Distilled 每天把 250+ 个 AI 公众号 / 海外信源的内容自动抓取、打分、用 LLM 蒸馏成"你真正该看的部分"。这个 API 把蒸馏结果以 JSON 提供,方便你二次开发。
5 大类,从社论到单篇拆解,字段丰富:
| 类别 | 内容 | 字段量 |
|---|---|---|
| 今日导读 | LLM 写的社论式综述(8 赛道)+ 蒸馏漏斗 + 实体动向 + 热词 | 50+ |
| 实体追踪 | 87 个 AI 主题:完整叙事 + 涨跌 / 热度 / 趋势 / 逐日 | 每个 ~12 |
| 每周洞察 | 跨文章的本周观察卡 | 14 |
| 信源画像 | 活跃信源:质量指数 / 风格标签 / 命中率 / 活跃度 | 50+ |
| 文章拆解 | 每篇:打分 / 等级 / 摘要 / 亮点 / why / 原文链接 | 每篇 12 |
注:以上是 AI 蒸馏的结论与元数据,不含公众号全文。
每天两刊:早刊 ~09:50、晚刊 ~20:50(北京时间)。流程是 250+ 信源 → 自动抓取 → 打分过滤 → LLM 蒸馏。API 返回当前最新刊,CDN 缓存 5 分钟。历史数据从接口上线日(2026-06-25)起逐日累积。
1) 找马奇诺领一个 token。2) 请求带上它:
curl -H "Authorization: Bearer dstl_你的token" \ https://distilled.makinote.cn/api/v1/today
就这样。↓ 下面是完整的鉴权、端点、字段和真实样本。
每个请求都要带 token(找马奇诺领)。两种方式:
# 方式 1 · Header(推荐) curl -H "Authorization: Bearer dstl_你的token" \ https://distilled.makinote.cn/api/v1/today # 方式 2 · query 参数 curl "https://distilled.makinote.cn/api/v1/today?token=dstl_你的token"
验不过返回 401。token 有时效(默认 1 年),过期也返回 401。响应统一包成 { "resource":"...", "date":"...", "data":{...} },CORS 已开。
| 端点 | 内容 | 大小 |
|---|---|---|
GET /api/v1/today | 今日导读社论 + 蒸馏漏斗 + 实体动向 + 热词 | ~12KB |
GET /api/v1/entities | 实体清单 + 叙事 + 涨跌/热度/趋势 | ~650KB |
GET /api/v1/insights | 每周洞察(跨文章观察) | ~9KB |
GET /api/v1/sources | 信源清单 + 画像 | ~108KB |
GET /api/v1/articles | 文章短摘要/亮点/打分/标题/来源/原文链接 | ~1.5MB |
每日两刊更新(早刊 ~09:50 / 晚刊 ~20:50),CDN 缓存 5 分钟。当前为最新刊,历史接口后续再加。
daily_digest | 导读社论。paragraphs[]:每段 lead(小标题)/text(正文)/entity;citations[]:引文(标题/来源/链接) |
funnel | 蒸馏漏斗:fetched/sources_active/avg_score/total_scanned/picks_count/est_read_minutes/est_saved_hours |
daily_brief | 一句话摘要 |
picks | 今日精选文章 |
trending | 热词 |
quality | 本刊质量指标 |
{
"date": "2026-06-25",
"edition": "am",
"daily_brief": "",
"funnel": {
"fetched": 54,
"sources_active": 35,
"avg_score": 44.9,
"total_scanned": 3826,
"issues_count": 141,
"picks_count": 7,
"est_read_minutes": 6,
"est_saved_hours": 2.7
},
"daily_digest": {
"paragraphs": [
{
"lead": "Claude · Anthropic工程重心在基础设施而非研究",
"text": "通过对1680份Anthropic工程师简历的分析发现,其核心人才画像并非研究科学家,而是具备平均12年经验的资深基础设施工程师,53%的人入职不到一年,最大人才来源是Google…",
"entity": "claude"
},
{
"lead": "Skills 玩法 · Loop Engineering与多Agent编排成新热点",
"text": "Loop Engineering通过将代码修复拆解为builder和checker两个专职Agent并由编排器循环调度,实现了全自动质量保障流水线[3]。一位开发者将火山引擎See…"
},
"…(共6条)"
],
"citations": [
{
"n": 1,
"link": "https://mp.weixin.qq.com/s?__biz=Mzg3Mzg5MjY3Nw==&mid=224752…",
"title": "以为Anthropic全是科学家?数据告诉你,他们要的是基础设施老兵",
"source": "AIGC开放社区",
"score": 82,
"minutes": 2
},
{
"n": 2,
"link": "https://www.latent.space/p/databricks",
"title": "Why the Frontier Ecosystem must be Open — Matei Zaharia and …",
"source": "Latent Space",
"score": 75,
"minutes": 3
},
"…(共17条)"
]
},
"trending": [
{
"keyword": "claude",
"source_count": 7,
"mention_count": 7
},
{
"keyword": "anthropic",
"source_count": 6,
"mention_count": 6
},
{
"keyword": "agent",
"source_count": 5,
"mention_count": 5
},
{
"keyword": "gpt",
"source_count": 5,
"mention_count": 5
},
{
"keyword": "自动化",
"source_count": 3,
"mention_count": 3
}
]
}entities[] | 每个实体:display(名)/type/narrative(叙事 summary+sections)/trend(趋势)/daily_counts(逐日)/today_count/today_top/latest/boards |
curated_ids | 精选(含完整叙事)实体的 id 列表 |
meta | 窗口/统计元信息 |
{
"window": 30,
"curated_ids": [
"zhipu",
"karpathy",
"qwen",
"claude",
"claude_code"
],
"entities": [
{
"display": "Agent",
"type": "Concept",
"narrative": {
"summary": "Harness生态快速成熟,本地Agent可用,模型迁移成本需警惕,新基准更真实。",
"sections": [
{
"topic": "Loop Engineering",
"summary": "拆分builder/checker双Agent,由编排器循环调度直到全量测试通过,实现代码修复自动化。已出保姆教程,可直接复用。",
"keywords": [
"Loop Engineering",
"Agent编排",
"质量保障"
],
"articles": [
{
"title": "How to design AI agent loops: schedules, goals, and subagents in Claude Code and…",
"score": 85,
"date": "2026-06-17",
"link": "https://www.lennysnewsletter.com/p/how-to-design-ai-agent-loops-schedules",
"summary": "设计 AI agent 循环(如心跳、定时、挂钩、目标四种类型),并确保循环具备工作树、技能、插件/连接器、子代理和状态跟踪五大要素。"
},
{
"title": "刚刚,全网爆火的Loop Engineering,保姆教程来了!",
"score": 77,
"date": "2026-06-25",
"link": "https://mp.weixin.qq.com/s?__biz=MzIyNjM2MzQyNg==&mid=2247723692&idx=1&sn=79bbf6…",
"summary": "使用两个专职Agent(builder负责代码实现与修复,checker仅检查与报告)并由编排器循环调度,自动修复代码直到所有测试、Lint、类型检查全部通过。"
},
{
"title": "说说 Claude Fable 5 上手感受: 1、自主执行的 Agency 很强,相比 Opus 4.8 更喜欢委派 subagent 干活;同时很多原来 O…",
"score": 77,
"date": "2026-06-11",
"link": "https://mp.weixin.qq.com/s?__biz=MzIwODA3MDUwOQ==&mid=2652776225&idx=1&sn=b154b2…",
"summary": "尝试用Claude Fable 5处理超长编程任务,定义清晰目标和验收标准后,放手让AI通过/goal、/loop等循环模式自主迭代运行。"
},
"…(共6条)"
]
},
{
"topic": "LIFE-HARNESS",
"summary": "北大提出的四层运行时干预框架,不修改模型权重,通过环境契约层和技能层等提升垂域Agent性能,训练后的harness可跨模型迁移,平均提升88.5%。",
"keywords": [
"LIFE-HARNESS",
"运行时框架",
"垂域Agent"
],
"articles": [
{
"title": "同一个模型,换套框架成绩差27%:SWE-bench分数到底谁说了算?",
"score": 88,
"date": "2026-06-17",
"link": "https://mp.weixin.qq.com/s?__biz=Mzg3Mzg5MjY3Nw==&mid=2247526485&idx=1&sn=921f63…",
"summary": "一项研究通过Claw-SWE-Bench基准揭示了SWE-bench评测中模型分数的巨大差异性。同一模型在更换测试框架(Harness)后,成绩最大可差27个百…"
},
{
"title": "Agent一定要训练吗?4层Harness让性能平均提升88.5%",
"score": 80,
"date": "2026-05-29",
"link": "https://mp.weixin.qq.com/s?__biz=MzIwMTc4ODE0Mw==&mid=2247720638&idx=2&sn=d76374…",
"summary": "论文《LIFE-HARNESS》提出一种新范式:不重新训练或微调LLM模型参数,而是通过在模型与环境的运行时接口(Runtime Harness)上增加环境契约…"
},
{
"title": "我把Seed 2.1 Pro塞进Claude Code,让它修我自己产品的bug",
"score": 73,
"date": "2026-06-24",
"link": "https://mp.weixin.qq.com/s?__biz=Mzg2OTA1OTAxNA==&mid=2247490934&idx=1&sn=4ed42a…",
"summary": "使用火山引擎的豆包大模型Seed 2.1 Pro接入Claude Code,利用其plan mode和子agent编排能力,自主修复开源项目中的真实bug。"
},
"…(共6条)"
]
},
{
"topic": "Gemma 4",
"summary": "在M2 Mac上用LM Studio运行gemma-4-26b-a4b模型,完成脚本重构、单元测试等Agent编码任务,本地Agent工作流变得简单。",
"keywords": [
"Gemma 4",
"本地推理",
"Agent任务"
],
"articles": [
{
"title": "端侧跑大模型,现在也太简单了",
"score": 77,
"date": "2026-06-24",
"link": "https://mp.weixin.qq.com/s?__biz=MzA3MzI4MjgzMw==&mid=2651040859&idx=2&sn=43a32b…",
"summary": "在本地Mac设备上运行Gemma 4等大模型,利用LM Studio和Pi框架完成代码重构、编写测试等Agent编程任务。"
}
]
},
"…(共7条)"
],
"lead": "Harness生态快速成熟,本地Agent可用,模型迁移成本需警惕,新基准更真实。",
"points": [
{
"text": "Loop Engineering和LIFE-HARNESS等多Agent harness系统可直接抄来用,实现自动化质量保障和垂域性能提升。",
"verdict": "可用",
"refs": [
{
"title": "垂域Agent狂飙88.5%!北大Life-Harness告别训练",
"link": "https://mp.weixin.qq.com/s?__biz=Mzk0MTYzMzMxMA==&mid=2247508434&idx=1&sn=04a69e…"
},
{
"title": "刚刚,全网爆火的Loop Engineering,保姆教程来了!",
"link": "https://mp.weixin.qq.com/s?__biz=MzIyNjM2MzQyNg==&mid=2247723692&idx=1&sn=79bbf6…"
}
]
},
{
"text": "端侧Gemma 4本地运行Agent编程任务,准确率可达云端75%,隐私场景已可用。",
"verdict": "可用",
"refs": [
{
"title": "端侧跑大模型,现在也太简单了",
"link": "https://mp.weixin.qq.com/s?__biz=MzA3MzI4MjgzMw==&mid=2651040859&idx=2&sn=43a32b…"
}
]
},
{
"text": "Seed 2.1 Pro接入Claude Code能修真实bug,与harness协同良好,可直接尝试。",
"verdict": "可用",
"refs": [
{
"title": "我把Seed 2.1 Pro塞进Claude Code,让它修我自己产品的bug",
"link": "https://mp.weixin.qq.com/s?__biz=Mzg2OTA1OTAxNA==&mid=2247490934&idx=1&sn=4ed42a…"
}
]
},
"…(共6条)"
]
},
"trend": {
"articles_7d": 107,
"articles_prev_7d": 72,
"direction": "up"
},
"today_count": 4,
"total_articles": 407,
"daily_counts": [
13,
15,
23,
"…(共30条)"
]
},
{
"display": "Claude",
"type": "Product",
"narrative": {
"summary": "开源编码模型性价比逼近Claude,Claude Tag开启团队Agent新范式",
"sections": [
{
"topic": "Claude Tag",
"summary": "Anthropic在Slack上线Claude Tag,可作为团队成员主动监控频道、等待数天的webhook依赖,并合并PR,内部已处理65%的PR。这是LLM…",
"keywords": [
"Claude Tag",
"Slack",
"Agent"
],
"articles": [
{
"title": "8B Agent追近Claude,ToolCUA让桌面Agent学会调度",
"score": 98,
"date": "2026-06-16",
"link": "https://mp.weixin.qq.com/s?__biz=MzIwMTc4ODE0Mw==&mid=2247720940&idx=2&sn=5c5230…",
"summary": "复旦大学与通义实验室发布ToolCUA工具,旨在解决桌面Agent在GUI操作和工具调用之间进行路径编排的难题。数据表明简单引入工具反而会降低模型成功率,如Cl…"
},
{
"title": "别光给Agent加Tool了,它根本选不明白!复旦×通义提出全新CUA训练范式",
"score": 78,
"date": "2026-05-31",
"link": "https://mp.weixin.qq.com/s?__biz=MzIzNjc1NzUzMw==&mid=2247893880&idx=3&sn=3801ec…",
"summary": "复旦大学和通义实验室提出ToolCUA训练范式,解决Agent在GUI操作与工具调用之间的路径选择难题。实验表明,在混合动作空间中直接给强模型加上工具调用反而会…"
},
{
"title": "说说 Claude Fable 5 上手感受: 1、自主执行的 Agency 很强,相比 Opus 4.8 更喜欢委派 subagent 干活;同时很多原来 O…",
"score": 77,
"date": "2026-06-11",
"link": "https://mp.weixin.qq.com/s?__biz=MzIwODA3MDUwOQ==&mid=2652776225&idx=1&sn=b154b2…",
"summary": "尝试用Claude Fable 5处理超长编程任务,定义清晰目标和验收标准后,放手让AI通过/goal、/loop等循环模式自主迭代运行。"
},
"…(共6条)"
]
},
{
"topic": "GLM-5.2",
"summary": "开源模型GLM 5.2在真实代码库中替代Claude Opus执行了代码审计、UI重构和45分钟自主bug修复,落地页重构首次即匹配设计系统,处理600万tok…",
"keywords": [
"GLM 5.2",
"开源模型",
"编码Agent"
],
"articles": [
{
"title": "斩获15个顶级零日漏洞:0G Lab联合新国立,北大和北邮团队构建的共识协议debug智能体框架",
"score": 83,
"date": "2026-06-11",
"link": "https://mp.weixin.qq.com/s?__biz=MzA3MzI4MjgzMw==&mid=2651038459&idx=1&sn=5135bb…",
"summary": "学习如何利用领域知识与大模型多Agent协同构建自动化漏洞检测框架,能够显著提升在工业级共识协议中发现深层零日漏洞的能力,降低安全审计的算力门槛。"
},
{
"title": "Why the Frontier Ecosystem must be Open — Matei Zaharia and Reynold Xin, Databri…",
"score": 75,
"date": "2026-06-25",
"link": "https://www.latent.space/p/databricks",
"summary": "为前沿AI生态系统构建开放、通用的元代理工具和统一数据存储层,以解决编码Agent间的可移植性、安全与支出控制碎片化问题。"
},
{
"title": "GLM 5.2: why I’m replacing Opus in Claude Code with this new model",
"score": 67,
"date": "2026-06-24",
"link": "https://www.lennysnewsletter.com/p/glm-52-why-im-replacing-opus-in-claude",
"summary": "开源权重模型GLM 5.2在真实复杂代码任务中展现出与顶级专有模型Opus相当的编码能力,同时大幅降低推理成本。这意味着开源编码模型在自主完成工程任务上的性价比…"
},
"…(共6条)"
]
},
{
"topic": "Seed 2.1 Pro",
"summary": "火山引擎发布豆包Seed 2.1 Pro,在Terminal Bench 2.1上与Opus持平,接入Claude Code后利用plan mode和子agen…",
"keywords": [
"Seed 2.1 Pro",
"豆包",
"编程Agent"
],
"articles": [
{
"title": "我把Seed 2.1 Pro塞进Claude Code,让它修我自己产品的bug",
"score": 73,
"date": "2026-06-24",
"link": "https://mp.weixin.qq.com/s?__biz=Mzg2OTA1OTAxNA==&mid=2247490934&idx=1&sn=4ed42a…",
"summary": "使用火山引擎的豆包大模型Seed 2.1 Pro接入Claude Code,利用其plan mode和子agent编排能力,自主修复开源项目中的真实bug。"
},
{
"title": "Claude Code 作者Boris:我已经不写 prompt 了,我写 loop",
"score": 70,
"date": "2026-06-04",
"link": "https://mp.weixin.qq.com/s?__biz=MzA5MTIxNTY4MQ==&mid=2461159971&idx=1&sn=e2bc19…",
"summary": "Claude Code的动态工作流允许用户通过loop逻辑让AI自行编排任务并执行,而非传统的手动编写prompt。这种方式提升了AI的自主性和效率,改变了人与…"
},
{
"title": "专访火山引擎谭待:模型好对MaaS来说是最重要的事,豆包2.1算「上牌桌」了",
"score": 67,
"date": "2026-06-24",
"link": "https://mp.weixin.qq.com/s?__biz=MzI2NDk5NzA0Mw==&mid=2248892981&idx=1&sn=7976fe…",
"summary": "MaaS商业成功的核心在于提供最顶尖的模型。火山引擎通过视频生成模型Seedance 2.0和代码模型豆包2.1 Pro证明自身实力,其中豆包2.1 Pro的编…"
},
"…(共6条)"
]
},
"…(共6条)"
],
"lead": "开源编码模型性价比逼近Claude,Claude Tag开启团队Agent新范式",
"points": [
{
"text": "GLM 5.2在45分钟自主修bug、落地页重构等任务上可替代Claude Opus,总成本仅3.36美元,可直接抄来降本",
"verdict": "可用",
"refs": [
{
"title": "GLM 5.2: why I’m replacing Opus in Claude Code with this new model",
"link": "https://www.lennysnewsletter.com/p/glm-52-why-im-replacing-opus-in-claude"
}
]
},
{
"text": "Seed 2.1 Pro接入Claude Code修了开源项目FanBox两个真实bug,完整利用harness编排且遵循原代码风格",
"verdict": "可用",
"refs": [
{
"title": "我把Seed 2.1 Pro塞进Claude Code,让它修我自己产品的bug",
"link": "https://mp.weixin.qq.com/s?__biz=Mzg2OTA1OTAxNA==&mid=2247490934&idx=1&sn=4ed42a…"
}
]
},
{
"text": "Claude Tag作为Slack群聊Agent可主动监控、等待webhook、合并PR,开启持久异步协作范式,但需审计和伦理边界",
"verdict": "观望",
"refs": [
{
"title": "[AINews] Claude Tag: Multiplayer, Proactive, Persistent Agents in Slack",
"link": "https://www.latent.space/p/ainews-claude-tag-multiplayer-proactive"
},
{
"title": "Claude 刚刚上线 “群聊Agent”:Karpathy盛赞的交互新范式,还是打工人的“数字监工”?",
"link": "https://mp.weixin.qq.com/s?__biz=MzU1NDA4NjU2MA==&mid=2247663896&idx=1&sn=3cf781…"
}
]
},
"…(共6条)"
]
},
"trend": {
"articles_7d": 53,
"articles_prev_7d": 54,
"direction": "flat"
},
"today_count": 6,
"total_articles": 216,
"daily_counts": [
5,
3,
13,
"…(共30条)"
]
},
{
"display": "大模型",
"type": "Concept",
"narrative": {
"summary": "",
"sections": [],
"auto_summary": "最近 30 天有 169 篇相关文章。值得做方向最多(6 篇)。最新一篇:直接从像素到单词:这个原生大模型统一单图、多图、视频和空间智能(06-24)。"
},
"trend": {
"articles_7d": 53,
"articles_prev_7d": 32,
"direction": "up"
},
"today_count": 3,
"total_articles": 169,
"daily_counts": [
13,
7,
7,
"…(共30条)"
]
}
]
}cards[] | 每张洞察卡:type(类型)/title/text(正文)/hook |
citations | 引文 |
{
"cards": [
{
"type": "洞察",
"title": "AI应用漏洞从框架蔓延至模型层,一个免费账户就能窃取百万应用数据",
"text": "[1] Dify 的 DifyTap 漏洞只需注册一个免费云账户,就能未经授权配置其他租户的追踪功能,完整窃取所有用户提问和 AI 回答,超过 100 万个应用受影响。[2] FFmpeg 的 PixelSmash 高危漏洞让攻击者用 50…",
"hook": ""
},
{
"type": "新苗头",
"title": "Agent 能力不再看答题分数,真实工作流通过率仅 8.6%",
"text": "[4] 伯克利等机构发布了 ALE 基准,要求 Agent 在 GUI 和 CLI 环境中完成 960 个来自真实从业者的长流程任务。当前最强 AI 系统在最难任务上的完整通过率只有 8.6%,主流系统平均 2.6%,失败主因是理解错误和方…",
"hook": ""
},
{
"type": "连点",
"title": "训练数据策略从堆量转向巧配,三个独立团队得出相似结论",
"text": "[6] 普林斯顿陈丹琦团队提出 WEBGRAPHMIX,用 50% 高介数中心性枢纽页面搭配 50% 长尾页面做预训练,在 1B 模型上准确率较随机采样提升 1.6 个百分点。[7] 腾讯混元发现 SFT 即使 loss 收敛,仍有 15.…",
"hook": ""
},
"…(共7条)"
]
}pinned_sources[] | 置顶信源:name/quality_index(质量)/hit_rate(命中率)/style(风格标签)/active_days/sparkline/articles |
today | 今日统计 |
highlights | 今日亮点(含代表文章) |
all_sources | 全部活跃信源画像 |
{
"today": {
"articles_total": 54,
"avg_score": 44.9,
"sources_active": 35,
"score_distribution": {
"90+": 1,
"80-89": 0,
"70-79": 2,
"60-69": 9,
"<60": 42
},
"top_source": {
"name": "SaaStr",
"avg_score": 60.7,
"count": 3
}
},
"pinned_sources": [
{
"name": "Simon Willison",
"quality_index": 60.1,
"hit_rate": 0,
"style": "严谨派",
"active_days": 26,
"articles_total": 68
},
{
"name": "Hugging Face Blog",
"quality_index": 59.3,
"hit_rate": 100,
"style": "严谨派",
"active_days": 17,
"articles_total": 32
},
{
"name": "Anthropic Cookbook",
"quality_index": 31.6,
"hit_rate": 100,
"style": "实操派",
"active_days": 3,
"articles_total": 5
}
]
}articles{} | 按原文链接为 key。每篇:title/source/date/score(打分)/level(收藏/精读/速览)/must_read/summary(短摘要)/highlights(亮点)/why_matters/weak_point/guide/category/scores |
score(0-100) 由 7 个维度(各 0-3)加权算出 —— novelty 新颖度 25% · depth 深度 25% · actionability 可操作性 20% · credibility 可信度 10% · timeliness 时效 10% · clarity 清晰 5% · signal_quality 信号质量 5%。scores 字段把 7 个分项都给你,可按自己偏好重新加权/筛选。{
"count": 1579,
"articles": {
"https://mp.weixin.qq.com/s?__biz=MzIwMTc4ODE0Mw==&mid=2247720940&idx=2&sn=5c523074a8cc6d328a48fa98d34e52d7": {
"title": "8B Agent追近Claude,ToolCUA让桌面Agent学会调度",
"source": "PaperWeekly",
"date": "2026-06-15",
"score": 98,
"level": "收藏",
"must_read": true,
"why_matters": "Agent开发者可直接使用ToolCUA-8B模型或借鉴其GUI-Tool混合训练思路,解决当前大模型Agent在长程桌面任务中因路径选择失当而频繁失败的痛点。",
"summary": "ToolCUA让8B桌面Agent在GUI与工具间自动调度",
"highlights": [
"发现简单融合GUI和工具调用会降低Agent成功率,8B模型通过ToolCUA训练后追近Claude。",
"提出Interleaved GUI-Tool Trajectory Scaling Pipeline,从已有GUI数据中批量合成混合操作轨迹进行训练。"
],
"weak_point": "clarity=2:方法论细节密集,阅读要求较高",
"guide": null,
"category": "AI/Tech"
},
"https://mp.weixin.qq.com/s?__biz=MzA5ODEzMjIyMA==&mid=2247736612&idx=3&sn=1bc120b8fed9b73a2ff5009b7d39ffee": {
"title": "腾讯混元&浙大:VLM 不只会聊天对话,还能原生输出像素级深度图",
"source": "AI科技评论",
"date": "2026-06-12",
"score": 98,
"level": "收藏",
"must_read": true,
"why_matters": "[AI开发者和机器人工程师] 可以直接利用已有的VLM架构获得稠密深度图,而无需集成庞大的外部深度模型。这将大大简化具身智能、空间计算、3D场景重建等系统的流程…",
"summary": "腾讯混元与浙江大学联合提出的DepthVLM,首次让视觉语言模型能以极低的算力成本原生输出像素级稠密深度图。其创新点在于不依赖外部模型,仅通过附加一个轻量级深度…",
"highlights": [
"DepthVLM在室内外深度估计任务上的平均性能达到了δ₁=0.876,超越了所有现有VLM和纯视觉模型标杆DepthAnythingV3。",
"该模型采用两阶段训练策略,先冻结骨干网络训练深度头,再端到端联合微调,成功平衡了几何预测精度与原有的图文对话能力。"
],
"weak_point": "clarity=2:技术细节充足,但篇幅冗长,精炼后可读性更强",
"guide": null,
"category": "AI/Tech"
},
"https://mp.weixin.qq.com/s?__biz=MzIwMTc4ODE0Mw==&mid=2247721067&idx=1&sn=a8a45c05f631722c62ca81ae16c01a97": {
"title": "刷屏数周后,陈德里的AutoResearch SKILL终于开源了",
"source": "PaperWeekly",
"date": "2026-06-19",
"score": 97,
"level": "收藏",
"must_read": true,
"why_matters": "[AI研究者]可以将AutoResearch框架作为起点构建自己的自动化科研Agent,利用其编排结构管理长周期复杂任务。[Agent开发者]能复用其状态管理和…",
"summary": "陈德里正式开源了AutoResearch SKILL,这是一个用于长周期自主科研任务的协议框架。框架通过状态保存、停滞检测和心跳守护三层编排结构,解决了AI A…",
"highlights": [
"AutoResearch解决了长任务中Agent自我迭代的三个核心问题:状态保存防止上下文丢失,停滞检测避免原地打转,心跳守护防止任务中断",
"paper-writing skill将科研写作分解为文献、结构、实验、图表和模拟评审五个子流程,并通过Quality Gates实现阶段性质量把控",
"第四篇self-play论文完成了16轮评审迭代和12次RL运行,包含285B参数实验,展示了验证信号噪声对self-play效果的定量影响"
],
"weak_point": "credibility=2:部分实验数据及结论源自项目自身披露,外部验证尚待补充。",
"guide": null,
"category": "AI/Tech"
}
}
}{
"resource": "today",
"date": "2026-06-25",
"data": {
"daily_digest": {
"paragraphs": [
{ "lead": "Claude · ...", "text": "...", "entity": "claude" }, ...
],
"citations": [ { "title": "...", "source": "...", "link": "..." } ]
},
"funnel": { "total_scanned": 3826, "sources_active": 35, ... },
"trending": [ "claude", "anthropic", "agent" ], ...
}
}