{
  "updatedAt": "2026-10-28T23:30:00+08:00",
  "timezone": "Asia/Singapore",
  "windowDays": 30,
  "halfLifeHours": 672,
  "items": [
    {
      "id": "security-audit-skill",
      "example": false,
      "importance": 5,
      "category": "open-source",
      "occurredAt": "2026-10-28T21:45:00+08:00",
      "ageHours": 1.6,
      "score": 4.992,
      "today": true,
      "whenLabel": {
        "zh": "截至 10 月 28 日",
        "en": "As of Oct 28"
      },
      "title": {
        "zh": "Cloudflare security-audit-skill 本周冲上 GitHub 趋势第一",
        "en": "Cloudflare security-audit-skill is #1 on GitHub weekly trending"
      },
      "summary": {
        "zh": "这是 Cloudflare 给编程代理用的多阶段安全审计技能：一个代理找漏洞，另一个独立的校验代理逐条复核。仓库 6 月创建，本周热度起来。截至 10 月 28 日，本周新增 16,388 star，总计 51,025。",
        "en": "Cloudflare's multi-stage security-audit skill for coding agents: one agent finds vulnerabilities and an independent verifier re-checks each. The repo was created in June and blew up this week. As of Oct 28: +16,388 stars this week, 51,025 total, #1 on GitHub weekly trending."
      },
      "why": null,
      "sources": [
        {
          "label": {
            "zh": "GitHub 仓库",
            "en": "GitHub repo"
          },
          "url": "https://github.com/cloudflare/security-audit-skill"
        }
      ],
      "tags": [
        "github",
        "agents"
      ]
    },
    {
      "id": "openai-eth-kakeya-qcd",
      "example": false,
      "importance": 5,
      "category": "research",
      "occurredAt": "2026-10-27T05:46:00+08:00",
      "ageHours": 41.6,
      "score": 4.79,
      "today": false,
      "whenLabel": {
        "zh": "10 月 27 日",
        "en": "Oct 27"
      },
      "title": {
        "zh": "OpenAI 称内部模型推翻了指数时间假设，并证明了 Kakeya 猜想和 QCD 存在性",
        "en": "OpenAI says internal models refuted the Exponential Time Hypothesis and proved Kakeya and QCD existence"
      },
      "summary": {
        "zh": "OpenAI 公布三项内部模型得到的结果：3-SAT 有次指数时间算法（即 ETH 不成立）；局部光滑猜想在所有维度成立，由此推出 Kakeya 猜想和 Stein 限制猜想；四维 QCD 与纯 Yang–Mills 理论严格存在。前两项公开了证明和 Lean 形式化，物理部分材料暂不公开，由 David Gross 团队整理。全部未经同行评审，ETH 结果由 Impagliazzo、Paturi、Williams 牵头审查。Anthropic 10 月 27 日也公开论文，称内部 Claude 模型在 10 月 16 日独立得到 ETH 不成立的结论。",
        "en": "OpenAI published three results from internal models: a subexponential-time algorithm for 3-SAT (so ETH is false); local smoothing in all dimensions, which implies the Kakeya and Stein restriction conjectures; and a rigorous existence proof for 4D QCD and pure Yang–Mills. Proofs and Lean formalizations are public for the first two; the physics materials are withheld while David Gross's group writes them up. None is peer-reviewed; Impagliazzo, Paturi and Williams are leading review of the ETH result. On Oct 27 Anthropic also posted a paper saying an internal Claude model independently found ETH false on Oct 16."
      },
      "why": {
        "zh": "如果成立，是理论计算机和数学的大事；目前都是厂商自述，未经评审。",
        "en": "If it holds up, it is a major result in theory and math. For now these are vendor claims without peer review."
      },
      "sources": [
        {
          "label": {
            "zh": "OpenAI 公告",
            "en": "OpenAI announcement"
          },
          "url": "https://openai.com/index/building-scientific-understanding-with-academic-communities"
        },
        {
          "label": {
            "zh": "Anthropic ETH 论文与 Lean 证明",
            "en": "Anthropic ETH paper and Lean proof"
          },
          "url": "https://github.com/anthropics/formal-math"
        }
      ],
      "tags": [
        "math",
        "lean",
        "unreviewed"
      ]
    },
    {
      "id": "anthropic-s1-ipo",
      "example": false,
      "importance": 5,
      "category": "funding",
      "occurredAt": "2026-10-26T21:00:00+08:00",
      "ageHours": 50.3,
      "score": 4.747,
      "today": false,
      "whenLabel": null,
      "title": {
        "zh": "Anthropic 公开递交 S-1 招股书，拟以 ANT 代码登陆纳斯达克",
        "en": "Anthropic files public S-1 for Nasdaq IPO under ticker \"ANT\""
      },
      "summary": {
        "zh": "10 月 26 日公开递交。招股书显示 2026 上半年营收 162.9 亿美元（同比 +1304%），净亏损 807 亿美元，其中约 720 亿来自可转债公允价值变动；截至 9 月 30 日净收入留存率约 900%。发行规模和价格区间尚未确定。",
        "en": "Filed Oct 26. The prospectus shows H1 2026 revenue of $16.29B (up 1,304%) and an $80.7B net loss, about $72B of it from convertible-note fair-value changes; net revenue retention was about 900% as of Sept 30. Share count and price range are not set yet."
      },
      "why": {
        "zh": "第一家公开完整财务数据的头部大模型公司，收入、毛利和算力承诺都有了官方数字。",
        "en": "The first frontier lab to publish full audited financials: revenue, margins and compute commitments are now on record."
      },
      "sources": [
        {
          "label": {
            "zh": "Anthropic 公告",
            "en": "Anthropic announcement"
          },
          "url": "https://www.anthropic.com/news/s1-sec"
        },
        {
          "label": {
            "zh": "SEC S-1 原文",
            "en": "SEC S-1 filing"
          },
          "url": "https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm"
        },
        {
          "label": {
            "zh": "CNBC 解读",
            "en": "CNBC coverage"
          },
          "url": "https://www.cnbc.com/2026/10/26/anthropic-ipo-live-updates-ai-s-1-filing.html"
        }
      ],
      "tags": [
        "anthropic",
        "ipo"
      ]
    },
    {
      "id": "kimi-k3-1-pro",
      "example": false,
      "importance": 5,
      "category": "model",
      "occurredAt": "2026-10-24T01:15:00+08:00",
      "ageHours": 118.1,
      "score": 4.427,
      "today": false,
      "whenLabel": {
        "zh": "10 月 24 日上线 · 权重预计 11 月 4 日前",
        "en": "Live Oct 24 · weights expected by Nov 4"
      },
      "title": {
        "zh": "Kimi K3.1 Pro 上线，权重预计 11 月 4 日前开放",
        "en": "Kimi K3.1 Pro launches; weights expected by Nov 4"
      },
      "summary": {
        "zh": "月之暗面在 Kimi K3（总参数 2.8T，激活 104B）上继续训练，100 万 token 上下文，已在 Kimi 应用和 API（kimi-k3.1-pro）上线。官方称权重 11 月 4 日前开放；截至 10 月 28 日晚，Hugging Face 上还没有。Artificial Analysis 新版 v5 指数中得 46 分，是最高的开放权重模型。",
        "en": "Moonshot built it on Kimi K3 (2.8T total, 104B active), with a 1M-token context. It is live in the Kimi apps and the API (kimi-k3.1-pro). Moonshot says weights will be released by Nov 4; as of Oct 28 evening they were not on Hugging Face. It scores 46 on the new Artificial Analysis v5 index, the highest open-weights model."
      },
      "why": null,
      "sources": [
        {
          "label": {
            "zh": "Kimi 博客",
            "en": "Kimi blog"
          },
          "url": "https://www.kimi.ai/blog/kimi-k3-1-pro"
        },
        {
          "label": {
            "zh": "Kimi 官方帖",
            "en": "Kimi post"
          },
          "url": "https://x.com/Kimi_Moonshot/status/2113680701770068067"
        }
      ],
      "tags": [
        "open-weights",
        "weights-pending"
      ]
    },
    {
      "id": "dessa",
      "example": false,
      "importance": 5,
      "category": "model",
      "occurredAt": "2026-10-24T01:02:00+08:00",
      "ageHours": 118.3,
      "score": 4.426,
      "today": false,
      "whenLabel": null,
      "title": {
        "zh": "Block 发布语音模型 Dessa，尚未开放下载",
        "en": "Block introduces the Dessa speech model; download is not open"
      },
      "summary": {
        "zh": "Dessa 是全双工语音到语音模型，只用人与人的对话音频从零训练（不用文本，也不用预训练模型），会打断、附和、发笑。Block 称它在一台 NVIDIA GB200 NVL72 上训练，可在高端 MacBook（M4 Max 及以上）本地运行。尚未发布：下载「即将推出」，早期使用只给一小群 AI 安全研究者；没有宣布开源或发布日期。一条 X 帖有 1349 万次浏览、6.9 万次喜欢。",
        "en": "Dessa is a full-duplex speech-to-speech model trained from scratch only on human conversation audio (no text, no pretrained model). It interrupts, backchannels, and laughs. Block says it was trained on one NVIDIA GB200 NVL72 and runs locally on a high-end MacBook (M4 Max and up). It is not released: download is “coming soon”, early access is only for a small group of AI safety researchers, and no open-source release or release date has been announced. One X post had 13.49M views and 69K likes."
      },
      "why": null,
      "sources": [
        {
          "label": {
            "zh": "Block",
            "en": "Block"
          },
          "url": "https://block.xyz/dessa"
        },
        {
          "label": {
            "zh": "CoinDesk",
            "en": "CoinDesk"
          },
          "url": "https://www.coindesk.com/tech/2026/10/24/block-s-new-voice-ai-sounds-eerily-human-sparking-awe-and-scam-fears"
        },
        {
          "label": {
            "zh": "The Verge",
            "en": "The Verge"
          },
          "url": "https://www.theverge.com/ai-artificial-intelligence/1017906/jack-dorsey-block-dessa-ai-voice-speech"
        },
        {
          "label": {
            "zh": "X 帖",
            "en": "X post"
          },
          "url": "https://x.com/jack/status/2113697902916751661"
        }
      ],
      "tags": [
        "voice"
      ]
    },
    {
      "id": "mistral-large-4-weights",
      "example": false,
      "importance": 4,
      "category": "model",
      "occurredAt": "2026-10-31T00:00:00+08:00",
      "ageHours": 0,
      "score": 4,
      "today": false,
      "whenLabel": {
        "zh": "预计约 10 月 31 日 · 尚未发布",
        "en": "Expected ~Oct 31 · not yet released"
      },
      "title": {
        "zh": "Mistral Large 4 权重预计月底放出，10 月 28 日晚还没上 Hugging Face",
        "en": "Mistral Large 4 weights are expected by month’s end and were not on Hugging Face as of Oct 28 evening"
      },
      "summary": {
        "zh": "1 万亿参数，激活 490 亿。Mistral 说过权重会在月底放出。社区有帖称 10 月 31 日。截至 10 月 28 日晚，Hugging Face 上还没有。把「今天已发布」的帖子当成未证实。",
        "en": "1T parameters, 49B active. Mistral said the weights would come by the end of the month. A community post claims Oct 31. As of Oct 28 evening it was not on Hugging Face. Treat “released today” posts as unconfirmed."
      },
      "why": {
        "zh": "这是即将发生的开权重，不是已经能下载的模型。",
        "en": "This is an upcoming open-weights release, not a model you can download yet."
      },
      "sources": [
        {
          "label": {
            "zh": "社区帖（称 10 月 31 日）",
            "en": "Community post (claims Oct 31)"
          },
          "url": "https://x.com/Nterlace/status/2115112229645463894"
        }
      ],
      "tags": [
        "expected",
        "open-weights"
      ]
    },
    {
      "id": "nemotron-3-5-transcribe",
      "example": false,
      "importance": 4,
      "category": "open-source",
      "occurredAt": "2026-10-28T00:00:00-07:00",
      "ageHours": 8.3,
      "score": 3.966,
      "today": true,
      "whenLabel": null,
      "title": {
        "zh": "英伟达开源 Nemotron 3.5 Transcribe：一个模型做转写、分说话人、翻译和音频问答",
        "en": "NVIDIA open-sources Nemotron 3.5 Transcribe for transcription, diarization, translation and audio Q&A"
      },
      "summary": {
        "zh": "总参数 34B、激活约 4.3B，支持 40 种语言、最长 1 小时录音、最多 8 个说话人。官方称英文 Open ASR 榜 9 个数据集平均 WER 4.71%，开放权重第一；单张 H100 80GB 可部署，OpenMDW-1.1 许可，可商用。",
        "en": "34B total, about 4.3B active; 40 languages, recordings up to an hour, up to 8 speakers. NVIDIA reports a 4.71% mean WER on the 9-set English Open ASR leaderboard, best among open weights. Runs on one H100 80GB; OpenMDW-1.1 license, commercial use allowed."
      },
      "why": {
        "zh": "会议转写、字幕这类场景多了一个能本地跑、可商用的开源选项。",
        "en": "A self-hostable, commercially usable open model for meetings and captions."
      },
      "sources": [
        {
          "label": {
            "zh": "Hugging Face 模型页",
            "en": "Hugging Face model card"
          },
          "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B"
        },
        {
          "label": {
            "zh": "英伟达博客",
            "en": "NVIDIA blog"
          },
          "url": "https://huggingface.co/blog/nvidia/nemotron-transcribe-granary"
        },
        {
          "label": {
            "zh": "Artificial Analysis 评测",
            "en": "Artificial Analysis eval"
          },
          "url": "https://x.com/ArtificialAnlys/status/2115459886045942161"
        }
      ],
      "tags": [
        "nvidia",
        "speech",
        "open-weights"
      ]
    },
    {
      "id": "aa-intelligence-index-v5",
      "example": false,
      "importance": 4,
      "category": "research",
      "occurredAt": "2026-10-28T10:11:00+08:00",
      "ageHours": 13.1,
      "score": 3.946,
      "today": true,
      "whenLabel": {
        "zh": "10 月 28 日",
        "en": "Oct 28"
      },
      "title": {
        "zh": "Artificial Analysis 智能指数 v5：六成换成私有题，开源与闭源差距拉大到 12 分",
        "en": "Artificial Analysis Intelligence Index v5: 60% private evals, open-vs-closed gap widens to 12 points"
      },
      "summary": {
        "zh": "指数 60% 改用私有数据集，新增 FrontierCode 2.0（按开源维护者是否会合并来打分）和 Terminal-Bench-Science。榜首 Claude Opus 5.5（max）58 分，GPT-6 Astra 57，GPT-6.1 Sol 55，Gemini 4 Argon 50。开放权重最高是 Kimi K3.1 Pro 46 分（权重未放出）；已放权重的 DeepSeek V4.1 Flash 1017 和 GLM-5.3 都是 32 分。",
        "en": "60% of the index now uses private datasets. It adds FrontierCode 2.0 (scored by whether open-source maintainers would merge the change) and Terminal-Bench-Science. Claude Opus 5.5 (max) leads with 58, then GPT-6 Astra 57, GPT-6.1 Sol 55 and Gemini 4 Argon 50. The top open-weights model is Kimi K3.1 Pro at 46 (weights not out yet); among released weights, DeepSeek V4.1 Flash 1017 and GLM-5.3 score 32."
      },
      "why": null,
      "sources": [
        {
          "label": {
            "zh": "Artificial Analysis 文章",
            "en": "Artificial Analysis article"
          },
          "url": "https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v5"
        }
      ],
      "tags": [
        "benchmark"
      ]
    },
    {
      "id": "anthropic-drug-development-mithril",
      "example": false,
      "importance": 4,
      "category": "research",
      "occurredAt": "2026-10-28T04:27:00+08:00",
      "ageHours": 18.9,
      "score": 3.923,
      "today": true,
      "whenLabel": null,
      "title": {
        "zh": "Anthropic 自己做药：收购一人药企 Mithril Medicine，公布与 6 家机构的药研合作",
        "en": "Anthropic starts its own drug programs, acquires one-person biotech Mithril Medicine"
      },
      "summary": {
        "zh": "Mithril 创始人一个人用 Claude 统筹 11 家外部实验室，不到 18 个月把一种罕见肌营养不良症候选药做到小鼠概念验证；若顺利，最早 2028 年底进入人体试验。合作方包括 BMS、Generate Biomedicines 等 6 家，目前大多仍处临床前。",
        "en": "Mithril's sole founder used Claude to coordinate 11 outside labs and took a rare muscular dystrophy candidate to mouse proof of concept in under 18 months; human trials could start as early as late 2028. Six partners include BMS and Generate Biomedicines; most work is still preclinical."
      },
      "why": {
        "zh": "「一人 + AI」公司被实验室直接收购，AI 从写代码走到管实验和供应链。",
        "en": "A one-person, AI-run company bought by a frontier lab: AI moving from code into running lab programs."
      },
      "sources": [
        {
          "label": {
            "zh": "Anthropic 原文",
            "en": "Anthropic post"
          },
          "url": "https://www.anthropic.com/research/early-progress-drug-development"
        },
        {
          "label": {
            "zh": "Anthropic 官方推文",
            "en": "Anthropic on X"
          },
          "url": "https://x.com/AnthropicAI/status/2115178675050307942"
        }
      ],
      "tags": [
        "anthropic",
        "biotech",
        "agents"
      ]
    },
    {
      "id": "genie-4",
      "example": false,
      "importance": 4,
      "category": "model",
      "occurredAt": "2026-10-28T00:00:00+08:00",
      "ageHours": 23.3,
      "score": 3.905,
      "today": true,
      "whenLabel": {
        "zh": "10 月 28 日 · 更广开放预计几周内",
        "en": "Oct 28 · wider access expected within weeks"
      },
      "title": {
        "zh": "DeepMind 发布 Genie 4 世界模型：带原生音频，可用文字实时改场景",
        "en": "DeepMind unveils Genie 4, a world model with native audio and real-time text steering"
      },
      "summary": {
        "zh": "实时生成可交互的 3D 世界，同时生成环境声、音乐和对话；可以边玩边用自然语言改天气、镜头或加入角色，单个场景可保持约 2 分钟。目前只对受信测试者开放研究预览；Google AI Ultra 用户和 Gemini API / AI Studio 合作方预计几周内开放。",
        "en": "It generates interactive 3D worlds in real time, with ambient sound, music and dialogue. You can change the weather, camera or add characters with plain text while playing, and a scene holds for about 2 minutes. It is a research preview for trusted testers; Google AI Ultra users and Gemini API / AI Studio partners are expected to get access within weeks."
      },
      "why": null,
      "sources": [
        {
          "label": {
            "zh": "DeepMind 博客",
            "en": "DeepMind blog"
          },
          "url": "https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/"
        }
      ],
      "tags": [
        "world-model"
      ]
    },
    {
      "id": "axiom-cfsg-lean",
      "example": false,
      "importance": 4,
      "category": "research",
      "occurredAt": "2026-10-27T22:34:00+08:00",
      "ageHours": 24.8,
      "score": 3.899,
      "today": false,
      "whenLabel": {
        "zh": "10 月 27 日",
        "en": "Oct 27"
      },
      "title": {
        "zh": "Axiom 用 AI 代理在 Lean 里形式化了有限单群分类定理",
        "en": "Axiom used AI agents to formalize the classification of finite simple groups in Lean"
      },
      "summary": {
        "zh": "这个定理的原证明分散在数百篇论文里。Axiom 用一群代理花 24 天写出约 4,310 万行 Lean，过程中记录了 34 处与已发表文献的重要出入，还得到一个新定理。代码以 Apache-2.0 开源，截至 10 月 28 日 118 star。目前还没有外部独立复核。",
        "en": "The original proof is spread across hundreds of papers. Axiom ran a swarm of agents for 24 days and produced about 43.1M lines of Lean, logging 34 notable deviations from the published literature and a new theorem. The code is open source under Apache-2.0, with 118 stars as of Oct 28. No independent rebuild yet."
      },
      "why": null,
      "sources": [
        {
          "label": {
            "zh": "Axiom 博客",
            "en": "Axiom blog"
          },
          "url": "https://axiommath.ai/research/the-enormous-theorem-formalized/"
        },
        {
          "label": {
            "zh": "GitHub",
            "en": "GitHub"
          },
          "url": "https://github.com/AxiomMath/CFSG"
        }
      ],
      "tags": [
        "math",
        "lean",
        "agents"
      ]
    },
    {
      "id": "google-tpu-on-oci",
      "example": false,
      "importance": 4,
      "category": "product",
      "occurredAt": "2026-10-27T00:00:00-07:00",
      "ageHours": 32.3,
      "score": 3.869,
      "today": false,
      "whenLabel": {
        "zh": "预计 2027 年上线",
        "en": "Expected 2027"
      },
      "title": {
        "zh": "Google TPU 首次进入别家云：Oracle OCI 预计 2027 年提供",
        "en": "Google TPUs come to Oracle Cloud, the first hyperscaler outside Google, from 2027"
      },
      "summary": {
        "zh": "Oracle 与 Google 10 月 27 日宣布，OCI 将提供第八代 TPU 8t（训练）和 8i（推理），2027 年起分多年上量。同日 The Information 报道 OpenAI 也计划购买 Google 的 TPU，双方未确认。",
        "en": "Oracle and Google said on Oct 27 that OCI will offer eighth-gen TPU 8t (training) and 8i (inference), ramping from 2027. The same day The Information reported OpenAI plans to buy Google TPUs; neither company confirmed."
      },
      "why": {
        "zh": "TPU 开始当商品卖，英伟达之外的算力选项变多。",
        "en": "TPUs are becoming a merchant product, widening compute options beyond Nvidia."
      },
      "sources": [
        {
          "label": {
            "zh": "Oracle 公告",
            "en": "Oracle announcement"
          },
          "url": "https://www.oracle.com/news/announcement/ai-world-oracle-brings-google-tpus-to-oracle-cloud-infrastructure-2026-10-27/"
        },
        {
          "label": {
            "zh": "路透：OpenAI 拟购 TPU（报道）",
            "en": "Reuters: OpenAI plans to buy TPUs (report)"
          },
          "url": "https://www.investing.com/news/stock-market-news/openai-plans-to-buy-googles-ai-chips-the-information-reports-4977342"
        }
      ],
      "tags": [
        "google",
        "tpu",
        "compute",
        "expected"
      ]
    },
    {
      "id": "gpt-6-1-terra",
      "example": false,
      "importance": 4,
      "category": "model",
      "occurredAt": "2026-10-27T02:06:00+08:00",
      "ageHours": 45.2,
      "score": 3.818,
      "today": false,
      "whenLabel": {
        "zh": "10 月 27 日",
        "en": "Oct 27"
      },
      "title": {
        "zh": "OpenAI 发布 GPT-6.1 Terra：API 价格比 GPT-5.6 Terra 低一半",
        "en": "OpenAI releases GPT-6.1 Terra at half the API price of GPT-5.6 Terra"
      },
      "summary": {
        "zh": "API 名称 gpt-6.1-terra，每百万 token 输入 1 美元、输出 5 美元。OpenAI 称它在 Terminal-Bench 4.0 上得 50.3%，GPT-5.6 Terra 为 21.5%（官方数据）。已进入 ChatGPT Work 和 Codex。",
        "en": "The API name is gpt-6.1-terra, at $1 input and $5 output per 1M tokens. OpenAI says it scores 50.3% on Terminal-Bench 4.0, vs 21.5% for GPT-5.6 Terra (vendor numbers). It is available in ChatGPT Work and Codex."
      },
      "why": null,
      "sources": [
        {
          "label": {
            "zh": "OpenAI 公告",
            "en": "OpenAI announcement"
          },
          "url": "https://openai.com/index/introducing-gpt-6-1-terra/"
        },
        {
          "label": {
            "zh": "OpenAI 官方帖",
            "en": "OpenAI post"
          },
          "url": "https://x.com/OpenAI/status/2114780711664398349"
        }
      ],
      "tags": [
        "api",
        "coding"
      ]
    },
    {
      "id": "rea",
      "example": false,
      "importance": 4,
      "category": "open-source",
      "occurredAt": "2026-10-25T00:00:00+08:00",
      "ageHours": 95.3,
      "score": 3.625,
      "today": false,
      "whenLabel": {
        "zh": "10 月 25 日",
        "en": "Oct 25"
      },
      "title": {
        "zh": "REA：让编程代理逆向二进制、APK 和 Electron 应用",
        "en": "REA lets coding agents reverse-engineer binaries, APKs, and Electron apps"
      },
      "summary": {
        "zh": "MIT 许可的命令行和 MCP 服务。10 月 25 日发布 v6.8.0。约 40 万 star。10 月 19–25 日按新增 star 计，是 GitHub 上 AI 项目第一。",
        "en": "An MIT-licensed CLI and MCP server. v6.8.0 shipped on Oct 25. About 400K stars. It was the #1 GitHub AI project by stars gained from Oct 19–25."
      },
      "why": null,
      "sources": [
        {
          "label": {
            "zh": "GitHub",
            "en": "GitHub"
          },
          "url": "https://github.com/morluto/rea"
        },
        {
          "label": {
            "zh": "网站",
            "en": "Site"
          },
          "url": "https://rea.tools"
        }
      ],
      "tags": [
        "agents"
      ]
    },
    {
      "id": "nanbeige-4-3-33b",
      "example": false,
      "importance": 4,
      "category": "model",
      "occurredAt": "2026-10-24T00:00:00+08:00",
      "ageHours": 119.3,
      "score": 3.537,
      "today": false,
      "whenLabel": {
        "zh": "10 月 24 日",
        "en": "Oct 24"
      },
      "title": {
        "zh": "Nanbeige4.3-33B：Apache-2.0 的 330 亿开源编程模型",
        "en": "Nanbeige4.3-33B, a 33B Apache-2.0 model for agentic coding"
      },
      "summary": {
        "zh": "330 亿参数，Apache-2.0，上下文 512K，面向代理式编程。截至 10 月 28 日，Hugging Face 趋势第 3，699 个 like，7,412 次下载。厂商报告 Terminal-Bench 2.1 为 80.9（未经独立核实）。",
        "en": "33B parameters, Apache-2.0, 512K context, aimed at agentic coding. As of Oct 28: Hugging Face trending #3, 699 likes, 7,412 downloads. The vendor reports Terminal-Bench 2.1 at 80.9 (unverified)."
      },
      "why": null,
      "sources": [
        {
          "label": {
            "zh": "Hugging Face",
            "en": "Hugging Face"
          },
          "url": "https://huggingface.co/Nanbeige/Nanbeige4.3-33B"
        }
      ],
      "tags": [
        "open-weights"
      ]
    },
    {
      "id": "anthropic-formal-verification",
      "example": false,
      "importance": 4,
      "category": "research",
      "occurredAt": "2026-10-23T06:01:00+08:00",
      "ageHours": 137.3,
      "score": 3.472,
      "today": false,
      "whenLabel": {
        "zh": "10 月 23 日",
        "en": "Oct 23"
      },
      "title": {
        "zh": "Anthropic 用 Claude 证明 123 个开源库的二进制不会崩溃",
        "en": "Anthropic used Claude to prove 123 open-source libraries' binaries won't crash"
      },
      "summary": {
        "zh": "与 Theorem 合作，Claude 为 123 个 OSS-Fuzz 库的 239 个模糊测试入口写出 Rocq 证明，约 1.8 亿行，在 RISC-V 二进制层面证明不会崩溃（有输入大小等前提）。部分证明已公开。",
        "en": "Working with Theorem, Claude wrote Rocq proofs for 239 fuzz entry points across 123 OSS-Fuzz libraries, about 180M lines, showing at the RISC-V binary level that they won't crash (under stated assumptions such as input size). Some proofs are public."
      },
      "why": null,
      "sources": [
        {
          "label": {
            "zh": "Anthropic 研究",
            "en": "Anthropic research"
          },
          "url": "https://www.anthropic.com/research/formal-verification"
        },
        {
          "label": {
            "zh": "公开的证明",
            "en": "Published proofs"
          },
          "url": "https://github.com/theorem-labs/oss-verify-proofs"
        }
      ],
      "tags": [
        "formal-verification",
        "security"
      ]
    },
    {
      "id": "gemini-4-argon",
      "example": false,
      "importance": 5,
      "category": "model",
      "occurredAt": "2026-10-13T00:00:00+08:00",
      "ageHours": 383.3,
      "score": 3.367,
      "today": false,
      "whenLabel": {
        "zh": "10 月 13 日",
        "en": "Oct 13"
      },
      "title": {
        "zh": "Google 推出旗舰模型 Gemini 4 Argon",
        "en": "Google launches flagship model Gemini 4 Argon"
      },
      "summary": {
        "zh": "面向编程和长程代理工作的新旗舰。输入和输出共用 100 万 token 窗口（单次回复可以超过 6.4 万，最高约 100 万）。Google 称它在 Vals Index 上排第一。介绍价为每 100 万输入/输出 token 2 美元 / 10 美元，到 11 月 30 日，之后为 4 美元 / 20 美元。相关报道把它描述成 Google 在取消 Gemini 3.5 Pro 以及 DeepMind 领导层变动之后的回击。",
        "en": "A new flagship for coding and long agentic work. One 1M-token window covers input and output (a single response can exceed 64K, up to about 1M). Google says it is #1 on the Vals Index. Intro price is $2 / $10 per 1M input/output tokens through Nov 30, then $4 / $20. A linked report frames it as Google's comeback after scrapping Gemini 3.5 Pro and a DeepMind leadership shake-up."
      },
      "why": null,
      "sources": [
        {
          "label": {
            "zh": "Google 博客",
            "en": "Google blog"
          },
          "url": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/"
        },
        {
          "label": {
            "zh": "报道",
            "en": "Report"
          },
          "url": "https://tech.yahoo.com/ai/gemini/articles/google-launches-flagship-ai-model-160147613.html"
        }
      ],
      "tags": [
        "gemini"
      ]
    },
    {
      "id": "deepseek-deepinfer",
      "example": false,
      "importance": 4,
      "category": "open-source",
      "occurredAt": "2026-10-20T00:00:00+08:00",
      "ageHours": 215.3,
      "score": 3.203,
      "today": false,
      "whenLabel": {
        "zh": "10 月 20 日",
        "en": "Oct 20"
      },
      "title": {
        "zh": "DeepSeek 开源线上服务用的推理框架 DeepInfer",
        "en": "DeepSeek open-sources DeepInfer, the inference framework behind its services"
      },
      "summary": {
        "zh": "Apache-2.0。目前支持 DeepSeek-V4.1-Flash，硬件为 NVIDIA SM100 和华为昇腾 950。不含 KV cache 系统、PD 分离和负载均衡，不是完整的服务栈。截至 10 月 28 日 23,723 star。",
        "en": "Apache-2.0. It currently supports DeepSeek-V4.1-Flash on NVIDIA SM100 and Huawei Ascend 950. It leaves out the KV cache system, PD disaggregation and load balancing, so it is not a full serving stack. 23,723 stars as of Oct 28."
      },
      "why": null,
      "sources": [
        {
          "label": {
            "zh": "GitHub",
            "en": "GitHub"
          },
          "url": "https://github.com/deepseek-ai/DeepInfer"
        }
      ],
      "tags": [
        "inference"
      ]
    },
    {
      "id": "comfyui-contest",
      "example": false,
      "importance": 3,
      "category": "open-source",
      "occurredAt": "2026-10-30T09:00:00-07:00",
      "ageHours": 0,
      "score": 3,
      "today": false,
      "whenLabel": {
        "zh": "截止 10 月 30 日 09:00（美西）",
        "en": "Deadline Oct 30, 09:00 PT"
      },
      "title": {
        "zh": "ComfyUI 开源应用赛 10 月 30 日截止",
        "en": "ComfyUI open-source app contest deadline is Oct 30"
      },
      "summary": {
        "zh": "奖金 2.5 万美元，设有开权重赛道。相关动态见 ComfyUI 的 OPENjutsu 帖。",
        "en": "Prizes are $25K, with an open-weights track. A related update is the ComfyUI OPENjutsu post."
      },
      "why": {
        "zh": "截止前后可能会出现一批开源的 ComfyUI 应用。",
        "en": "A wave of open ComfyUI apps is likely around the deadline."
      },
      "sources": [
        {
          "label": {
            "zh": "ComfyUI 帖",
            "en": "ComfyUI post"
          },
          "url": "https://x.com/ComfyUI/status/2114843030364795231"
        }
      ],
      "tags": [
        "contest"
      ]
    },
    {
      "id": "openbot-desktop",
      "example": false,
      "importance": 3,
      "category": "open-source",
      "occurredAt": "2026-10-28T23:10:00+08:00",
      "ageHours": 0.2,
      "score": 3,
      "today": true,
      "whenLabel": null,
      "title": {
        "zh": "OpenBot：把 Codex、Claude、Grok 当成常驻同事的本地桌面工作台，一天涨星 1000+",
        "en": "OpenBot, a local-first desktop for persistent Codex/Claude/Grok teammates, gains 1,000+ stars in a day"
      },
      "summary": {
        "zh": "每个代理有独立工作区和会话，支持任务队列、文件互传、内置浏览器和代理之间互发消息，也接 Gemini、Cursor、Cline 等 CLI。仓库约 3000 星，GitHub 日榜当天 +1053。注意：仍是开发预览版，代理默认全权限运行；许可为非商用 PolyForm Noncommercial。",
        "en": "Each agent gets its own workspace and session, with task queues, file transfer, an embedded browser and agent-to-agent messaging; also drives Gemini, Cursor and Cline CLIs. About 3,000 stars, +1,053 on the GitHub daily trending list. Still a dev preview: agents run with full access by default; PolyForm Noncommercial license."
      },
      "why": {
        "zh": "多代理「同事化」的开源尝试，涨星速度在同类里很突出。",
        "en": "An open attempt at treating many agents as teammates; unusually fast star growth."
      },
      "sources": [
        {
          "label": {
            "zh": "GitHub 仓库",
            "en": "GitHub repo"
          },
          "url": "https://github.com/nightly-labs/openbot"
        }
      ],
      "tags": [
        "agents",
        "open-source",
        "github-trending"
      ]
    },
    {
      "id": "magpie",
      "example": false,
      "importance": 3,
      "category": "open-source",
      "occurredAt": "2026-10-28T21:45:00+08:00",
      "ageHours": 1.6,
      "score": 2.995,
      "today": true,
      "whenLabel": {
        "zh": "截至 10 月 28 日",
        "en": "As of Oct 28"
      },
      "title": {
        "zh": "magpie：在本机互转 OpenAI、Anthropic 和 Gemini 接口",
        "en": "magpie, a local gateway between the OpenAI, Anthropic, and Gemini APIs"
      },
      "summary": {
        "zh": "本地 LLM 网关，监听 127.0.0.1:3425，说 OpenAI、Anthropic、Gemini 的接口并能互相转换。可以让 Claude Code 用 Kimi、让 Codex 用 DeepSeek。额度用尽时自动换账号。MIT。截至 10 月 28 日有 24,419 star。这里没有单独的发布日期。",
        "en": "A local LLM gateway on 127.0.0.1:3425. It speaks the OpenAI, Anthropic, and Gemini APIs and converts between them, so Claude Code can use Kimi and Codex can use DeepSeek. It switches accounts when a quota runs out. MIT license. 24,419 stars as of Oct 28. No separate release date was recorded."
      },
      "why": null,
      "sources": [
        {
          "label": {
            "zh": "GitHub",
            "en": "GitHub"
          },
          "url": "https://github.com/yetone/magpie"
        },
        {
          "label": {
            "zh": "网站",
            "en": "Site"
          },
          "url": "https://usemagpie.ai"
        }
      ],
      "tags": [
        "gateway"
      ]
    },
    {
      "id": "deeptutor",
      "example": false,
      "importance": 3,
      "category": "open-source",
      "occurredAt": "2026-10-28T21:45:00+08:00",
      "ageHours": 1.6,
      "score": 2.995,
      "today": true,
      "whenLabel": {
        "zh": "10 月 28 日",
        "en": "Oct 28"
      },
      "title": {
        "zh": "港大 DeepTutor：开源的个性化 AI 导师",
        "en": "HKU DeepTutor, an open-source personalized AI tutor"
      },
      "summary": {
        "zh": "香港大学的开源个性化 AI 导师。截至 10 月 28 日，当天新增 2,376 star，总计 47,336。",
        "en": "HKU's open-source personalized AI tutor. As of Oct 28: +2,376 stars that day, 47,336 total."
      },
      "why": null,
      "sources": [
        {
          "label": {
            "zh": "GitHub",
            "en": "GitHub"
          },
          "url": "https://github.com/HKUDS/DeepTutor"
        }
      ],
      "tags": [
        "education"
      ]
    },
    {
      "id": "less-ai-looking-ui",
      "example": false,
      "importance": 3,
      "category": "open-source",
      "occurredAt": "2026-10-28T21:45:00+08:00",
      "ageHours": 1.6,
      "score": 2.995,
      "today": true,
      "whenLabel": {
        "zh": "10 月 28 日",
        "en": "Oct 28"
      },
      "title": {
        "zh": "让编程代理少做出「一看就是 AI」界面的设计技能在涨",
        "en": "Design skills that push coding agents toward less AI-looking UI are climbing"
      },
      "summary": {
        "zh": "截至 10 月 28 日：hallmark 当天新增 2,109 star，总计 40,092；impeccable 本周新增 6,713 star，总计 91,473。",
        "en": "As of Oct 28: hallmark +2,109 stars that day, 40,092 total; impeccable +6,713 stars this week, 91,473 total."
      },
      "why": null,
      "sources": [
        {
          "label": {
            "zh": "hallmark",
            "en": "hallmark"
          },
          "url": "https://github.com/Nutlope/hallmark"
        },
        {
          "label": {
            "zh": "impeccable",
            "en": "impeccable"
          },
          "url": "https://github.com/pbakaus/impeccable"
        }
      ],
      "tags": [
        "design",
        "agents"
      ]
    },
    {
      "id": "minimax-mcode-stack",
      "example": false,
      "importance": 3,
      "category": "open-source",
      "occurredAt": "2026-10-28T10:17:00+08:00",
      "ageHours": 13,
      "score": 2.96,
      "today": true,
      "whenLabel": null,
      "title": {
        "zh": "MiniMax 开源 MCode Stack：团队自己写 MiniMax Code 时用的工作流和技能",
        "en": "MiniMax open-sources MCode Stack, the workflows and skills its team uses to build MiniMax Code"
      },
      "summary": {
        "zh": "15 个可直接安装的 SKILL.md 技能和清单，讲何时拆子代理、长命令等多久、长会话怎么不降级、每轮上下文放什么。MIT 许可，10 月 28 日刚发布。",
        "en": "15 installable SKILL.md skills and checklists covering when to spawn subagents, how long to wait on long commands, keeping long sessions sharp and what goes into context each turn. MIT license, released Oct 28."
      },
      "why": {
        "zh": "大厂把内部编程代理用法直接开源，比「最佳实践」帖子实在。",
        "en": "A lab publishing its real internal agent workflows, not generic tips."
      },
      "sources": [
        {
          "label": {
            "zh": "GitHub 仓库",
            "en": "GitHub repo"
          },
          "url": "https://github.com/MiniMax-AI/mcode-stack"
        }
      ],
      "tags": [
        "minimax",
        "skills",
        "coding-agents"
      ]
    },
    {
      "id": "liquid-lfm3",
      "example": false,
      "importance": 3,
      "category": "model",
      "occurredAt": "2026-10-27T00:00:00-07:00",
      "ageHours": 32.3,
      "score": 2.902,
      "today": false,
      "whenLabel": {
        "zh": "10 月 27 日",
        "en": "Oct 27"
      },
      "title": {
        "zh": "Liquid AI 发布 LFM3：约 2B 参数的开放权重小模型家族",
        "en": "Liquid AI releases LFM3, a family of ~2B open-weight small models"
      },
      "summary": {
        "zh": "共用一个 1.7B 基座，有文本、视觉、音频、全模态等六个模型，文本模型 256K 上下文，面向本地运行的代理。官方称 1.7B 版 SWE-bench Verified 43.4%。许可证是 LFM Open License，年收入 1,000 万美元以上的公司商用受限。",
        "en": "Six models share a 1.7B base, covering text, vision, audio and omni, with 256K context for the text model, aimed at agents running locally. Liquid says the 1.7B model scores 43.4% on SWE-bench Verified. It uses the LFM Open License, which limits commercial use for companies with over $10M in revenue."
      },
      "why": null,
      "sources": [
        {
          "label": {
            "zh": "Liquid AI 博客",
            "en": "Liquid AI blog"
          },
          "url": "https://www.liquid.ai/blog/lfm3"
        },
        {
          "label": {
            "zh": "Hugging Face",
            "en": "Hugging Face"
          },
          "url": "https://huggingface.co/collections/LiquidAI/lfm3"
        }
      ],
      "tags": [
        "open-weights",
        "small-models"
      ]
    },
    {
      "id": "ling-3-1-flash-1019",
      "example": false,
      "importance": 3,
      "category": "model",
      "occurredAt": "2026-10-27T13:46:00+08:00",
      "ageHours": 33.6,
      "score": 2.898,
      "today": false,
      "whenLabel": {
        "zh": "10 月 27 日",
        "en": "Oct 27"
      },
      "title": {
        "zh": "蚂蚁开源 Ling-3.1-flash-1019：总参数 560B、激活 25B，MIT 许可",
        "en": "Ant's inclusionAI releases Ling-3.1-flash-1019: 560B total, 25B active, MIT license"
      },
      "summary": {
        "zh": "混合线性注意力 MoE，100 万 token 上下文，比上一版强化了推理和代理能力。官方称 Artificial Analysis 指数 v4.3 得 44 分（上一版 41）。SGLang 和 vLLM 可直接部署。",
        "en": "A hybrid linear-attention MoE with a 1M-token context, post-trained for better reasoning and agent work. inclusionAI reports 44 on the Artificial Analysis index v4.3, up from 41. Runs on SGLang and vLLM."
      },
      "why": null,
      "sources": [
        {
          "label": {
            "zh": "Hugging Face",
            "en": "Hugging Face"
          },
          "url": "https://huggingface.co/inclusionAI/Ling-3.1-flash-1019"
        }
      ],
      "tags": [
        "open-weights"
      ]
    },
    {
      "id": "unlimited-ocr-2",
      "example": false,
      "importance": 3,
      "category": "model",
      "occurredAt": "2026-10-26T00:00:00+08:00",
      "ageHours": 71.3,
      "score": 2.787,
      "today": false,
      "whenLabel": {
        "zh": "10 月 26 日",
        "en": "Oct 26"
      },
      "title": {
        "zh": "百度 Unlimited-OCR-2：面向长文档的开源 OCR",
        "en": "Baidu Unlimited-OCR-2, an open-source OCR model for long documents"
      },
      "summary": {
        "zh": "33 亿 MoE，做长文档 OCR，MIT 许可。截至 10 月 28 日，Hugging Face 趋势第 12，300 个 like，8,357 次下载。",
        "en": "A 3.3B mixture-of-experts model for long-document OCR, MIT license. As of Oct 28: Hugging Face trending #12, 300 likes, 8,357 downloads."
      },
      "why": null,
      "sources": [
        {
          "label": {
            "zh": "Hugging Face",
            "en": "Hugging Face"
          },
          "url": "https://huggingface.co/baidu/Unlimited-OCR-2"
        }
      ],
      "tags": [
        "ocr"
      ]
    },
    {
      "id": "cursor-computer-use",
      "example": false,
      "importance": 3,
      "category": "product",
      "occurredAt": "2026-10-23T23:59:00+08:00",
      "ageHours": 119.3,
      "score": 2.653,
      "today": false,
      "whenLabel": {
        "zh": "10 月 23 日",
        "en": "Oct 23"
      },
      "title": {
        "zh": "Cursor 代理现在可以操作电脑界面了",
        "en": "Cursor agents can now operate desktop UIs"
      },
      "summary": {
        "zh": "代理能点击、输入、截屏，驱动 iOS 模拟器等桌面应用，测试自己写的功能。自托管机器上加 --computer-use 开启，支持 macOS 和 Linux。",
        "en": "Agents can click, type and take screenshots to drive desktop apps such as the iOS Simulator and test what they built. On self-hosted machines, enable it with --computer-use; macOS and Linux are supported."
      },
      "why": null,
      "sources": [
        {
          "label": {
            "zh": "Cursor 官方帖",
            "en": "Cursor post"
          },
          "url": "https://x.com/cursor_ai/status/2113661506634305834"
        },
        {
          "label": {
            "zh": "Cursor 文档",
            "en": "Cursor docs"
          },
          "url": "https://cursor.com/docs/cloud-agent/self-hosted/computer-use"
        }
      ],
      "tags": [
        "agents",
        "coding"
      ]
    },
    {
      "id": "jev-and-kev",
      "example": false,
      "importance": 3,
      "category": "research",
      "occurredAt": "2026-10-01T00:00:00+08:00",
      "ageHours": 671.3,
      "score": 1.501,
      "today": false,
      "whenLabel": {
        "zh": "2026 年 9–10 月",
        "en": "Sep–Oct 2026"
      },
      "title": {
        "zh": "Jev 用校准概率代替文本，Kev 是它的开源再实现",
        "en": "Jev outputs calibrated probabilities instead of text; Kev is an open reimplementation"
      },
      "summary": {
        "zh": "TypeSafe 的 Jev 是封闭的「决策模型」，输出校准过的概率而不是文本，用来替代「让大模型当裁判」。已有开源再实现，例如 Kev。时间范围是 9 月到 10 月，没有更精确的单日。",
        "en": "TypeSafe's Jev is a closed “decision model” that outputs calibrated probabilities instead of text, as an alternative to using an LLM as a judge. Open reimplementations exist, including Kev. The reported window is September–October, with no single day."
      },
      "why": null,
      "sources": [
        {
          "label": {
            "zh": "Kev 介绍",
            "en": "Kev introduction"
          },
          "url": "https://jaredpalmer.com/blog/introducing-kev"
        },
        {
          "label": {
            "zh": "排行榜",
            "en": "Leaderboard"
          },
          "url": "https://huggingface.co/spaces/multimodalart/jev-decision-index"
        }
      ],
      "tags": [
        "decision-model"
      ]
    }
  ]
}