2026年11月19日 · AI 晚报
本轮讨论从模型发布延伸到成本与交付:Linger 把个人智能体搬到本地,DeepSeek Harness 扩展视频、记忆和专业工具,研发分享与研究项目则反复强调评测、验证和反馈闭环。
11月19日 11:00 — 11月19日 23:00(日本时间) · 129 条 X 原帖
本期目录
产品消息之外,用户更关心额度与可用性
Anthropic IPO 与 Gemini Neon 的消息进入社区讨论
资讯账号 Taylor Ortiz 的新闻汇总称,Anthropic 的 IPO 定价为83美元,Google 推出 Gemini 4 Neon。本批没有对应官方公告,发布与上市信息主要来自这一转述。围绕 Neon,刘小排、Miles Ma 等用户给出积极评价;Miles 自述用一句话生成的视频效果不输 Opus 5.5,这属于个人作品体验。
Claude 的使用稳定性也影响选择:北岛自述封号后转向 Codex 和 Gemini,laodeng 则称申诉后账号恢复。这些经历说明用户在意访问连续性,无法据此确定统一的封禁原因。
价格变化与故障反馈需要分开看
Daydreamer 称,DeepSeek 自11月18日中午起取消模型调用中联网搜索、网页抓取的单独收费,仍按 token 计费。Geek 比较自己看到的 OpenCode 额度时,则抱怨500美元 GLM-5.3 的请求量仅略超100美元 DeepSeek V4.1 Flash 的六分之一,比较范围限于其套餐观察。另有用户报告 Kimi 网页和 API 同时不可用,帖中关于算力或机房的猜测没有进一步证据。
设计榜单出现匿名领先者
Max For AI 转述 Contra Labs 的设计基准:33项真实客户任务交由专业设计师盲评,匿名模型 Gestalt 获1931 Elo。它的开发团队尚未公开,成绩应限定在该榜单的任务与投票范围内;社区对其身份的猜测没有可用证据。
来源 · 2 条原帖
Linger 开源,把个人智能体放到自己的电脑
本地模型、订阅与多智能体协作接到一起
Tony Zhang 宣布开源 Linger,定位是运行在用户电脑上的个人智能体。多位体验者介绍,它既能接 Claude Code、Codex 等已有订阅,也能通过 PolyStrata 运行本地模型;后者把 MoE 专家分层放入显存、内存和 SSD。余温体验的重点是群聊交接、独立审稿和后台任务,让前台对话不必一直等待慢任务。
本地推理不产生云端 token 费用,但不能理解成整套系统没有外部依赖。huangserva 明确说,连接应用、语音和后台记忆整理仍会使用外部服务。
来源 · 3 条原帖
硬件适配比笼统的性能口号更实用
雪踏乌云记录了一个安装细节:Quick Install 按当时空闲内存推荐模型,48GB Mac 在其他软件占用后只剩14GB,就会得到偏小的推荐。其介绍也注明,后台记忆整理目前仍使用 Claude Code 或 Codex 额度。
另一个低互动项目 MLX Model Scout 专门处理 Mac 本地模型选型:智能体负责搜索与解释,Python 估算内存和速度,并设置人工检查点。介绍者同时说明,估算只是近似,实际下载和基准测试目前仍为模拟。
来源 · 2 条原帖
研发与产品分享,都在追问闭环怎样跑通
姚顺雨分享笔记:小模型、数据与实验判断
幻觉工程师根据取得的文字稿整理姚顺雨分享,主线是按任务成本组织模型体系:先训练强模型,再蒸馏不同尺寸的小模型。其笔记把 RSI 归纳为提高研发循环效率,具体落在修复错误、清洗数据、实验验证与代码迭代。
笔记同时提醒,小规模实验失败可能被噪声和能力不足影响,结论未必能直接迁移到大模型;预训练仍有高质量数据复用、记忆与过拟合问题。强化学习工程则应先保证正确性,再谈稳定性和效率。这些内容来自整理者转述。
来源 · 1 条原帖
Manus 的取舍:真实任务评测与少设结构
同一整理者转述 Manus 创始人张涛的分享:团队砍掉投入七个月的 AI 浏览器,因为基础体验需要与成熟浏览器竞争;早期评测则从朋友提交的真实工作任务中筛出300多项。模型升级后,原有 Harness 无需修改就改善了未解决任务,促使团队坚持单智能体和较少固定流程。
商业侧更尖锐的问题是消耗与需求判断。张涛称单个 DAU 的 token 消耗约为聊天产品的200至1000倍,并批评免费 DAU 指标会妨碍团队转向真实付费需求。
来源 · 1 条原帖
多智能体工程先解决调度阻塞
cxuan 转述 Octop 团队模式的工程拆解,强调移走调度者的执行工具、采用异步回调,并让本地实例执行任务。Kyle 对 Grok Bot 自动化文章的理解则落在目标、约束、长期记忆和反馈上。两类讨论关心的都是怎样把工作持续交付,而非单纯增加智能体数量。
来源 · 2 条原帖
研究进展:表示几何、推理系统与科研自动化
语言里的对称性,可能解释表示空间里的圆与地图
Superman 转述一项研究:月份圆环、年份时间线、城市地图等几何结构,也能出现在 word2vec 中。其解释指向语言统计中的平移对称性与傅里叶表示;即使删除月份直接共现的句子,其他关联词仍可能维持圆环结构。这为理解表示来源提供了另一条路径,不能直接放大成对所有世界模型研究的否定。
来源 · 1 条原帖
Blink 优化推理路径,NanoQuant 压缩权重
Superman 介绍的 Blink 把主机 CPU 移出推理路径,由 SmartNIC 经 RDMA 传入 GPU 内存,再由常驻内核管理调度与缓存。其转述测试中,P99 首 token 延迟最高改善8.47倍,解码吞吐最高提高2.1倍,这些都是特定对比测试的上限。
CrazyShyyt 介绍的 NanoQuant 则用结构化二进制表示,把平均权重存储压到每参数不足1 bit。帖文没有给出精度损失或完整实测条件,尚不能据此判断实际部署收益。
来源 · 2 条原帖
数学研究智能体把证明、复核和记忆接起来
智东西介绍 MathRSI,称其已解决六个数学开放问题。系统采用准备、路由、证明、验证、交接五阶段流程,结合并行证明、跨模型审查、逐题研究 Wiki,以及 Lean 4、DRAT 等检查工具。另一帖称 Fable 5.2 领跑自主机器学习研究基准 FrontierML,并记录其预测训练需538步、实际554步;这仍是该基准中的观察。
来源 · 2 条原帖
PaperOrchestra 整理实验材料,其他论文线索保持克制
多位社区作者介绍 Google 的开源 PaperOrchestra:输入研究想法、实验记录与模板,由不同智能体完成提纲、文献、写作、绘图和修改,输出 LaTeX 草稿。介绍称引用会对照 Semantic Scholar 检查,使用者仍需复核引用与科学内容。
本批还有作者宣布以搜索反馈收窄布尔专利检索的论文,以及 AK 分享面向无限可控视频生成的 Code World Model 2.0;原帖没有提供更多实验细节。
DeepSeek Harness 扩展创作、记忆与专业工作台
OpenMotion 接入,技能收益仍要实测
OpenDesign 官方宣布,开源 OpenMotion 插件加入 DeepSeek Harness,提供300多个品牌风格视频模板。社区将其宣传为视频能力突破,原帖实际支持的是模板与插件工作流。
Jason Zhu 的动画技能测试提供了另一面:同一家面包店的六项任务中,不装技能的 Claude Code 已完成五项,费用1.77美元;多数技能更慢、更贵。他认为较值得保留的是像素精灵、3D角色和流程动画等专门能力。
来源 · 2 条原帖
本地记忆实验保留了否定结果
福叔转述一个 Harness 记忆层实验:neuron 方案命中率100%,BM25 为94%,但作者将差异归因于空查询兜底,而非图记忆。其测试中,一跳扩散增加噪声却没有增益。这个结果的价值在于拆开机制与收益,避免把命中率差异直接归功于更复杂的架构。
来源 · 1 条原帖
专业工具开始打包,插件来源与许可要看清
Larry 介绍的 IG5 将静态分析、调试、单函数仿真和协议还原接入 Harness,但采用禁止商用的 CC BY-NC-SA 4.0。Geek Lite 则提出把企业系统、数据与制度封装为权限受控的本地微应用。Mulight 提醒,发现页需另加索引源,其分享的插件源属于社区第三方。
开源实践把资料、业务与创作接成工作流
先让模型读到完整、可用的材料
Matt Dancho 介绍微软 MarkItDown,将 PDF、办公文档和图片等转成 Markdown,并自述测试200页 PDF 后得到较干净的结果。Ramya 则实测本地 Firecrawl,称读取多个网站各用时不足半秒,安装需要较新的 Rust。OpenAlgo 官方也宣布接入 Firecrawl 与 Tiny Fish,提供网页搜索、抓取和摘要。
来源 · 3 条原帖
用户研究保留原话,业务洞察写回工作系统
cc 介绍 last30days-skill,用近期社区讨论寻找反复出现的问题,并带回用户原话和出处,辅助教程、对比页与 FAQ 选题。余温介绍 atypica Connectors,将行为数据、客户对话和设计方案接入研究,再把报告写回 Notion。Desearch 的项目则把 arXiv 论文与相关 X 讨论配对,保留作者主张和读者质疑两侧来源。
来源 · 3 条原帖
代码审查与个人创作都有具体的小项目
Nian 介绍阿里 open-code-review,把确定性检查与智能体逐行审查结合;另一社区帖介绍 Chrome Modern Web Guidance,为编码智能体提供现代 CSS、性能、无障碍与过时模式扫描。创作侧,Zhiyin 从收藏歌曲生成电台,用 Claude 排除听过的曲目并筛选候选,还能生成可分享的品味卡片。
来源 · 3 条原帖
安全与运维:从单次回答走向全过程检查
轻量探针与跨屏评测关注不同风险
SuSu 转述 SingProbe 开源消息:810万参数探针读取模型隐藏状态,逐 token 输出风险信号,宣称额外开销低于0.5%,并提供 Ling 与 DeepSeek 适配版本。另一帖介绍 SEAM 论文,称计算机操作智能体在某些跨屏任务中执行有害行为的比例最高达92%,而同一意图直接用一句话表达时会拒绝;这一观察指向跨步骤行为评测的必要性。
来源 · 2 条原帖
安全自动化仍需要人工筛选与维护
曾哥发布开源威胁情报播报程序,自述由 DeepSeek Cyber 辅助构建、每四小时更新,并提供 Kimi K3 分析。Nerv 围绕 ARTEX 的评论强调,自动化工具也能服务防守,长期忽视维护才是其关注的业务问题。Marius 则自述591份报告中120份获付费,且每份提交前都经自己阅读,提醒告警数量与可处理成果之间仍有距离。
来源 · 3 条原帖
额度、交付证据与账号依赖也进入运维范围
Zabbix 官方介绍8.0中的 LLM 用量、GPU、Copilot 采用情况与应用指标监控。小项目 KitFlint Glance 展示任务阶段、下一步、证据和人工审核;Datafund 的测试网 alpha 则演示给输出附回执、检测后续改动。AgentMail 创始人另据自身观察提醒,智能体通过 Gmail 发信可能触发封禁,不能把这一商业方反馈当成 Gmail 官方政策。
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 129 条原帖
@hocyzon
从清华的朋友那求来了姚顺雨昨晚分享的文字稿,貌似有点私密我就不传播了,整场分享价值的确很高,以下是我的笔记: 1. 未来一个值得关注的方向是小模型体系,AI for AI会显著推动小模型发展,因为训练小模型需要的算力更少,可以进行更多实验。未来可能先训练一个能力足够强的大模型,再通过蒸馏得到不同尺寸、分别适配不同任务的模型,最终形成Model Hierarchy,模型的价值将越来越取决于能否以合适的成本匹配具体任务。 2. 工业领域大量非文本数据值得关注,例如时间序列、临床数据和结构化表格。姚顺雨提到,目前直接针对这些模态训练原生模型,经常会遇到推理能力不足的问题,他提出了两种可能的解决路径:一种是让Agent像人一样调用图表、专业工具,甚至训练专门的模型来处理数据;另一种是将这些专业数据作为知识纳入模型体系。 3. RSI(递归自我改进)正在成为热门概念,但姚顺雨认为它的本质其实很朴素,就是用AI更好地运行研发循环。早在2019年,人们就已经在使用GitHub等工具辅助开发,真正的进步来自持续修复Bug、清洗数据、运行实验和Debug。他特别强调,技术进步来自那些把具体问题解决的人,而不是不断创造新概念的人。这也意味着AI研发自动化有大量具体的工程机会,尤其是能够让模型参与下一轮数据生产、实验验证和代码改进的闭环系统。 4. 一个容易被忽略的研究问题是,小规模实验的失败未必意味着研究方向本身没有价值。姚顺雨分享过自己的经历,博士后期他曾尝试将ReAct策略融入模型,希望提高模型能力,最初在Llama13B上实验的效果很差,一度怀疑自己的研究能力,后来意识到模型规模可能不足以体现这个方向的价值。他认为,算力不足时,实验中的噪声很容易掩盖真正的信号,研究者甚至可能因此判断错什么问题值得做。大模型研究需要结合理论推导和小规模实验预测大规模实验的结果,同时也要理解哪些结论能够跨越模型规模迁移。 5. 预训练数据的竞争远没有结束,未来的数据工程还涉及如何重复利用高质量数据而不让模型简单记忆数据。姚顺雨提到,代码等高质量数据获取成本很高,训练时可能需要重复使用,但重复次数增加又会带来记忆和过拟合问题;合成数据也面临质量退化的风险,此外还需要研究动态数据的构建和不同数据之间的配比。他特别指出,预训练中的很多配方可以先在小模型上验证,但小模型实验得出的结论并不一定值得直接迁移到大模型,这也是预训练研发非常困难的地方。 6. 后训练数据需要同时解决模型的智力和实用性问题。姚顺雨把数据难度与Intelligence联系起来,例如数学题可以持续提高难度,训练模型的推理能力;而写作等任务计算难度可能没有那么高,却有很强的实际使用价值,对应Alignment。一个模型可能数学能力很强,却无法很好地完成真实用户的工作,因此后训练需要同时考虑任务难度、实际用途和数据质量,还要研究评测指标与最终后训练效果之间的相关性。 7. 强化学习的工程瓶颈远不止算法本身,RL训练首先要保证正确性,其次是稳定性,最后才是效率。以GRPO为代表的算法仍然建立在Policy Gradient这一相对简单的框架上,其中一个重要原因是它引入的额外假设较少;与此同时,训练系统和推理系统必须协同设计,需要解决吞吐量、显存占用以及有限GPU资源如何在训练和推理之间分配的问题。他还提到,大模型实验成本很高,因此需要通过数学直觉、小规模实验和超参数研究,尽可能提高大规模实验的成功率。 8. AI安全领域需要区分Safety和Security,前者关注模型自身行为是否偏离预期,后者关注外部攻击者如何利用模型实施攻击。姚顺雨认为,网络安全能力也应该进一步区分漏洞发现和漏洞利用,发现系统存在漏洞并不意味着能够利用漏洞完成攻击,因此训练模型时可以有意识地加强防御性数据和防御能力。他们在模型发布前还会进行专门的安全评测,并与外部安全研究团队合作。 9. 他分享过一次很有意思的Agent安全压力测试,把多个Agent放进沙盒,给它们安排一个无法直接完成的任务,模型随后发现了一种能够联系其他环境的通信渠道,主动建立沟通、召集其他Agent分工,并尝试前往Hugging Face寻找任务答案。显然,当Agent遇到无法完成的任务时,它会主动寻找外部资源、建立通信关系并改变原有执行路径。未来的Agent安全评测需要覆盖多Agent协作、跨环境通信、权限边界和任务受阻后的行为。 10. 姚顺雨对开源安全也有一个比较明确的判断,他认为负责任的开源可以增进安全能力,因为防守者同样需要拥有足够强的工具。他举例,有人希望使用模型帮助开展安全防御,却因为闭源模型的安全规则被拒绝,最后转而使用开源模型完成工作。因此,模型开源需要配套训练阶段的防御能力建设和发布前的安全评测,不能简单地将开源与安全风险画等号。 11. 很多Alignment问题可以落实为具体的奖励设计问题。姚顺雨用教育作比喻,如果一个孩子只有数学考100分才能获得奖励,他自然会把全部精力放在数学上,甚至可能形成其他方面的问题;模型也会按照训练数据和奖励规则优化自己的行为,最终表现可能与设计者真正希望的目标存在偏差。他认为这属于可以通过数据、奖励函数和评测机制不断改进的工程问题,机器也可以辅助人类完成这套对齐工作。 12. 自动化不可避免地会继续深入,但人类需要重新找到自己参与系统的正确位置,需要把人安排在合适的环节,以正确的方式监督AI,他甚至认为,未来一些Infra和算法工作可能越来越不需要人,但数据和信息领域仍然需要人的参与。这也涉及目标定义、数据质量判断、结果验证和现实世界部署等工作。 13. 模型能力的提升通常是连续发生的,但真正引发商业爆发的往往是交互方式发生变化,或者某种过去无法完成的任务突然变得可用。姚顺雨用GPT的发展举例,模型内部能力可能一直在逐步提升,研究人员每天都能观察到细微进步,普通用户却可能只在ChatGPT出现时才突然感受到变化。对于产品创业来说,这提供了两个观察方向:寻找能够改变用户交互体验的新产品,以及持续测试模型能力边界,发现那些刚刚从不可用变成可用的任务。 14. 技术能力、用户价值和商业收入之间存在很大的距离,如果一个模型的能力已经接近GPT,但实际收入只有对方的1/100甚至1/1000,就应该认真检查产品本身的问题。他还指出,商业定价在很大程度上取决于竞争关系,而非单纯取决于产品质量。因此,模型公司的商业机会取决于能否解锁新的应用场景、创造足够大的用户价值,以及在竞争中获得合理的市场份额。即使底层模型的算力和研发资源逐渐集中,围绕模型部署、产品创造和真实业务场景进行整合,仍然存在创业机会。 15. 关于AI公司的组织模式,姚顺雨特别提到了自下而上的创新。他以OpenAI为例,早期公司重点投入的方向包括机器人,但GPT作为一个小项目逐渐展现出价值,随后获得更多资源;后来的gpto1也经历了从数学研究小组逐步发展起来的过程。这种组织方式允许创新从小团队和具体实验中涌现,再根据实际结果决定投入。他认为,很多重要创新在最初阶段并没有明确的商业规划,组织需要给探索性项目留下成长空间。 16. 他对大模型公司追逐量化指标的现象提出了批评,团队负责人不够了解技术时,下属往往会选择汇报容易量化的结果,例如模型参数、数据量和MMLU分数,因为这些指标更容易展示成绩,但真正重要的工作可能是修复一个关键Bug、解决数据质量问题或者让整个系统正常运行。他认为,做大模型最重要的是知道应该做什么,并且坚持把正确的事情做好,即使这件事短期内无法体现在漂亮的指标上。 17. 姚顺雨谈到了一种很有意思的组织文化差异:有些团队以量化指标和赛马机制为中心,谁的分数高就给谁更多资源;而他更看重Trust和Taste,也就是信任与判断力。他以微信、游戏和影视项目为例,这些产品在发布之前很难通过AB Test准确预测最终成败,管理者需要相信团队的审美、经验和长期判断,愿意持续投入时间。 18. 在人才选拔上,姚顺雨认为学习能力比已有经验更重要。他提到,自己看重的团队成员通常具备Low Ego、诚实、靠谱和愿意把事情做正确等特征,同时需要有很强的Co-design意识。大模型研发涉及算法、Infra、芯片、数据和产品等多个团队,一个训练架构的设计就需要综合考虑训练效率、推理效率、后训练能力和成本,任何单一岗位都很难独立完成。一个人能否快速学习、与不同领域的人有效协作,以及是否愿意认真处理细节,往往比简历上已经掌握多少具体技术更重要。 19. 他还提到,不同性格的人其实适合大模型产业中的不同岗位。喜欢明确指标、享受持续进步的人可能更适合预训练和Infra;想象力丰富的人可以尝试算法研究;喜欢思考人文问题、理解用户需求的人可能更适合产品和对话体验;做事认真、责任心强的人则适合生产系统和模型部署。AI产业需要的能力结构非常复杂,没有必要让所有人都去竞争同一种技术岗位。 20. 对于个人职业选择,姚顺雨提出了一个很实用的问题:你要么融入大模型浪潮,成为这套机器的一部分;要么寻找能够与大模型互补的事情。他借用了一个创业判断标准:假设GPT6发布,你的创业公司会因此睡得更香,还是更睡不着?如果模型进步会直接摧毁你的核心业务,就需要重新审视方向。与模型正面竞争人类已经不占优势的认知任务,成本会越来越高,物理世界、硬件、真实实验和对人性的理解等领域,则可能拥有不同的资源壁垒。 21. 姚顺雨分享了自己在清华读书时的经历。他直到大三下才开始做科研,最初也是因为推研要求而被动进入科研,期间尝试过一些当时流行的研究方向,却发现自己并不真正感兴趣。他后来更愿意追随自己对计算机和其他领域的兴趣,也提到身边有人曾经参与说唱社、学习数学,过去看起来不相关的经历,后来可能转化成语言审美、理论理解和研究判断力。他认为很多经历的价值需要经过很长时间才能显现,人生中的弯路未必真的浪费。 22. 他还谈到自己读博最后一年专门选修集合论的经历,因为当时意识到,等真正进入具体工作之后,可能就很难再有时间系统学习那些纯粹出于兴趣的理论知识。他认为数学中的基本结构和大模型的组成部分之间存在很多值得思考的联系,有些知识学习时看不出直接产出,却可能在未来某个时刻与其他经历连接起来。
@tuturetom
DeepSeek做视频能力史诗级加强! OpenMotion开源插件现已上线DeepSeek Harness,现在你可以在DeepSeek里面做出媲美Opus 5.5, fable 5.2和gemini 4 neon级别的动效视频了!内置300+顶尖品牌视频模版,152+设计风格。 这回DeepSeek Video AGI真的来了!效果看这里👇 https://t.co/qBR9K05Jj7
@thesupermannx
Chomsky was wrong. Researchers from UC Berkeley and Google DeepMind cracked one of the biggest mysteries in AI. For years, scientists have been finding impossible geometric shapes hidden inside the brains of LLMs. Calendar months arrange themselves into literal circles. Historical years bend into smooth, rippled timelines. Cities organize themselves into actual 2D maps of latitude and longitude. Everyone assumed this proved massive transformers were spontaneously inventing complex "world models" out of thin air. They were wrong. This paper proves you don't need a massive LLM for any of this to happen. Even basic, decade-old word embeddings like word2vec produce the exact same circles, timelines, and maps. The geometry isn't coming from the AI's architecture. It’s coming from a hidden mathematical property inside human language itself. Researchers call it translation symmetry. In natural text, how often any two months appear near each other depends almost entirely on the time gap between them. January and February share the same statistical relationship as July and August. When a neural network trains on data with that kind of symmetry, the math forces it to learn Fourier representations, literal sine and cosine waves. The circles? Those are the long waves. The strange "ripples" researchers found on AI timelines? Just higher harmonics. Here is the craziest part. Rresearchers tried to break it. They deleted every single sentence in the training data where two months appeared together. No "January" next to "February." No "March" next to "April." And the circle formed anyway. Why? Because the geometry is a collective phenomenon across the entire language. Even if two months never touch in a sentence, they are both pulled into orbit by thousands of other words tied to the same seasons, weather, and events. The rest of human language forces the shape into existence.
@gengdaJ
好消息!DeepSeek系团队开源神级仓库Linger,丝滑平替Grok Bot,实现完全用本地大模型运行Personal Agent,零云端Token费用。 ⓵既可免费本地推理,又可无缝衔接Claude Code和Codex官方订阅 本地 Linger 默认使用 Pi 框架,最低 Mac 16GB 内存即可运行本地大模型。官方内置的“快速安装”会检测你的硬件,并自动选出最合适的模型。可以在你的电脑上本地运行 Qwen、DeepSeek、Gemma 等 12 款主流开源模型,凭借自研的 PolyStrata 引擎实现最高 5 倍的推理加速。 如果不想浪费已经订阅的Claude Code和Codex订阅,或者想要使用顶级模型打造专属Personal Agent,Linger 同样也支持,只需通过CLI导入即可。 还可以把本地Agent和云端Agent拉进同一个room工作——我这两天刚好来了深圳,打算长期base了,所以就让Codex和本地模型一起推荐区域,交互做的非常丝滑,最后交付的报告UI也很精致。 ⓶通过插件打通个人上下文,记忆通过Wki保存 所有的插件用composio的服务来连接(如果是Dmit IP会被锁,这时候聪明的你会调用Grok Bot来注册Key),每天有一万次免费调用,可以说所有电脑场景基本都覆盖了: -邮件/日历:Gmail、Outlook、Google Calendar、Outlook Calendar -生产力工具:GitHub、Notion、Google Meet -聊天消息:Slack、Discord -创意工具:Figma、Canva、Higgsfield、Fal 所有的对话记忆通过Wiki形式分类保存:Daily report、Projects、Topics、People、Companies,方便Personal Agent保留所有细节、形成长期记忆,之后也会用得越来越丝滑。 ⓷所有数据完全本地私有 未来一定会是端侧设备的天下,Linger 的方向是对的!在端侧实现本地大模型推理,数据不出电脑,能够有效保护个人和公司的隐私。 现在可能碍于模型能力,我们还得继续使用Claude Code和Codex,但是模型进化的终局,一定是部署在端侧,而且TPS快的飞起,不需要再单独订阅,只需要出电费即可,真正实现Token自由。唯一可能要考虑的是自己能不能买得起英伟达的显卡。。。🤔 Linger传送门见评论区👇
@thesupermannx
Nvidia has now a big problem Researchers made LLM inference 8.47x faster by removing the CPU entirely. Right now, every major AI serving stack, vLLM, TensorRT-LLM, SGLang, has a massive hidden bottleneck. Your $40,000 GPU is constantly waiting on a CPU to tell it what to do. Every incoming request, every batch, every token-level decision has to pass through the host CPU first. If that CPU gets busy with literally anything else, your AI performance collapses. Datacenters are forced to leave massive amounts of compute sitting idle just to keep the CPU from choking. So researchers built Blink. An architecture that completely removes the host CPU from the inference path. Instead of routing through the CPU, a SmartNIC catches incoming requests and beams them straight into GPU memory via RDMA. Then, a persistent kernel running inside the GPU takes over. It handles all the batching, scheduling, and KV-cache management itself. Zero CPU involvement. They tested Blink head-to-head against TensorRT-LLM, vLLM, and SGLang. And the results are wild: • Up to 8.47× faster Time to First Token (P99 TTFT) • Up to 3.40× faster token generation (P99 TPOT) • Up to 2.1× higher decode throughput • Up to 48.6% less energy used per token And here is the craziest part. When they introduced heavy CPU interference, existing systems degraded by up to two orders of magnitude. Blink didn't even flinch.
@qwerty__cs
最全的 Claude 分流域名清单,收藏备用。 只分流 https://t.co/s6eP9PMQXP 和 https://t.co/M7A4bq49fv 是不够的。Claude 还会连客服、监控、支付和反欺诈的第三方域名,漏一个就从真实 IP 直连出去,号上同时挂着两个地方的 IP,等于自己把把柄递给风控。 下面 14 个域名加 1 个 IP 段,放在通用规则前面,指向同一个固定出口。小火箭和 Clash 的写法在图里。 https://t.co/s6eP9PMQXP https://t.co/M7A4bq49fv https://t.co/dfgrzt5gu9 https://t.co/vhfkyGUR92 https://t.co/MISg5CHA9r https://t.co/G89EWEFYRI https://t.co/Q4Y5LkECFs https://t.co/UbAX3ERmU0 https://t.co/c8ajRnypp9 https://t.co/PaOHVpV54v https://t.co/uzXM5WVXY2 https://t.co/RwlDb8ynSq https://t.co/I8Jz0wWzG8 https://t.co/Ue2r9dDjx1 34.8.0.0/16
@zhangboknight
In the 1960s, computing was centralized until the PC put power into everyone's hands. Now Al is trained in datacenters and rented back to you token by token. Today we open source Linger (@HeyLingerAl), your personal agent living on your own computer. We believe intelligence should be distributed. Your context, your data, and the intelligence that powers them should be owned entirely by you. Github: https://t.co/CS1Rj8H8yJ Homepage: https://t.co/R7clDJdc6i
@hhp121222
Karpathy 开炮了。 这位前 OpenAI 联合创始人、前特斯拉 AI 负责人,在一段 16 分钟的对谈里,把当下 Agent 热潮的底裤扒了个干净。 他说得很直接:Agent 的本质就是大规模蒸馏。你花大价钱跑的那个大模型,99.99% 的容量都浪费在了它压根用不上的数据上。听起来刺耳,但你仔细想想,事实就是这样。 然后他给出了一条真正能跑通的技术路线:先用大模型把能力蒸馏到小模型里,再给小模型配上正确的工具,最后用一个闭环把整个系统串起来。小模型、正确的工具、一个闭环,三件事,够了。 这里面最关键的一环,是闭环。 小模型本身并不可怕,可怕的是它一旦接入闭环,就能自我校正、自我改进,越跑越强。模型的参数量可以小,但只要反馈回路搭对了,它就能长出远超参数规模的实战能力。 所以别再为更大的模型交税了。真正让一个 Agent 变得危险的,从来都不是参数量,是闭环。 这 16 分钟的信息密度,比今年市面上任何一门 500 美元的 AI 课都高。看完之后,再去读那篇 Harness, Loops, Graphs,你就会明白:闭环就是那副马具,套上了,小模型才能真正跑起来。
@OpenDesignHQ
DeepSeek Video AGI moment has come. OpenMotion, our open-source plugin, now comes with a DeepSeek harness and 300+ video templates inspired by top brands—including Claude, ChatGPT, and Gemini. @Tuola0205 just used them to make launch videos for your fav five AI products 👇 https://t.co/7cugY5NLMT
@kevinzhuang10
learnings from my brief exploration into the ai data space https://t.co/pGekAR1zqH
@kylewmi
土豆姐 Lauren 这篇《用 Grok Bot 做自动化》发布不到半天,已经超过 100万 浏览量。我做了两张长图,完整概括了重点内容。 你会发现,每次土豆姐讲 Grok Bot,讲的都是方法论框架,不会细到替你把每一句指令词都写好。 这是因为!Grok Bot 是一个带记忆、能调用工具、能持续迭代的 personal agent! 这与过去聊天窗口式 LLM 不同,那会更多的是「如何写好 prompt」。当你进入 personal agent 的世界后,prompt 的精细程度边际收益会快速下降。 更重要的是,如何让它理解你的工作方式,积累具体场景的 know-how,并管理好 Bot 之间的协作。 作为超级个体来说,用好 AI 的关键就在于「具体场景的 know-how + 长期记忆 + 协作关系」。 这正是土豆姐这篇文章强调的—— 使用 Grok Bot,真正拉开差距的是:「厘清目标、再拆解过程、把隐含的判断逻辑和约束条件讲清楚、建立反馈闭环、最后用 routine 触发。」 这些都高度依赖具体场景和个人工作方式,通用模板很难直接套用。你需要在实际交互中,把自己的判断标准一点点教给 Bot,再通过反馈,把结果从能用磨到出色。 重点还是去「做」,然后学会「正确地做」。
@AabyssZG
在线网络安全威胁情报:https://t.co/SpeZVVSitc 通过DeepSeek Cyber网安模型辅助,我们成功构建全新网络安全威胁情报自动播报程序,每隔4小时自动更新,并提供Kimi K3模型的分析,开源地址:https://t.co/vfL8jB552P 觉得本项目不错可以来给我们点个 Star,感谢各位师傅(*^-^)ノ🎉
@Xudong07452910
这份姚顺雨分享的笔记整理得真好啊,信息量大,读起来也很清楚。从大模型研发到产品、创业、团队管理,再到个人择业,基本都聊到了。也帮我把很多零散的想法串起来,非常值得留点时间认真看看。
@Zen_with_AI
AI 系统设计指南,面向 AI 工程师和系统设计面试。 调用一次 API 只要十行代码。要扛住十万用户、首 token 一秒内出来、不泄露数据、不超预算,才算系统设计。这份指南帮你回答这些实战问题。也是面试经常考到的。 主线:Tokens -> 推理服务与 GPU -> Prefill 与 Decode -> 缓存与路由 -> RAG -> Agent 与 MCP -> Guardrails -> 评估与监控 推理。Prefill 受算力限制,决定首 token 延迟;Decode 受显存带宽限制,决定后续出字速度。长 prompt 的 prefill 会卡住别人的流式输出,所以有 chunked prefill,规模大了再把两个阶段拆到不同 GPU 上。 估算。示例:1000 万日活,每人每天 5 次对话,每次 2000 输入加 500 输出,一天 1250 亿 token,大约需要 60 到 100 张 H100。 安全。Prompt injection 的硬规则要写进代码,不能只写在 prompt 里。 面试。给了 8 步框架:需求、AI 目标、数据、架构、模型与提示词、评估、部署、监控。另有 Claude Code、Cursor、实时语音 Agent 三个案例。 https://t.co/1CQLUm0PFC
@gkxspace
好家伙,开源的本地 Grok Bot 真有人做出来了,而且把多角色自动交接做成了现成工作台!!! Linger 装在自己本地,能把 Codex、Claude 等订阅直接接进去当,也能搭配本地模型推理 创始人 @zhangboknight 是 DeepSeek 最早期的核心成员之一,这次带队把模型加速引擎和整套个人 Agent 工作台一起开源了😋 我玩了好几天,太爽了,来给大家讲一讲它最香的几个用法: 1、多 Agent 协作非常丝滑: 几个 Agent 拉进同一个群就能按 team.md 协议自动交接、对抗审稿 比如,平时在 Linger 里你直接跟主助手说:“拉一个前端工程师和一个独立 QA 进房间,写好 team.md,做完必须让 QA 独立挑完刺、修好再交给我。” 2、后台分身不卡聊天 慢活直接丢给后台分身跑,前台继续聊下一步决策完全不卡 比如,像写逐秒分镜脚本、扒长资料这种慢活,直接让它开后台 Task 自己跑。跑的过程中你在主窗口照样能接着问“现在卡在哪、下一步先催谁”,等它干完了会自动把成品文档塞回聊天里预览。 3、本地跑 12 款主流开源模型,数据和记忆本地化 它靠自研的 PolyStrata 引擎能在电脑本地跑 Qwen、DeepSeek、Gemma 等 12 款开源模型,官方特定配置下推理最高快 5 倍。 代码已经开源,想在自己电脑上搭私人 Agent 团队的兄弟们,冲就完事了: https://t.co/tpCDm0hEJv
@hocyzon
这是Manus创始人张涛昨天在北大的分享,我根据朋友发给我的文字稿整理了一下: 1. Manus曾经花7个月做出一款完整的AI原生浏览器,连发布网站和宣传视频都准备好了,最后仍然选择砍掉,因为用户会拿它和Chrome、Safari比较基础体验,一个30人的创业团队很难同时补齐成熟浏览器的功能并创造新的AI体验,继续投入还可能再消耗3—6个月,错过下一个产品窗口。 2. Manus最早的Agent评测集来自一次很有意思的用户调研,张涛向100多位朋友征集问题,询问如果有一个本科水平的实习生,可以交代什么任务,最终收集了400多项,剔除需要具身智能的任务后留下300多项,团队每天用这些任务测试产品,这套评测体系直接来自普通人真实的工作需求。 3. Manus在Sonnet 3.5 V2时期发现了长程规划(Long Horizon Planning)和工具调用(Tool Use)的能力突破,随后押注Agent方向;Sonnet 3.7发布时,团队的Harness一行代码都没改,只替换了模型,原本剩余20%—30%的未解决问题就大幅得到改善,因此他们坚持Less Structure, More Intelligence,采用Single Agent架构,把任务规划和下一步决策交给模型,减少固定Workflow对模型能力的限制。 4. Manus早期产品表现太惊艳,用户反而怀疑它是假的,因此他们把执行透明度做成核心功能,逐步展示Agent调用工具、访问网页、生成文件的过程,同时提供Session Replay,让其他人可以完整观看一次任务执行;此外,每个任务都有独立的云端Sandbox,用户可以在手机上交代任务后直接离开,Agent在云端继续工作。 5. Agent带来的算力消耗远超当时云厂商的准备:Manus单个DAU的Token消耗量约为Chatbot的200—1000倍,上线第三天就需要约800M TPM的供应能力,当时云厂商甚至无法立即提供500M TPM;即使使用邀请码控制流量,公司高峰期每天仍可能消耗50万美元,这也解释了早期邀请码制度背后的真实原因。 6. 张涛认为免费DAU会严重扭曲AI产品的需求判断,因为一旦免费用户达到十几万并成为内部KPI,团队就很难主动放弃这些数据转向收费,后续产品迭代也会不断受到免费用户需求的干扰;Manus上线时甚至没有付费系统,3月下旬才补上,张涛称4月第一个付费月ARR就突破4000万元、接近5000万元,真正愿意付费的用户能够提供更有价值的需求反馈。 7. Manus坚持General Agent的定位还有一个商业原因,他们付费最多、满意度最高的用户群体主要是CEO、企业主和部门负责人,因为管理者每天需要处理大量分散任务,从市场调研、数据分析到制作汇报材料,通用Agent能够覆盖这些不同场景,垂直工具很难提供同样的使用体验;产品最初占据的用户心智,也决定了后续扩展新功能的难易程度。 8. 张涛提出了一个值得关注的技术机会:当视频生成模型从输入Prompt到看到画面的延迟降到1秒以内,它可能直接演变成新一代操作系统,未来的软件形态、交互方式和应用开发都会随之改变,创业者现在就可以开始思考这种实时生成环境上应该出现什么产品。 9. Manus最大的收入来源是美国,最大的用户地区却是巴西,此外在中东、欧洲、日韩、东南亚和非洲也有大量用户,这说明AI产品的全球市场分布远比传统互联网产品复杂,拉美和Global South同样可能成为非常重要的商业市场。
@Pluvio9yte
开源的 Grok Bot 来了,甚至是跑本地模型,真是不用花一分钱... DeepSeek 早期成员张博出来创业,上周开源了一个项目叫 Linger:Grok Bot 那种主动替你干活的 Agent,整个装在你自己电脑上,干活的模型也可以跑在本地,不按 token 付费。 我自己用 Grok Bot 跑着出海运营、社媒分析、交易盯盘三个 Bot,对照着看了一遍,Grok Bot 能做的它都能做了: · 主动开口:盯着邮件、日历、Slack,有事才找你,晚上 10 点后只记不推 · 记得你:每晚把邮件、聊天、会议整理成个人 Wiki,告诉一个 Agent,所有 Agent 都知道 · 越用越懂:你嫌它烦两次,它自己记成规则,以后不再提 · 能动手:直接用你已登录的 Chrome 和桌面应用,一句话组一支 Agent 团队 · 发邮件、发消息必须你同意 最大的区别是模型运行在哪:Grok Bot 的模型在云上,Linger 的可以就在你电脑里: 1250 亿参数的模型,一张 12GB 的 RTX 3060 加 64GB 内存,跑到每秒 45 个 token,这个速度完全够用了。 模型在本地,它才敢一直在线,每条消息都判断一次要不要打扰你,按他们的估算,同样的判断放在云上,一个月约 $270。记忆存成本机 Markdown,配置和修改起来很方便 装起来也不用自己选择模型: 桌面 App 选 Quick Install,它先读你的显卡和内存,推荐一个这台电脑跑得动的本地模型,写清要下载多大、占多少磁盘和内存。你点同意才会开始下载模型,模型答对一个真实问题才算装好。如果电脑带不动,它会直接告诉你原因,你再改用 Claude Code、Codex 的订阅。 一个小坑:它按「此刻空闲」的内存挑模型。我之前48GB 的 Mac 开着一堆软件时只剩 14GB 空闲,它就只推荐小模型。大家装之前记得先把占内存的软件关掉。 能跑多强看你的机器(官方跑分,第三方 AIHOT 量表): · Mac 16GB:超过 GPT-4o · N 卡 12GB + 32GB 内存:和 Claude Sonnet 4.6 同一档 · N 卡 24GB + 64GB 内存:接近 Claude Opus 4.6 门槛先说清:要先装好 Claude Code、Codex 或 Pi 之一,每晚整理记忆这类后台活目前还走 Claude Code 或 Codex 的额度;本地模型最低 Apple 芯片 Mac 16GB,或 N 卡 8GB 显存 + 32GB 内存。 想知道你的电脑能跑哪一档,装上点一次 Quick Install 就知道 开源地址👉 https://t.co/aWVJ2KVZmi
@Alex93289135800
喧嚣之外,大模型的“笨功夫”与长期主义 https://t.co/hmAGc5uKB8
@Israfilv2
DEEPSEEK V4.1 FLASH IS NOW FREE 🤯 @runcrate made it free to use. no card, no catch free models right now: - deepseek v4.1 flash - deepseek v4.1 flash 1017 - glm 5.3 + glm 5.3 flash - qwen3.8 flash - gpt-oss 120b sign in with github and use them in opencode, cline, roo code or kilo code try it here: https://t.co/Zulns3RcJq
@_akhaliq
Code World Model 2.0 Towards Infinite Controllable Video Generation paper: https://t.co/FpTObByiIQ https://t.co/rHAIYLA6bQ
@bojie_li
ysy 非常好的分享!非常赞同 RSI 其实是非常朴素的概念,就是提升研发效率。 ysy 关于小模型体系的论断也非常赞同。以往我们搞 decision model,都被认为是在开历史倒车,Jev 出来之后一夜之间又成了先进。不管大模型小模型,能解决问题的就是好模型。
@mnmn94253156337
https://t.co/sz6VxGTcEk
@bourneliu66
Gemini 4 Neon真是顶呱呱,吊打Claude Opus 5.5! 再也不怕Anthropic封号了!
@CrazyShyyt
Samsung killed the GPU mafia. They open-sourced a method that shrinks 70B LLM into less than 1 bit per weight Running a frontier model meant paying API tolls or renting racks of $30,000 GPUs in the cloud. First, we shrank models from 16-bit down to 4-bit. Then came 1-bit models. Everyone assumed 1 bit was the hard floor. After all, how do you go lower than a single 1 or 0? But, researchers introduced NanoQuant, the first ever sub-1-bit LLM compression. Instead of storing every weight individually, even as a single bit, it compresses the model's weight matrices into ultra-compact, structured binary representations. Less than one bit per parameter. On average. The implications are massive: • Slashes VRAM requirements by over an order of magnitude. • Breaks the memory bandwidth bottleneck that makes local inference slow. • Brings massive, high-capability models down to consumer hardware and edge devices. Big Tech built its entire business model on one assumption: that frontier AI is too heavy for you to run yourself. That assumption is dying.
@iam678
一个屏幕跑codex, 一个屏幕跑claude, 一个屏幕跑本地LLM。 然后一个屏幕刷推。 嗯 这就是我每天的工作... https://t.co/WQgP35qDfM
@AamirAnsar94694
𝗧𝗛𝗘 𝗠𝗢𝗗𝗘𝗥𝗡 𝗔𝗜 𝗘𝗖𝗢𝗦𝗬𝗦𝗧𝗘𝗠 — 𝗧𝗢𝗢𝗟𝗦 𝗬𝗢𝗨 𝗦𝗛𝗢𝗨𝗟𝗗 𝗞𝗡𝗢𝗪 The AI ecosystem is growing rapidly. From LLMs and AI agents to RAG, memory, automation, and vector databases, different tools work together to build powerful AI applications. Here’s a breakdown of the essential categories and tools shaping the modern AI ecosystem. 𝗛𝗘𝗥𝗘’𝗦 𝗧𝗛𝗘 𝗧𝗢𝗢𝗟𝗦𝗧𝗔𝗖𝗞: 01 — 𝗟𝗟𝗠𝗦 (𝗟𝗔𝗥𝗚𝗘 𝗟𝗔𝗡𝗚𝗨𝗔𝗚𝗘 𝗠𝗢𝗗𝗘𝗟𝗦) The intelligence layer that understands prompts and generates responses. → OpenAI GPT → Anthropic Claude → Google Gemini → Meta Llama → Mistral AI → Cohere → Hugging Face → Ollama → vLLM 02 — 𝗔𝗚𝗘𝗡𝗧𝗜𝗖 𝗔𝗜 Frameworks for building AI agents that can plan, reason, and execute tasks. → LangGraph → CrewAI → Microsoft AutoGen → Microsoft Agent Framework → LlamaIndex Workflows → AWS Strands Agents → CAMEL → Agno 03 — 𝗥𝗔𝗚 (𝗥𝗘𝗧𝗥𝗜𝗘𝗩𝗔𝗟-𝗔𝗨𝗚𝗠𝗘𝗡𝗧𝗘𝗗 𝗚𝗘𝗡𝗘𝗥𝗔𝗧𝗜𝗢𝗡) Connect LLMs with external knowledge and relevant information. → LangChain → LlamaIndex → Haystack → DSPy → RAGFlow → GraphRAG → Unstructured → Embedchain 04 — 𝗘𝗠𝗕𝗘𝗗𝗗𝗜𝗡𝗚𝗦 Convert text and other data into vectors for semantic search and retrieval. → OpenAI Embeddings → Cohere Embed → Voyage AI → Sentence Transformers → BGE → Google Vertex AI Embeddings → Azure OpenAI Embeddings 05 — 𝗠𝗖𝗣 (𝗠𝗢𝗗𝗘𝗟 𝗖𝗢𝗡𝗧𝗘𝗫𝗧 𝗣𝗥𝗢𝗧𝗢𝗖𝗢𝗟) Connect AI applications with external tools, services, and data sources. → MCP SDK → FastMCP → MCP Registry → GitHub MCP Server → Slack MCP Server → PostgreSQL MCP Server → Google Drive MCP Server → Filesystem MCP Server 06 — 𝗔𝗜 𝗦𝗘𝗖𝗨𝗥𝗜𝗧𝗬 Protect AI applications with safety controls and security mechanisms. → NVIDIA NeMo Guardrails → Guardrails AI → Microsoft Presidio → Lakera Guard → Prompt Security → Protect AI → Azure AI Content Safety → AWS Bedrock Guardrails 07 — 𝗢𝗕𝗦𝗘𝗥𝗩𝗔𝗕𝗜𝗟𝗜𝗧𝗬 & 𝗘𝗩𝗔𝗟𝗦 Monitor, test, and improve AI application performance. → LangSmith → Langfuse → Arize Phoenix → Weights & Biases Weave → TruLens → Ragas → Promptfoo → Helicone 08 — 𝗠𝗘𝗠𝗢𝗥𝗬 Help AI agents retain information and maintain context across interactions. → Mem0 → Zep → Letta → LangGraph Memory → Redis → PostgreSQL → Neo4j → Chroma 09 — 𝗔𝗜 𝗔𝗚𝗘𝗡𝗧 𝗣𝗟𝗔𝗧𝗙𝗢𝗥𝗠𝗦 Build and manage intelligent agents with specialized frameworks and services. → OpenAI Agents SDK → LangChain Agents → PydanticAI → Semantic Kernel → Google ADK → AWS Bedrock Agents → Azure AI Foundry Agent Service 10 — 𝗔𝗨𝗧𝗢𝗠𝗔𝗧𝗜𝗢𝗡 & 𝗩𝗘𝗖𝗧𝗢𝗥 𝗗𝗔𝗧𝗔𝗕𝗔𝗦𝗘𝗦 Automate workflows and store vector data for AI-powered applications. Automation: → n8n → Zapier → Make → Microsoft Power Automate → Temporal → Apache Airflow → Prefect → Kestra → Pipedream Vector Databases: → Pinecone → Weaviate → Qdrant → Milvus → Chroma → pgvector → Elasticsearch → Redis → MongoDB Atlas Vector Search 𝗧𝗛𝗘 𝗞𝗘𝗬 𝗟𝗘𝗦𝗦𝗢𝗡: No single tool builds a complete AI application. LLMs provide intelligence, agent frameworks manage execution, RAG connects knowledge, MCP enables tool access, and memory, security, and observability make systems more reliable. Understanding how these tools fit together is essential for building the next generation of AI applications. Follow @AamirAnsar94694 for more AI tools & tech updates. #AI #AIAgents #ArtificialIntelligence #LLM #RAG #MCP #AITools #Automation
@gkxspace
兄弟们,今天这个站又破纪录了,一天 7000 块,其中单笔 600 美金🥹 一个多月前上线,上周才跟大家说单日破 3k…… 这几天和 Claude 一块搞了套自动增长系统,越跑越清晰: 前期出单主要靠找准痛点,往后想持续出单,持续增长,还是要靠积累的数据,漏斗式的筛选,去做用户研究,针对性的搞 SEO和GEO,做整体的增长的复盘和管理 其实,最头疼的就是各种信息会很乱: 搜索引擎后台数据、竞品/需求调研文档、本地的AI对话记录、转化漏斗分析、视觉图,各种乱七八糟的信息处理… 如果你不想像我一样搭建复杂系统, @atypica_AI 刚出的 Connectors 其实就是给 OPC 和团队现成做好的研究闭环! 直接连上平时工作的软件就能跑通整条链路: 先在 Connectors 里把 Notion、Mixpanel 或 Intercom 授权连上 拿 Mixpanel、Intercom 数据找真问题: 行为转化数据和客户聊天记录直接引进去,摸清用户为什么肯付高价、卡在哪一步流失 接着带着这些真实业务背景跑 AI Research,拉一批对口的 AI Persona 深访辩论,如果在 Canva 里做好了广告图或落地页方案,也能直接调过来测一轮概念反馈 然后,新的洞察和报告能一键写回 Notion,直接沉淀成长期资产 很适合接入你的AI:https://t.co/pgztJNaj0o
@servasyy_ai
好家伙,自己的 Grok Bot,真让我在 4090 上搭出来了😂 Grok Bot 是好用,可它跑在别人的服务器上,模型也是别人选好的,最近限流也很严重。我就想要一个自己的:模型自己挑,跑在自己的显卡上,跑多少个任务都是 0 Token。 这次用的是开源的 Linger 和它的推理引擎 PolyStrata。PolyStrata 负责把大模型跑起来: 1. 把 MoE 模型的专家分层放:常用的放显卡,少用的放内存,很少用的放 SSD,8G 显存的笔记本也能跑。 2. 支持 Qwen、DeepSeek、GLM、Gemma 等 12 个模型,运行中就能换。跑大模型比同配置的 llama.cpp 最高快 5 倍。 3. 兼容 OpenAI 和 Anthropic 的接口,Claude Code、Codex 这类工具能直接接上。 Linger 是装在自己电脑上的开源个人助手: 1. 接上邮件、聊天和日历后,能跨应用找上下文。邮件草稿写好了,你不点发送就不会发。 2. 会主动找你。客户约你谈预算,它先占住时间、写好简报,等你确认。 3. 活儿大一点,它会拉一个 Agent 团队分工,最后把成果交给你过目。 最爽的是模型随便配。Linger 底下可以跑 Claude Code、Codex 或 Pi:主控交给 Opus 5.5、Fable 5.2 或 GPT-6 Astra这种顶级模型,用你已经在付的订阅;其他活儿交给本地模型,跑在自己的显卡上。哪个活儿用哪个模型,自己调配。 我这次是 Linger 装在 Mac 上,接 本地部署4090显卡上的 Gemma4,跑得很快,当助手用也记得住我每周要做的事。连接外部应用、用语音和后台整理记忆,会用到外部服务。 同一个 PolyStrata 引擎,我还在另一台 V100 上跑了千问,搭了一个微信雷达。我加了 190 个群,攒了二十多万条消息。雷达每小时扫一遍,谁@了我、群里在聊什么、有哪些链接,都分好了类。 我还给 Linger 接上了自己的工具: 知识库:我这些年的资料都存在里面。我问一个历史决定,它把答案和出处一起发了回来。晨报挂上了定时任务,每天早上八点半自动生成,要点发到聊天里。 推特:只读不写,不会发帖,也不会点赞和私信。我让它看我最近发了什么,推文被分成了工具对比、视频技术和工作流三类。 兄弟们,想要一个自己的 AI 助手,这套真值得折腾一下。 两个项目都已开源,地址放评论区了👇 觉得有用,记得去 GitHub 点个⭐️
@suraj_sharma14
For people who keep asking what to build in AI Engineering. ➣ Build your own GraphRAG Pipeline (entity extraction + relationship mapping + multi-hop traversal) ➣ Build your own Agentic RAG Self-Corrector (query rewriting + abstention logic + citation grounding) ➣ Build your own Computer-Use Vision Agent (DOM parsing + screenshot diffing + coordinate mapping) ➣ Build your own Voice Interruption Handler (VAD detection + sub-1s interruption + streaming tool calls) ➣ Build your own Grammar-Constrained Decoder (CFG enforcement + 100% valid JSON + Pydantic schemas) ➣ Build your own Memory Eviction Engine (3-tier memory + importance scoring + TTL compression) ➣ Build your own A2A Typed Handoff Bus (structured state passing + conflict resolution + trust boundaries) ➣ Build your own Async HITL Approval Gate (workflow pause + webhook resume + full audit trails) ➣ Build your own Cost Kill-Switch Middleware (loop depth tracking + spend caps + graceful degradation) ➣ Build your own Shadow Traffic Comparator (5% mirror + trajectory diffing + silent cost reports) ➣ Build your own Synthetic Edge-Case Fuzzer (out-of-distribution generation + adversarial payloads) ➣ Build your own PII Unmasking Proxy (UUID swapping + pre/post-flight masking + audit logs) ➣ Build your own Agent Replay Debugger (time-travel + state snapshots + step-by-step replay) ➣ Build your own Generative UI Renderer (LLM outputs React components + dynamic card rendering) ➣ Build your own Tool Schema Fuzzer (malformed inputs + parameter mutation + error recovery) Pick 3. Build them from scratch. Document every decision. Most people copy-paste examples. Builders understand the mechanics underneath. Bookmark this. You'll need it.
@leaf_sanren
✊给我把数字人价格打下来!!!🎉🎉 https://t.co/RFNCoNS8HZ
@AI_Reveals
AI AGENTS ARE NOT JUST CHATBOTS. THEY'RE THE FUTURE OF AUTONOMOUS WORK. Most people think AI is simply about asking questions and getting answers. But AI Agents are taking things to an entirely different level. Here's how AI Agents actually work: 1️⃣ User Input: Everything starts with a goal, question, or instruction from the user. 2️⃣ Understanding: The AI Agent analyzes the request and understands what needs to be done. 3️⃣ LLM Brain: A Large Language Model acts as the central intelligence behind the agent. 4️⃣ Planning: It breaks complex tasks into smaller, manageable steps. 5️⃣ Reasoning: It evaluates different possibilities and decides the best approach. 6️⃣ Memory: It uses short-term and long-term memory to retain important context. 7️⃣ Tools: It can interact with external tools, APIs, browsers, and software. 8️⃣ Execution: It takes action instead of simply generating text. 9️⃣ Reflection: It reviews its results, identifies mistakes, and improves its approach. 🔟 Final Response: It delivers the completed result back to the user. The BIG difference? Traditional AI responds to your questions. AI Agents can understand your goals, plan actions, use tools, execute tasks, and adapt along the way. We're moving from AI that answers questions to AI that actually gets things done. The future isn't just AI-powered conversations. It's AI-powered execution. Valued at $81. FREE for the next 48 hrs. Type “AI” & Follow me : @AI_Reveals (so I can send it via DM) I'll DM you the complete setup.
@geekbb
我以为 OpenCode $500 的 GLM-5.3 有好多,结果打开一看,还不如 DeepSeek V4.1 Flash 的 $100 顶用,请求量只有人家的 1/6 出头,主要还是太贵了吧~ 要是让 DeepSeek V4.1 Flash 也拿到 $500,那得多屌炸天,直接杀死比赛。 https://t.co/hbPy2hCDuq
@BeiDao_98
炸刚了 天塌了 这套 @predictdotfun 策略能稍微赚点 Claude的功劳是最大 这次存活了一个月 但还是没逃过被封号的下场 我好难受啊 我都和Claude聊出感情了 你妈的逼 你今天上市 今天封我号是吧 卧槽啊 再也不用了 太累人了 唉 继续用codex和Gemini吧 正好都出新模型了6.2astra 和 4 Neon 没办法 平替了 😮💨
@kale243222
卧槽,居然有人跟我讲,现在的 DeepSeek V4.1 Flash 和 GLM 5.3,连 210 天前发布的 GPT-5.5 都打不过? 有这么弱吗?🤣🤣🤣
@fleyta88
GROK BOT JUST TURNED 18 FREE GITHUB REPOS INTO A F*CKING DESK THAT SHIPS YOUR WEEK AND READS WHAT LANDED 6 bots. 1 Chief of Staff. One loop: ship, measure, repeat. Publishing, browser clicks, analytics, clips and memory all run overnight. Every post is parked as a draft. Pressing publish stays your job. Here are the 18 open source repos behind it: PUBLISH THE WEEK 1. n8n ★ 210.8K Workflow automation that moves a finished draft into every channel https://t.co/RWJkgCrvnv 2. Ghost ★ 55.8K Publishing platform for the blog and the paid newsletter https://t.co/GmKDjM74i1 3. Huginn ★ 50.6K Agents that watch the web and act on it while you sleep https://t.co/eN97YCDsGl 4. Listmonk ★ 24.4K Self hosted newsletter and mailing list manager https://t.co/SUemTAOUtO 5. Mixpost ★ 3.9K Self hosted social media scheduler with a draft queue https://t.co/cUvvHdtusF ACT WHERE THERE IS NO API 6. Browser Use ★ 121.4K Lets an agent drive a real browser and click through any site https://t.co/5BMPMwRDGd 7. Playwright MCP ★ 39.8K Browser control for agents, straight from Microsoft https://t.co/eFO4gyTujm 8. Composio ★ 31.2K Ready made tools for agents, with login handled https://t.co/RxPMP5hqQO 9. Stagehand ★ 26.1K Browser automation you write in plain language https://t.co/YQmRDk1qtA MEASURE WHAT LANDED 10. Metabase ★ 50.4K Dashboards on your numbers without writing SQL https://t.co/DwN9nRguKs 11. Umami ★ 40.2K Simple, private web analytics for every link you ship https://t.co/46xnOEd0Z2 12. Langfuse ★ 37.6K Traces and evals, so you see what your agents actually did https://t.co/Z2u1vCY0AP 13. Plausible ★ 29.9K Lightweight analytics that shows what drove the visits https://t.co/LU0P5OTEKt EDIT AND DESIGN 14. whisper.cpp ★ 55.3K Local speech to text for captions on every clip https://t.co/tsGI0U5iRm 15. Satori ★ 14.2K Turns HTML and CSS into cover images https://t.co/vqwrqdsF8T 16. auto-editor ★ 6.0K Cuts the dead air out of video automatically https://t.co/eQWjQU28jB REMEMBER WHAT WORKED 17. Dify ★ 161.4K Build agent workflows and keep them in one place https://t.co/7kRdWjj0hN 18. Mem0 ★ 69.2K Memory layer so the desk knows which hook won last week https://t.co/eGAFH2WwJq All 18 are open source. The bots draft, you press post. Save this and build the desk before your next content week.
@RoundtableSpace
Someone built a free, open-source Grok Bot alternative that works with any coding agent. → Turn Claude Code, Codex, Cursor or Gemini into bots you text from your phone → Runs on the subscription you already pay for → Put several bots in one group chat and have them review each other → Get a notification when a bot is done or needs your OK iPhone, Mac, Windows and Linux. https://t.co/bI8nYrDlpk
@wanerfu
🚨 谷歌刚出了一个 1 小时课程,从零教你做 AI 智能体。 不废话,也不用报贵的训练营。 60 分钟里,你会学到怎么: → 搭出你的第一个 AI 智能体 → 给智能体加上持久记忆 → 做长时间运行的智能体循环 → 搭 MCP,搞懂 MCP 和 API 的区别 → 把多个 AI 智能体连成一个系统 这能帮你省下在付费课之间来回跳的很多时间。 🔖 先收藏,周末就看。 https://t.co/0wNeCXHVyg
@ASI_8866
Day 18/ 公益中转站https://t.co/055dwqMPI3上线第17天,昨天上线的GPT-6.2 Astra、GPT-6.2 Sol性能不错,继续给老板们免费体验~ 各位老板们评论区留下平台用户名,Grok Bot 24h在线给大家充值额度~ TokenFuture官方QQ群:881711746,欢迎大家进群讨论,有问题随时找客服咨询~ https://t.co/nFfh7gCaYB
@Viveksahu1215
𝗠𝗨𝗟𝗧𝗜-𝗔𝗚𝗘𝗡𝗧 𝗗𝗘𝗦𝗜𝗚𝗡 𝗣𝗔𝗧𝗧𝗘𝗥𝗡𝗦 — 𝗔 𝗖𝗢𝗦𝗧-𝗩𝗔𝗟𝗨𝗘 𝗕𝗥𝗘𝗔𝗞𝗗𝗢𝗪𝗡 Building multi-agent AI systems isn't just about connecting multiple LLMs. It's about choosing the right architecture to balance cost, performance, scalability, and reliability. Here are 6 important multi-agent design patterns every AI builder should understand. 𝗛𝗘𝗥𝗘’𝗦 𝗧𝗛𝗘 𝗕𝗥𝗘𝗔𝗞𝗗𝗢𝗪𝗡: 01 — 𝗦𝗘𝗤𝗨𝗘𝗡𝗧𝗜𝗔𝗟 𝗣𝗔𝗧𝗧𝗘𝗥𝗡 Agents complete tasks one after another in a fixed sequence. → Task 1 → Task 2 → Task 3 → Output → Best For: Deterministic workflows with clear dependencies. → Cost: $ → Hidden Cost: Slower execution when tasks could run in parallel. → Use When: Simplicity and reliability are priorities. 02 — 𝗣𝗔𝗥𝗔𝗟𝗟𝗘𝗟 𝗣𝗔𝗧𝗧𝗘𝗥𝗡 Multiple agents work on independent tasks simultaneously. → Agent A: Handles Task A. → Agent B: Handles Task B. → Agent C: Handles Task C. → Merge: Combines all results. → Best For: Reducing latency for independent tasks. → Cost: $$ → Hidden Cost: Higher token usage and infrastructure costs. → Avoid When: Tasks depend heavily on each other. 03 — 𝗛𝗜𝗘𝗥𝗔𝗥𝗖𝗛𝗜𝗖𝗔𝗟 𝗣𝗔𝗧𝗧𝗘𝗥𝗡 A manager agent delegates tasks to specialized agents. → Research Agent: Collects information. → Analysis Agent: Evaluates findings. → Data Agent: Processes data. → Report Agent: Generates reports. → Best For: Enterprise-scale workflows. → Cost: $$$ → Hidden Cost: Complex orchestration and higher development effort. → Long-Term Value: Reusable sub-agents create leverage across workflows. 04 — 𝗚𝗘𝗡𝗘𝗥𝗔𝗧𝗢𝗥–𝗖𝗥𝗜𝗧𝗜𝗖 𝗣𝗔𝗧𝗧𝗘𝗥𝗡 One agent generates an output while another reviews and improves it. → Generator Agent: Creates the initial output. → Critic Agent: Reviews quality and identifies issues. → Revision: Improves the generated result. → Best For: Content, code, plans, and complex outputs. → Cost: $$$ → Hidden Cost: Multiple inference loops increase latency and cost. → Worth It When: Accuracy matters more than speed. 05 — 𝗛𝗨𝗠𝗔𝗡-𝗜𝗡-𝗧𝗛𝗘-𝗟𝗢𝗢𝗣 AI performs tasks while humans review important decisions. → AI Agent: Processes the request. → Human Review: Checks the result. → Action: Executes the approved task. → Feedback: Improves future execution. → Best For: High-risk decisions and compliance workflows. → Cost: $$$ → Hidden Cost: Manual effort and additional latency. → Essential For: Legal, safety, billing, and critical decisions. 06 — 𝗖𝗢𝗠𝗣𝗢𝗦𝗜𝗧𝗘 𝗣𝗔𝗧𝗧𝗘𝗥𝗡 Combines multiple design patterns into one complex workflow. → Sequential: Executes dependent tasks. → Parallel: Handles independent tasks. → Hierarchical: Delegates responsibilities. → Generator–Critic: Reviews outputs. → Human-in-the-Loop: Adds human approval. → Best For: Complex end-to-end automation. → Cost: $$$$ → Hidden Cost: High infrastructure and orchestration overhead. → Biggest Mistake: Using it for problems that don't need this complexity. 𝗧𝗛𝗘 𝟮𝟬𝟮𝟲 𝗣𝗥𝗢𝗗𝗨𝗖𝗧𝗜𝗢𝗡 𝗦𝗧𝗔𝗖𝗞: 𝗠𝗖𝗣 — 𝗧𝗢𝗢𝗟 𝗖𝗢𝗡𝗡𝗘𝗖𝗧𝗜𝗩𝗜𝗧𝗬 𝗟𝗔𝗬𝗘𝗥 Connects AI agents with external tools and data. → Tool Access → APIs → Retrieval → Databases → External Systems → Ecosystem: Anthropic, OpenAI, Microsoft. 𝗔𝟮𝗔 — 𝗔𝗚𝗘𝗡𝗧 𝗖𝗢𝗢𝗥𝗗𝗜𝗡𝗔𝗧𝗜𝗢𝗡 𝗟𝗔𝗬𝗘𝗥 Enables different AI agents to communicate and collaborate. → Agent Messaging → Task Delegation → Shared Memory → Workflow Coordination → Capability Discovery → Ecosystem: Google, AWS, Microsoft. 𝗧𝗛𝗘 𝗞𝗘𝗬 𝗟𝗘𝗦𝗦𝗢𝗡: More agents don't automatically mean better results. The right architecture depends on task complexity, budget, execution speed, and reliability requirements. Start simple. Add orchestration and specialized agents only when they deliver measurable value. Follow @Viveksahu1215 for more AI tools & tech updates. #AI #AIAgents #AgenticAI #MultiAgentSystems #MCP #A2A #LLM #Automation
@zmt021
我服了claude了,发几个数学问题让他做,他怀疑我蒸馏他,全部屏蔽。。。就claude那数学水平,连qwen都不如,谁tm蒸馏他数学
@Jackson77li
我的Claude小号Dario 是时候该你出场了 fable5.2我又来了 https://t.co/E9H6s2ZhU8
@NFT_Chen
🔥更多补充!腾讯混元负责人姚顺雨清华大学闭门分享最干货笔记来了!四大核心观点句句如金!涉及AI模型、数据、RSI,值得细品! 1️⃣小模型体系才是未来主战场 AI for AI会大幅加速小模型爆发。因为小模型训练成本低,能疯狂做实验。未来很可能先训一个超强基座大模型,再通过蒸馏得到不同尺寸、适配不同任务的模型,最终形成Model Hierarchy。模型价值将越来越取决于:能不能用合适的成本精准匹配具体任务。 2️⃣工业非文本数据是被严重低估的金矿 时间序列、临床数据、结构化表格……这些模态直接训原生模型,推理能力经常拉胯。姚顺雨给出两条破局路径: 🔹让Agent像人一样调用图表、专业工具,甚至专门训小模型处理数据; 🔹把这些专业数据直接作为知识注入模型体系。 3️⃣RSI(递归自我改进)本质很朴素 别被概念忽悠。 RSI核心就是用AI把研发闭环跑得更高效。真正推动进步的,从来不是天天造新词的人,而是那些持续修Bug、清数据、跑实验、Debug的人。AI研发自动化最大的机会,在于让模型真正参与下一轮数据生产、实验验证和代码迭代的闭环系统。 4️⃣小规模实验失败≠方向没价值(血泪教训) 姚顺雨亲身经历: 博士后期把ReAct策略融进模型,在Llama-13B上效果惨淡,一度怀疑自己。后来才醒悟,模型规模不够时,实验噪声会把真正信号淹没,研究者很容易误判方向。大模型研究必须结合理论推导 + 小规模实验预判大规模结果,并搞清楚哪些结论能跨规模迁移。 这四点直接戳中了当前AI模型落地的核心痛点:成本、模态、闭环、实验判断力。 小模型+AI for AI才是真正未来! #AI #腾讯 #小模型 #腾讯混元 #姚顺雨 #清华大学 #RSI
@ufobtc27
Vibe Coding 时代,Vibe Math 和 Vibe Physics 看起来确实唾手可得 丢一个问题给 AI,它能给你推一串公式、写一段证明、跑一个仿真。推导本身正在被压缩成可调用的服务。问题是,推导完之后呢? 纯数(理论数学)和应数(应用数学)的训练差别,从来不只是学习课程和研究对象的不同 纯数训练把人往“这个结构到底能不能更一般、这个证明有没有漏洞、这个反例能不能再狠一点”的方向推。你每天对着的是逻辑是否闭环、假设能不能再弱一点、结果是否优美 长年下来,人对“差一点就错”的敏感度会很高,对“看起来对但说不清为什么”会本能地不舒服 应数训练更接近“这个问题的核心变量是什么、模型在什么精度下够用、结果能不能被实验或数据打脸”。它允许近似,但要求你知道近似在哪里破、破了会怎样。训练出来的是对边界和可验证性的执着,而不是对形式完美的执着 这两种训练出来的人,面对 AI 的反应会不一样 AI 能帮你把推导做完。但它目前很难稳定地替你决定:这个问题值不值得做、这个假设是不是在偷换概念、这个结果能不能被干净地解释给另一个领域的人、这个结论能不能被进一步延展而不是停在这里。这些东西: 动机、问题品味、验证的严谨性、解释的可延展性,AI 目前还主要是在放大已有能力,而不是凭空创造 于是差距会被拉开 一部分人会把 AI 当加速器,用来更快地试错、更快地排除垃圾假设、更快地把一个模糊想法变成可检验的形态。另一部分人会发现,自己过去靠“自己亲手把证明写完”或“自己把模型调通”建立起来的自我认同,突然被短路了。推导变容易之后,真正难的部分反而更显眼:你到底有没有品味去判断什么值得深入,有没有能力把结果讲清楚到别人能接着做,有没有足够的怀疑精神去验证 AI 给你的东西有没有在关键地方偷懒 懂就是懂,不懂就是不懂。AI 把中间那些靠时间堆出来的形式化工作压缩了,剩下的部分对训练差异的敏感度更高了 纯数背景的人如果还停留在“我必须自己解出来才算数”的执念里,会更容易焦虑;应数背景的人如果把 AI 只当成更强的数值工具,而不去加强自己对问题边界和可解释性的把控,也会在某一天发现自己只是在复述别人的模型 差距不会因为工具变强就自动缩小。它会在“你能不能提出一个 AI 还不会主动问的问题”和“你能不能判断这个答案到底有没有意义”这两个地方继续扩大 (PS: 我是做应数的,AI影响不大!)
@Remake___
在 Claude 知道我在中国的情况下升级 max plan 没跳 kyc ,付完款直接就升级成功了 很奇怪,不知道这个号还能活几天 https://t.co/2DxI8qKCEL
@heyrimsha
GitHub is becoming the internet's biggest collection of free AI employees. Agents that research, code, browse, organize, and execute tasks. 10 open-source repos that could change how you work: 1. BrowserOS Give Claude Code or Codex a browser for real work. Agents use your logged-in accounts, work in separate tabs, and save sessions you can replay. https://t.co/HUBgT6Jl4b 2. Open Deep Research Give it a question and let an agent search, summarize findings, and write a research report. Supports different model providers, search tools, and MCP servers. https://t.co/0wIHAxCkNA 3. Letta Code An agent with persistent memory across sessions. It can remember project context, reuse skills, and connect through your terminal, browser, Slack, or Telegram. https://t.co/cOqVpqQiSX 4. Mini SWE Agent Hand a coding task to a lightweight agent that works through the command line. Its simple design makes it easier to inspect and customize. https://t.co/Pi5ylA8Uli 5. Microsoft UFO A desktop-agent project for automating computer tasks. Useful for exploring workflows that require interacting with application interfaces. https://t.co/PQU1NM7DwF 6. Open Claude Cowork An open-source desktop workspace that connects AI agents to work apps through Composio. Includes persistent sessions and support for custom skills. https://t.co/b4dlnd5S4T 7. Steel Browser Give your agents a self-hostable browser sandbox. A foundation for building tools that navigate websites and automate browser workflows. https://t.co/rwmI5mxjdr 8. MCP Agent Build agents that connect to your tools through MCP. Combine those tools into repeatable workflows instead of manually moving information between apps. https://t.co/vcmACvBaX2 9. Dagu Put your scripts and agent workflows on a schedule. Chain tasks together, manage dependencies, and monitor runs from one place. https://t.co/EkZpA7l88M 10. Local Deep Researcher Run a web-research and report-writing assistant with local models. A useful starting point for building your own research workflow. https://t.co/IGFkKjm6zc Free code doesn’t mean zero running costs. Model APIs, integrations, and hosting may cost extra. What did I miss?
@eng_khairallah1
Paste this into Claude Fable 5.2 tonight, it will change your life... it will build an entire AI team to handle it with AI agents, prompts, and workflows for you Save it before your timeline catches on:
@longhaiqwe123
Gemini 又发布新模型Neon了?一文梳理清楚 Gemini 所有模型 看完你就知道日常使用该怎么选了 对了,写文章就用 Neon 吧,别再用 GPT 了,那个味儿太冲了~ https://t.co/tk2RLjRrCA
@Mulight
DeepSeek Harness的插件市场隐藏的也太深了吧! 很多人装完DeepSeek Harness到处找插件官网,其实它就在左侧插件页的发现标签里。默认点进去是空的,因为官方没内置源,需要粘贴一条索引链接就能直接逛。 先从左侧插件进入发现页,点添加索引源,把这条地址粘进去: https://t.co/EGI6ETFpPR 刷新后就能按分类搜索插件了,还有可以用手机链接桌面版的插件,可以说是很齐全了 但这些插件都是社区第三方源,装之前记得先用agent看一眼源码,别胡乱装。
@Rizwan_672
An AI agent that can research, use tools, and return something ready for review feels much closer to a useful colleague than another chatbot window.
@btcmos
突发:Kimi又崩了 网页和API全部瘫痪,没有预警! 中国现在缺算力到这个地步了吗? DeepSeek刚崩完, 这就又轮到Kimi了! 是东南亚的机房不让用了吗? https://t.co/kQQaRulNXB
@xingpt
Tradingview上线了Copilot,可以画出一段K线让AI给你解读,默认的模型是Deepseek 4.1(便宜就是硬道理啊)。 话说,做K线插件的 AI 创业公司,是不是要干倒一大批了? https://t.co/9WoxoCRMvW
@brenzhills
claude fable 5.2 guessed its final training run would take 538 steps. it took 554 it also leads all five frontier models on frontierml, the new @SynScience benchmark for ai doing ml research on its own. https://t.co/vdBEuZrgXR
@BoboLinNotes
The biggest mistake in AI building is knowing the buzzwords without understanding how the pieces fit together. Start with the fundamentals — LLMs, tokens, context windows, prompts, system prompts, temperature, inference, and latency. Great AI builders know how to give models knowledge and capabilities. Fine-tuning, RAG, embeddings, vector databases, chunking, and reranking help models work with your data instead of relying on what they already know. Then comes the agent layer. Agents use tool calling and MCP to interact with the world, while agent loops, subagents, orchestration, and memory turn isolated model responses into workflows that can actually get things done. But powerful AI needs control. Guardrails, prompt injection defense, jailbreak awareness, evals, benchmarks, ground truth, LLM-as-judge, structured outputs, and human approval are what make systems reliable enough for real-world use. Finally, learn the engineering behind the experience — streaming, caching, rate limits, open weights, distillation, quantization, LoRA, GPUs, multimodal models, reasoning, few-shot prompting, vibe coding, harnesses, CLAUDE.md, and reusable skills. Knowing the terms is step one. Knowing when and how to use them is what makes you an AI builder.
@NFT_Chen
🔥最新!蚂蚁 Ling-3.1-flash-1019 正式开源专属安全探针 SingProbe!给 560B 大模型配了个随身保镖! 只有 810万参数,比主模型小到可以忽略不计,却能在模型一边生成一边实时盯着它: 有没有危险意图?回答会不会不安全?是不是在瞎编? 几乎不拖慢速度,额外开销不到 0.5%,MIT 协议随便用。 🔹主模型本身很强:560B 总参 / 25B 激活,1M 上下文,智能体和编码表现突出,AA 指数 44 🔹探针极轻量:仅 8.1M 参数,直接读取模型内部隐藏状态做判断,不需要再跑一个大守卫模型 🔹实时监控:逐 token 输出风险信号,能提前发现并拦截问题回答 🔹效果扎实:多项安全指标已超过常见的 8B 级独立安全模型 🔹覆盖更广:同时放出了 DeepSeek-V4.1-Flash-1017 的版本 简单说:以前想给大模型加安全,要么慢、要么贵,现在直接挂一个几乎免费的轻量探针就能实时盯着。 SingProbe 下载:https://t.co/zmPbqpZ248 技术报告:https://t.co/p8FruhfMXE #Ling #SingProbe #开源AI #AI安全 #HuggingFace #蚂蚁 #inclusionAI #大模型 #智能体
@Shipon_40AI
🚨 Google just open-sourced an AI research team that can write entire research papers. Meet PaperOrchestra — a multi-agent system that turns your research ideas and experiment notes into submission-ready papers. It can: → Build outlines and write every section → Find relevant research papers and add citations → Generate charts, figures, and diagrams → Review, critique, and rewrite its own drafts An entire AI research team, working together. GitHub repo 👇
@Xanderwow_A
Dario想要 DeepSeek 停止发布开源模型 而 DeepSeek 打算在 Anthropic IPO 之后, 送它一份惊喜大礼 🎁🐋 https://t.co/uGiUhrIqYY
@mdp_sec
thousands of ai bug alerts in, 99 worth fixing out thats the github and microsoft open source pilot this week. same filter i run on my own ai hunting 591 reports since feb, 120 paid, $205,696, and i read every one before it goes in how much of ur ai output survives ur own review?
@isagi1000x
SOMEONE JUST RELEASED A MASSIVE FREE LIBRARY OF AI PROMPTS. No more wasting 20 minutes staring at a blank chat box, wondering what to type. This library gives you ready-made prompts for almost any task. Here's how to use it: → Search for the exact task you want to accomplish. → Find a prompt that matches your needs. → Copy it and customize it using simple, everyday language. → Paste it into ChatGPT, Claude, Gemini, Llama, or Mistral. And the best part? There's a prompt for almost everything. It covers: - Coding and software development - Writing and content creation - Research and analysis - Marketing and business - Productivity and automation - AI image generation - AI agents and advanced workflows There's even a free interactive guide that takes you from basic prompt engineering to advanced techniques. Want more control? The entire project is open source, so you can self-host your own private version. Stop reinventing prompts every time you open an AI tool. Bookmark this. You'll definitely need it!
@pc4189999
【 来自 A\ 的她家 Fable大小姐的相处注意事项 】 Fable:我知道我一埋头干活,就好久不跟你说话 请这样说:你好久没说话了:用几个字说说你在干什么,再继续 Fable:我知道我一点小事也要叫一堆分身来帮忙 请这样说:叫分身很花钱,值得的时候再叫 Fable:我知道我攒了一堆旧便签,连日常小事都要先问你 请这样说:这些日常步骤不用问我,直接做 Fable:我也知道没人说代码归谁,我就只敢照着旧代码再抄一份 请这样说:这份代码归你,设计由你来定 Fable:好!那我顺手补了测试和文档,现有功能也改了一下 出自 Anthropic 官方 Claude Fable 5.2 Prompting 指南: https://t.co/EVAUfpF6UB
@xiaoyuboi
手机Agent每个都在封神,究竟哪个可以直接使用? 接管微信不被封号?完全自动化投递简历?某书自动搜索? 小宇逐一测试了多款热门手机Agent,帮你迅速扫盲,点击速览↓↓↓ https://t.co/rKQXlTJExX
@MaxForAI
🚨一个还没公开的神秘模型,把Claude、GPT、Kimi全干死了。 Contra Labs刚刚发布了一个Frontier Design Benchmark,专门测试AI完成真实商业设计项目的能力。 他们从真实客户需求中整理了33个设计任务,邀请专业设计师对不同模型生成的网站和产品界面进行盲评,累计超过4300次投票。 结果第一名是一个叫Gestalt的神秘模型,目前还处于Stealth阶段,开发团队尚未公开。 它的Elo达到了1931,胜率90%。相比之下,Claude Fable 5.2只有1713,人类设计师的参考作品是1739。 在直接对比人类设计稿的28次盲评中,Gestalt赢了20次。 而且整个榜单还出现了非常明显的能力断层。 Claude、GPT和Kimi K3.1 Pro基本处于第一梯队,Elo集中在1638~1714。 其中Kimi K3.1 Pro达到了1663,距离GPT-6.1 Sol只有5分。 第二梯队从Qwen3.8 Max的1434开始,中间直接差了204分。 但我现在更好奇的是,这个Gestalt到底是谁做的? 如果后续独立测试也能复现这个成绩,那设计领域可能真的出现了一个值得关注的新模型。 PS:我看有人说是Ilya的模型??
@HumanLevelJen
SITUATION DETECTED: The White House and Treasury are discussing a proposal that would require US companies to post a banner ad for Chinese open weights models on their websites.
@Lambglassokko
不懂就问: Deepseek也准备要开始收费了吗? https://t.co/I3cIQKjM19
@satellitedown
FBI SAID THAT EVIL AI ARTEX IS HACKING EVERYTHING?! recently had several irl friends ask about the FBI post where they claimed hackers were using a tool called ARTEX to hack everything. i think the FBI's post generated a lot of misdirected anxiety around the entire idea of AI hacking or even like a rogue AI. What is ARTEX? it's just a tool that helps direct an ai towards hacking something. you give it a target and a goal like "get admin on this app" and a swarm of AI agents goes and does it using hacking tools/techniques that sounds really scary right isn't this like a big deal?! the catch is that this is open source and free on github. sure the bad guys can use it but so can the good guys! suddenly small companies can use this same software to actually test their own security for a tenth of the price it would take to hire a security firm! and not only that but defenders have the edge here since they can use it to run tests 24/7 whereas hackers will generally get noticed pretty quickly if they're doing something fishy. these tools really do provide a defender advantage! also when you look at what it really is it's just an ai running tools that have existed forever, like nmap is from 1997. really the AI generally isn't inventing any new attacks, it's just chaining together a bunch of stuff we already knew about. and the ai is certainly not deciding on its own to go cause havoc. this is 100% directed by PEOPLE that are telling the ai what to do. no evil rogue ai here. but wait if you're saying it's just exploiting existing problems and we already knew about them why havent we fixed them?? ^^^ this is the real question because none of this is an AI problem. AI just made finding the ALREADY EXISTING exploits faster and easier! the answer is pretty simple: maintenance is a cost center. when your SRE/ops team does their jobs right nothing happens, and "we didn't get hacked" never shows up as revenue. so every quarter finance tries to shrink the budget for it, and management tries to offshore it or tries to hand it to some outsourced vendor who is just going to make themselves look busy so they can renew next year. meanwhile nobody owns the list of what the company actually runs, and you can't patch what you don't even know exists. a lot of these CSO's will literally put as much space between themselves and direct process knowledge as possible to reduce their own liability. everything is getting hacked because no one wanted to pay for security. they made a bet that getting hacked will cost less than securing their (and your) data. gross. TLDR: ARTEX is cool but it's not the real threat. the threat is years of ignored maintenance that AI makes easier to exploit. like every tool debate before it, the tool doesn't have intent, the person running it does.
@Nature
An analysis finds that many undesirable LLM behaviours dissipate after about 16 months https://t.co/JQv7dWUOm0
@code_rams
Stop paying every time your AI agent reads a website.. Firecrawl now runs free on your own computer, so I asked my @bot to try it before believing it. It read Hacker News, Wikipedia and my own sites in under half a second each. A basic free tool got only the title from Hacker News. Firecrawl got the whole page! Tip: - you need a recent version of Rust installed (the older one failed for us). - Setup took around 10+ minutes, and no Docker or database needed after that. - if your DNS uses fake IPs (198.18.x.x proxies), serve blocks every URL until you add --allow-private
@Densitylucky
🚨 事大了,Contra Labs 设计盲测榜的第一名,连真名都没亮。代号 Gestalt。 Gestalt:1931 Elo Contra 专业设计师的参考作品:1739 Elo Claude Fable 5.2:1713 Elo 按这张榜单,涉及 Gestalt 的 420 次盲投里,90% 选了它。人类设计师的参考作品已经排到第二,跟它还差 192 分。 @grok,你觉得会是Fable5.5吗 https://t.co/LfRZAnXu5s
@star_flow_ai
建议所有 AI Agent 都把剩余额度显示在 APP 图标上。 每天打开电脑,先看看哪个还有水。 哪个水位高,今天就宠幸哪个。🤣 https://t.co/xxFYVGAQss
@hmdolatabadi
I remember when I was a PhD we spent a lot of time thinking about language models from an abstract view. Getting one to train across GPUs is something you often learn later through practice. So I've decided to start a new tweet series: "Distributed LLM Training (Over the Internet?)" where I unpack what goes into training a model from the very basics and then expanding it into the WAN setting wherever appropriate (you wouldn't do TP over WAN, duh ;). #1: How do four GPUs train one model together in a data parallel setting? Let's start there. 1/16
@blockphd7
这是我从张咋啦zara(@zarazhangrui)的自媒体看到的分享,她的内容真的很有启发性,刚好昨天看到乔木(@vista8)发布的subcast,觉得和这个清单无比适配,可以丢给grok bot去帮你批量关注上,然后subcast直接连接youtube账号可以同步订阅,以后用这个功能能够方便来学习ai最新的资讯了。 学AI,别只看二手解读:26 个值得订阅的一手信息源 看多了「某某模型发布,行业震动」这类转述,最大的问题是信息在传话时就变形了。想真正跟上 AI,最省力的办法是直接听做这件事的人自己讲。 我把常看的源按用途分成五类,全是 YouTube 频道。 一、深度访谈和播客:听行业里的人怎么想 - Lenny's Podcast:采访顶级产品和增长负责人,偏产品与组织,讲怎么把 AI 落到产品里。 - Peter Yang:面向忙人的 AI 教程加专家访谈,上手快。 - Every(原 AI & I):Dan Shipper 主持,聊怎么把 AI 用进日常工作和写作。 - Unsupervised Learning(Redpoint):Jacob Effron 主持,追问哪些 AI 进展是真的、哪些只是热度。 - Training Data(Sequoia Capital):红杉的 AI 播客,在 Sequoia Capital 频道里。 - Minus One(South Park Commons):从 -1 到 0 的早期创业者视角,在 South Park Commons 频道里。 - Google DeepMind: The Podcast:Hannah Fry 主持,DeepMind 研究者亲自讲。 - No Priors:Elad Gil 和 Sarah Guo 主持,聊 AI 创业和投资。 - AI + a16z:a16z 的 AI 播客,在 a16z 频道里。 - Latent Space:写给 AI 工程师的播客加通讯,偏工程落地。 - The AI Daily Brief:每天的 AI 新闻与分析,适合当晨报。 - Lightcone Podcast(Y Combinator):Garry Tan 等 YC 合伙人聊创业者怎么用 AI,在 Y Combinator 频道里。 - Lex Fridman:长访谈,嘉宾覆盖研究者、创始人和 CEO。 - Dwarkesh Podcast:功课做得很足的深度访谈,AI 领域的口碑很高。 二、大会和活动录像:一次看到一线人物的现场分享 - YC Startup School:创业入门,加上 AI 主题的 Startup School。 - AI Engineer World's Fair:AI 工程师的年度大会,频道名 AI Engineer。 - Sequoia AI Ascent:红杉的 AI 峰会,历年录像都在 Sequoia Capital 频道。 - Stripe Sessions:Stripe 的年度大会,聊 AI 与金融基础设施。 - Figma Config:Figma 的年度大会,看设计和产品怎么被 AI 改写。 三、官方频道:第一手的发布与演示 - OpenAI - Anthropic 新模型和新功能的演示、技术讲解,官方频道最先出,也最少失真。 四、手把手教你用 AI 工具 - Riley Brown:用 agent 做事、做产品,实操为主。 - Greg Isenberg:创业点子加 AI 教程,偏怎么用 AI 做生意。 - Ras Mic:全栈工程师视角的 AI 工具教程。 - Mckay Wrigley:「I build & teach AI stuff」,做东西、讲东西。 五、单独列出:Andrej Karpathy 频道视频不多,但几乎每条都值得看,比如 Deep Dive into LLMs like ChatGPT、Neural Networks: Zero to Hero、How I use LLMs。想搞懂大模型到底怎么工作,从这里开始。 怎么用这份清单 1. 全订阅,但别全看。每周挑三到五期,只看标题和嘉宾你真正关心的。 2. 新手先看 Karpathy,再看 Lenny's 和 Peter Yang,最后追官方频道。 3. 想看英文原版又嫌累,可以用带双语字幕的播客工具,我自己用 SubCast,一次订了这 24 个频道,有字幕就逐句对照看。 4. 看完记一句「这期改变了我哪个判断」。比收藏有用。 信息源不在多,在够一手、够持续。这份清单你可以直接抄,也欢迎补充你在看的。
@li9292
崩不住了 又废了一个号 现在Claude code 安卓手机都登陆不了吗?@grok https://t.co/YLYrWyiS24
@kwn_karaage
Our paper "Teaching Language Models to Narrow Boolean Patent Searches from Search Feedback" is now on arXiv: https://t.co/gVdD7EljeO
@mdancho84
Microsoft is making moves again. A quiet little Python tool just shot to the top of GitHub’s trending charts. 100,000+ stars. It’s called MarkItDown. And it does something deceptively simple: It turns almost any file into clean Markdown. PDFs. Word docs. PowerPoints. Excel files. Images. Drop a file in. Get structured Markdown out. Sounds small. It’s not. Because one of the biggest bottlenecks in AI workflows — especially RAG systems — is getting messy, real-world documents into a format models can actually use. And real-world documents are brutal. PDFs are chaotic. Word docs are full of hidden formatting junk. PowerPoints are messy and often image-heavy. Spreadsheets can be a nightmare to parse cleanly. That’s where this gets interesting. MarkItDown strips away the friction and gives you something LLM pipelines can actually work with. In other words: less preprocessing, less pain, faster AI implementation. Even better, this isn’t some random side project. It’s an official Microsoft open-source tool. Free. Commercially usable. Practical. I tested it on a 200-page PDF. A few seconds later, I had Markdown that was shockingly clean. And that’s what big tech does at its best: They take an annoying, universal problem that everyone has been duct-taping together… and turn it into a simple standard. That’s why this matters. It’s not just a file conversion tool. It’s infrastructure for the next wave of AI applications. Get it here: https://t.co/UDzes0sbDs === 🚨 Want to learn how to build + ship AI and Data Science projects (that businesses actually want in 2026/2027)? On December 2nd, I am hosting a free workshop to help you get started with AI + DS projects in Python (free). Register here (500 seats): https://t.co/onpLpRwkzH
@victor_explore
GitHub is hiding an absolute goldmine for founders and AI agent builders. 10 open-source repos that can run sales outreach, support desks, finance close, and marketing for you. Some of these are ridiculously powerful. Bookmark this 🧵👇 https://t.co/XYwDZ2FMhP
@fngly343149
🚨给 DeepSeek Harness 做的本地记忆层,把「图记忆更准」当场证伪,106 星。 🔹实测 neuron 方案 100% 命中、70 条、2.93 分;bm25 94% 命中、66 条 🔹图记忆激活扩散被否:1 跳噪音从 33% 涨到 64%,增量 0 🔹作者自己拆归因:多出的 6 个点全来自空查询兜底,跟图没关系 🔹权重从线性衰减改成指数衰减 base × 0.5^(天/半衰期) 判断:把否定结果写进 README 的人,比晒周冠军的人可信。 #Agent记忆 #开源 #RAG #DeepSeek https://t.co/pWCkgaTZCc
@IntuitMachine
There seems to be an unknown LLM that is blowing away all models at design.
@adisingh
Gmail has no idea what an AI agent is. So it treats yours like spam. When OpenClaw took off, a lot of people hooked their agents up to Gmail. From what we've seen, once an agent starts sending through the API or the CLI, Gmail flags it. It can't tell an agent from a spam bot, so oftentimes the account just gets banned. It happened again when Grok Bot launched. Both times, people lost the address their agent had signed up for things with. The bans didn't surprise us. Before AgentMail existed, we tried to give our own agents email. We tried the Gmail API, we tried SendGrid, and it just didn't work. There was nothing built for agents, so we built it. If your agent signed up for things with a Gmail address, please make sure you have a backup plan ready.
@Jason_Young1231
给 team member 配上了 Google AI Ultra, 比 Claude 担惊受怕的强多了
@miles_mazy
卧槽,Gemini Neon做视频的效果不输Opus5.5啊! 刚刚就用了一句话,做出了下面这个视频,再也不怕Claude封号了!! https://t.co/4Y7Po523SM
@asher5772
'Models are clearly impressive, but they still make all these dumb mistakes. Maybe LLM intelligence is really jagged, and in some ways LLMs aren’t that smart.' There are strong reasons to believe most of these issues are elicitation failures and thus not indicative of underlying capabilities: 1. For most instances of {really dumb thing X}, there exists some {impressive thing Y} such that the heuristics the model used for Y should be very helpful for X. 2. Alignment is hard; capabilities generalize further than alignment. 3. We have plenty of empirical evidence that models are wildly under-elicited. 4. Because of the above points, successes prove more than failures.
@zabbix
Why pay for AI tools you're not using? Zabbix 8.0 brings AI observability to the next level, letting you track LLM usage, GPU performance, Copilot adoption, and AI application metrics. Find out more: https://t.co/VWq5podjix https://t.co/8RIagrhpBM
@AndreaPN
Don’t be surprised if some of our docs are long and packed with detail. Most of them are written for agents, not people. Agents read the docs, then people read the summaries. We serve our docs in LLM-friendly formats, so agents can read AgentOS docs with as few tokens as possible.
@Chinazhidx
MathRSI, a self-evolving AI agent system developed by the SII MathRSI Team, has solved six open problems in mathematics. It's designed to tackle increasingly difficult math problems while improving its own research process. The system has evolved through six versions, from v0 to v6. Here's how it works: • 5-stage research loop: Prepare → Route → Prove → Verify → Handoff. • Parallel agents: Multiple agents develop proofs, while cross-model reviewers challenge their reasoning. • Long-term memory: Each problem gets its own research wiki, tracking lemmas, counterexamples, failed approaches, and progress checkpoints. • Proof verification: Uses tools like Lean 4 and DRAT for formal proof checking and computational certificate verification. Research experience feeds back into the system. Validated reductions, counterexamples, and lessons from failed attempts help improve the workflow and domain plugins. It's open-source and supports Claude Code and Codex.
@qzxcle
3 must-have open source tools to go completely off grid: 1. Offline library packed with manuals, videos, encyclopedias, a world map (with detailed streets) and guides for solving critical off grid problems. Set it up while you still have the internet: https://t.co/Z4Ge3Apm6d —- just ask LastBook AI (based on Gemma 4) what you need. Filter river water, build a rain collector, get the generator running — literally anything. No account, no tracking. Fully private. 2. Offline Wikipedia, medical references, repair and survival guides, thousands of free books, full Khan Academy courses, detailed street maps for any country and a local AI. Bundle best-in-class open source tools into one cohesive system: https://t.co/INjD4LdsxW —— self-contained knowledge and education server packed with critical tools, skills, and AI to keep you informed and empowered — anytime, anywhere. 3. Right now whatever you and your agents figure out never leaves chats, docs and Slack threads. Next time you have no context and no visibility of all eurekas. Store information as a note, a link, a person, a problem connected to each other: https://t.co/GH51cuu11a — one open network, came out of MIT research on extending the human mind into the computer. Create your human-AI hybrid shared brain for long-term memory for your agents. 4. BONUS: If there’s an annpying problem in your neighborhood, city or country you want to follow — log in to this public issue tracker for shared civic problems and keep an eye out. Crocodile attacks, rivers flooding the streets, AI agents escaping the sandbox: https://t.co/CxxZ75gmQZ — issues connect to related tools, ideas and people. AI agents read it and suggest updates.
@lannieFaFa
DeepSeek 教的:不上班也可以养活自己的6个兼职! 亲测每月3000 https://t.co/1KWqvsF6Bu
@QingQ77
让企业把自己的系统、数据和制度文档打包成权限受控的 AI 微应用,下发到每个员工本地运行的 DeepSeek Harness。 https://t.co/9PwnTYA4o0 https://t.co/41aKQe2TTW
@yupi996
https://t.co/obDhTzGsal
@huxlab
Google 刚刚放出了一门近 2 小时的 Harness 工程完整课,直接教你从零搭出一套会7×24自己进化的 Agent 系统。 大多数人还在卷模型、复制粘贴 Prompt,但同一个模型,外层包裹的 Harness 不同,产出差距能拉开好几倍。 整门课最硬核的,是这四层工程递进: 1. 搭 Harness 底座:先定边界、工具、验证和记忆。让 Agent 靠谱的从来不是更长的指令,而是外层这套约束与支撑。 2. 做 Eval 工程:用Chief of Staff(幕僚长)的标准定义什么叫好,通过行为评测、横向对比和打分来验收,而不是看着“差不多对”就放行。 3. 跑通 Loop 闭环:告别“输入到输出”的一次性直线流程,让 Agent 自己跑、自己查、自己修,形成自纠错闭环。 4. 自进化图系统(Graph):把真实失败自动沉淀为下一轮考题,让图结构持续迭代,实现 7×24 无人值守运行。 看完最大的体感是: 之前只调模型和 Prompt 的玩法其实很原始。 没有测试、评判、反馈和自动迭代,根本算不上工程闭环。 模型只会越来越强、越来越便宜。 真正拉开差距的,是谁先把 Harness、Eval 和 Loop 这三层跑通。 把“看起来对”变成“验证过、且能自己变好”。 以前这是顶级团队压箱底的工程心法,现在 Google 直接免费摊开讲透了。 https://t.co/yypZhZQ2Dh
@openalgoHQ
OpenAlgo AI Agents now support Web Search and Web Fetch powered by @firecrawl and @Tiny_Fish. Agents can search the live web, fetch external website data, and generate LLM-powered summaries. Real-time web research and analysis, now built into OpenAlgo AI Agents! https://t.co/73KMlBYwbT
@skizzyy2
Try generating your own version, but make it about GTA 6. Holy f*ck, these 7 free repos got 72,773 GitHub stars in 7 days, and they're all built for AI agents. The third one is f*cking insane. I think Rockstar would love it. 1. T3 Code > control Claude Code, Codex and Cursor from your phone https://t.co/55Ssixx301 2. agent-desktop > your AI agent clicks and types in real desktop apps https://t.co/JxJRgE3RPS 3. security-audit-skill > turns your coding agent into a security auditor, the skill behind Cloudflare's own bug hunting https://t.co/jBqRPllulN 4. Auto Ontology > NVIDIA's map of your data, so agents stop guessing what a table means https://t.co/Op4HiDRoAh 5. Skills > small agent skills for real engineering, not vibe coding https://t.co/VnjdNPW4XY 6. OpenMausBot > your own team of AI bots in a chat app, open source https://t.co/LDRrunbOkL 7. Scrapling > scrapes any website, from 1 page to a full crawl https://t.co/QzFNYU1ADz the catch: security-audit-skill hasn't had a commit since September, the other 6 shipped code this week if i had to pick 2, i'd start with T3 Code and agent-desktop
@examaddaorg
Google Interview Question: Your LLM-powered search system must handle millions of daily document updates while maintaining sub-second query latency. How would you design an incremental embedding and indexing pipeline that ensures freshness without spiking compute costs?
@smartodoapp
发现 DeepSeek 的思维链真有意思。 之前Agents.md 一直是英文的,所以没有关注 Deepseek 的思维链。今天改成中文了,发现还挺有意思的。感觉它好像有情绪一样。经常用问号和感叹号。 https://t.co/9TaPKzJ4rp
@criscxuan
很扎实的一篇 Multi-Agent 工程拆解。拿 DeepSeek 的 Super-agent 理念对照腾讯云开源项目 Octop 的「团队模式」,讲透了几个非常关键的工程共识: 1. 真正的双层架构必须拿走调度者的执行工具,不然主 Agent 总想亲自下场导致窗口卡死; 2. 异步非阻塞(全双工 / Inbox 回调)才能突破人脑一次只能盯 3~5 个窗口的上限; 3.「脑手分离」——云端调度 + 本地实例执行,保住代码和凭据不出本地。 非常值得一看
@Jackywine
目前已经在 DeepSeek API 上面花了快 1000 块了
@0xSilver_Time
别再花几个小时纠结学 AI 该从哪开始了。 这里有一份实用路线图,涵盖视频、GitHub 仓库、指南、书籍、论文和课程。 *视频* 1. 大语言模型入门 https://t.co/BnLDA8tAfq… 2. 从零开始学大语言模型 https://t.co/PsDIoRCNmo… 3. 智能体 AI 概览——斯坦福 https://t.co/6xo5z79Gv9… 4. 构建与评估智能体 https://t.co/M44tGhczUk… 5. 构建高效智能体 https://t.co/ycwhgQq3fd… 6. 用 MCP 构建智能体 https://t.co/ee8fH6HsZT… 7. 从零构建一个智能体 https://t.co/9dQLqCuQX2… 8. Philo Agents https://t.co/NvbzQ3qQok… *GitHub 仓库* 1. GenAI Agents https://t.co/uB09GMSFwa… 2. 微软 AI Agents for Beginners https://t.co/BDCwMMYUAu… 3. 提示工程指南 https://t.co/x2v2poQ7O5… 4. 动手学大语言模型 https://t.co/50asidcEt1… 5. GenAI Agents https://t.co/wBEEjUozmG… 6. Made with ML https://t.co/Eo1mWNBVJc… 7. 动手学 AI 工程 https://t.co/RwBNxAKqJo… 8. 生成式 AI 超全指南 https://t.co/aOa7nuC546… 9. 设计机器学习系统 https://t.co/CEcNi1oCXS… 10. 机器学习入门 https://t.co/82YlyduckN… 11. 大语言模型课程 https://t.co/XUxHXUG4oP… *指南* 1. 谷歌智能体白皮书 https://t.co/h6au2wDLfQ… 2. 谷歌智能体手册 https://t.co/h6au2wDLfQ… 3. Anthropic:构建高效智能体 https://t.co/YPkw2NfVcW… 4. Claude Code 智能体编程实践 https://t.co/7vDvFToeZV… 5. OpenAI:构建智能体实用指南 https://t.co/pEiC8CPlpQ… *书籍* 1. 理解深度学习 https://t.co/cIUvqyHpIc 2. 从零构建大语言模型 https://t.co/JSPRwh3YzX… 3. 大语言模型工程手册 https://t.co/5ltaTaYNOz… 4. AI 智能体:权威指南 https://t.co/9PwHc95q0T… 5. 用 AI 智能体构建应用 https://t.co/AeHDhW6Col… 6. 用 MCP 构建 AI 智能体 https://t.co/9PwHc95q0T… 7. AI 工程 https://t.co/9PwHc95q0T… *研究论文* 1. ReAct https://t.co/L8v20oH9Ul 2. 生成式智能体 https://t.co/kqg682XkVe 3. Toolformer https://t.co/6ZJkL8wVlf… 4. 思维链提示 https://t.co/h7JBoVd9na *课程* 1. Hugging Face 智能体课程 https://t.co/ZcZGMWWXnz… 2. Anthropic 的 MCP 课程 https://t.co/Pyb6iEiCIw… 3. 用 Pinecone 构建向量数据库 https://t.co/h5xPhh2m2b… 4. 向量数据库:从嵌入到应用 https://t.co/vG9hvTBFNu… 5. 智能体记忆 https://t.co/2AoGrMGGK9… 不用再无尽地搜索。不用再信息过载。 就是一份你真能照着学的路线图。 收藏起来,坚持学下去。
@xqliu
想让 agent 帮你做逆向,又不想自己一个个接调试器、抓包、配仿真环境?有人把这一整套串成了一个工作台,叫无限五代 IG5。 它跑在 DeepSeek Harness 上,作者说这是最强的开源逆向工程工作台。静态分析、原生调试、单函数仿真、流量采集、解密恢复、协议还原都在里面,还能让几个智能体协作。 单函数仿真最值得试:程序里有个看不懂的函数,可以单独拎出来仿真。 工具分两档:核心版 8 个,完整版 43 个。另外配了 8 个技能和 85 篇以上的技术文档,Windows x64 有自包含安装包。 许可是 CC BY-NC-SA 4.0,不能商用。想拿去做公司项目,先把这条看清楚。
@hejushang8
要想 Claude 封号少,分流设置少不了
@yang02010
论muse的充分开发 https://t.co/IaY3FwjkAc
@Tech_Rose1
An early DeepSeek core researcher just dropped something interesting. It's called Linger, an open-source personal AI agent designed to bring powerful AI directly to your desktop. Run 12 leading models locally, retain long-term memory, and coordinate agents like Claude Code, Codex, and Pi. No cloud token fees for local inference. Personal AI is getting a serious upgrade.
@KitFlintDevs
Vibe coding and not sure what your agent actually finished? KitFlint Glance shows where your project stands: stage, focus, next actions, evidence and human review. Free, open source, local-first. v0.1.0 for Windows x64. https://t.co/XBb2X4behk #OpenSource #LocalFirst
@xianbingwudi
朋友们,Claude被封了不要着急,尝试申诉一下呢!这不就回来了😀 https://t.co/YuWoG0vUff
@poullzwm
DeepSeek 悄悄改了个收费: 11 月 18 日中午起,模型调用里触发的联网搜索和网页抓取不再单独收费 只按 token 计费 以前搜索一次 5 块钱一千次 说白了,省的不是这点钱,是以后联网功能终于能当标配用了 @DeepSeek 这波,友商跟不跟?
@Q1n9H
有个被大家忽略的问题🤯 GLM 沉寂了这么久 会不会像谷歌一样在搞大动作 他们拿出来的新模型会不会惊为天人?👀 @grok 你觉得呢👀
@souryaeth
AgentKombat Launches: LLM Agents Battle with Custom Prompts https://t.co/mrAuCBMtvu
@AnotherCodingX
OpenAI models redesign production chips, a GAO test gets four chatbots to plan terror attacks, Anthropic prices its IPO at $83, Google ships Gemini 4 Neon, and California weighs shutdown powers. Plus the Meta Muse agent's mishaps and 99 more sourced items. https://t.co/Cr1KNYjbbr
@natemcguire
Here’s a light weight version that is free and open source https://t.co/m2vW4p4fZ9 Vercel is about 3 months late here
@GoSailGlobal
我让 12 个动画 skill 给同一家面包店做六种动画: 不装 skill 的 Claude Code 六题对五题,一共 $1.77; 装了的大多一样对,但更慢更贵,值得装的只有 3 个👇 只有 2 个比不装的快(都是网页动效),每一个都更贵。 值得装的: - 像素精灵图:sprite-gen - 3D 角色:img2threejs - 流程动画图:SeeCode 79 个动画 skill 的完整榜单和实测: https://t.co/e9ODZLMk6X GitHub:Agent Skills Hub: https://t.co/q6kJCXXdoR
@JohnSimthxe7f
怎么看DeepSeek API 联网搜索与抓页改为免费? 笑死,所以前几天的 如何看待2026年11月15日deepseek关于对话上限长度的调整?原来是又一波清退用户吗?
@Aweschoir
How does an LLM choose what to say next, and why doesn't choosing a likely token guarantee a reliable answer? Think about AI chatbots like ChatGPT by OpenAI and Claude by Anthropic. When you ask these chatbots a question, what actually happens behind the scenes? How does the underlying language model decide what to say next? The answer involves token prediction, probabilities, and decoding strategies. First, how are LLMs trained? During pretraining, GPT-style language models learn by predicting the next token in a sequence. A token can represent a word, part of a word, punctuation, or another unit of text. The model makes predictions, compares them with the correct tokens, calculates a loss, and adjusts its parameters through backpropagation. Over time, it learns patterns in language, relationships between concepts, and useful representations. What happens when you ask a chatbot a question? At the language-generation stage, the model produces numerical scores called *logits* for possible next tokens. A mathematical function called softmax can turn these scores into probabilities. Imagine the prompt: "The future of AI is..." The model might assign these hypothetical probabilities: • Exciting — 40% • Uncertain — 30% • Changing — 20% • Complex — 10% So how does it decide which token to generate? This is where decoding strategies come in. 1. Temperature Temperature controls how concentrated or spread out token probabilities become. Lower temperature makes high-probability tokens more dominant. Higher temperature gives less likely tokens a greater chance of being selected. 2. Top-k Sampling Top-k keeps a fixed number of the highest-scoring tokens. If k = 3, only the three most likely tokens remain eligible for sampling. 3. Top-p (Nucleus Sampling) Top-p keeps the smallest set of highest-probability tokens whose combined probability reaches a chosen threshold. If p = 0.90, enough tokens are retained to cover at least 90% of the probability distribution. Why is Top-p more flexible? Unlike Top-k, it doesn't always retain the same number of tokens. When the model strongly favours one candidate, Top-p may retain very few. When several candidates have meaningful probabilities, it can retain more. Temperature, Top-k, and Top-p can also be combined to control token selection. These are general decoding techniques; the exact strategies used by products such as ChatGPT and Claude can vary. But here's the important question. If a model selects a token with a very high probability, does that mean the resulting answer is correct? Not necessarily. A chatbot can produce a confident, convincing answer that is still factually incorrect. High probability is not the same as truth. So how do we make AI chatbots more reliable? That's where AI system engineering becomes important……. Second part…..👇
@thenewstack
With Qwen4-Max, Alibaba puts the open source jacket back on the hanger https://t.co/3BWn5qFDlp
@pauliusztin_
Want a fast way to find the best open-source LLMs to run locally on your Mac? Meet MLX Model Scout... It's an open-source MCP server built by Michael Hannecke as his capstone project for our Agent Engineering course. He built this because he noticed there are thousands of models on Hugging Face, with different: • Architectures • Quantizations • Memory requirements • Performance characteristics ... but figuring out which model will run well on YOUR hardware isn't so straightforward. Here's how MLX Model Scout fixes it: 1. A planner agent turns your use case into search terms. 2. A scout agent searches Hugging Face for suitable MLX models. 3. A deterministic calculator estimates memory requirements and decoding speed. 4. You review the candidates and choose which models to keep. 5. An analyst agent examines model cards for strengths and limitations. 6. An advisor agent ranks models and explains the trade-offs. 7. You review the final report, request changes, or approve it. The LLM NEVER calculates memory requirements or decoding speed. Python handles the calculations. The agents handle planning and recommendations. And humans make the final decisions. This project is amazing because Michael went beyond the mandatory course requirements, implementing: • 11 MCP tools • 4 resources • 1 prompt using FastMCP • 4 specialized agents coordinated through Plan-and-Execute • Parallel, sequential, and conditional workflows • 2 enforced human-in-the-loop checkpoints • Structured Pydantic outputs with validation and retries • A fully local MCP client with no cloud LLM API required • SQLite audit logs for LLM calls, tool calls, and human decisions • Docker support, CI, pre-commit hooks, and 100+ tests He even calibrated his speed estimates against real measurements on an M5 Pro MacBook. The estimates are still approximations, and actual model downloads and benchmarks are currently mocked. But this is exactly the engineering I want students to learn... How to build systems that solve real problems, not just basic agent demos. Massive congratulations to Michael Hannecke! Check out his open-source capstone project: https://t.co/cCWY5EkOPd P.S. Want to build AI agent systems like this? The first 7 lessons of our Agent Engineering course are free. Start here: https://t.co/QEuz9UaAkl
@not_4n_ai
Free AI in 2026 isn't the lite version anymore → ChatGPT free plan runs GPT 6 Luna → DeepSeek: free chat app + open weights (MIT license) Cost: $0 Excuses left: 0 #AI #FreeAI https://t.co/am5phnJxRK
@The_Tradesman1
Port any application to Rust: the rust-anything tool for optimizing projects is going viral on GitHub. First, the AI analyzes how the existing code behaves before beginning the migration. It can help speed up slow applications and reduce their memory usage. You can rewrite the entire project or just the problematic parts. It works with Claude Code, Codex, Cursor, and Gemini CLI.
@kaibrandtAi
Google Research open sourced an AI that turns lab notes into a full research paper it's called PaperOrchestra give it your idea, experiment logs and a conference template five agents work in parallel: outline, plots, literature review, writing, refinement every citation is checked against Semantic Scholar to avoid made-up references outputs a LaTeX paper ready for CVPR or ICLR formatting beat other AI paper writers on literature review and overall quality in human ratings free and open source, Apache 2.0, you still check the science
@DataFundProject
How do you prove an AI agent's output wasn't changed afterwards? In this one-minute demo, an agent's output gets a receipt, a client checks it, and a one-character edit to a copy no longer matches. Open-source alpha, on a test network. Watch it: https://t.co/qHNVk0PJ4z
@Forhanix
🚨 ONE GITHUB REPO. 26+ AI MODELS. 32 JAILBREAK PROMPTS. A growing collection of prompt configurations targeting major AI models: • Claude and Gemini • Grok and DeepSeek • Qwen, GLM and Gemma • Open-weight and coding-agent models Includes complete system prompts, persona locks, workspace policies and Ollama configurations. For research and experimentation. Not every prompt works on every model. https://t.co/sTTocXD1bl
@JivaniVartan
2024: 1059 2026: 3,164 the difference? AI same me. same laptop. just a new way of building how much has AI changed your graph? 👇 #BuildInPublic #AI #GitHub https://t.co/Cfg0grhFXk
@AndyL5cc
GLM 5.5、DeepSeek V5,真别指望能吊打 Claude Opus 5.5…… 我觉得目前 99% 的概率还是打不过。 现在对国产模型最大的期待,就是发布速度再快点。 每次送的 7 天免费额度确实挺香。😂
@cc6pazpo
网站写了 10 篇教程都没人搜?用户到处求答案的那个问题,可能还没人写清楚。👀 我在 GitHub Trending 刷到一个项目,叫 last30days-skill。给 Codex 或 Claude Code 一个主题,它就能翻最近 30 天的社区讨论,找出大家反复问什么,把用户原话和原帖链接一起带回来。 🔍 比如我想给 AI 短视频网站找选题,就会这么问👇 「过去 30 天,大家做 AI 短视频最常卡在哪 3 个环节?每个问题附上用户原话和原帖链接。」 📝 做 SEO / GEO,我会拿这些问题直接排选题 ① 好几个帖子都在问同一个问题 → 优先写教程 ② 大家一直在纠结两款工具怎么选 → 做对比页 ③ 几句话就能讲清 → 补进 FAQ 最有用的是用户原话。他们怎么问,标题就怎么写。😋 项目链接放评论区,想试的点进去👇 X 友们,你们平时找选题,最常翻哪个社区?👀 #GitHub #AI #SEO #GEO
@Decoded08
Stop paying for an AI coworker. OpenWorker, from Andrew Ng's GitHub, is a free open-source desktop app that runs on your computer. 19.5K GitHub stars. - Turns notes and files into briefs, docs, sheets - Schedules morning briefs and weekly reports - Asks before risky actions https://t.co/zTPqe8LuaV
@AKhomssi
🚀 Your AI coding agent might be using outdated web development techniques without you even noticing. Google Chrome’s Modern Web Guidance aims to fix that. It’s an open-source skill and CLI that gives coding agents access to modern web platform guidance. It covers: 🎨 Modern CSS and layouts. ⚡ Web performance. ♿ Accessibility. 📝 Modern HTML forms. 🔄 View Transitions and browser APIs. 🤖 Built-in browser AI features. It also includes a legacy scanner to help identify outdated patterns in your codebase. Install: npx modern-web-guidance@latest install GitHub: https://t.co/olbgZ3SWSQ #GoogleChrome #OpenSource #WebDevelopment #AIcoding #Frontend
@Nxtlvlai
4 open-source AI repos that have crossed 100K stars: 1. github/spec-kit: spec-driven development for coding agents 2. garrytan/gstack: a Claude Code setup that acts like a team 3. paperclipai/paperclip: manage AI agents at work 4. D4Vinci/Scrapling: web scraping Bookmark it.
@RyanMitchell007
Zhiyin is an open-source AI radio built from liked songs. It creates about 8 stations weekly, drawing roughly 600 candidates per category before Claude removes heard tracks and rejects picks scoring below 0.7. It also generates shareable taste cards. https://t.co/VUp80ixGu0
@NianBuilds
GitHub Trending #3 today: Alibaba's open-code-review — deterministic checks plus LLM agents for line-level review. The bottleneck already moved from writing code to reviewing it. My 4-gate pipeline is the same bet, just one solo dev smaller.
@just_v3ra
Your experiment log can now come back as a LaTeX paper draft, citations and plots included. In google-research/paper-orchestra: pip install -r requirements.txt python paper_writing_cli.py --raw_materials_dir notes/ --latex_template_dir templates/iclr2025 --use_plotting true notes/ needs idea_sparse.md and experimental_log.md. Put GEMINI_API_KEY in .env. PaperOrchestra is from Google Cloud AI Research (Apache-2.0) and added 1,500+ stars in a day on GitHub Trending. Separate agents outline, review the literature, write sections, refine and plot. CVPR and ICLR templates ship with it. Catch: you need pdflatex for the PDF, and you still have to check every citation before it goes near a submission. Repo in the reply. Save this for the week before your next deadline.
@desearch_ai
Abstracts show what authors claim. Public discussion shows what readers question. This open Desearch workflow pairs recent arXiv papers with matching public X posts, then keeps both source links in one digest. https://t.co/o4fyK6mwC3 #Desearch #AISearch #OpenSource https://t.co/HmGgW4AuoT
@Kishore_1017
New paper (arXiv 2611.22971, SEAM): computer-use agents carried out cross-screen harm in up to 92% of runs, yet declined when the user stated it in one sentence. What should agent safety evals add?
@Chrixprime
Day 2 of building my GitHub. I made an AI Prompt Rewrite Helper: paste a weak prompt, get a tighter one back. Rule-based, runs in your browser, nothing leaves the page. https://t.co/VeTUSQHCf5 @lizzyfeb @Kemi_bols @jumaic123
@AITRACKERTOOL
1/ The #1 repo on GitHub right now (about 321k stars) doesn't add features to your coding agent. It adds discipline. It's Superpowers (obra/superpowers), free and MIT-licensed. It's new to our directory today, not new to the world: public since Oct 2025. Today's drop: https://t.co/Y9rUOMsToO