{
 "date": "2026-10-17",
 "headline": "美国开源阵营一周三连发：Reflection 用 Apache 2.0 放出 501B 的 Beam",
 "lede": "继周四 Meta 开放 Muse Spark Open 之后，Reflection 周五发布首个模型 Beam。Artificial Analysis 给的智能指数是 30，在美国开源模型里排第 2；Anthropic 同一天发了两篇“Claude 做科研”的成果。",
 "sample": true,
 "url": "https://ai.van2go.net/2026-10-17/",
 "count": 14,
 "sections": [
  {
   "key": "labs",
   "title": "主流大模型厂商在干嘛",
   "items": [
    {
     "title": "Reflection 发布首个模型 Beam（Apache 2.0）",
     "summary": "501B 总参数、23B 激活，开放权重可商用；Hugging Face 已上架，Transformers、vLLM、SGLang 都支持。",
     "url": "https://huggingface.co/blog/reflection-beam",
     "src": "huggingface.co",
     "tags": [
      "Reflection",
      "开源权重",
      "MoE"
     ],
     "metric": null,
     "unverified": false,
     "hot": true
    },
    {
     "title": "Anthropic：Claude 自主设计量子电路，部分比已发表最优方案好一个数量级",
     "summary": "给出 8 个化学、优化、物理模拟、数论领域的电路需求和自动评分器，Claude 基本独立完成设计，能明显降低实用量子计算所需的资源。",
     "url": "https://www.anthropic.com/research/designing-efficient-quantum-circuits",
     "src": "anthropic.com",
     "tags": [
      "Anthropic",
      "AI for Science"
     ],
     "metric": null,
     "unverified": false,
     "hot": false
    },
    {
     "title": "Anthropic：用 Claude 解开 ECDSA 曲线种子来源之谜",
     "summary": "几十年来一直有人怀疑 NIST 椭圆曲线的种子有问题，Anthropic 的结论是“只是时间戳”，不需要采取任何行动。",
     "url": "https://www.anthropic.com/research/ecdsa-curve-origins",
     "src": "anthropic.com",
     "tags": [
      "Anthropic",
      "密码学"
     ],
     "metric": null,
     "unverified": false,
     "hot": false
    },
    {
     "title": "Anthropic 调查：大众和 AI 从业者对 AI 进展的预期差距",
     "summary": "39% 的美国成年人、91% 的 Anthropic 员工认为 2030 年前 AI 能做出诺奖级发现；对自己工作被 AI 接手的预期，两边差距更大。",
     "url": "https://www.anthropic.com/research/ai-expectations-gap",
     "src": "anthropic.com",
     "tags": [
      "Anthropic",
      "经济"
     ],
     "metric": null,
     "unverified": false,
     "hot": false
    }
   ]
  },
  {
   "key": "evals",
   "title": "主流评测 / 数据公司在干嘛",
   "items": [
    {
     "title": "Artificial Analysis：Beam 智能指数 30，美国开源模型第 2",
     "summary": "排在 Muse Spark Open（39）之后、Nemotron 3.5 Super VL（26）之前；每个任务约用 26k 输出 token，只有同分段对手的一半。",
     "url": "https://artificialanalysis.ai/articles/reflection-beam",
     "src": "artificialanalysis.ai",
     "tags": [
      "榜单",
      "智能指数"
     ],
     "metric": "AA 30",
     "unverified": false,
     "hot": true
    },
    {
     "title": "Epoch AI：开放权重模型在网络安全能力上落后 6–7 个月",
     "summary": "基于 Epoch 能力指数（ECI），比较开放模型和闭源模型在网络安全任务上的差距。",
     "url": "https://epoch.ai/data-insights/open-closed-cyber-eci-gap",
     "src": "epoch.ai",
     "tags": [
      "Epoch",
      "网络安全"
     ],
     "metric": null,
     "unverified": false,
     "hot": false
    },
    {
     "title": "UniPat：PaperBenchX，测 AI 能否跨学科复现论文",
     "summary": "评估 Agent 能否重建已发表的实验、运行各学科的专业软件、重新得出可验证的结果（10 月 8 日发布，近期热度回升）。",
     "url": "https://unipat.ai/blog",
     "src": "unipat.ai",
     "tags": [
      "UniPat",
      "基准"
     ],
     "metric": null,
     "unverified": false,
     "hot": false
    }
   ]
  },
  {
   "key": "github",
   "title": "GitHub 上有什么好玩的",
   "items": [
    {
     "title": "morluto/rea",
     "summary": "用 Agent 逆向一切：从应用行为一路拆到原生二进制。连续两天 Trendshift 第 1。",
     "url": "https://github.com/morluto/rea",
     "src": "trendshift.io",
     "tags": [
      "Agent",
      "逆向"
     ],
     "metric": "热度 419",
     "unverified": false,
     "hot": true
    },
    {
     "title": "debpalash/VoiceStudio",
     "summary": "完全本地运行的开源 ElevenLabs 替代：声音克隆、视频配音、听写、有声书，支持 646 种语言。",
     "url": "https://github.com/debpalash/VoiceStudio",
     "src": "trendshift.io",
     "tags": [
      "TTS"
     ],
     "metric": "热度 96",
     "unverified": false,
     "hot": false
    },
    {
     "title": "MadsLorentzen/ai-job-search",
     "summary": "基于 Claude Code 的求职框架：评估岗位、改简历、写求职信、准备面试，全在本机运行。",
     "url": "https://github.com/MadsLorentzen/ai-job-search",
     "src": "trendshift.io",
     "tags": [
      "Claude Code"
     ],
     "metric": "热度 47",
     "unverified": false,
     "hot": false
    },
    {
     "title": "K-Dense-AI/scientific-agent-skills",
     "summary": "177 个科研 Agent Skills，外加 100 多个科学数据库，覆盖生物、化学、医学。",
     "url": "https://github.com/K-Dense-AI/scientific-agent-skills",
     "src": "trendshift.io",
     "tags": [
      "Skills",
      "科研"
     ],
     "metric": "热度 32",
     "unverified": false,
     "hot": false
    }
   ]
  },
  {
   "key": "x",
   "title": "X 上有什么好玩的",
   "items": [
    {
     "title": "@hanrui_w：Nebius 24 小时内首发三款美国开源模型",
     "summary": "Reflection Beam、Nemotron 3.5 Super VL、Muse Spark Open 都在发布当天上线 Nebius Token Factory。",
     "url": "https://x.com/hanrui_w/status/2111107154619101681",
     "src": "x.com",
     "tags": [
      "推理平台"
     ],
     "metric": "♥ 44",
     "unverified": false,
     "hot": false
    },
    {
     "title": "@KadinYu：“Claude 30 分钟破解 90 年代密码学悬案”",
     "summary": "把 Anthropic 当天发的两篇研究（ECDSA 曲线来源、量子电路）打包成一条热帖。",
     "url": "https://x.com/KadinYu/status/2111235084359700677",
     "src": "x.com",
     "tags": [
      "Anthropic"
     ],
     "metric": null,
     "unverified": false,
     "hot": false
    },
    {
     "title": "@maga_musk：Muse Spark Open 拿下 Code Arena 美国开源第一（1511 分）",
     "summary": "分数来自帖子本身，还没在 Arena 官方榜单上核实。",
     "url": "https://x.com/maga_musk/status/2111336964930539808",
     "src": "x.com",
     "tags": [
      "榜单"
     ],
     "metric": null,
     "unverified": true,
     "hot": false
    }
   ]
  }
 ]
}