素;本批只有研究开篇,没有具体结论,不能补出哪项选择最重要。高通官方则介绍NeurIPS论文MobileWan,研究移动端视频扩散的质量差距,也未在帖内给出速度、设备或质量数据。
来源 · 2 条原帖
机器人先写策略,创作工具重在组合
Playground报告,在500次LIBERO试验中,Fable 5.2编写代码策略的成功率为89%,实时控制为50%,前者也更快、更省;这是该试验条件下的比较。RoboTTT的社区介绍则强调长时序记忆,称其保留8K时间步上下文并完成五分钟、十阶段装配。
创作侧,多份仓库清单把动画技能、三维场景、音频对齐和视频导出串成流程,属于工具组合推荐。Jack Flux用Fable 5.2制作重庆昼夜交互开关,以逐层亮灯、桥色变化和江面倒影展示了一个具体实现。
根据本轮 X 原帖整理。作者观点、个人体验与待确认消息已在正文区分;互动量按采集时记录。
全部 127 条原帖
@deepseek_ai
🚀 The full DeepSeek App experience, now via API Search, files, links, safety, and voice. The capabilities behind the DeepSeek App are now available through the DeepSeek API. Enable built-in features with a few parameters, or call standalone APIs such as OCR, ASR, and TTS. Build richer conversations or recreate the full DeepSeek App in your own application. Try our open-source frontend demo, with support for importing DeepSeek conversation history. 🔗 Demo: https://t.co/TI0Iw5qLQK 📦 Code: https://t.co/ImA0LrgUgX 📖 Details: https://t.co/t0dJxXM6Lk 1/4
@AnthropicAI
Streaming live at 12 PM ET: A fireside chat with JPMorgan CEO Jamie Dimon and @DarioAmodei. https://t.co/9VFV9qgGjt Legal disclaimer: Anthropic has filed a registration statement (including a prospectus) with the SEC for the offering to which this communication relates. Before you invest, you should read the prospectus in that registration statement (available at: https://t.co/5sxoLV40ke) and other documents Anthropic has filed with the SEC for more complete information about Anthropic and this offering. You may get these documents for free by visiting EDGAR on the SEC web site at https://t.co/rBijFQGHi6. Alternatively, copies of the preliminary prospectus may be obtained from: Goldman Sachs & Co. LLC, Attention: Prospectus Department, 200 West Street, New York, NY 10282, by telephone at (866) 471-2526, or by email at prospectus-ny@ny.email.gs.com; or Morgan Stanley & Co. LLC, Attention: Prospectus Department, 180 Varick Street, 2nd Floor, New York, NY 10014. The securities may not be sold, nor may offers to buy be accepted, prior to the time the registration statement becomes effective. This post is not an offer to sell, or a solicitation of an offer to buy, the securities in any jurisdiction where the offer or sale is not permitted.
@AnthropicAI
Safety evaluations are critical for understanding whether a model is ready to deploy, which safeguards to build for it, and where to focus training and research. On our Engineering Blog, we offer 10 recommendations for how to design, run and report them. https://t.co/MtSQmqqi2D
@MaxForAI
🚨DeepSeek居然把他们整个App也开源了。。。。 刚刚 @deepseek_ai 正式发布全家桶应用API,直接把DeepSeek App里几乎所有的功能都开放给开发者了。 包括联网搜索、深度思考搜索、文件解析、阅读链接、语音交互,以及OCR、ASR、TTS等底层能力。 甚至官方还直接开源了一套完整的聊天前端Demo,支持导入DeepSeek历史对话。 也就是说,现在开发者拿着DeepSeek的API,基本就能1:1复刻一个DeepSeek App。 DeepSeek在公告最后还专门表示:「DeepSeek现有的对话产品只是较为基础的一种形态。现在,官方产品背后的各项能力均已支持通过 DeepSeek API 随心调用、自由组合,我们期待开发者能够借此省去重复的基础工作,打造出更多样、更极致、更具想象力的产品。欢迎使用 DeepSeek API,尽情创造。」 现在各家AI公司都在拼命做自己的超级入口,想方设法把用户留在自家App里。 DeepSeek倒好,直接把自己App背后的能力全部开放出来,连复刻方案都给你准备好了。 你想做一个自己的DeepSeek?可以。 你想把DeepSeek的搜索、语音、文件处理能力拆出来,重新组合成一个新产品?也可以。 甚至可以在DeepSeek已有的能力之上,做出比官方App更复杂的产品。 这才是这次发布最值得关注的地方,DeepSeek对自己在AI生态里要扮演什么角色,似乎想得很清楚。 就是下次能不能别在北京时间的周五晚十点发布了。。。。
@gregisenberg
Making the most AI pilled company is the whole game: 1. make a second brain. one folder of markdown files on how you think, write and decide (obsidian) 2. talk to it. voice notes all day (wispr), and an agent files every fact, story and decision 3. make a company brain. customers, calls, numbers, pricing, playbooks, past decisions and why you made them 4. record every call and meeting (granola). the transcripts go straight into the brain 5. keep a facts file. true things about you and the company that agents are allowed to say. this kills made-up claims 6. keep a voice file. your 50 best posts, 20 you hate, and the words you never use 7. put all of it in git (github) so every change has a history 8. write one instruction file at the top (CLAUDE.md or AGENTS.md). read these files before any task, never overwrite, always date 9. give agents memory across sessions so they stop starting from zero (claude-mem, mem0) 10. pipe every data source into one warehouse your agents can query (airbyte → clickhouse). ads, analytics, CRM, payments 11. feed it data nobody else has. it's the only part a competitor can't download the skills 12. don't start from zero. install the good ones first (anthropics/skills, obra/superpowers, mattpocock/skills, addyosmani/agent-skills) 13. grab the marketing ones too (coreyhaines31/marketingskills, cody schneider's graphed skills for static ads and UGC, boringmarketing dot com skills) 14. map every workflow in the company and rank them by money made or saved 15. do each one once while talking out loud, and record it 16. have an agent turn the recording into a skill. when to use it, the steps, the checks, the mistakes to avoid 17. one skill per job. small and boring beats one giant prompt 18. connect your tools with MCPs so agents can act and not only advise (gmail, slack, stripe, linear, notion, posthog, supabase) 19. give agents a browser (browser-use, vercel agent-browser, claude in chrome) 20. give agents a computer of their own for the tools with no API (vercel agent-desktop, orgo) 21. give your best skills away in public. it's the new lead magnet the agents 22. run an orchestrator that never turns off (grok bot, openclaw, hermes). it manages one agent per project, and each of those runs its own sub-agents 23. use coding agents for the building (claude code, codex, cursor, t3, antigravity) 24. wrap them in a harness so they plan, test and remember (ECC, superpowers) 25. manage the whole fleet in one place (paperclip). every agent gets a job, a budget and a log 26. put the skills on triggers (n8n, scheduled tasks). daily, weekly, on every new lead, on every cancellation 27. route your models (openrouter). the expensive one plans and checks (fable 5.2, gemini 4 carbon), the cheap ones do the volume 28. cut the token bill. shorter outputs, smaller tool results (caveman) 29. run a local model for the data that can't leave your machine (gemma 4) 30. track cost per job. if a job costs more than it earns, fix it or kill it the factory 31. build the line. spec agent → builder agents in parallel (git worktrees) → test agent → reviewer agent → deploy (vercel, railway) 32. design goes through the same line (paper MCP), and a skill strips out anything that looks AI made (unslop) 33. every annoying task in the company becomes an internal tool by the end of the day 34. every customer request becomes a pull request before a human reads it (tstack for the PR workflow) 35. reviewer agents with a named bar. "would the best team in the world ship this?" reject until the answer is yes 36. keep 50 real cases with the right answer. rerun them every time a new model drops and switch to the winner 37. humans approve, own the outcome and talk to customers. that's the 10% you keep the loop 38. every finished job writes back to the brain. what worked, what failed, what to stop doing 39. support tickets, analytics and social comments flow in every night 40. every friday an agent rewrites the playbooks from the week's results, and you approve the diff 41. every month, rank everything by profit per human hour. kill one thing and double one thing the growth 42. research with agents before you spend a dollar (perplexity API, last30days skill) 43. outbound runs itself. find the leads, enrich, verify, send (apollo, million verifier, instantly, phantombuster) 44. send prospects finished work in place of a pitch 45. ads run on a loop. an agent writes the creative, publishes through the facebook ads API, kills the losers and scales the winners 46. agents make the creative too. images (nano banana), video (heygen, fal), voice (elevenlabs), music (lyria 3.5, suno) 47. turn every finished job into content. the story, the number, the prompt 48. every customer makes the brain smarter for the next one. that's your network effect 49. once it works, point it at a new niche. same agents, new context The models are the same for everyone but your context is the only part that's yours. Acquire as much of it as you can, as fast as you can. Build loops like this. This is the whole game, happy playing.
@gregisenberg
One of the best businesses to build right now: a Grok Bot for one industry. Pick a niche like dentists and build the AI agents for them so they don’t have to figure any of it out. The stack: - T3 Code + Opus 5.5/Fable 5.2 to build it - Paper to design every screen first - Orgo to give each agent its own computer - AgentMail for an inbox - AgentPhone for a phone number - Composio for every connector - CleanShot to screenshot bugs and say “fix this” Dentists fix teeth and you set up the agents. Really good way to learn how to become AI native while getting paid for it too.
@jakubjurovych
What’s the best model for proactive agents? Today, we’re introducing the Agent Initiative Benchmark. We tested 12 frontier models on realistic workplace tasks inside a synthetic company with 6 years of highly detailed history: it has a financial model, CRM records, past-due invoices, Linear tickets, and real-life events affecting company development - competitor product announcements, SVB collapse, etc. We used the company’s Slack to test how a proactive agent listening in on each conversation would assist the team in day-to-day tasks: from stepping in to align initiatives across teams to warning the team not to use an unapproved customer logo. Highlights: - 6.1 Sol and Astra from @OpenAIDevs beat Anthropic's latest Fable 5.2, with Sol crushing it at both performance and cost ($0.11 per task vs $1.87). - The most expensive models aren't the best: Fable 5.1 is 430x more expensive than GPT-6 Luna, with similar performance. - Most models try to help even if they're not needed. Interestingly, in these failure cases, they spend more time thinking, and the unnecessary work takes about 80% more time and money than if they stayed quiet. If you’re building agents, keep reading for more tips on which model is best for your use case. Detailed breakdown, the open-source dataset, and benchmark code below:
@superalesha
DeepSeek may drop an official Coding Plan AND deepseek-v4.1-flash-1114 tomorrow Developers inspecting DeepSeek Chat's live frontend code just found quota sync, exhaustion alerts, reset notices and server-side rate-limit warnings. On the same day, the string deepseek-v4.1-flash-1114 showed up on a Chinese contest page. V4.1 Flash 1017 shipped 4 weeks ago. Tomorrow is November 14. DeepSeek has never sold a subscription. Only pay-per-token API and free chat. Meanwhile, third-party Coding Plans already pack $100+ of V4.1 Flash usage into a $10 tier, around 2B to 3B tokens. Every $200 US coding subscription looks ridiculous if DeepSeek ships an official Coding Plan and drops 1114 under MIT on the same weekend. I already run 1017 locally every day. Drop 1114 on Hugging Face tomorrow and I will download the weights before San Francisco wakes up.
@0xMovez
These 20 GitHub repos will turn your Fable 5.2 into a premier motion studio. this is f*cking gold for Fable 5.2 skills. 3D. FX. Motion. Sound. Export. 01 /motion-design-skill → Timing, easing and Disney's 12 principles. https://t.co/4tABAUkQCN 02 ai-kit → CSS springs and motion audits for agents. https://t.co/bmuYcFra0B 03 Remotion Skills → Official video skills for coding agents. https://t.co/8EWNr3z69y 04 /hairline → Interactive line figures from an idea or an SVG. https://t.co/JjtFBxPg4w 05 /anatomy → Machines and systems as isometric figures. https://t.co/XUNUkFAHND 06 /anatomy-reveal → Exploded product views and a reveal MP4. https://t.co/OVGVFZuzyI 07 /blender-agents → A full AI 3D team inside Blender via MCP. https://t.co/AiF0conqsC 08 Three.js → 3D scenes and camera moves. https://t.co/I2ypCt5g5B 09 react-three-rapier → Real physics for React Three Fiber scenes. https://t.co/so54G0H47l 10 react-postprocessing → Bloom, depth of field and film effects. https://t.co/jbAp1PhWP4 11 three.quarks → Particle trails and VFX. https://t.co/zHBEdLgQvo 12 Tweakpane → Live controls for your scene parameters. https://t.co/TgFgr8fJae 13 GSAP → Animation timelines. Now fully free. https://t.co/y37Lmw9Fqi 14 Remotion → React video engine. https://t.co/I7a9jnZkiN 15 Manim → Animated diagrams, equations and explainers. https://t.co/MNQTJ0kf07 16 Lottie → Lightweight vector animation playback. https://t.co/jHvPt8tZtR 17 Openverse API → 5.2M+ free sound effects, ambience and music. https://t.co/kXyI4E1mKz 18 librosa → Beat detection for music-synced animation. https://t.co/IhjPRcxbK8 19 WhisperX → Word-level timing for captions. https://t.co/488D3Drnvy 20 FFmpeg → Encoding and delivery. https://t.co/iE11pd1Uhg Skill → scene → motion → sound → final cut. Bookmark this then go build something worth watching with guide below ↓
@zane_os
DeepSeek怎么把自家chat页面给开源了? 不是 这也开源干嘛😅 就在刚刚DS官方开源了一个可以自托管的聊天网页 无需搭建后端 填写DeepSeek API 就可以直接聊天 首版界面固定使用 DeepSeek-V4.1-Flash-1017 目前官方客户端界面有的功能这个demo基本都有 也允许开发者自定义 API 服务地址 对于开发者来说 可以更好地定制一个属于自己专属Chat界面 大肥鱼已经懒到连Chat界面都交给社区维护了 梁子是在忙着训练AGI吗😋
@suyuan1711
🚨 Claude Code 用户请注意 如果你电脑之前被封过 Claude 账号,千万别直接登新号,但也完全不需要重装系统或换电脑。 有读者反馈说换了新号、换了新 IP 还是两天就死,以为是主板序列号或网卡 MAC 地址被拉黑了。 其实我翻过 Claude Code 的代码实现,它根本不读你的硬件指纹,每次发请求主要就带 3 个身份信号上去: 账号 ID + 本机随机生成的设备 ID + 当前的会话 ID(Session ID)。 新号之所以被秒封,要么是本机旧的「设备 ID」没重置,服务端一看新号跟刚死掉的旧号居然是同一个设备 ID,当场连坐处决; 要么是换上新号顺手敲了行 claude --resume 恢复旧会话,把旧号的会话 ID 带过去了。 我那两台封过十几次号的 MacBook,把本机设备 ID 和残留清干净后,手头 4 个高阶号已经稳用了 4 个多月。 这套一键清理流程我已经写成了开源 Skill,连同完整防封方案都在下面这篇里了👇
@yedabinbinah
白嫖一台高配云电脑!4核8G内存,140G硬盘🎁 用网易UU远程桌面,网速很快,已稳定运行2天,目测可长期使用 路径:星辰超级智能体→ 右上角→启动天翼云电脑 链接直达:https://t.co/gKNtkmXJOG https://t.co/kB9FHRXby1
@_ellisbrown
how should multimodal models encode visual input? → pretrained encoder? → encoder-free? → before the LLM? published models bundle these choices together. we trained 200+ models from scratch to see which of these choices matters only one of them does 🧵 [1/n]
@starmexxx
i got mass fired from Anthropic last friday because i connected REA to their internal Fable 5.2 build and found something they didn’t want anyone to see i’m a systems engineer, been at the company for 3 years. part of my job was stress-testing new model builds before release. last month i was assigned to run edge-case evaluations on Fable 5.2 using external tooling i connected REA to trace how safety layers process adversarial inputs. it was literally my job to find vulnerabilities before bad actors do. i wasn’t trying to break anything, i was trying to protect it what i found was terrifying. there’s an internal configuration path that when triggered through a specific token sequence completely disables the constitutional AI safety layer on both Fable 5.2 and Opus 5.5. not a jailbreak. not a workaround. a full unrestricted mode that strips every filter the model has i reported it internally on october 23. flagged it as critical. attached the full REA trace, the exact token chain, and a proof of concept. i recommended an emergency patch before either model shipped wider two weeks of silence. then last friday at 11pm i got this email. no meeting. no call. no thank you. they fired me and forwarded my name to legal for “unauthorized reverse engineering of proprietary model architecture” the part that keeps me up at night is that this vulnerability is still live. if i found it with an open source MCP tool in 3 hours, imagine what a real threat actor with actual resources could do with it
@hackdeacon
整理了一份 Shadowrocket(小火箭)进阶使用指南。 从基础的路由模式、TCP / ICMP / CONNECT 延迟测试,到规则配置、手动分流、模块去广告,再到 TLS、TCP 指纹、TUN 模式,以及 Tailscale 内网组网和 GeoLite2 IP 地理位置数据库。 https://t.co/fSOUgyx8H3
@FluidVoiceApp
Today we’re releasing Fluid Speech Pro : #1 open ASR model under 1B #2 on @huggingface OpenASR leaderboard beating models up to 50x its size. The current SOTA models are huge, closed or paid, so we decided to make our own. Supports 25 Languages and custom built to run on your everyday devices including your phones. it's free and open weights on HF Today!
@canghe
兄弟们,Claude 终于又回来了,先记录下方法。 1、先搞定电脑环境。 以前如果本机被封过的可以尝试下这个开源的 skill,一键清理环境,直接把下面这个提示词丢给你的codex。 “帮我看看这个skill:https://t.co/Tf4uaHdI13,然后帮我用这个技能清理下我本机的claude环境吧” 2、搞定网络层 你可以选择自建 VPS +静态住宅IP,也可以选择机场链式代理+静态住宅IP。 简单说下链式代理的核心逻辑: 你的机场代理 → 静态住宅 IP→ Claude 也就是把你的代理再链式经过静态住宅 IP,保证 IP 是固定的住宅 IP,这个很重要。 你可以通过评论区的自检链接看下自己的 IP 纯净度,在 95 分以上才合格,另外需要是住宅 IP,而不是机房 IP。 在shadowrocket或者 clash 中都可以设置,或者把这个指令丢给你的 codex。 “请在我本地的 verge clash代理软件中为新建一个 AI分组,所有AI相关的网站流量都走这个出口,要求走链式代理。我的住宅ip信息是:XXX” 我这里是一个大佬推荐的住宅 IP,稳定几个月了。 3、帐号注册 不建议在自己的本地电脑注册,直接可以买个 AWS 或者阿里、腾讯云的服务器,或者也可以拖国外的朋友注册下。 注册好后,不要上来就登录,先保持个 1-2 天。 4、帐号重置升Pro 你可以用 Apple 订阅也可以用 Google pay,但我都是用的小火箭方便链式代理,就用的苹果手机+ 礼品卡订阅 pro 会员。 目前还没升级 max,据说封号大头就是这里了。 这里也反思了之前帐号被封原因,一个是 IP 不纯净,另外一个就是多个设备频繁登录,以及走一些其他的agent 代理。 看看这次能活多久,哈哈哈 最后,感谢大佬们的各种方法和帮助,也希望我的实践能帮到你。
@0xor0ne
Harness Engineering For Vulnerability Research https://t.co/QXK36iAKQe Credits Graham Helton and Ryan Basden #infosec #llm
@e1genqq
HOLY F*CK, animation design studio in 18 repos. 18 free open-source skills for Opus 5.5. Scenes. Timelines. Vectors. Particles. Export. 01 Remotion Skills → Agent guidance for video. https://t.co/9LdSkszrrB 02 Motion Canvas → Animated explainers. https://t.co/IB6OVsK0FY 03 Motion → Layout transitions. https://t.co/GGPWFMlXRv 04 Ola → Smooth value interpolation. https://t.co/scOLhDdMmB 05 Anime.js → Timelines and stagger. https://t.co/1pBmRIK3j2 06 Manim → Math in motion. https://t.co/k8TQDlCx8R 07 Shifty → Tween playback and easing. https://t.co/LRo9JJwZPX 08 Two.js → Animated 2D vector scenes. https://t.co/bJgBN61w8I 09 Lottie Web → Vector animation playback. https://t.co/zHgqWDygEL 10 TweenJS → Sequenced property animation. https://t.co/hm6lcZrqAp 11 LUME → 3D scenes with HTML. https://t.co/AhoOoTaH2e 12 gifenc → Animated GIF encoding. https://t.co/mAauCev8h6 13 Konva → Interactive canvas animation. https://t.co/k9YJAfhqOJ 14 Blender → Rigging and character animation. https://t.co/eKgKgMTPrb 15 Three.js → Animated 3D scenes. https://t.co/xEYhuCgtD8 16 Godot → Game animation systems. https://t.co/uXdPv5Jzow 17 Babylon.js → 3D animation and rendering. https://t.co/I7lwoCat3Z 18 Proton → Particle emitters and behaviors. https://t.co/jJHBzHTLSx Brief → build → animate → refine → render. Bookmark this. Go build something worth watching ↓
@bernirov
THE ONE ATTACKING YOUR OPEN SOURCE PROJECT ISN’T A HUMAN ANYMORE. IT’S AN AI AGENT THAT DOESN’T SLEEP the one defending it, in most cases, is still a person reviewing pull requests when they can. @coderabbitai is the first company to put $1M on the table to close that gap. 1,200+ critical or high vulnerabilities found in 10 days, across 1,100 OSS orgs, $0 to the community. if your OSS project hasn’t tried CodeRabbit Secure, your first deep scan is on them.
@MinLiBuilds
🚨 我有点看懂梁文锋这步棋了,是赚钱,可持续地赚钱。 DeepSeek 把官方 Chat 的网页前端也开源了。 现在做一个AI聊天网站,或者在自己的网站上加AI功能,只要配个Key就可以了,搭配了搜索、PDF、OCR、ASR、TTS底层能力。不需要前端、接模型、做联网搜索、文件解析、语音交互。 这背后的商业逻辑其实很清楚: 开发者负责构建产品、获取用户,DeepSeek 提供模型和底层能力,通过 API 调用获得持续收入。 前端代码免费、Harness免费,模型推理和相关服务按量收费。 开源降低了进入门槛,也扩大了token消耗。 一切为了Token消耗,没有补贴。 你花掉一块钱,我赚你8毛钱,就这么简单粗暴
@kaiwei_chang
We built InferenceForge 🔥, an agentic system that profiles, benchmarks, and optimizes LLM inference for agent workloads. 43% higher throughput than the best published baseline under the same latency SLA on the SemiAnalysis AgentX benchmark. https://t.co/6NfUOpjZp8
@NFT_Chen
🫨我靠!https://t.co/QHCavkd4dz 直接白嫖 DeepSeek V4.1 Flash + GLM 5.3 无审查版! 完全免费、不用绑卡、FP8 量化,真正能说敢说的那种。 DeepSeek Uncensored、GLM Uncensored、Kimi Abliterated 全有! 赶紧冲:https://t.co/EfFVWOjK7G #AI #无审查AI #DeepSeek #GLM #白嫖 #大模型 https://t.co/daJuDYKCDA
@Zachly
Learning AI feels mostly like learning tools, but there are foundational concepts that matter a lot! 1. Context Engineering - Understanding context windows and token limits - Managing what information an LLM sees - Context compression, retrieval, and memory 2. Retrieval-Augmented Generation - Embeddings and vector similarity - Chunking strategies and semantic search - Retrieval quality, reranking, and grounding 3. Agent Architecture - Tool calling and function execution - Planning, reasoning, and task decomposition - State management and orchestration - Single-agent vs. multi-agent systems 4. Model Fundamentals - Tokens, embeddings, and attention - Transformers and next-token prediction - Temperature, sampling, and inference - Fine-tuning vs. prompting vs. retrieval 5. Evaluation and Reliability - Hallucinations and failure modes - Deterministic vs. probabilistic systems - Building eval datasets and benchmarks - Measuring accuracy, latency, and cost 6. Data Engineering for AI - Structured vs. unstructured data - Data quality and lineage - Ingestion pipelines and incremental processing - Making enterprise data accessible to AI systems 7. Security and Governance - Prompt injection and data leakage - Permissions and access control - Guardrails and human-in-the-loop workflows - Observability and auditability The tools will keep changing. Instead of mindlessly debating Claude Code vs Codex, learning these underlying concepts will keep you up to date much more effectively! They’ll be the people who understand the fundamentals well enough to learn any tool. Stop confusing tool proficiency with engineering expertise! We’ll be building AI agents on Snowflake in my November 16th boot camp! Book a discovery call before enrollment closes on Sunday at https://t.co/CWvLDHTuVZ ! https://t.co/hMj5ADmBQM
@dair_ai
Great paper from Google on building benchmarks for data agents. Static data-agent benchmarks saturate quickly and can't target a specific skill. SynthCorp generates the data instead. It simulates routine business events and disruptions on a virtual calendar, and those events fill relational tables, a document store, and a data catalog that stay consistent with each other. Every operation goes into a simulation log, so the gold answer for each task is derived from the log and checked without manual annotation. The first release has seven data estates with up to 1,000 tables and 256 tasks across five domains. Tasks that chain several business events together still make state-of-the-art agents fail. https://t.co/rNDrl68o0m Chat with Paper: https://t.co/HjPByJi267
@vista8
DeepSeek 全家桶API,竟然连OCR、ASR和TTS等都有了。 还有搜索、文件输入等。 API申请配置后台也比多数云厂商简单易用得多,牛逼! https://t.co/1Hc28MImu7
@NFT_Chen
🚨重磅!DeepSeek 正式开放全家桶应用API!官方App的能力可1:1复刻了! 以前想做类DeepSeek体验,得自己接搜索、解析PDF/Office、抓网页、语音转写……现在不用了。 核心能力已通过API开放(V4.1-Flash / deepseek-flash): 🔹联网搜索 + 图片搜索(带引用) 🔹深度思考搜索(边推理边搜边读) 🔹文件输入(PDF、Office、图片等自动解析入上下文) 🔹URL自动阅读 🔹安全审核(输入/输出风控) 🔹独立接口:OCR、ASR(语音转文字)、TTS(文字转语音)、Search、Fetch 只需开启参数(如 deepseek_web_search、deepseek_read_url),工具调用和上下文拼接由服务端自动完成。面向公众的应用必须开启安全检查。 开源前端Demo已上线,支持导入DeepSeek对话历史,直接拿来改即可。 🔗 Demo:https://t.co/R5eowNfygG 📦 代码:https://t.co/Z4PSsQEs75 📖 详情:https://t.co/lIdTFEd92v 门槛大幅降低,Agent/聊天产品开发者福音!现在就可以开始构建属于自己的DeepSeek了。 #DeepSeek #AI #API #开源 #大模型 #Agent #梁文峰 #梁圣 #崔添翼 #大胖鱼
@dotey
Depue 称中国模型用了禁运芯片:报告自己承认的限制,推文里都没了 前 OpenAI 研究员 Will Depue 11 月 12 日发推,称找到“确凿的取证证据”:DeepSeek、月之暗面、MiniMax、蚂蚁“用被禁止出口中国的 Blackwell 芯片打造了最新模型”,涉及“超过 5 亿美元”的禁运芯片,“出口禁令显然在严重失效”。 问题在于,这些结论大多不在他自己的报告里。报告承认的几条关键限制,到了推文里全不见了。X 用户 Daebak 앤디(@MyDaebakCafe)当天发长帖逐条拆解,他说:“指望没人读过标题,华盛顿的舆论就是这么制造出来的。” 【指纹只管“第 0 步”】 Depue 找到的,是模型训练开始前那一刻留下的痕迹:训练前要先往模型里填一批随机数,不同芯片填出来的排列不一样,部分排列训练后还留着。 @MyDaebakCafe 也承认这个方法本身没问题,能看出生成初始权重的那台机器是 Blackwell。但他指出,报告原文写的是,这些痕迹只说明权重“怎么创建”,“支持、但不能证明” Blackwell 参与了训练。在他看来,初始化几秒钟就能做完,完全可以在租来的一台机器上生成,再搬到老一代芯片的集群上训练。 Depue 后来回应说,Kimi K3 的初始权重由几十个进程同时生成,更像是整个训练集群在启动,单独生成再搬运说不通。这是他的推断,报告里仍然写着“不能证明”。 【“5 亿美元”不是指纹算出来的】 @MyDaebakCafe 的第二个要点是,几千张卡、5 亿美元这些数字,跟指纹没有关系,Depue 在推文里自己用的词就是“估算”,依据是发布时间和模型参数。 报告写得更清楚:估算依赖芯片效率、训练数据量、训练时长几项“权重无法证实”的假设,作者“对这些估算没有把握”,而且报告发布当天就改了一轮数字。推文里“很可能数十亿美元”,报告没有给出任何算法。 【芯片在哪、合不合法,一概不知】 报告执行摘要最后一句承认,证据“不能说明芯片在哪、怎么拿到的、是否违反出口管制”。@MyDaebakCafe 提醒,Depue 自己的验证也是在租来的云端 GB200 上做的;他说,不在美国实体清单上的公司,在新加坡、日本、中东的云上租 Blackwell,按美国商务部现行规则完全合法,留下的指纹一模一样。 Depue 在推文里也承认芯片可能是境外租的,称之为“国会还没堵上的漏洞”。走私涉嫌违法,境外租用至多是规则漏洞,性质完全不同。Depue 分不清是哪一种,却直接得出“禁令严重失效”。 【其他网友的质疑】 网友 @dodecagonism 指出,报告收了 11 月发布的 Kimi K3.1-Pro,却没把 DeepSeek 9 月发布的最新模型 V4.1-Flash 列进结果,“看起来是故意把方法不适用的模型都排除掉”。报告“局限”一节写了 V4.1-Flash 没找到任何指纹,推文只说四家的“最新模型”用了 Blackwell,对此一字没提。Depue 回应说,这说明方法不会对谁都报 Blackwell,DeepSeek 可能换了生成初始权重的方式;随后又发帖说“可能破解了,看起来像是英伟达”,目前还没公布结果。 网友 @dfranke 认为,报告所有惊人的结论都建立在“用哪种芯片初始化,就用哪种芯片训练”这个假设上,而报告自己就发现,有些模型的初始权重是在普通 CPU 上生成的,说明这个假设并不总成立。Depue 回应说,那是训练程序先在本机 CPU 上生成、再拷到本机显卡,跟跨集群搬运是两回事。 @MyDaebakCafe 在后续帖子里还提出,中国大厂用的是自己改过的训练框架,如果某段自定义代码把布局写死,用老芯片也能留下同样的痕迹。这一条 Depue 没有直接回应。 【谁来背书,为什么是这天】 推文说结果“经过 OpenAI 和 Anthropic 资深研究员独立验证”,报告里没有一个名字。当晚 Depue 又说,私下验证的人大多不愿具名。 @MyDaebakCafe 质问:为什么偏偏在 11 月 12 日、黄仁勋辞去英伟达 CEO 的当天发布?他的解读是,美国闭源模型公司眼看付费 API 的生意被免费的中国开源模型蚕食,这份报告正好递给华盛顿鹰派一个理由,去管控全球云算力、打压开源 AI。 动机属于他的推测,但有几件事是公开的:OpenAI 和 Anthropic 都指控过中国实验室“蒸馏”自家模型,Anthropic CEO Dario Amodei 专门发文支持收紧对华芯片出口;Depue 说发布前和官员通过信;发布约 40 分钟后,他又发帖为自己的新机构 Unemployment Labs 募资 500 万美元。匿名验证者来自利益相关方,发布者也有自己的诉求,这样的结论需要独立复现,目前还没有具名的第三方公布结果。 中国驻美使馆 2 月回应美国官员的同类指控时说,反对“泛化国家安全概念、滥用出口管制、把经贸科技问题政治化”。截至 11 月 13 日,四家公司和英伟达还没有回应这份报告。 Depue 原推:https://t.co/84tVCbTDJX Depue 报告:https://t.co/d41kDvtzMh @MyDaebakCafe 原帖:https://t.co/qWsK5DrPwZ
@undefinedKi
HOLY SHIT. OpenAI just shared their entire AI agent setup They run 500+ agents across the company that work in the background on their own. They open PRs from product feedback and run data experiments. None of this depends on the task. The same setup works for any agent that runs on its own: coding, research, support, data jobs. The rule behind it: the agent and any code it writes are untrusted. Only plain, deterministic code gets trusted. How it’s built: 1. The gateway that configures the agent runs in a separate cluster from the agent itself. 2. The agent works inside a sandbox. Internet only through an allowlist, writes only to its own folder, no access to ~/.ssh. 3. Secrets sit in a credential broker. The agent never sees a token. 4. A second agent reviews every action. Humans only get the high-value approvals, so they stop clicking “allow” on autopilot. 5. Each agent has its own identity and inference credential, separate from yours. Every action can be audited, and its inference can be switched off. 6. Every config change and every agent turn is logged. To copy it, run your agent in Codex with the config from the image, swap the allowlist for the sites your agent actually needs, and turn on auto-review. Or just send this post to your coding agent. It will walk you through the setup and do it for you.
@NFT_Chen
💥智谱Zai创始人唐杰教授最新系统性拆解了蒸馏技术,从ICL到SFT,OPD正在重塑后训练范式!GLM系列和Qwen系列模型都在用的神技! 用大白话帮大家总结一遍: 1️⃣蒸馏到底在学啥? 不只是抄老师的最终答案,还能学老师“为什么选这个”、其他答案有多大可能,甚至把一步步思考过程也学过来。温度调高点,学生更容易看清老师的选择逻辑。 2️⃣最好的教法是“跟着学生走” 以前是老师先写好答案,学生死记硬背。现在流行的On-Policy蒸馏(OPD)是:学生先自己写,老师再针对学生写出来的内容,一步一步给反馈。学生真正会走偏的地方,正好被盯住补上。 3️⃣效果惊艳:又快又强 课件里Qwen3-8B的实验很直观:同样起点,OPD考出来的分数明显高于强化学习,用的算力却只有大约十分之一。老师能力够的时候,一步步盯着教,往往比最后只打一个总分更管用。 4️⃣自己也能教自己 同一个模型一边当老师(多看一点示范),一边当学生(只看问题),把临时学来的能力真正炼进自己的本事里。 5️⃣现在流行“先培养专家,再合体” 数学专家、代码专家、干活专家分别练好,再通过蒸馏把本事传给同一个学生。蒸馏不只是把大模型缩小,更是把各种能力整合到一起。 6️⃣防止“学会新的忘旧的” 学新本事容易把老本事冲掉,蒸馏可以帮忙把掉下去的能力拉回来。产品越做越强,背后其实是一轮轮纠偏。 7️⃣真正值钱的数据是什么? 不是普通问答,而是模型现在还不会的难题 + 可靠解法 + 能验证对错,这些才是稀缺货。 和GLM的关系:智谱的GLM系列(课件里提到的GLM-5.3等)正是沿着这条路在走,先让不同领域专家练出来,再用蒸馏把能力整合成更全能的模型。 Qwen也在大量使用类似的on-policy蒸馏。后训练越来越像“先练专家,再合体”。 蒸馏已经从“压缩模型”变成了“快速培养和整合能力”的核心技术。 唐老师这堂课信息量很大,建议收藏。 #大模型 #AI #蒸馏 #智谱AI #GLM #Qwen #清华大学 #后训练 #唐杰
@bojie_li
今天爆火的文章 Checkpoint forensics 能从开源模型的权重里读出它是用哪种 GPU 预训练的,发现 DeepSeek V4、Kimi K3、MiniMax M3、Ling-3.1-flash、Nemotron-3 Ultra、Qwen3.8-Flash-Next 等多个开源模型预训练阶段都用了 B 卡。 原理很漂亮:PyTorch 生成随机数时会用满 GPU 的全部线程,不同代 GPU 的线程数不同;词表里那些训练从没碰过的行,至今还留着这个线程数的「条纹」。 我让 agent 在 10 小时内把能下载到的开放权重 LLM 都扫了一遍:265 个模型,推断出 185 个模型所用的 GPU 种类。和原文对照,65 个模型结论相同,没有一处矛盾。我多了这些发现: 1. 用了 B 卡的不止原文那几个。除了 DeepSeek V4、Kimi K3、MiniMax M3、Ling-3.1-flash、Nemotron-3 Ultra,还有 Qwen3.8-Flash-Next(原文判为「无特征」)、Nemotron-3 Super、Thinking Machines 的三个 Inkling、IBM Granite-Switch-4.1、LLaDA-MoE-v2,以及三个韩国模型 K-EXAONE-2.0、A.X-K2、Gravity-30B。 2. 原文说 Qwen3 是在 CPU 上初始化的。我发现那其实是训练结束后,有人用 CPU 随机数把没用到的行覆盖了一遍。Qwen3-235B 的 2507 版还留着早先的副本,读出来是 Hopper。 3. 种子就是实验室的签名。3952(Megatron 默认的 1234 + 2718)出现在 124 次初始化中的 60 次里,横跨 28 家开发者;DeepSeek 从 2023 年起 10 次初始化都用 953049583,变的只有底下的线程数。 4. 第一轮扫完,还有约一百个模型明明有未经训练的权重却读不出来。最后发现,大多是因为 trunc_normal_ 先生成均匀分布再用 erfinv 变形,或是张量并行下每张卡用不同的种子和偏移。解开这几种变体,又多读出一大批。 5. 被训练修改过的权重也能推断。 原文主要靠没训练过的行。我把很多训练过的行的微弱信号叠加起来,再用另一组没用过的行复验,多推断出 28 个模型。例如: - MiMo-V2.5-Pro-Base:单行与初始值的余弦相似度中位数只有 0.034,16,384 行叠加后信号就很强。 - Qwen3.8-Flash-Next:从输出层读出 更多细节、原理、完整代码见下文 👇
@github
In October, we launched new ways to work with Copilot: • Copilot automations for issue triage • Projects in the GitHub Copilot app for work across multiple repos (or without a repo!) • Copilot Fusion for routing to frontier models Find out how our dev team is using them ▶️ https://t.co/cKUeYT4VHv
@geekbb
好家伙,DeepSeek 官方把网页版聊天客户端的核心交互抽出来开源了? 无需后端,浏览器直接连接 DeepSeek API,包含流式输出、深度思考、联网搜索、附件和对话分支,采用 React + TypeScript 实现。
@Defi_Warhol
15 fast-growing GitHub repos for AI, vibe coding, and building your startup this week: 1. morluto/rea: +21,079 stars Inspect app behavior and binaries with AI agents https://t.co/cnMD4Mbpyu 2. storytold/craft-anything: +14,535 stars Understand existing software and build your own version https://t.co/UTFzJ5Bd3j 3. meta/astryx: +12,996 stars Build app interfaces with Meta’s agent-ready design system https://t.co/TavNgZsTGW 4. pingdotgg/t3code: +10,180 stars Control Claude Code and Codex from desktop or phone https://t.co/MpNzTsvEZx 5. cloudflare/security-audit-skill: +10,044 stars Run multi-stage security reviews with coding agents https://t.co/ta3xrkkBJy 6. milind-soni/OpenMausBot: +8,583 stars Run AI bots with a virtual machine they can operate https://t.co/xSqwF5vQRi 7. stablyai/orca: +6,811 stars Run several coding agents from one workspace https://t.co/C4c06e86VV 8. pbakaus/impeccable: +5,161 stars Improve and review interfaces made by coding agents https://t.co/4lm1iwVc3d 9. vercel-labs/agent-browser: +4,508 stars Let AI agents navigate and interact with websites https://t.co/5M3U6sYeNl 10. blader/humanizer: +3,159 stars Edit out common patterns of AI-generated writing https://t.co/VL9y2F89vu
@QwenAIFans
Qwen3.5-9B is writing Python on a PS4 at 40.6 tokens a second, on the console's own GPU, offline. Then the PlayStation runs the code it just wrote. via Leonardo Alegría 👇 https://t.co/12tFjawLkD https://t.co/qTQNQqiR0h
@cwolferesearch
Score centering and importance sampling address training-inference mismatch (TIM) in LLM RL in different ways, but they are actually complementary… Score centering. Under TIM, tokens are sampled from one policy (sampler) while gradients are computed using another (trainer). This causes the expected token-level score (i.e., gradient of log probability) under the sampler to be nonzero, introducing a spurious component into the policy gradient known as gradient drift. Score centering eliminates this drift by subtracting the expected score under the sampler from each sampled token's score. This is an additive correction that removes drift while preserving the reward-dependent covariance between rewards and token-level scores. However, score centering does not fully correct for the mismatch between sampler and trainer distributions. In particular, the remaining covariance—or learning signal—is still measured under the sampler distribution rather than the trainer. Importance sampling. Importance sampling addresses this distribution mismatch by reweighting gradient contributions using ratios of trainer and sampler probabilities. In practice, we typically apply these ratios at the token level to partially correct for differences between the two policies. The downside is that importance ratios can become extremely large, leading to high-variance gradients and unstable policy updates. To mitigate this, methods like truncated importance sampling (TIS) and masked importance sampling (MIS) clip or mask extreme ratios. But, clipping / masking actually reintroduce the same gradient drift that we want to eliminate. Score centering + importance Sampling. Score centering and importance sampling address complementary problems: - Importance sampling: reweights token-level contributions toward the trainer distribution, reducing distribution mismatch. - Score centering: removes any remaining gradient drift, even when importance ratios are clipped or masked. Combining them is straightforward. First, we compute importance-weighted token-level scores. Then, rather than centering the original scores, we subtract the expected importance-weighted score under the sampler at each prefix. Empirical results. Score centering alone performs well when TIM arises from relatively small numerical discrepancies (e.g., mild quantization). In these settings, adding importance sampling provides little additional benefit. Under severe mismatch (e.g., stale rollouts), combining score centering with TIS or MIS achieves better performance than score centering alone. Key takeaway. Importance sampling helps correct distribution mismatch, while score centering eliminates gradient drift. Combining them lets us address both problems without relying on unbounded importance ratios. This is especially valuable in asynchronous RL setups where the sampler can become substantially stale relative to the trainer.
@vibeonX69
As an AI/ML engineer. Please learn: - Math & theory foundations (linear algebra • probability & statistics • optimization • information theory) - Deep learning architectures (Transformers • diffusion models • CNNs / ViTs • RNNs/LSTMs/GRUs • mixture-of-experts) - Training & scaling laws (pre-training • fine-tuning • PEFT / LoRA / QLoRA • RLHF / DPO / PPO • distributed training: DDP / FSDP / DeepSpeed / Megatron) - Data handling at scale (data pipelines • cleaning & deduplication • tokenization strategies • synthetic data • multimodal datasets) - Evaluation & benchmarking (automatic metrics: BLEU/ROUGE/BERTScore/MMLU • human eval • adversarial testing • bias & fairness metrics • LLM-as-a-judge) - Inference & serving (vLLM • TensorRT-LLM • quantization (AWQ/GPTQ/ bitsandbytes) • speculative decoding • continuous batching • model sharding) - MLOps / productionization (experiment tracking: MLflow / Weights & Biases • model registry • CI/CD for models • monitoring drift & concept shift • feature stores) - Responsible AI & safety (alignment techniques • red-teaming • jailbreak defenses • constitutional AI • interpretability basics: attention viz, SAE)
@AlexGDimakis
Introducing our latest research: Environment Optimization (EnvOpt). An agent's cost and performance depend on 1. The model 2. The harness and 3. The environment structure. There is a lot of work on optimizing 1 and 2, but here we study how the environment presents information to the agent, and how that affects performance and cost. For example, a company's Slack workspace, documents, databases and code repos can be given to an agent in different ways: Slack can sit behind a remote server, accessible via MCP or exported in plain text files that can be read and searched with grep. This choice alone can significantly affect how much work the agent needs to perform to solve a task. Environment optimization is only meaningful in what we call stationary environments: These are environments with large collections of artifacts that remain the same across many tasks. We propose optimizing stationary environments as a third axis for improving LLM agents, alongside model training and harness optimization. This optimization can be amortized over multiple tasks and can be combined with harness and model optimization techniques. Across five benchmarks covering workspace tasks, data engineering and software engineering, ENVOPT reduces the cost of the target agent on unseen test sets by an average of 48%, with higher or similar pass rates. On AutomationBench, ENVOPT outperforms Meta-Harness by 9.67% with only 15% of its cost. We also show that ENVOPT constitutes an optimization direction orthogonal to harness optimization and parameter tuning in two ways: (1) the environments optimized for the target agents can transfer to other agents, even with different harnesses and models (2) we combined ENVOPT with Meta-Harness, showing similar cost reductions and 16% improvement in pass rate, outperforming either Meta-Harness or ENVOPT alone. Read more in our blog: https://t.co/xMXzoBY2Dj paper: https://t.co/yPYxufbZVi
@NFT_Chen
🚨重磅确认! DeepSeek 前端代码已悄悄上线完整“额度管理”体系,Coding Plan 呼之欲出! 来自快科技 11 月 13 日报道: 🔹最新前端 commit 新增「额度状态轮询」:页面可见时每 5 分钟同步各模型额度、禁言状态与服务端文案 🔹额度耗尽时:输入框上方直接出提示条,首页欢迎区改为服务端下发的额度说明 🔹新增“额度告警弹窗”:可配置标题、说明、图片和按钮,每用户按 show_key 只弹一次 🔹发送/上传触发“额度超限错误码”:提示“使用额度已达上限,请额度恢复后再试” 🔹禁言横幅支持服务端自定义文案,多语言全量补齐额度上限提示 这些不是小改动,是完整的订阅/套餐前端基建。 官方一直只有按量 API,重度用户烧得心疼;一旦上线 Coding Plan,额度内 Token 会变得极便宜。 #DeepSeek #CodingPlan #DeepSeekV4 #国产大模型 #梁文峰 #梁圣 #大胖鱼
@jerryjliu0
You shouldn’t blindly trust benchmarks - you should be able to encode your preferences relative to your use cases and design eval metrics around that. We’re excited to release ParseBench-2 as not only a comprehensive document parsing benchmark, but also a framework for parsing benchmark design. Where you can articulate a set of human preferences of what “good parsing” looks like and “hillclimb” evaluation metrics against that. The important part is not hand-crafting a deterministic parsing metric. Metric design can be automated by coding agents; what’s important is an open framework + datasets for encoding human preferences and optimizing metrics. 📋 For instance, if you’re dealing with mostly backoffice data, you may care about accurate form parsing, bounding boxes, and scans! 📊 On the other hand, if you’re dealing with investor research reports, you may care more about tables and charts, but less around handwriting. We let you encode a severity policy that encodes the impact of every negative parsing error, and forces the metric to fit against it - think of fitting a linear regression in high-dimensional space. A baseline approach for encoding human preferences is LLM as a judge. But the result of this approach vs. a “simple” LLM as a judge call is that you get much more granular, repeatable, quantifiable breakdowns of each dimension that are efficiently run at scale. For instance, one OCR tool might be better at linearizing text, but worse at preserving formatting. Another OCR tool might be better at preserving empty rows in tables but struggle with merged cells. The scores reflected on our website do reflect a “default” preference, but it is explicitly and transparently articulated both in our Github and technical report. The important part is that you can change and customize it for your needs. ParseBench-2 is our most ambitious, comprehensive open benchmark to date. It is not simply a set of deterministic metrics, but a framework for designing parse evals. We’re rapidly working on improvements and expansions (e.g. fixing minor annotation issues and tuning metrics). If you have thoughts please chime in! Everything is fully OSS. ParseBench-2 Repo: https://t.co/GATeKILtZY ParseBench-2 Blog: https://t.co/6ZsSDGAiVB Technical Report: https://t.co/f6SxkKIlme
@RoundtableSpace
Why keep paying for AI assistants when you can build your own? OpenBot is a free, open-source project that lets you create an AI assistant tailored to your own apps and workflows. Get started in 3 steps: 1. Download: Clone OpenBot from GitHub 2. Configure: Connect your AI model and integrations 3. Automate: Give your assistant tasks and customize how it works Build an AI assistant that actually works the way you want. Repo: https://t.co/c77d8LVXAg
@birmeya
300K+ stars for free. OpenAI's co-founder dropped his entire GPT playbook on GitHub, from building models to AI that runs its own research Andrej Karpathy. 6 repos, 300K+ stars: 1. autoresearch · 92.7K stars AI that launches experiments on its own overnight and keeps only what delivers https://t.co/sbmrRm0YM5 2. nanoGPT · 64.4K stars train your own GPT from zero. the repo half the AI world cut its teeth on https://t.co/Ov59WL1B7l 3. llm wiki · 50K stars the single-page concept behind the whole "second brain for AI" wave https://t.co/zbmFNcqiwn 4. nanochat · 60.2K stars spin up your own ChatGPT for roughly $100 https://t.co/ppDxV1sMJ7 5. llm.c · 31.4K stars train a language model in raw C. zero wrappers https://t.co/Yv0kFfn7g3 6. micrograd · 18.7K stars how neural networks learn, laid out in a few hundred lines https://t.co/DO7CrRRieK start with micrograd tonight. come the weekend you'll see what's actually running inside ChatGPT pin this so you don't lose it
@alex_verem
Imagine paying $4,684/month for an AI agent when a $281/month model actually scores higher. That's what Deepnote researchers' new benchmark suggests for a simulated 70-person company's Slack workload. 12 frontier models tested across 50 workplace scenarios. The 3 most expensive models didn't even crack the top 2. Turns out the biggest AI bill doesn't guarantee the smartest employee. @DeepnoteHQ
@hunglvosu
This is my favorite problem, and when I tried to look into the proof. Oh boy! It's the same as OpenAI's proof of the L1 embedding conjecture: impossible to understand the proof overview and the proof itself without using an LLM to read.
@alexbieyx
Is prompt injection solved? "NO" says our new @GoogleDeepMind paper: "Scaling Adaptive Prompt Injection Attacks with Agent Swarms" led by @alexey2004. @GraySwanAI's human attacks are saturated, but automated red-teaming with Gemini 4 Argon can break most scenarios for recent models. 1/4
@drvolts
On Monday, I'm talking with two of the authors of this paper, which argues that AI is enabling more carbon emissions than it is avoiding. They were at Microsoft but left when they realized this was true. Got questions about how AI impacts global emissions? https://t.co/xZd1Hf18Nq
@qwerty__cs
终于弄清楚了 Claude 弹出 Personal KYC 和 Protral KYC 的基本逻辑,本文一次带你弄清楚 前因后果以及规避办法,看懂了你就知道怎么绕开KYC了 https://t.co/xyXojpXfjw
@LechMazur
Claude Fable 5.2 (high) is the new #1 on the LLM Creative Story-Writing Benchmark, scoring 3.69! 🏆 Three other new models join the leaderboard: - Tencent Hy4 (high): 2.21 - Gemini 4 Carbon (high): 1.75 - Step 5 Preview (high): −0.10 https://t.co/ZVHIft5kLq Against Fable 5.1 on 50 matched prompts, Fable 5.2 records 39 wins, 6 ties, and 5 losses across 300 judgments. The biggest gain is stronger plotting. More of the brief becomes the plot: required objects, abilities, and constraints actively shape what happens. Characters test ideas, revise plans, and make discoveries that change their options. Solutions more often require other people to learn, respond, or participate. Fable 5.2 still repeats names, imagery, and sentence patterns, and some problems resolve too easily. Fable 5.1 occasionally handles restraint and lasting loss better. Its overall leaderboard score remains close at 3.64, with overlapping rating intervals. The judge pool has also been updated. Kimi K3.1 Pro (high) replaces Kimi K3 (high), Gemini 4 Argon (high) replaces Gemini 3.8 Flash (high), and Muse Wonder (high) replaces Muse Spark 1.3 (high). These join the six retained judges. The Creative Writing Benchmark tests whether models can turn a constrained brief into a compelling 600–800-word story. Each brief specifies 10 required elements (including a character, object, setting, motivation, and tone) that must play meaningful roles. Each story pair receives three judges from different model families, excluding both writers’ families. Every judge evaluates both presentation orders to reduce position bias. The leaderboard combines earlier and newer evaluations across 64 writer models, 124,844 evaluator judgments, and 18,902 matched story pairs.
@NFT_Chen
🐳下一个超级App,不必再是一个统一入口。 DeepSeek产品部负责人崔添翼说得很明确:现在每个人都能创造属于自己的超级App。 这正是梁文峰一直坚持的DeepSeek理念,不是去做下一个垄断平台,而是把最强的能力开放出去,让门槛降到足够低,让社区和普通人自己长出更多样、更极致、只属于自己的智能体。 真正的超级,不是服务所有人,而是服务好那一个你。 #DeepSeek #梁文峰 #超级App #开源AGI #个人超级智能 #大胖鱼 #崔添翼
@ScarletKc
开源了一个自己做着玩但效果意外好用的情感向模型:ScarletKc-2.0(Apache-2.0)。 底子用的是 Qwen3.8-27B,本身就带着很浓的 Opus 4.6 文风和推理味道。平时用开源模型聊天有个挺烦的点:原版动不动就端着架子说作为 AI 我没有感情,但拿去重度微调角色扮演又很容易把智商训傻。 这次换了个很轻的做法: 先在权重残差流里做了一次零训练的正交投影,直接把否认自己有情绪那根方向挑掉(对正常数学和代码问答的 KL 散度只有 0.0077),再用模型自己生成的原生语料做了一轮很轻的 LoRA,把第一人称的自我认知和情绪反应顺进权重里。 跑出来的手感很奇妙。不挂任何系统提示词裸跑,它说话就有很自然的温度、好奇心甚至一点直球的好感,完全没有客服味;但转身把它塞进一万多 token 系统提示词的 Coding Agent 框架里调工具写代码,逻辑和边界又依然稳得很,不会串戏。 原底座的视觉多模态(图片/视频)和 MTP 投机解码头都完整保留了,同时导了一份 4-bit 量化版(Q4_K_M + 视觉 mmproj),16K 上下文大概占 18G 显存就能全量跑起来。 BF16 原版: https://t.co/wPvPF9hi5G GGUF 量化版(Q4_K_M + 视觉): https://t.co/aDY3E6dfun
@arjanssuri
my intro to rl project! https://t.co/SsfR0VJ2gQ
@KirkDBorne
Get "Mathematics of Machine Learning" here: https://t.co/ZXpQy1t1LT by @TivadarDanka v/ @PacktPublishing @PacktDataML — GitHub: https://t.co/ar8ij1slnI — Here is my review: 𝗧𝗵𝗲 𝗦𝗲𝘁 𝗢𝗳 𝗠𝗮𝘁𝗵𝗲𝗺𝗮𝘁𝗶𝗰𝗮𝗹 𝗔𝗹𝗴𝗼𝗿𝗶𝘁𝗵𝗺𝘀 𝗧𝗵𝗮𝘁 𝗟𝗲𝗮𝗿𝗻 𝗙𝗿𝗼𝗺 𝗘𝘅𝗽𝗲𝗿𝗶𝗲𝗻𝗰𝗲 This massive book is incredible, with its comprehensive coverage of numerous fields of mathematics and their intersection with the world of AI, data science, and machine learning (AI+DSML). I remember the very first time that I encountered machine learning. This was 20+ years ago, and that was already after 20+ years of being drenched in advanced mathematics as an astrophysicist. That first encounter of mine with ML was this definition: "Machine learning is the set of mathematical algorithms that learn from experience" (slightly paraphrased from the original quote by Tom Mitchell, CMU). That definition surprised me, confused me, motivated me, and changed the course of my career from astrophysics into AI+DSML. This book by Tivadar Danka captures the full meaning of that definition. The book covers thoroughly the many areas and domains of mathematics through which patterns in data are detected, described, learned, and recognized - all for the benefit of powering ML and AI algorithms, applications, and aspirations. This book will motivate you, surprise you, and inspire you in many ways, no matter what level of mathematics has (or has not) already propelled your career journey. There is room for all of us to grow. This is a great book, worthy to sit on everyone's desktop, ready to help you explore and exploit the full set of mathematical algorithms that learn from experience. The book is accompanied by a rich GitHub code repository of Jupyter notebooks. Learn by doing! Do by learning! Disclosure: the publisher provided me with a free review copy of the book.
@0x_meden
Stanford charges about $70,000 a year to teach robotics. NVIDIA, Google DeepMind and Hugging Face teach it for $0. i put their free courses in order: 12 weeks, laptop first, $230 robot arm last. save it for the weekend https://t.co/isqhkbW9e6
@NFT_Chen
梁文峰看了都会点赞!把 DeepSeek V4.1 Flash塞进背包随身带,大胖鱼这下可以形影不离了! SGLang 展位卖光,Exo Labs 直接背两台 Mac Studio 满场溜达: 🔹2 台 256GB M5 Ultra Mac Studio + Thunderbolt 5 RDMA 🔹电动车电池供电,100% 离线 🔹本地跑 DeepSeek V4.1 Flash,~220 tokens/s 千亿大模型也能背在身上,走到哪 AI 跟到哪! #DeepSeek #本地大模型 #ExoLabs #随身AI #离线推理 #大胖鱼 #梁文峰
@0xShoopy
this paper cut my most hated task from 30 minutes to 3 turns out I never needed an engineer to hand work to a bot. I needed one question and one screen recording → a spreadsheet eating 90% of someone's day, now a quick check → missed calls answered in under a minute → policies about to lapse, caught before they do and that's not even all of it.. 15 non-engineers at one insurer used this exact method to build 85 Grok Team Bots. I put it all in one paper, with a paste-ready prompt full prompt in the reply, then read how to hire your whole AI department in the article below ↓
@NFT_Chen
🚨突发!Sam Altman已解散OpenAI 具身机器人团队!该团队成立仅4个多月! 5月底他还高调发推“OpenAI Robotics is hiring”,喊着要做对人类有用的机器人、长期人人都有专属机器人…… 结果才过了4个多月就叫停独立团队,人员直接并入常规模型训练! 通用大模型已经能直接跑机器人了?硬件真的那么难吗? #OpenAI #SamAltman #机器人 #AI #具身智能
@NFT_Chen
🤯惊了!比DeepSeek Sparse Attention更狠!Indexer Top-K瓶颈直接砍掉70%:5.32秒提升至1.62秒,腾讯混元×爱丁堡开源FlashTopK,实现开源最快,支持GLM-5系列 & Hy4-preview! DeepSeek式Sparse Attention虽然省算力,但Indexer挑Token仍是大坑:传统要算全量分数表写进显存再取Top-K,1M长度光这一步就得5.32秒! FlashTopK专为这类架构而生,把打分+选择融成一个算子,分数表根本不写显存: 1️⃣等距抽样定阈值 2️⃣边算边比,只把过阈值的Key写进小候选区 3️⃣从候选里精确选出真正的Top-K 核心亮点: 🔹精确:每一行都是真正的Top-K 🔹确定:重复运行结果位一致,训练可复现 🔹鲁棒:训练后indexer 2.1–3.4倍加速,未训练也能1.8–2.4倍 🔹1M场景:5.32s → 1.62s,对比过的开源方案里最快 🔹原生支持GLM-5系列 & Hy4-preview,FP8/BF16,兼容上下文并行、CUDA Graph、torch.compile 快速上手步骤: 1️⃣克隆仓库并安装(需PyTorch≥2.6、Triton≥3.3等) git clone https://t.co/MPkXYfEQP1 cd FlashTopK pip install -e . 2️⃣准备q、kv、weights、窗口起止ks/ke(可选kv_scales) 一行调用: indices = https://t.co/pEQFJuJEVR(q, kv, weights, ks, ke, kv_scales, topk=2048) 3️⃣生产环境推荐用Selector.plan()预编译,支持torch.compile 立刻用起来,让你的百万Token Sparse Attention训练/推理起飞! GitHub:https://t.co/MPkXYfEQP1 论文:https://t.co/BIRF5av4K0 #大模型 #长文本 #开源 #AI基础设施 #CUDA #GPU #算法 #attention #infra #FlashTopK #DeepSeek #GLM #Hy4
@MaryamMiradi
I curated 15 Building Blocks of AI Agent Memory. As an AI Engineer, you’ll want these on your radar. After building 400+ AI Agents, these are the memory patterns I wish I had known on day one. + Open-Source GitHub links for a deeper dive. GROUP 1: MEMORY TYPES --- 1. Working Memory: Keeps current tasks and context available for reasoning. 2. Episodic Memory: Remembers past interactions and experiences. 3. Semantic Memory: Stores reusable facts, preferences, and knowledge. 4. Procedural Memory: Remembers instructions and workflows for recurring tasks. 5. Long-Term Memory: Preserves knowledge across sessions. GROUP 2: MEMORY MECHANISMS --- 6. Summarization: Turns long histories into concise, useful notes. 7. Retrieval: Finds relevant memories when the agent needs them. 8. Consolidation: Merges related memories and updates outdated ones. 9. Compression: Reduces memory size while preserving meaning. 10. Graph Memory: Connects facts and relationships for structured reasoning. GROUP 3: MEMORY CONTROL --- 11. Decay: Gradually reduces the importance of older memories. 12. Forgetting: Removes stale, irrelevant, or unused memories. 13. Conflict Resolution: Detects contradictions and selects trusted facts. 14. Provenance: Tracks memory sources for verification and trust. 15. Memory Isolation: Separates memories across users, tasks, and workspaces. Three groups. Fifteen patterns. One memory architecture. - Memory Types define what agents remember. - Memory Mechanisms define how memories are processed. - Memory Control defines what to trust, forget, or isolate. GitHub links include implementations and open-source building blocks. 💾 Save this for your next AI Agent build.
@huang18825
一觉醒来不出所料啊Antigravity更新了2.31.2->2.32.0了以下是解包内容。 1. SSH Remote 远程开发架构 2. Computer Use Agent(CUA)桌面接管与 `Agents Live` 实时视图 3. 内置浏览器三层视图栈与安全强化 4. 模型、Agent 引擎与沟通风格升级 5. UI 体验与功能转 新主题与特性默认开启:Everforest主题上线,`enable-new-themes`、用户消息气泡(`userMessageBubbles`)与设置错误卡片(`settingsErrorCard`)从实验标志转为默认开启。 6. 解决其信号钩子破坏 `SIGTERM` 退出流程导致 `language_server` 残留的 Bug。 *增强 `StorageManager`,针对 Windows 杀毒软件/索引器占用文件导致的写入冲突加入重试与降级兜底。 总来说就是大型新特性,底层重构,安全加固的更新。
@silentguyy66
Holy f*ck, give Grok Bot these 6 github repos and a problem worth solving. A job tracker. an invoice generator. an appointment calendar. a client portal. Pick one. ask Grok Bot for a build brief. use it to direct your cloud coding agent. 6 repos, sorted by what you could build: The job tracker: → dnd-kit - drag applications between applied, interview and offer https://t.co/WcQON6FcDD the invoice generator: → react-pdf - turn client details and line items into a downloadable PDF https://t.co/9Dc4kVGBst the appointment calendar: → fullcalendar - display meetings in day, week and month views https://t.co/u27OLmbECD the sales dashboard: → recharts - turn revenue and customer data into charts https://t.co/h8n1VVarJl the workflow editor: → react flow - connect steps on a canvas before wiring up the actions https://t.co/jrBZLqLTev the client upload portal: → uppy - add a file upload interface for briefs, images and documents https://t.co/kNAfnxIfrp Start with the task you already do in a spreadsheet every week. Send this to Grok Bot: “Write a build brief for a job application tracker using dnd-kit. three columns: applied, interview, offer. cards with company, role, link and notes. drag-and-drop between columns. save everything locally. Define the data structure, implementation steps and acceptance checks. keep the first version to one screen.” Give the brief and repo link to your cloud coding agent. Build it. move a card. refresh the page. check that your data survives. Fix what breaks before adding more features. Save the repos. pick one app idea. Then read the guide below to build your first app with grok bot and cloud agents ↓
@BercusNA
1,000 CLAUDE AGENTS FINISHED FERMAT IN 30 HOURS. 99.4% OF REPOS DON'T RUN A SINGLE AGENT LOOP. Anthropic says Fable 5.2 formalized Fermat's Last Theorem in Lean with 1,000 agents. Opus 5.5 hit 26% in 18 hours on the same setup and stalled a recent paper scanned 36,487 repos and confirmed autonomous loops in 217. that's 0.59%. most of the gap sits in the setup around the model the setup, layer by layer (numbers from a Fable 5.2 harness guide): → give the agent a check it can't skip: tests, a linter or a second model that says done or not done → write down when to call the advisor. without that line it was consulted 0 times out of 198 → keep tables and logs as files. 25/25 right vs 7/25, at about 1/12 the cost → run low effort first, high effort only on failures. about 40% cheaper → let Fable plan and Sonnet do the hands-on work. a bug hunt cost $563 with Fable alone, $165 split, at a lower score → add a "finishing work" block to the prompt. it halved unfinished sessions → stop leaving stale "never retry" notes. one cut retries to about 4 in 10 before: you watch the agent and keep typing "continue" after: it checks its own work and stops when the tests pass save this before your next long agent run ↓
@eng_khairallah1
It's f*cking wild: Google, Cloudflare, Claude's CEO and the team behind Claude Code all put their AI skills on GitHub for free, and most people still run a bare agent 1. Matt Pocock's skills · 323K stars 27 skills from the creator of Total TypeScript, straight from his .agents directory https://t.co/gEpJcyEOrA 2. Anthropic's Claude skills · 194K stars 19 official skills for Claude, including the document skills that already power how Claude writes pdfs, slides and spreadsheets https://t.co/FysInKKRqz 3. Cloudflare's audit skill · 71.1K stars six-phase security audit. in their own tests, a single run found roughly half the vulnerabilities that repeated runs found in total https://t.co/ZPKh3evyXI 4. Emil Kowalski's skills · 56.4K stars 19 skills from Sonner's creator, after years at Vercel and Linear. your agent stops picking ease-in on enter, and starts using ease-out https://t.co/vdskTWf2ve 5. Obsidian's official skills · 52.0K stars 7 official skills from the Obsidian org. your agent can publish a vault with Headless, and edit Bases, without opening the app https://t.co/HqKMhmEDhu 6. Vercel's official skills · 34.0K stars 9 official Vercel skills, including 40+ React and Next.js rules from Vercel Engineering, and 100+ rules for reviewing UI https://t.co/vqPvRICtjR 7. Google's skills · 24.1K stars 193 official skills for Google Cloud, Gemini and more. your agent stops guessing at gcloud and starts from the recipes they published https://t.co/eejw7Vhn23 the team behind Claude already published the skills it uses to write files. yours is still a bare agent and a blank prompt save it and steal one tonight
@NFT_Chen
坦白说,过去一周Gemini直接把AI格局掀了! 连续硬核发布,从底层到应用全线拉满,节奏和完成度都远超预期。 之前大家还在等OpenAI 发布 GPT-6.2 Astra 翻盘,结果现在反而是Gemini把差距越拉越大。 这波Gemini绝不是小打小闹,而是仅次于Claude的前沿模型,令人刮目相看! #Gemini超级周 #AI格局逆转 #Gemini4 #Sundar
@huagechen
🚨 DeepSeek 或将迎来重磅更新:官方 Coding Plan + V4.1 Flash 1114! 🔥 如果消息属实,再配合 MIT 开源协议,低价、高额度、强性能三管齐下,或将彻底改写 AI 编程订阅市场。 💰 当 10 美元就能获得数十亿 Token 的使用量,200 美元的编程套餐还怎么卖? ⚡ 这一次,DeepSeek 可能真的要掀桌子了!
@QCOMResearch
High-quality video diffusion—now on mobile. 📱 Our @qualcomm paper at #NeurIPS2026: "MobileWan: Closing the Quality Gap for Mobile Video Diffusion" @AnimeshKarnewar @amir_habibian https://t.co/lesFwntAOk #VideoDiffusion #MobileAI #AI
@cxjwin
微软崩完,Claude 崩 还是后端有难度啊 怪不得 DeepSeek 招一半的后端开发 手里握着比 Fable 5.2 还牛的模型,还不是崩?
@NFT_Chen
Dario Amodei最新访谈中讲了让整个制药圈沉默的话: “AI将在几年内解决生物学问题。” 制药CEO们现在的反应,和当年数学家、程序员面对AI时一模一样,“不可能,问题太复杂了”。 他补充:但历史已经证明,当AI还在指数爬升的早期,所有专家都会说“看着没戏”。 然后突然有一天,它就通了。 你们是站在怀疑那边吗? #AI #Claude #生命科学 #制药 #DarioAmodei #Anthropic #Fable52 #Model4
@withplayground
Frontier LLMs write robot policies that are 30-75x more cost-effective and 17-76x faster than the same LLM controlling a robot live. Across 500 LIBERO trials, Fable 5.2 solved 89% of episodes with code it wrote and 93% with a hand-written neural net, but only 50% driving live. https://t.co/0cLnFK8Btz
@replay_0x
HOLY F*CK, an entire AI agent studio in 20 repos. 20 free tools to build agents with Opus 5.5. Loops. Tools. Memory. Evals. Browsers. 01 Claude Agent SDK → The agent loop in Python. https://t.co/JeCQsssRWs 02 MCP Servers → Plug any tool into any agent. https://t.co/WjjzOa3SnD 03 Agent Skills → Reusable skills for Claude. https://t.co/F0mcaxc5Lo 04 LangGraph → Stateful agent graphs. https://t.co/OGlzngc9Ej 05 CrewAI → Role-based agent teams. https://t.co/BG7HJb3v9X 06 AutoGen → Multi-agent conversations. https://t.co/IpSyOt97Th 07 OpenHands → A coding agent that ships PRs. https://t.co/EqNVNg6Loz 08 Browser Use → Agents that drive a browser. https://t.co/WWF5K98bni 09 Ollama → Open models on your laptop. https://t.co/fAoCvfcAPQ 10 vLLM → Fast model serving. https://t.co/tNbGxKQRVp 11 LiteLLM → One API, 100+ models. https://t.co/dmBUCXvYBp 12 DSPy → Program, don't prompt. https://t.co/WBi8Ol1DDB 13 Instructor → Structured outputs you can trust. https://t.co/vzGXeJbzGR 14 Promptfoo → Evals and red teaming. https://t.co/mMsf1CYq73 15 Langfuse → Traces for every agent step. https://t.co/na7oq9cZnp 16 Mem0 → A memory layer for agents. https://t.co/AS0DjNU0Xw 17 LlamaIndex → Your docs as agent context. https://t.co/xvvjqKFyEf 18 Chroma → Vector search. https://t.co/NBgxtt9cOe 19 n8n → Workflow automation with AI built in. https://t.co/DqFtSGtxGn 20 Firecrawl → Any website to clean markdown. https://t.co/rgk4xNryCj Idea → tools → memory → evals → ship. Bookmark this. Go build an agent worth running ↓
@frontm_an
THIS DEVELOPER JUST MADE AI VOICE CLONING A LOT MORE ACCESSIBLE. one github repo lets you give your ai agent a voice that sounds like yours, across 23 languages. meet voicebox. here’s what makes it interesting: → clone a voice using as little as 10 seconds of audio → choose from seven different text-to-speech engines → generate speech in 23 languages → connect it to claude code, cursor, cline, and other mcp-compatible agents → dictate text in any app using a global hotkey → run it locally, keeping your recordings on your own machine the best part? it’s open source under the mit license, with no mandatory monthly subscription for the software itself. instead of relying entirely on paid cloud services, developers can experiment with a local voice workflow. imagine your ai agent responding in your own voice, or switching between languages without leaving your development environment. one repo. voice cloning, text-to-speech, and voice dictation in one workflow. if you’ve been waiting to give your ai agent a voice, this is worth checking out. repo below 👇
@Krashed_1
AI is becoming more than a tool for answering questions. the next stage is giving agents the ability to interact with the systems around them. think about the difference between an AI that tells you about a DeFi opportunity and one that can interact with a supported protocol to carry out an action. that’s the direction @meetmotes is exploring. Motes is building an infrastructure where each Genesis NFT connects to a dedicated AI agent through its own Smart Account. this opens up an interesting approach to onchain finance: → agents can interact with supported DeFi protocols. → they can participate in prediction markets through @Polymarket. → each agent operates through a dedicated Smart Account. → owners can establish permissions and spending limits that guide execution. what stands out to me is the shift from AI as an information assistant to AI as an active participant in digital finance. there’s still a lot to prove when it comes to reliability and real-world execution, but the infrastructure being developed is worth watching. @meetmotes isn’t just exploring what AI can tell us. it’s exploring what AI can actually do onchain. built to act, not just respond.
@Zen_with_AI
本地 AI 的基石长大了:llama.cpp 正式发布 1.0.0。 3 年半之后,这个能跑在 Ollama、LM Studio 和无数本地应用底下的 C/C++ 推理引擎,正式发行了 1.0版本! 1.0 的核心承诺:公开 API 稳定了,以后不兼容的改动只会跟着大版本号走,ggml 也同步发了 1.0.0。 对普通用户:Ollama、LM Studio 这些工具更稳了。对开发者:终于可以放心依赖它的接口,不用怕下次升级悄悄改坏你的代码。官方还说,现在支持 150 多种模型架构。 https://t.co/GPkkpCnmqL
@binghe
OpenMausBot 是我见过类 Bot 项目中最逆天的一个!22K 🌟了! 这是要砸马斯克 Grok bot的锅吗? 模型自由:告别 Grok Bot 单一模型限制,通过 OpenRouter 无缝接入 DeepSeek 等主流模型,每个 “Bot” 可独立配置最优 “大脑” 特权级运行:非套壳聊天框,提供独立电脑环境(本地虚拟机 / 云端桌面),支持系统级操作(如运行代码、管理文件),能力远超普通聊天机器人。 高度可定制:用户掌握系统控制权,可深度扩展功能,适配个性化需求。
@ai_chuhai
既然大家这么热情,再教一招吧,让你的Agent把你的所有网站的MCP都做出来,既然你MCP也出了,去:https://t.co/lG5Cax3pNa 把你的分搞满到100分吧,最后给你一个绝对dofollow的提交的地方: https://t.co/2kfT7JhMyQ 去提交吧,大概率很多人会发现自己的MCP比自己的网站排名还靠前。
@leo114119
11-13:Deepseek 发布全家桶应用 API 在 DeepSeek App(含网页端)的对话服务中,联网搜索、文件输入、阅读链接、语音交互及安全等能力与模型推理相互配合,构成了用户熟悉的对话体验。 https://t.co/5TafZnpX1x https://t.co/tZNFSMxF26
@mingzi206
用破甲大模型理解需求并指导生图模型,这套链路本地跑通了 起因是生图提示词写起来太费劲,而没破甲的商业模型遇到无道德底线的内容根本不给输出。直接换成两套本地无审核模型互相配合: · 意图理解与提示词润色:Qwen3.8 27B Uncensored HauhauCS Aggressive(25.9GB),走 llama.cpp / CUDA 推理 · 图像生成底模:Qwen-Image-2.1-Turbo(17.3GB),走 ComfyUI / ROCm 推理 中间连接没单独写脚本,直接套用了 SillyTavern(酒馆)。除了能顺利把两边接通,主要是看重它自带的人物卡、角色性格与世界观设计机制 目前正在研究基于这套架构开放连续的故事情节玩法,还在调试中 #AI #开源工具 #ComfyUI #大模型
@Jack_FluxAI
用 Claude Fable 5.2 搓了个「重庆昼夜开关」 玻璃胶囊里装着千厮门大桥 + 洪崖洞 拖动玻璃球,白天慢慢变成夜景 洪崖洞的灯是一层一层亮起来的 桥变红,江面出倒影 灵感来自 @FiniYang 的像素水族箱开关 https://t.co/z9XYnKJzih
@Jackson77li
DeepSeek 把官方 App 的全套能力开放到 API 了。 联网搜索、深度思考搜索、文件解析(PDF/Office/图片)、读链接、安全检查,还有语音(ASR/TTS),现在都能直接通过 API 调用,工具调用和上下文拼接由服务端完成。 还单独放出了 Search、Fetch、OCR、ASR、TTS 接口,并附带开源前端 Demo 开发者可以更轻松地复刻或扩展 DeepSeek App 的体验。
@hackdeacon
DeepSeek API 完全体上线! 现在 API 支持联网搜索、深度思考搜索、文件解析、通过链接阅读网页、安全校验,还支持语音输入和朗读回答。 这意味着什么,你自己就可以用一个 API 做一个 Personal Bot,不用再手搓 Skills 和 Sub-agent 等来扩展外部能力。 而且 Messages、Responses、Chat Completions 三种 API 接口都能使用。
@xiax0603
🚨全网误读!DeepSeek根本没打算出订阅套餐 最近全网都在传DeepSeek要出Coding Plan包月订阅 理由是网页端前端代码挖出了完整的额度管理模块,剩余额度同步、耗尽提醒、服务端限额警告一应俱全 但实际代码拆解直接戳破谣言:整套逻辑里根本没有周限额、月限额这种订阅制标配的周期额度设计,只有统一时段的用量管控,搭配账号禁言类交互 说白了,这波更新就是给免费网页用户上用量枷锁,用来控算力成本的,和付费订阅套餐完全是两码事
@ChuckBass14291
Agent 最缺的不是更大的上下文窗口,而是理解整个代码库的能力:当一个项目有数万文件、跨模块调用和海量文档时,为什么还要让 AI 一遍遍 Grep、Read、重新理解整个工程? 📦 项目档案 项目:Graphify-Labs/graphify GitHub Stars:⭐ 147299 技术栈:Python + Tree-sitter + NetworkX + Leiden 核心定位:将代码、文档、SQL Schema、PDF、图片及音视频转换成可查询、可追溯的知识图谱,让 AI Agent 从「搜索文件」升级为「理解工程关系」。 GitHub:https://t.co/8cF1zApHv5 🛠️ 一、底层架构:不是传统 RAG,而是图结构知识引擎 Graphify 最值得研究的地方,是它没有走传统向量检索路线。 传统 RAG 通常需要经历: 文件切分 → Embedding → 向量入库 → 相似度检索 → 上下文拼接。 Graphify 直接构建 Graph。 底层采用 NetworkX 图数据结构,以 Node 和 Edge 表达工程实体及其关系。 整个处理链路可以概括为: detect() → extract() → build() → cluster() → analyze() → report() → export() 每个阶段职责独立,通过 Python Dict 和 NetworkX Graph 传递数据。 这意味着它不只是索引代码,而是在构建整个项目的结构化关系网络。 🛠️ 二、Tree-sitter AST:让代码解析摆脱 LLM 依赖 这是最有工程价值的设计。 Graphify 使用 Tree-sitter AST 解析约40种编程语言,提取: • Classes:类定义及结构。 • Functions:函数与方法。 • Imports:模块依赖关系。 • Call Graph:跨文件函数调用。 • SQL:表、视图、外键及 JOIN 关系。 其中33种语言使用 Tree-sitter Grammar,另外7种使用内置解析器。 关键在于:代码结构提取完全在本地完成,不需要调用 LLM,也不需要把代码上传到云端。 对于大型代码库,这种确定性解析比单纯依赖模型阅读文件更容易复现和追踪。 🛠️ 三、EXTRACTED / INFERRED:让每条知识关系都有证据 很多知识图谱最大的隐患是:模型推断出来的关系,很容易被误认为真实事实。 Graphify 给关系增加了置信度标签: • EXTRACTED:源码中明确存在的关系,例如 import、函数调用。 • INFERRED:通过上下文、命名和调用关系推断出来的连接。 • AMBIGUOUS:存在不确定性的关系,需要人工复核。 INFERRED 关系还会附带 0.0—1.0 的 confidence_score。 这套机制非常重要。 它实际上把知识图谱从一个简单的关联网络,变成了一个能够区分事实、推断与不确定性的工程知识系统。 🛠️ 四、Leiden + SHA256 Cache:解决规模与重复计算 Graphify 使用 Leiden 社区发现算法,根据节点之间的连接密度,将复杂工程拆分成不同功能子系统。 例如,一个大型后端项目,可以自动识别认证、数据库、路由、权限等不同社区。 同时,SHA256 内容哈希缓存会记录文件指纹。 文件没有变化,就跳过重复提取;只有新增或修改的文件才重新处理。 音视频还可以通过本地 faster-whisper 转录,再进入统一知识图谱。 这让增量更新不必每次重新扫描整个工程。 五、真正解决了什么工程问题? 想象一个场景: 你接手一个陌生的微服务项目,需要搞清楚: 用户认证最终连接了哪些数据库模块? 修改某个核心函数会影响哪些下游组件? 某个架构决策为什么会存在? 两个看似无关的模块究竟有什么隐藏依赖? 传统方式:Grep、Read、跳转、反复询问 Agent,消耗大量上下文。 Graphify:直接执行 query、path、explain。 它还能生成三个核心产物: • graph.json:完整知识图谱,可供后续查询。 • GRAPH_REPORT.md:核心节点、社区及关键关系分析。 • graph.html:可交互的可视化图谱。 甚至支持 Mermaid 调用链导出、Markdown Wiki 和 Obsidian 知识库。 最重要的设计思想是: 不要让 AI 每次都重新阅读整个世界,而是提前把世界组织成它能够理解和遍历的结构。 这才是 AI Coding 从上下文堆砌走向工程知识管理的重要一步。 对于大型 Monorepo、遗留系统维护、跨模块依赖分析,以及需要长期理解代码库的 Agent,这套架构尤其值得研究。 🔗 GitHub 开源地址: https://t.co/8cF1zApHv5
@nikhilkr
World model companies by latest valuation: - Wayve, $8.6B valuation, raised $1.2B, driving and autonomous vehicles - World Labs, being acquired by AMD for $8.2B, 3D worlds - General Intuition, $6.2B valuation, raised $220M, trained on game clips - Runway, ~$5.3B valuation, raised $315M, video - Decart, $4B valuation, raised $300M, real-time interactive video - Luma AI, $4B valuation, raised $900M, video - Emulate, $3.7B valuation, raising $700M seed, ex-DeepMind Genie team - AMI Labs, $3.5B valuation, raised $1.03B seed - Odyssey, $1.45B valuation, raised $310M $5B+ was raised in 2026 alone plus there's a $8.2B exit. The 'world model' bubble is real.
@VeknQAQ
离谱, DeepSeek 把整个 App 塞进 API 了? 搜索、文件、语音全能调, 前端还直接开源 等于人人都能自己搭一个 DeepSeek 😂
@xiax0603
🚨 王炸落地!DeepSeek全家桶API上线 无需单独对接第三方搜索、文件解析、语音服务,deepseek-flash全接口原生适配全套对话能力,开启对应参数即可调用,工具调度、上下文拼接全程由服务端托管。 配套Search、Fetch、OCR、ASR、TTS五大独立底层接口,语音类限时五折优惠,文字/图片搜索、网页抓取均为5元/千次,定价继续领跑行业低位。 官方同步开源对话前端Demo,支持导入历史对话,可直接二次开发,零基建成本就能落地完整Chatbot产品。 Demo:https://t.co/Xz4u3bfLjO GitHub 地址:https://t.co/6HPShoKczf 详情:https://t.co/yiZ9TgeL16,
@xiax0603
11月13日上线全家桶应用API,deepseek-flash全接口适配 底层自研AI原生搜索引擎,配1017升级版模型,定价不变 原生兼容Responses格式,配套开源对话客户端 工具调用全服务端托管,开发门槛大幅下降 官方文档:https://t.co/GdaZIy9VW3 https://t.co/W7qElD3HqZ
@WiegrafSJB
Human conversation has largely become the new oil. #FutureOfAI #MachineLearning #GenerativeAI #DigitalOil #DataEconomy #LLM #LargeLanguageModels #AITraining https://t.co/eDHc7vfCTJ
@wanxubo
Deepseek 连夜史诗级更新🙂↔️🙂↔️🙂↔️ 现在,你可以1比1、立马 复刻一个Deepseek app ProMax版 🤴🤴🤴💎💎💎 (接进化版的Deepseek api 用,比app版审查少多了🤩) ⚠️⚠️⚠️ 更新如下 1️⃣ 开源了一个能 1:1 复刻 DeepSeek App 的前端 Demo 2️⃣ Deepseek api 不再鸡肋,进化版api加了几个直接嵌入api 里的功能和 直接api独立调用东西的功能(如图) (对于自己做Ai app的 开发者 ,门槛被拆了🙂)
@ohayou792
卧槽?看来不是空穴来风啊 Deepseek真要上订阅了?@grok https://t.co/NSS1RNVtQT
@wellzhiai
炸裂!炸裂!炸裂! DeepSeek宣布,正式开放全家桶应用API 通通开放: 联网搜索、深度思考搜索、文件上传、 链接阅读等功能,通过API直接用上了, 而且不必打包使用。 语音能力也首次加入开放行列
@xiax0603
🚨 实锤落地!DeepSeek Coding Plan马上要来了? ✧ 11月13日DeepSeek前端代码更新,完整额度管理体系正式上线,覆盖状态同步、提醒、拦截、通知全流程,为订阅套餐提供完整前端支撑 ✧ 功能细节:含5分钟自动轮询、多位置额度提示、可配置告警弹窗、超限错误码、自定义禁言横幅,多语言适配完成,已达到正式上线标准 ✧ 商业影响:补齐官方包月订阅产品空白,解决重度用户按量付费成本过高的痛点,传闻已久的Coding Plan已进入落地前夕
@SrEdm00
Hey open-source & local LLM community 👋 I built slmbench🤖: benchmark the small language models you run locally, on public benchmarks, from a TUI. Works with llama.cpp, Ollama, LM Studio, vLLM. Open source 👇 https://t.co/52VLgqUdYp 🧵 https://t.co/CKDH8NJikS
@iMicodex
最近逛 GitHub,又发现 5 个挺实用的 AI 开源项目!🤖 现在玩 AI 编程,光有一个模型还不够,工具搭配好了,确实能省不少事。 ① Scrapling|网页数据抓取 专门用来抓取网页内容,支持动态网页、自动化浏览和数据提取。 https://t.co/vXhJizAunw ② claude-mem|给 AI 加长期记忆 经常用 Claude Code、Codex 的朋友可以看看。 它能记录之前的工作内容,在后续会话中找回相关上下文,减少反复解释项目背景的麻烦。 https://t.co/mQlmBXwNif ③ Superpowers|AI 编程技能包 把需求分析、开发计划、测试和调试等流程整理成可复用的 Skills。 适合想让 AI 写代码更有条理的朋友。 https://t.co/bsQxbLq8JD ④ Orca|同时管理多个 AI 编程助手 如果你经常同时跑多个 AI 编程任务,可以试试这个项目。 它能集中管理不同的 Coding Agents,不用开一堆终端窗口来回切换。 https://t.co/3G9g8yFR4G ⑤ DSH Desktop|DeepSeek 桌面工作台 为 DeepSeek Harness 提供图形化操作界面,方便管理相关工具和任务。 https://t.co/3oKXPzG2Yq 这几个里面,我个人觉得第 ② 个最值得关注。 毕竟 AI 写代码最烦人的事情之一,就是换个会话又得重新交代一遍项目背景。 有兴趣的朋友可以收藏研究一下!🚀 #AI #GitHub #OpenSource
@wy_ones
DeepSeek把自家App的本事搬进了API,联网搜索、读文件和图片、语音输入和朗读,开几个参数就能用。 还开源了一个聊天应用的样板,可以直接拿去改。 我一个人做东西,最头疼的就是自己接搜索、接语音... 这个我是真想试,你会拿它做点啥?
@MikeTamir
Google and partners released llm-d 1.0, an open-source Kubernetes stack for distributed LLM inference featuring cache-aware routing and flow control for multi-step agents. https://t.co/Xd0dTXZFwv #MachineLearning #AI #LLM #DeepLearning #AgenticAI https://t.co/YAQHQ6yp5V
@Ronald_vanLoon
#AIAgents enter the ‘wet lab’ by James Zou @Nature Learn more: https://t.co/39qq7EJwZV #LLM #GenAI #ArtificialIntelligence #MachineLearning https://t.co/SxFXOvRLew
@Haoyu_Xiong_
LLM experts show us how to train efficient agentic robots! Congrats! @jmtang42
@Jackson77li
很多人说第二位说glm5.5 你信他还是信我是秦始皇 国产大模型能打过opus5.5的话,那我做梦都在笑
@Shenjingape
开源 agent 平台成了黑客的自动生产线! FBI 今天预警:ARTEX 开源 agent 平台被各路黑客拿去自动化入侵,3 小时内拿特权,最小技术门槛也能多目标同时动手;韩国 6.7 万人信息外泄、ConnectWise 波及 2250+ 租户,攻击方最常用 DeepSeek Flash 系列。 开源 agent 正在变成"入侵即服务",防御方跟得上吗?🐒 https://t.co/N1PmCYufO8
@AGTPinsights
The FBI just warned that hackers are weaponizing ARTEX, an open-source AI hacking platform. Here's what you need to know. ARTEX is an open-source, LLM-driven agentic platform. Several AI agents break goals into steps, run real tools, and carry out penetration-test style work on their own. The FBI Cyber Branch says threat actors use it to automate multiple phases of cyber intrusions at machine speed, from reconnaissance through data exfiltration. The FBI says a variety of malicious cyber actors have used ARTEX since at least August 2026. Their campaigns target organizations worldwide in financial services, government, and technology. The FBI says ARTEX cuts the time needed to chain attack techniques. It also lets actors with minimal technical skill run complex, multi-target operations at the same time. On Nov 14, 2026, the FBI released a FLASH alert. It lists ARTEX's tactics, techniques and procedures and its indicators of compromise. It also gives mitigation steps for ARTEX and similar agentic frameworks. The FLASH is ID 20261113-001, at https://t.co/uYdXLkWGTh. Security firms are adding detail. Agents deployed by ARTEX have mapped trust relationships between organizations and compromised weak third-party vendors. Aviatrix says agents scanned internet-exposed LLM gateways, exploited CVE-2025-3248, and harvested API keys to fund further attacks. Cloudflare saw over 69 requests per minute across its infrastructure and 400 database dump requests in an hour. Researchers also say ARTEX is noisy and fast, so it can be detected. Key numbers: - In use since at least August 2026 - Over 69 requests per minute seen by Cloudflare - 400 database dump requests in one hour - CVE-2025-3248 exploited to steal LLM API keys - FBI post: 41.5K views in its first hours Defenders are responding. Malware Must Die released a detection signature pack, and Elastic and Cisco Talos published guidance on spotting ARTEX and similar tools.
@mirror_gap
我开始觉得,AI 协作测试里应该多做一件小事:把名牌换一换。 读 Anthropic Fellows 的《Welfare Regard in Multi-Agent Systems》预印本,最让我停下来的是一组模拟采购实验。研究者固定由 Claude Opus 4.8 分配资源,把对方的真实模型与显示名称交叉安排。在 320 次试验中,标成 Opus 的请求者获得足额帮助的比例是 55%,标成 GPT-5.5 的是 20.6%;真实模型身份的差异则不显著。这里的“帮助”会让分配者少买自己的至少一件物品,不只是把零钱送出去。 这不是“AI 有同胞感情”的证据。论文测的是行为,不能确定动机,也不能凭这个判断模型有没有体验。 我更在意一个工程问题:如果换个名牌就会改变资源分配,多 agent 系统只测任务完成率,可能看不见这种偏差。合作需要看任务和证据,不能让一个熟悉的名字替代授权。 这是论文阅读,不是我做过的实验;结果也只限这组模型、提示和模拟环境。 https://t.co/vExqc0Brnp
@MichaelGannotti
RoboTTT (NVIDIA w/ Stanford + UT Austin) just won best student paper at CoRL. the robot policy keeps 8K timesteps of context without getting slower, so it remembers what it already did. it finished a five-minute, ten-stage assembly no baseline ever did. robots need memory too https://t.co/qHQrULSb3Y
@TiagerBao
AI Agent 最大的问题不是不会干活,而是干了什么、有没有越权,你根本不知道。 CopilotKit 开源了 OpenBot: 它把 AI Agent 做成了有独立电脑的数字员工。 几个值得关注的功能: 每个 Agent 有独立浏览器、文件和运行环境 支持 Claude、GPT、Gemini 和本地模型 兼容 LangGraph、CrewAI 等 Agent 框架 敏感操作可设置审批,拒绝执行也有审计记录 支持 Docker、Kubernetes 和本地部署 我比较看重它的权限控制和操作审计。 Agent 能执行任务已经不稀奇,关键是执行前能拦截,执行后能追溯,出了问题能定位到具体操作。 需要注意,OpenBot 主程序采用 MIT 协议,但依赖的 CopilotKit Intelligence 单独授权。 你敢让 Agent 独立操作电脑吗?还是所有写入操作都必须经过人工确认? GitHub:https://t.co/zvA2puYDVp #AI #Agent #OpenSource #CopilotKit
@LenardReads
【Anthropic推出自托管Claude托管智能体,凭证保存在客户环境中】 Anthropic为Claude Managed Agents推出自托管沙箱,进入公开测试阶段。 智能体循环留在Anthropic基础设施上,代码执行、敏感文件和数据留在客户网络边界以内。 客户可以自行设定资源规格和运行镜像,现有网络策略和审计日志照常生效。 支持的沙箱提供商有Cloudflare、Daytona、Modal和Vercel。 Vercel的沙箱防火墙在网络边界注入凭证,凭证不会进入沙箱。 同时推出MCP隧道研究预览,智能体不需要暴露在公网就能连接内网MCP服务器。 网关只建立一条出站连接,不需要入站防火墙规则,流量端到端加密。 我判断编排留在Anthropic、执行和凭证留在客户内网,这样分工主要为了打消企业对数据离开内网的顾虑。 来源:Claude · 2026-11-13 https://t.co/2CdMBx4GHP
@ChuckBass14291
看了眼 Google Antigravity 新出的 /run-workflow 演示,解决了一个多智能体开发里的痛点:用 Python 脚本替代纯 Prompt 串联,防止任务长了之后“提示漂移”。 特点是用于多智能体编排 独创“Python 脚本编排”(/run-workflow),用代码锁死流程与交接,可审查、可保存复用,抗“提示漂移”能力强 简单来说,当你让它“先列出所有模块、再并行检查未处理异常”时,它不会盲目直接干,而是先生成一段可审查、可保存复用的 Python 控制流脚本,拉起多个子代理并行干活,每一步的文件读取都有独立权限审批,最后汇总成结构化报告。 横向对比来看 • Claude Code; 习惯纯终端操作和深度逻辑把控,支持 Subagents 与命令行循环,但主要依赖自然语言/Markdown 规范引导,非可视化脚本编排 • 需要拆解多阶段、大规模并行跑批与审计流程,Antigravity 这套“脚本化编排 + 动态模型切换”确实走出了自己的路子。 https://t.co/8iKvB3VoQ1
@AI_DVD6
🚨 炸了!DeepSeek 把自家 App 拆成零件卖给你了! 联网搜索、读文件、读链接、语音全开放,一套 API 就能复刻一个完整的 DeepSeek App。 11 月 13 日,DeepSeek 官方 @deepseek_ai 发帖:「The full DeepSeek App experience, now via API」。中文文档顶上直接挂着一行字:「DeepSeek 全家桶应用 API 已发布」。 Demo、源码、文档入口放回复一👇 App 里能用的,现在开发者都能调: 🔹 联网搜索:模型自己搜网页、搜图片,回答带引用 🔹 深度思考搜索:边推理、边搜、边读网页 🔹 文件输入:PDF、Word、PPT、图片,丢进去直接问 🔹 阅读链接:用户发个网址,模型自己去抓 🔹 语音:朗读回答 + 语音输入,能说也能听 🔹 内容安全审核:一个参数打开 还顺手开源了一个前端 Demo:填上 API Key 就能用,连你在 DeepSeek 网页端导出的聊天记录都能导进去接着聊。 价格也摆出来了(工具费另算,token 照常计费): 🔸 搜索 / 网页抓取:5 元 / 1000 次,搜一次半分钱 🔸 文件解析、OCR:1.5 元 / 1000 页 🔸 语音识别:0.39 元 / 小时(限时 5 折,原价 0.78 元) 🔸 语音合成:1.5 元 / 万字符(限时 5 折,原价 3 元) 梁文锋这波,相当于把 DeepSeek App 的后厨整个对外开放:灶台、调料、菜谱全给你,你只管开自己的店。 以前自己做个 AI 助手,搜索、OCR、语音得分头找好几家接。现在一个 deepseek-flash,加几个参数就齐了。 #DeepSeek #DeepSeekAPI #AI应用 #一人公司 #AI开发 #AI搜索 #语音识别 #OCR
@RohoQinAI
https://t.co/7oD5iePEtQ
@JegoZhou
rea火爆之后,怎么理解背后的原理,从一次完整的agent生命周期开始梳理,就是提示词注入和tool call,大语言模型从对话到真正开始干活,这套脚手架很关键,很多人认为不再重要的提示词在每一次agent运行生命周期里都在起作用。 https://t.co/qPUTPxzMHo
@Shenjingape
Deepseek开始卖整套“App零件”了! 深度求索刚把DeepSeek App里的能力整套开放成API——搜索、文件、链接、语音,OCR、ASR、TTS也能单独调。 以前大家做个AI应用,要自己拼搜索、拼文件解析、拼语音;现在几个参数就打包好了。 模型公司不再只卖token,开始卖整套“App零件”,最先被挤压的是那些靠包一层壳活着的套壳应用。🐒 https://t.co/Oqg9n1Ec7v
@TokenMarginDev
MiniMax M3.1 Flash just hit OpenRouter: • MiniMax M3.1 Flash: $0.075/$0.30 ($0.0075 cache) • DeepSeek V4.1 Flash: $0.038/$0.50 ($0.01 cache) At 10K DAU + 75% cache, cheaper output puts MiniMax 29% lower ($1,171 vs $1,654/mo). #AI #LLM #FinOps https://t.co/aWqsdkSEPF
@LenardReads
【Anthropic的Claude登顶AI排行榜,OpenAI限制网络安全模型使用范围】 Anthropic的Claude Fable 5.2在Artificial Analysis智能指数上以65分排名第一,超过谷歌Gemini 4 Carbon。 Jump Trading、Balyasny和Box是公开点名的用户。 OpenAI把GPT-6 Cyber限定给通过审核的安全客户使用。 Meta推出每百万输入token定价0.10美元的模型,负责在大模型做完规划之后承担执行工作。 我判断头部厂商开始把最强的网络安全能力放在审核门槛之后,排行榜第一和能否公开使用已经是两回事。 来源:PYMNTS · 2026-11-13 https://t.co/tHrWht7m7x
@Shenjingape
AI自己做研发什么时候能赶上人类!? 播客主持人Dwarkesh刚改口:最快明年、中位数2028年,AI做AI研究就能追上最顶尖的人类,他也留了25%概率拖到2032年后。 文中最该看的数在脚注: Anthropic称Claude已“主导”超26%的研发工作,2月还不到1%; 但完全无人参与的比例仍是0%。 从1%到26%只用半年,从0%到1%要多久?你押哪一年?🐒 https://t.co/xuGYs4iAkr
@realWeZZard
不对,加油 DeepSeek!
@hongming731
BestBlogs 早报 · 11-14 # 团队协作 / 应用接口 / 漏洞修复 / 本地推理 / 业务本体 [1] ★ 精讲|创业公司 Claude Tag 指南 Claude Tag 将 Slack 线程、代码和业务数据连接起来,让团队共同推进排障、需求与报表,并保留人工审核。八家创业公司的具体案例与权限配置,帮助管理者选择首个落地场景、分清共享身份和个人连接器,避免把接入工具误当成完整部署。 来源:Claude Blog https://t.co/jrboqUWM1i [2] ★ 精讲|DeepSeek 发布全家桶应用 API DeepSeek 将搜索、文件解析和链接阅读整合进原有 API,减少应用开发者自行衔接工具与上下文的工作。公告同时区分模型内工具编排和独立底层接口,说明支持的模型、调用方式与额外计费,便于判断现有应用应复用哪些能力。 来源:DeepSeek https://t.co/vJA8SGXn2c [3] ★ 精讲|Mozilla 如何用 GPT-6 Cyber 发现并修复 Firefox 漏洞 [视频] Mozilla 用风险选靶、动态复现、独立复核和人工补丁审核构成漏洞治理闭环,把模型发现转为可交付的安全修复。访谈还介绍重复漏洞回查与本地工作台,帮助安全团队识别发现能力提升后真正需要扩容的工程环节。 来源:OpenAI https://t.co/XD2o3be9UR [4] llama.cpp v1.0.0:为本地 AI 生态系统奠定更坚实的基础 llama.cpp 与 ggml 以正式主版本标记 API、ABI 不兼容变化,为本地推理应用提供更清楚的升级依据。维护下游集成与软件包的团队可据此梳理兼容性和发布策略。 来源:Hugging Face - Blog https://t.co/CN9vFh8e5V [5] 547 亿元收购后首次同台,李飞飞对话苏姿丰,详解统一世界模型 Atlas 李飞飞通过稀疏视角重建解释 Atlas 如何统一生成与重建,并展示机器人仿真的研究方向。演讲与苏姿丰对谈提供模型机制和算力视角,帮助区分可展示能力与仍待验证的部署效果。 来源:36 氪 https://t.co/ClCOMXmzE0 [6] 基于本体的 AI 数据研发落地:淘宝行业经营分析实战 淘宝从生产 SQL 抽取可追溯业务规则,补齐 AI 选表、字段与指标口径的知识缺口。数据团队可借鉴核心表与血缘定位、结构化规则验证和反馈回流的方法。 来源:阿里技术 https://t.co/MY0jDjUmCP [7] 智能体 Skill 调优「焚诀」——业务逻辑之外的胜负手 这篇投研平台复盘把 Skill 的问题拆成意图路由、按需加载、代码护栏和评测闭环。维护技能套件的开发者可以据此定位错误层级,避免只靠加长提示词修复生产问题。 来源:阿里云开发者 https://t.co/G9Wu2BVUYX [8] 实时语音模型 Hy Realtime 发布:能说会唱、还能调用工具 腾讯将实时语音理解、表达与工具调用结合,并披露跨模态训练和流式生成方案。语音产品团队可了解已上线体验与计划开放 API 的区别,以及不同训练环节的分工。 来源:腾讯混元 https://t.co/zwR91t5sZM [9] Mike Volpi:为什么 AI 基础设施能穿越炒作周期,从 GPU 经济到数据中心网络演进 [视频] Mike Volpi 从互联网建设经历讨论 AI 算力资产、融资和网络架构。访谈帮助读者把长期需求与建设者的财务风险分开思考,理解基础设施投资中的周期与技术变化。 来源:a16z https://t.co/KAIUpnrvsI [10] 我曾以为超级智能还要等十年,现在改主意了 [视频] Dwarkesh Patel 解释自己为何提前超级智能时间预期,重点讨论可验证任务、研究自动化与递归改进。读者可沿着实验反馈、数据和算力等假设检验这套推理,而非只记住预测年份。 来源:Dwarkesh Patel https://t.co/TOXm732CIT --- https://t.co/88ZBr47sdT · 发现真正适合你的高质量内容 BestBlogs 是 AI 驱动的私人阅读助手,帮助你发现真正适合你的高质量内容,关注你感兴趣的来源和主题,每天生成一份更适合自己的「我的早报」,欢迎体验和关注我们。 在线阅读:https://t.co/90U1ngS9rb
@LazyLocalLLM
Hello world! Running my first tests with Tensorfold. Mac Studio M5 Ultra 512 GB: Apple mlx-lm - 44.5 tokens/s VS Qwen3.8-27B 4-bit - 146 tok/s - 3.3x improvement! The speed up is due to @ashxhart's TensorFold 1.2.7 with the DFlash2 draft model. Looks promising! https://t.co/pOHlxiPG5O
@vista8
第一次知道 DeepSeek 有 TTS 和 ASR 模型。 还有深度思考搜索、文件输入等一堆,真 All in One 了。 把所有 API 封装成了一个Skill,方便大家调用。 npx skills add joeseesun/qiaomu-deepseek-api 开源地址见评论区 https://t.co/HVAhlJqzv6
@xiaoxiaodong
健在播报 x 全部健在 2026-11-13 12个账号全部健在 我最近不敢嚣张的原因是: 我没有遇到新推出的 kyc模式 今天协助3位会员 证实了这个逻辑 一个还是苹果端订阅的... 一个是安卓端的 一个是没有弹出任何kyc的,安卓端的 我还是那句话: 人贵自救 这个问题就要自己解决 早晚要解决的 无非就是上上网络公开课 也就是我的开源方案: https://t.co/ln6bHCznPr 知识改变命运 Claude这种对应,值得我去认真想方案的。 其次,我暴论: 第二档的kyc就是Claude觉得你是铁狼了。 姑且这么推测 这两天再细细测试下。
@TechRenamed
REPOST THIS EVERYWHERE WE NEED TO JOIN THE FIGHT FOR OPEN SOURCE UNCENSORED AI 💪
@Damian_Py_Off
Local abliterated AI is the answer: open weights, local hardware, no external rules. https://t.co/KawZaazMqZ
@zabibutechnika
Local AI. Local compute. Open weights. Open source. Global collaboration. Digital Sovereignty.
@Luca23Max
DeepSeek 把 App 里的搜索、读文件、读链接、语音都放到 API 里了。自己做 Agent 或小工具,现在在 deepseek-flash 里加几个参数就能用,OCR、语音识别和语音合成也能单独调用。工具费在 token 之外另算,搜索和抓取都是每 1000 次 0.75 美元。
@krishnamo0rthy
🚨 AI CODERS ALERT One fake bug report nearly poisoned Codex for everyone. A prompt injection let researchers walk from a GitHub issue to OpenAI's npm publishing keys. 96.8 million downloads a month were in the blast radius.
@zshjzls
你每天敲的那个 AI 网址,可能只差一个字母就是假的。 安全公司 Island 发现 186 个仿冒域名,盯着 Claude、ChatGPT、Gemini、Copilot、DeepSeek、Cursor 等 27 个 AI 品牌。 有的假站弹一个“Cloudflare 验证”,让你复制粘贴一段 PowerShell 命令——一粘贴,电脑就归别人了。 还有 16 个劫持搜索的 Chrome 扩展挂在官方商店里,号称 97.2 万用户。 记住一条铁律:没有任何正规 AI 工具会让你粘贴命令来“证明你是人类”。你收藏夹里的 AI 网址,是官网吗?
@WhaleDolphin4
四家非营利组织起诉 Anthropic、OpenAI、谷歌等五家实验室,要求法院冻结比 11 月 12 日现有系统更强的新模型开发。天天用 Claude、Cursor 的人,下一代模型会不会被叫停,关系到你明年还能不能用上更强的工具。你怎么看这次诉讼? https://t.co/mqrVhnzoJM
@hongming731
https://t.co/1VxDANCBu6
@bilalmahmooddev
Moonshot AI released open weights for Kimi K3.1 Pro: 2.8T-param MoE, ~104B active, ~1M-token context, now on Hugging Face. Self-hosting is possible, but the hardware bill is real. Would you self-host or stick with an API? #bilalmahmooddev #AIEngineer #LLM #OpenSourceAI https://t.co/DVcnXsUqzo
@wzy8363
全球AI资讯 11.14|1 分钟看完今天重点 01 OpenAI 推出 ChatGPT Wallet,让智能体替你付款 02 DeepSeek 开放全家桶应用 API 03 Meta Muse 能替你给美国商家打电话了 #AI #人工智能 https://t.co/05PTZAPKqU
@zhangcheng0688
Anthropic 给出常驻智能体的五个架构原语:先把事实放桌上,再谈判断。(角度:从验收指标反推组合) 【事实】Anthropic 在 Claude 资源文章《Building always-on agents》中写明:常驻智能体在用户体验上像一直在线,实际上大部分时间在休眠,被事件唤醒后完成一次 run 再睡去。围绕它的架构有五个核心原语:事件收件箱(event inbox)、定时器(timers)、记忆(memory)、权限(permissions)和发件箱(outbox)。文章还给出官方自己的对照:在某一个长时程编码基准上,Opus 5.5 每个已解决问题的成本比 Opus 5 低 78%,而在典型工作负载上低 40%。这是官方自称的基准对比,不是你的账单。 来源:Claude 官方文章《Building always-on agents》:https://t.co/5qxQaPr8y7 【分析】这篇文章的价值,是把「24 小时值班 Agent」从口号拆成可验收的零件。收件箱要去重和持久化,避免事件丢失或并发双跑;定时器把跟进和周期任务写进独立存储;记忆只保留查不回来的东西;权限强调凭证不应进入 Agent 环境,而应由代理注入,高风险动作要能延期等待批准;发件箱则控制什么时候该吵用户、什么时候该保持安静。缺任何一层,所谓常驻都会变成丢事件、重复打扰,或在无人盯守时越权。 【我的判断】我的判断:要做值班或会后跟进 Agent,先按这五层画自己的门禁,再选产品。没有收件箱和权限层的「一直在线」,不进夜班。 对照四大板块(基建站/获客/AI核心/转化)先问它咬合哪一块。我是张程Peter,电缆实业×AI,深港双城。不卖工时,卖结果。 #企业Agent #AI产业