{
  "date": "2026-08-04",
  "title": "早报｜2026-08-04：新能源规划强调消纳能力，“8 万亿 token”统计口径存疑",
  "url": "https://daily.yisiliu.io/20260804/",
  "summary": "新能源规划把消纳与调节能力放在更前位置；大模型调用量需区分免费调用、付费调用和独立用户。",
  "sections": [
    {
      "id": "judgment",
      "title": "今日判断",
      "items": [
        {"title":"电力政策","summary":"两部门规划提出，2030 年全国新能源利用率保持在 90% 左右。接下来要看储能、输电和需求响应能否按进度跟上。"},
        {"title":"个人金融","summary":"四家国有大行已用明示表列出正常履约时的年化成本上限，具体口径以银行明示表为准。"},
        {"title":"中文互联网","summary":"“一天 8 万亿 token”来自社区对单一开发者入口的转述，现有材料没有原始统计页，不能据此推算用户数、收入或全网份额。"},
        {"title":"城市夜场","summary":"上海和杭州各有两场当天晚间活动，时间、场馆和售票入口明确。"}
      ]
    },
    {
      "id": "macro",
      "title": "宏观、金融与政策",
      "items": [
        {"title":"新型电力系统规划把新能源利用率目标定在 90% 左右","summary":"国家发展改革委、国家能源局印发的规划提出，到 2030 年非化石能源发电量占比达到 50%，全国新能源利用率保持在 90% 左右。","why_it_matters":"储能、输电和需求响应的配套进度会影响项目回报。","source_name":"国家发展改革委、国家能源局","source_url":"https://www.ndrc.gov.cn/xxgk/zcfb/tz/202608/t20260803_1406883.html","tags":["能源","政策"]},
        {"title":"个人贷款进入综合融资成本“一表明示”阶段","summary":"央视网财经转述《中国证券报》调研称，四家国有大行披露正常履约情况下相关个人贷款的年化成本上限为 6%，具体口径应以银行明示表为准。","why_it_matters":"贷款产品比较应回到统一的年化成本口径。","source_name":"央视网财经","source_url":"https://finance.cctv.com/2026/08/02/ARTIdzt1Eb0IAtSVp5V43yQl260802.shtml","tags":["个人金融","银行"]},
        {"title":"四个核电项目共八台机组获核准","summary":"新浪财经报道，四个核电项目、合计八台机组获核准；行情动因仍需结合后续招标和订单公告判断。","why_it_matters":"核准是产业链资本开支的前置条件，但不是收入确认。","source_name":"新浪财经","source_url":"https://finance.sina.com.cn/jjxw/2026-07-31/doc-iniktkcc1182086.shtml","tags":["核电","产业"]}
      ]
    },
    {
      "id": "hotspots",
      "title": "中文互联网热点 / 讨论场",
      "items": [
        {
          "title": "“8 万亿 token”热搜：单一入口调用量不等于用户规模",
          "summary": "部分社区帖子转述称，OpenCode 单日处理了约 8 万亿个与 DeepSeek V4 Flash 有关的 token；现有材料未包含原始统计页，数字及统计范围无法独立核验。",
          "discussion": "微博样本主要放大数字规模和国产模型竞争叙事；知乎可见条目延伸到定价与现金流问题。36 氪文章提供同期国产模型发布背景，不构成对 8 万亿数字的验证。",
          "near_model": {"name":"单位经济","field":"产业经济学","analysis":"边际收入、推理成本、免费额度占比和付费留存决定规模能否变成可持续业务；调用量本身不能说明毛利。"},
          "far_model": {"name":"客运周转量","field":"交通统计","analysis":"乘客人数、乘车次数和人公里不可互换；长上下文、智能体循环和重试同样会放大 token 总量。"},
          "model_boundary": "两种框架都无法替代原始平台披露，也回答不了模型在真实任务中的正确率、延迟和迁移成本。",
          "next_variables": "原始统计页、付费 token 占比、结算单价、推理成本、去重独立会话和连续留存。",
          "source_name": "微博热搜",
          "source_url": "https://s.weibo.com/weibo?q=%23DeepSeek%E4%B8%80%E5%A4%A9%E6%B6%88%E8%80%97%E4%BA%868%E4%B8%87%E4%BA%BF%23",
          "sources": [
            {"name":"微博热搜","url":"https://s.weibo.com/weibo?q=%23DeepSeek%E4%B8%80%E5%A4%A9%E6%B6%88%E8%80%97%E4%BA%868%E4%B8%87%E4%BA%BF%23","role":"讨论与注意力样本"},
            {"name":"知乎讨论","url":"https://www.zhihu.com/question/2067611934703216182","role":"社区讨论样本"},
            {"name":"36 氪","url":"https://36kr.com/p/3922353035433603","role":"同期产品背景，不验证调用量"}
          ]
        },
        {
          "title": "新型电力系统规划：电源扩张之后，瓶颈落在消纳",
          "summary": "国家发展改革委、国家能源局印发的规划提出，到 2030 年非化石能源发电量占比达到 50%，全国新能源利用率保持在 90% 左右。",
          "discussion": "微博讨论主要围绕规划目标展开，知乎上的问题更多涉及行业结构和 A 股相关板块。社区链接只能说明议题受到关注。",
          "near_model": {"name":"最大流—最小割","field":"运筹学","analysis":"网架和调峰能力构成容量瓶颈；这是简化视角，实际消纳还受潮流稳定、储能时序和负荷曲线限制。"},
          "far_model": {"name":"剧场换景调度","field":"舞台管理","analysis":"演员数量之外，舞台空档、布景转换时间、工作人员和提示序列共同决定演出吞吐；它提示电量存在时间错配。"},
          "model_boundary": "网络流定位空间瓶颈，换景调度补上时间错配；成本分担仍要回到电价、辅助服务和容量补偿制度。",
          "next_variables": "新增储能调用小时、充放电次数、辅助服务收入，以及虚拟电厂结算身份。",
          "source_name": "国家发展改革委、国家能源局",
          "source_url": "https://www.ndrc.gov.cn/xxgk/zcfb/tz/202608/t20260803_1406883.html",
          "sources": [
            {"name":"国家发展改革委、国家能源局","url":"https://www.ndrc.gov.cn/xxgk/zcfb/tz/202608/t20260803_1406883.html","role":"事实与政策原文"},
            {"name":"微博议题","url":"https://s.weibo.com/weibo?q=%23%E6%96%B0%E5%9E%8B%E7%94%B5%E5%8A%9B%E7%B3%BB%E7%BB%9F%E5%BB%BA%E8%AE%BE%E5%8D%81%E4%BA%94%E4%BA%94%E8%A7%84%E5%88%92%E5%8F%91%E5%B8%83%23","role":"讨论与注意力样本"},
            {"name":"知乎讨论","url":"https://www.zhihu.com/question/2067656780985921835","role":"社区讨论样本"}
          ]
        }
      ]
    },
    {
      "id": "local-life",
      "title": "上海、杭州今天可以做什么",
      "summary": "周二只列 8 月 4 日当天场次；上海偏脱口秀与爵士，杭州偏亲子舞台剧与乐队现场。"
    },
    {
      "id": "ai",
      "title": "AI、科研与开发者工具",
      "summary": "三篇论文分别关注增量分词、企业文档抽取评测和语言模型智能体超参数优化；工具样本关注小模型部署与会话检索。"
    },
    {
      "id": "threads",
      "title": "今天可以继续追的线索",
      "items": [
        {"title":"电力消纳","summary":"看各省储能调用小时和辅助服务结算规则；若容量增长而调用率长期不升，规划仍停在容量配置。"},
        {"title":"大模型调用量","summary":"看 OpenCode 是否公开去重会话、付费占比与连续七日留存；若只剩 token 总量，无法支撑商业判断。"},
        {"title":"贷款成本明示","summary":"看股份行、消费金融公司是否同步披露综合年化上限；若附加费用仍留在表外，新规的可比性就会被削弱。"}
      ]
    }
  ],
  "items": [
    {"title":"新型电力系统规划把新能源利用率目标定在 90% 左右","summary":"两部门规划提出，2030 年全国新能源利用率保持在 90% 左右。","why_it_matters":"评估储能和输电项目时，还要看实际调用与结算规则，不能只看新增容量。","source_name":"国家发展改革委、国家能源局","source_url":"https://www.ndrc.gov.cn/xxgk/zcfb/tz/202608/t20260803_1406883.html","tags":["能源","政策"]},
    {"title":"个人贷款进入综合融资成本“一表明示”阶段","summary":"央视网财经转述《中国证券报》调研称，四家国有大行披露正常履约情况下相关个人贷款的年化成本上限为 6%。","why_it_matters":"贷款比较应看统一年化成本口径。","source_name":"央视网财经","source_url":"https://finance.cctv.com/2026/08/02/ARTIdzt1Eb0IAtSVp5V43yQl260802.shtml","tags":["个人金融","银行"]},
    {"title":"“8 万亿 token”热搜：单一入口调用量不等于用户规模","summary":"社区转述的单一入口调用量缺少原始统计页，不能直接换算成独立用户、收入或市场份额。","why_it_matters":"免费额度、自动化循环和长上下文都会放大 token 总量。","source_name":"微博热搜","source_url":"https://s.weibo.com/weibo?q=%23DeepSeek%E4%B8%80%E5%A4%A9%E6%B6%88%E8%80%97%E4%BA%868%E4%B8%87%E4%BA%BF%23","tags":["DeepSeek","AI economics"]},
    {"title":"TokTier: Exact Stateful Tokenization for Agentic LLM Serving","summary":"论文报告，在其测试范围内，增量分词修复约需 0.5—1.1 毫秒。","why_it_matters":"分词也会成为智能体长会话的前端延迟来源。","source_name":"arXiv","source_url":"https://arxiv.org/abs/2607.29678v1","tags":["agent serving","tokenization"]}
  ],
  "sources": [
    {"name":"国家发展改革委、国家能源局","url":"https://www.ndrc.gov.cn/xxgk/zcfb/tz/202608/t20260803_1406883.html"},
    {"name":"央视网财经","url":"https://finance.cctv.com/2026/08/02/ARTIdzt1Eb0IAtSVp5V43yQl260802.shtml"},
    {"name":"新浪财经","url":"https://finance.sina.com.cn/jjxw/2026-07-31/doc-iniktkcc1182086.shtml"},
    {"name":"微博热搜","url":"https://s.weibo.com/weibo?q=%23DeepSeek%E4%B8%80%E5%A4%A9%E6%B6%88%E8%80%97%E4%BA%868%E4%B8%87%E4%BA%BF%23"},
    {"name":"知乎讨论","url":"https://www.zhihu.com/question/2067611934703216182"},
    {"name":"36 氪","url":"https://36kr.com/p/3922353035433603"},
    {"name":"秀动","url":"https://www.showstart.com/event/list"},
    {"name":"arXiv","url":"https://arxiv.org/list/cs.AI/recent"},
    {"name":"Cloudflare Blog","url":"https://blog.cloudflare.com/smaller-faster-safer-models/"},
    {"name":"Product Hunt","url":"https://www.producthunt.com/products/inventory"}
  ],
  "local_life": {
    "shanghai": [
      {"time":"2026-08-04T19:30:00+08:00","title":"SpicyComedy·Norah/71/丹妮/油百万领衔脱口秀","venue":"SpicyComedy（新天地时尚一期）","price":"¥120 起","why":"工作日晚间脱口秀场次。","source_name":"秀动","source_url":"https://www.showstart.com/event/304903"},
      {"time":"2026-08-04T20:30:00+08:00","title":"Echoes Of Inspiration","venue":"上海 JZ CLUB","price":"¥1 起，实时票档见售票页","why":"JZ Club 晚间现场。","source_name":"秀动","source_url":"https://www.showstart.com/event/306025"}
    ],
    "hangzhou": [
      {"time":"2026-08-04T19:30:00+08:00","title":"大型沉浸式亲子舞台剧《重返侏罗纪》","venue":"杭州临平大剧院","price":"¥30 起","why":"室内亲子舞台剧，适合家庭观众。","source_name":"秀动","source_url":"https://www.showstart.com/event/304450"},
      {"time":"2026-08-04T20:00:00+08:00","title":"跳山羊「瀑布」2026 TOUR｜杭州站","venue":"UN!","price":"¥100 起","why":"独立乐队巡演杭州场。","source_name":"秀动","source_url":"https://www.showstart.com/event/304223"}
    ]
  },
  "papers": [
    {"title":"TokTier: Exact Stateful Tokenization for Agentic LLM Serving","summary":"论文报告，在其测试范围内，增量分词修复约需 0.5—1.1 毫秒。","why_it_matters":"分词也会成为智能体长会话的前端延迟来源。","source_name":"arXiv","source_url":"https://arxiv.org/abs/2607.29678v1","tags":["agent serving","tokenization"]},
    {"title":"ExtractBench: A Benchmark for Schema-Guided Enterprise Document Extraction","summary":"同时评估字段准确率、长记录完整性、出处定位和实际成本；长列表截断是重要风险。","why_it_matters":"企业文档评测需要计入漏记录和可追溯出处。","source_name":"arXiv","source_url":"https://arxiv.org/abs/2607.29677v1","tags":["document AI","benchmark"]},
    {"title":"AgentHPOBench: A Benchmark For Evaluating LLM Agents as Sequential Hyperparameter Optimizers","summary":"论文把语言模型智能体与传统 HPO 基线置于统一流程下比较，并报告现有智能体在持续迭代与稳定进步方面仍有明显限制。","why_it_matters":"它提供可比较基准，但未证明语言模型智能体已是更好的调参器。","source_name":"arXiv","source_url":"https://arxiv.org/abs/2607.29626v1","tags":["agents","HPO","benchmark"]}
  ],
  "generated_at": "2026-08-04T07:48:00+08:00"
}