{ "text": "Meta开源Muse glimmer30B\nMeta開源Muse Glimmer 30B,最值得看的不是參數也不是跑分,而是它把大模型競爭從雲端大廠的機房重新推回普通人的電腦桌上。一個30B級別開放權重面向本地Agent的模型放在今天這個節點,意思很明確。AI不能只是一張雲服務帳單,也不能永遠被少數公司關在黑盒裡。這件事要放在Meta的整體敘事裡看。\n扎克伯格最近把話說得很直白\n人工智能的主要\n目的應該是發明創造\n不應該只是自動化替代\n超級智能如果真的會到來\n那它不應該只掌握在幾家公司手裡\n而應該盡量廣泛地分發給\n普通人開發者和小團隊\nMuse Glimmer就是這個敘事\n落到產品上的一顆棋子\nMuse Glimmer 30B的重點不是\n挑戰閉源模型的全部能力\n而是給本地AI留出一條現實路線\n現在很多模型動不動\n就是幾千億上萬億參數\n聽起來很震撼\n但普通用戶根本跑不起來\n他們需要昂貴GPU\n需要雲端推理\n需要訂閱和調用額度\nMeta選擇30B這個體量\n就是在說\n模型不一定越大越接近未來\n能部署能修改能嵌進\n工作流才更接近真實使用\n這和LLAMA時代的路線一脈相承\nMeta一直在開放模型上押注\n只是中間經歷過搖擺\n當OpenAI Anthropic Google把高端模型\n能力不斷往閉源體系裡收\nMeta反而重新把開放權重當成差異化\n它沒必要在每一項榜單上贏\n它要贏的是開發者心智\n誰能让更多人拿去改\n拿去跑拿去做產品\n誰就能在AI底層生態裡留下位置\nMuse Glimmer還帶有Execute Torch和PTE\n這樣的邊緣部署意味\n这个细节很重要\n它不是只给服务器看的模型卡\n而是在暗示手機筆記本\n邊緣設備本地工作站都\n可能成为AI agent的运行环境\n未来AI agent\n不一定每一步都要回云端\n他可以在本地看文件讀\n圖片調工具處理私人數據\n再把真正复杂的\n任务交给云端大模型\n这种路线最大的价值是隐私和成本\n用戶把本地文件工作資料聊天\n記錄圖片和代碼交給雲端模型\n永远会有心理门槛\n企業更敏感\n尤其涉及客戶數據合同\n研發資料和內部流程\n一個能在本地跑的模型\n就算能力不是最强\n設備而變得有价值\n设备而变得有价值\n很多真實場景不需要\n世界最聰明的模型\n只需要一個足夠聰明\n足夠便宜足夠可控的助手\n這也是開源模型真正\n打閉源模型的地方\n閉源模型擅長提供最強通用能力\n開源模型擅長被改造成具體工具\n小團隊可以微調它\n企業可以把它部署到內網\n開發者可以把它接\n進自己的agent框架\n硬體廠商可以把它塞進邊緣設備\n閉源模型賣的是能力\n開放模型賣的是可塑性\n兩者不是同一場比賽\nMeta這次還把數據中心社區\n基金放在同一個敘事裡\n說明它知道AI擴張正在\n遇到新的政治成本\n數據中心需要電力、土地\n水、輸電設施和社區許可\n過去科技公司建機房\n地方政府往往歡迎\n因為它帶來投資和稅收\n現在情況變了\n居民開始擔心電費上漲\n環境壓力和基礎設施被佔用\nAI越龐大越不可能只用\n技術語言解釋自己\n10億美元社區基金表面上是補償\n背後是AI基建時代的通行費\nMeta想告訴當地居民\n數據中心不是只把電和水抽走\n也會把錢、崗位和公共服務帶回來\n警察、消防員、社區成員被點名\n是一種很現實的政治溝通\nAI公司已經發現\n光說技術改變世界沒用了\n真正被影響的人要看到具體回報\nMuse Glimmer和社區基金看似是兩件事\n其實對應同一個問題\nAI權力會不會過度集中\n模型集中在雲端\n算力集中在數據中心\n收益集中在少數公司\n成本卻擴散給社區和消費者\n這套結構如果不調整\n反彈會越來越大\nMeta現在講開放、講本地、講回饋社區\n就是在給自己的AI\n擴張換一種合法性\n這不代表Meta突然變成公益組織\n開源也不是純粹善意\nMeta的商業邏輯很清楚\n它沒有像OpenAI那樣把\n訂閱模型做成第一入口\n也沒有Google這樣的搜索默認位置\n它最擅长的是平台\n分发广告和开发者网络\n开放模型能削弱\n闭源模型的收费能力\n降低竞争对手的护城河\n同时让更多应用围绕\nMeta的模型生态生长\n如果一个开放模型足够好\n大量开发者就会围绕它做\n工具插件微调和部署方案\n哪怕Meta不直接收每一次调用费\n它也能影响标准框架和生态方向\n安卓当年不是靠系统授权费赚钱\n而是靠开放系统占住移动入口\nLLAMA和Muse Glimmer也有类似逻辑\n用开放打破别人的封闭利润池\n再在更大的生态里寻找收益\nMuse Glimmer的30B体量还有一个现实意义\n它让AI从演示回到产品\n很多AI发布会喜欢展示极限能力\n仿佛模型能解决所有复杂问题\n但真正落地时\n企业关心的是成本延迟\n部署安全稳定可维护\n一个30B模型\n如果能在普通设备或\n消费级GPU上完成稳定任务\n它的商业价值可能比一个只能在\n云端高价运行的巨型模型更直接\n尤其是agent场景\nagent不是一次聊天\n而是一连串动作\n理解目标拆任务\n读文件调用工具\n检查结果\n失败后重试\n这里面很多步骤并不需要最强模型\n文件分类表格检查界面操作\n图片理解简单代码修改本地搜索\n都可以由较小模型承担\n把这些步骤放到本地执行\n云端只处理最难的推理\n整个系统成本会明显下降\n这也解释了为什么\n本地模型会越来越重要\nAI的下一阶段不是谁\n能回答得更漂亮\n而是谁能更便宜地完成更多动作\n一个公司每天调用百万次模型\n如果每一步都走最贵的闭源API\n成本会很快失控\n本地模型可以承担大量\n低风险重复性隐私敏感任务\n它不必取代旗舰模型\n只要把旗舰模型从低\n价值任务里解放出来\n就已经很有意义\n对普通创作者和小公司来说\n这条路线尤其关键\n大公司可以买算力\n签云合同\n雇安全团队\n小团队没有这个条件\n如果未来所有AI能力都\n绑定高价订阅和云端调用\n小团队会永远站在门外\n开放权重模型\n让他们至少有机会在\n本地搭一个能跑的系统\n哪怕不完美\n也能开始改造自己的工作流\n当然开源模型也有风险\n开放权重意味着能力会扩散\n好的用途会扩散\n坏的用途也会扩散\n安全过滤、滥用检测\n责任归属都会变复杂\n闭源公司常用安全作为封闭理由\n不能说完全没有道理\n但如果只因为风险就把能力锁死\n最后得到的是少数公司\n控制所有AI基础设施\nMeta现在押的是另一种答案\n风险需要治理\n但不能用风险作为永久垄断的理由\nMuse Glimmer真正要证明的\n不是30B能不能击败所有大模型\n而是开放路线能不能在AI\n商业化最激烈的时候继续存在\n闭源路线说\n最强能力必须集中\n才能安全可靠可收费\n开放路线说能力必须分发\n社会才不会被少数入口控制\n两条路线都会存在\n但它们代表完全不同的未来\n接下来要看的不是一两次榜单成绩\n而是开发者是否真的使用它\n有没有人把它部署到本地agent\n有没有企业拿它做内部流程\n有没有硬件厂商把它放进设备\n有没有开源社区围绕\n它做微调和工具链\n如果这些事情发生\nMuse Glimmer的意义就不\n只是一个模型发布\n而是Meta在重新争夺AI\n基础设施的话语权\n这场竞争还有一个更大的背景\nAI正在从模型能力竞争\n转向分发方式竞争\nOpenai想把模型做成超级应用\nGoogle想把AI塞回搜索和安卓\n苹果想把AI放进硬件和系统\nMeta则想让开放模型\n成为开发者默认底座\nGlimmer只是其中一块拼图\n但它指向的方向很清楚\n让AI不止存在于云端\n也存在于每台电脑每个\n应用每个本地工作流里\n所以这次发布不能只\n看成Meta又开源了一个模型\n它更像一次路线宣言\nAI如果只能在数据中心运行\n用户就永远是租客\nAI如果可以在本地运行\n用户才有机会成为主人\nMuse Glimmer 30B 的价值就在这里\n它可能不是最强的模型\n但他在提醒整个行业\n未来的AI不应该只属于机房\n也应该属于桌面\n这里还要看Costello对AI投资的判断\n他认为Openai和Anthropic\n不是AI故事的终点\n而只是开头\n未来会出现更多\n千亿美元万亿美元级公司\n影响科学医疗能源\n教育制造这些基础领域\n这个判断放到 Muse Glimmer 30B 上\n就能看见Meta的算盘\n如果未来有大量垂直AI公司出现\n他们不一定都愿意从闭源API开始\n科学研究就是一个典型场景\n实验室需要处理论文图像\n代码仪器数据和内部记录\n很多资料不能随便\n上传到外部云服务\n一个本地多模态模型\n就算不是最强\n也可以承担文献整理\n实验记录搜索图表理解\n工具调用和初步假设生成\n它不会替代科学家\n但会把很多低价值搜索\n和整理动作压低成本\nMeta很清楚\n闭源模型越强\n外界对模型集中化的担心就越重\n一个公司如果控制模型\n控制分发\n控制价格\n控制安全规则\n开发者就会变成依附者\n今天接口价格可以接受\n明天规则可能变\n今天调用额度够用\n明天业务增长后成本可能翻倍\n开放模型给开发者留了一条退路\n这条退路本身就是谈判筹码\n这也是为什么开放权重会压低行业利润率\n权重会压低行业利润率\n只要有足够好的开源替代品\n闭源公司就很难对中低端任务收太高价格\n低端任务收太高价格\n旗舰模型仍然可以卖高价\n但大量日常任务会被本地模型和开放模型吃掉\n本地模型和开放模型吃掉\nMeta不靠模型订阅赚钱\n反而适合推动这种价格下行\n他要让AI变成空气\n然后在空气里做平台\n有人会问30B模型真的够吗\n这个问题要看任务\n写一段复杂法律意见\n可能不够\n做一个本地客服分类器够\n通读一本手册并回答设备维修问题\n可能够整理图片和文字生成初稿\n可能够控制一组本地工具跑流程，也可能够\n可能也够\nAI产品不需要每次都\n派最强模型上场\n就像公司不会让总裁\n去处理每张报销单\n真正成熟的AI系统\n很可能是多模型协作\n小模型负责本地\n便宜低风险高频动作\n中型模型负责复杂判断\n旗舰模型负责少数关键推理\nMuse Glimmer的位置就在\n第一层和第二层之间\n它不抢所有舞台\n但可以进入很多工作流\n这个位置一旦铺开\n使用量会非常大\n所以Muse Glimmer看起来是技术事件\n背后其实是分发事件\nAI行业未来最大的战场\n不只是模型谁更聪明\n而是谁能接触更多开发者\n更多设备\n更多工作流\n更多真实业务\n开放权重提供了一种复制速度\n一个模型被下载后\n可以在无数环境里变成不同产品\n闭源模型再强\n也必须等用户来到它的入口\n当然开放模型要赢也不容易\n开发体验必须好\n文档必须清楚\n工具链必须稳定\n推理速度必须能接受\n硬件适配不能太痛苦\n如果一个模型开源了\n但部署起来像拆炸弹\n开发者很快就会放弃\nMeta这次把exec torch等部署线索放进来\n说明它知道只开源权重还不够\n真正要争的是落地链路\n安全问题也不能回避\n一个能在本地运行\n能调用工具\n能处理多模态输入的模型\n如果被错误使用\n也会带来自动化滥用\n开放社区需要评估\n紅隊使用限制和透明報告\nMeta不能一邊強調開放\n一邊把風險全丟給社區\n開放路線要走得久\n必須同時建立開放治理\n但閉源也不是天然安全\n閉源模型的能力集中後\n外界更难知道它如何訓練\n如何過濾\n如何拒絕\n如何偏向某些服務\n安全不是封閉和開放的簡單選擇\n而是透明度、責任和可驗證性之間的平衡\n可驗證性之間的平衡\nMuse Glimmer的價值之一就是讓\n更多研究者能直接測試模型\n而不是只能相信廠商聲明\nMeta設立社區基金\n某種意義上也是在\n承認AI基建的外部成本\n數據中心建在某個地方\n當地居民可能面對電價用水\n噪音、土地和稅收分配問題\n科技公司過去常把\n這些問題當成地方溝通\n現在必須當成核心戰略\n因為沒有社區許可\n再強的模型也需要機房供電\n把這兩條線合起來\nMeta的新敘事就完整了\n一邊用開放模型說AI能力要分散\n一邊用社區基金說AI\n基建收益要回流\n這不一定全是理想主義\n但他比單純喊超級智能更聰明\n他知道AI行業如果繼續只\n展示估值、參數和算力\n公眾反彈會越來越強\n未來真正有殺傷力的Muse Glimmer應用\n可能不会出现在发布新闻里\n而会出现在很小的工作场景里\n一个设计师用他本地整理素材\n一个律师事务所用它检索内部文件\n一个工厂用它读设备图片\n一个个人开发者用它\n搭自己的桌面agent\n每个场景都不震撼\n但加起来就是生态\n这也是开放模型最强的地方\n它的价值不是一次性释放\n而是被别人不断重新发明\n闭源模型像一座中央电站\n开放模型像一批可移动工具箱\n中央电站功率更大\n工具箱更容易进入角落\nAI要真正改变社会\n不能只靠中央电站\n也需要工具箱\n所以Muse Glimmer不是Meta对闭源\n模型的一次正面冲锋\n而是一次侧翼包抄\n它不说自己在所有能力上第一\n他说自己更容易被拿走\n更容易本地跑\n更容易进入普通设备\n这个策略如果成功\nMeta会在AI入口战争里拿到\n一个很难被关闭的位置\nExecute Torch的存在\n说明Meta对端侧部署有长期准备\n端侧AI不是把云端模型硬塞进手机\n而是重新考虑模型大小\n延迟内存占用电池和隐私。Muse Glimmer\n如果能和端侧框架形成完整链路\n就会成为Meta连接硬件厂商的桥\n以后手机眼镜耳机笔记本桌面设备\n都可能需要一个能本地\n处理多模态任务的模型底座\n这对开发者很有吸引力\n闭源API最舒服的地方是省事\n最痛的地方是不可控\n接口变更价格变更限流地区\n限制内容政策都可能影响产品\n开放模型麻烦一点但可控\n对长期做产品的人来说\n可控往往比短期省事更重要\nMuse Glimmer如果性能足够稳定\n就会被放进很多不愿\n完全依赖云端的产品里\n企业采用AI时\n常常不是被能力卡住\n而是被审批卡住\n法务问数据去哪\n安全团队问权限怎么管\n财务问每月调用成本\n业务部门问延迟能不能接受\n开放本地模型可以\n同时回答几个问题\n数据可以留在内网\n成本可以固定\n延迟可以降低\n权限可以自己定义\n它不是完美答案\n但比纯云端更容易\n通过某些企业流程\nMeta还可以借开放\n路线给监管传递信号\n闭源超级模型越强\n监管越担心权力集中和黑箱决策\n开放模型让更多\n研究者能审计能力和风险\n至少表面上更符合\n广泛分发的公共叙事\n对一家掌握社交平台和\n广告系统的公司来说\n这种姿态很重要\n它需要证明自己不是\n把AI变成下一层垄断\n但市场不会因为姿态就买单\n开发者最后只看三件事\n够不够好\n跑不跑得动\n出问题能不能修\nMuse Glimmer需要在真实任务里证明自己\n比如本地图片理解、文档问答\n桌面自动化工具调用失败恢复\n只要这些能力能稳定\n解决一批中等难度问题\n30B就不再是小模型\n而是刚好够用的生产力模型\nAI行业正在出现一种新分工\n最强闭源模型负责天花板\n开放模型负责地板和中间层\n天花板决定行业想象力\n地板决定普及速度\n没有天花板\nAI缺少突破\n没有地板\nAI只会变成少数公司的高价服务\nMuse Glimmer的意义就在地板和中间层\n他让更多人能踩上去\n扎克伯格讲发明创造而不是自动化\n其实是在修复AI的公共形象\n过去两年\n很多人听到AI就想到裁员替代\n数据中心耗电和大公司集中全力\nMeta要把话题拉回\n创造力和个人赋能\nMuse Glimmer适合这个叙事\n因为它看起来不像一个\n只服务企业裁员的工具\n而像一个可以被\n个人拿走的创造工具\n当然这种叙事也要接受现实检验\n如果开放模型最后主要\n被大公司拿去降低成本\n普通人没有明显受益\n那赋能就会变成漂亮话\n真正能证明Meta的\n是它有没有让个人\n開發者、小企業、學校\n研究者用更低成本\n做出过去做不了的东西\n开源的价值不能停在下载量\n要落到新作品和新产品上\n未来一年\nMuse Glimmer 最值得觀察\n的不是发布热度\n而是二次开发热度\n有没有高质量微调版本\n有沒有中文代碼、醫療、法律\n教育设计方向的社区版本\n有沒有人把它做成一鍵本地 Agent\n有没有硬件厂商围绕它\n优化驱动和推理速度\n这些才是开放模型真正的生命迹象\n如果这些生态长出来\nMeta就会完成一次很聪明的布局\n它不需要每个用户直接打开Meta产品\n也不需要每次推理都走Meta服务器\n只要大量AI工具底层\n运行的是Meta开放模型\n它就会成为看不见的基础设施\n基础设施的力量往往比\n一个热门应用更持久\nMeta的优势是\n它不急着把模型本身变成利润中心\n它可以允许模型价格下降\n因为它真正赚钱的地方在\n广告社交关系和平台分发\nOpenai必须证明订阅和API能覆盖成本\nAnthropic必须证明企业安全\n和高端能力能卖出价格\nMeta则可以用开放模型\n削弱对手收费能力\n这种打法很狠\n如果未来个人设备都能跑中型模型\nAI产品形态也会变\n很多應用不再需要把\n所有請求傳到伺服器\n而是把模型嵌進本地功能裡\n寫作軟體可以本地整理草稿\n剪輯軟體可以本地識別素材\n瀏覽器可以本地總結網頁\n開發工具可以本地讀取專案\n雲端仍然存在\n但本地會承擔更多基礎動作\n所以這件事最後要落到一個判斷\nAI的未來不會只有一種入口\n雲端旗艦模型會存在\n本地開放模型也會存在\n前者負責極限能力\n後者負責普及和控制權\nMeta現在壓的是後者\n而且押得很聰明\n它不一定贏走所有收入\n但可能贏走很多預設選擇\n如果把它放進更長週期\nMuse Glimmer\n代表的是AI商品化之後的下一步\n模型能力會越來越像基礎能力\n真正差異會轉移到數據\n場景部署體驗和生態\nMeta提前把模型底層打開\n是在承認模型本身遲早會變便宜\n既然便宜不可避免\n不如主動讓它變便宜\n再去控制更大的生態入口\n這對閉源公司是一種壓力測試\n閉源模型如果只是略強一點\n卻貴很多\n很多企業會改用開放模型\n閉源模型必須在關鍵\n任務上強到無法替代\n或者在產品體驗上省心得足夠明顯\n否則開源模型每進步一步\n閉源模型的中低端收入就少一塊\nMuse Glimmer也會推動AI Agent\n從玩具走向基礎設施\n以前很多agent演示看起来很聪明\n但实际运行成本高延迟长\n隐私风险大\n把一部分能力放到本地之后\nagent可以更安静的处理日常任务\n它不需要每次都上云\n不需要每一步都花钱\n也不需要把所有私人资料交出去\n這就是它的賣點\nMeta不是在送一个模型\n而是在告诉市场\nAI的权力不一定只能往云端集中\n普通设备也可以有智能\n小团队也可以有底座\n开发者也可以有退路\n这个故事比单纯\n参数发布更有传播性\n因为它集中了\n很多人对AI垄断的焦虑\n所以缪斯glimmer值得被单独拿出来讲\n不是因为它\n一夜之间改写大模型格局\n而是因为它把一个被\n忽略的问题摆到台面\nAI普及到底靠更大的模型\n还是靠更容易被部署的模型\n亿元巨头会继续冲天花板\n但真正改变日常工作的往往\n是那些能被普通人装进电脑\n切入流程\n反复修改的东西", "language": "zh", "duration": 1455.937625, "segments": [ { "start": 0.0, "end": 2.033, "text": "Meta开源Muse glimmer30B", "source": "source_caption" }, { "start": 2.033, "end": 28.0, "text": "Meta開源Muse Glimmer 30B,最值得看的不是參數也不是跑分,而是它把大模型競爭從雲端大廠的機房重新推回普通人的電腦桌上。一個30B級別開放權重面向本地Agent的模型放在今天這個節點,意思很明確。AI不能只是一張雲服務帳單,也不能永遠被少數公司關在黑盒裡。這件事要放在Meta的整體敘事裡看。", "chunk": 0, "language": "zh", "avg_logprob": -0.15083410414002782, "no_speech_prob": 7.302859705848874e-12, "compression_ratio": 1.1764705882352942, "source": "whisper_gap_repair", "text_original": "Meta開源Muse Glimmer 30B,最值得看的不是參數也不是跑分,而是它把大模型競爭從雲端大廠的機房重新推回普通人的電腦。一個30B級別開放權重面向本地Agent的模型放在今天這個節點,意思很明確。AI不能只是一張雲服務帳單,也不能永遠被少數公司關在黑盒裡。這件事要放在Meta的整體敘事裡看。", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 28.0, "end": 30.3, "text": "扎克伯格最近把話說得很直白", "source": "source_caption", "text_original": "扎克伯格最近把话说得很直", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 30.866, "end": 32.033, "text": "人工智能的主要", "source": "source_caption" }, { "start": 32.033, "end": 34.033, "text": "目的應該是發明創造", "source": "source_caption", "text_original": "目的应该是发明创造", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 34.233, "end": 36.266, "text": "不應該只是自動化替代", "source": "source_caption", "text_original": "不应该只是自动化替代", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 36.6, "end": 38.866, "text": "超級智能如果真的會到來", "source": "source_caption", "text_original": "超级智能如果真的会到来", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 39.0, "end": 42.166, "text": "那它不應該只掌握在幾家公司手裡", "source": "source_caption", "text_original": "那它不应该只掌握在几家公司手里", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 42.433, "end": 44.866, "text": "而應該盡量廣泛地分發給", "source": "source_caption", "text_original": "而应该尽量广泛的分发给", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 44.866, "end": 47.566, "text": "普通人開發者和小團隊", "source": "source_caption", "text_original": "普通人开发者和小团队", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 48.2, "end": 50.066, "text": "Muse Glimmer就是這個敘事", "source": "source_caption", "text_original": "Muse Glimmer就是这个叙事", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 50.066, "end": 52.1, "text": "落到產品上的一顆棋子", "source": "source_caption", "text_original": "落到产品上的一颗棋子", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 52.3, "end": 54.6, "text": "Muse Glimmer 30B的重點不是", "source": "source_caption", "text_original": "Muse Glimmer 30B的重点不是", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 54.6, "end": 56.766, "text": "挑戰閉源模型的全部能力", "source": "source_caption", "text_original": "挑战闭源模型的全部能力", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 56.866, "end": 60.066, "text": "而是給本地AI留出一條現實路線", "source": "source_caption", "text_original": "而是给本地AI留出一条现实路线", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 60.2, "end": 61.8, "text": "現在很多模型動不動", "source": "source_caption", "text_original": "现在很多模型动不动", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 61.8, "end": 64.0, "text": "就是幾千億上萬億參數", "source": "source_caption", "text_original": "就是几千亿上万亿参数", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 64.166, "end": 65.366, "text": "聽起來很震撼", "source": "source_caption", "text_original": "听起来很震撼", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 65.4, "end": 67.5, "text": "但普通用戶根本跑不起來", "source": "source_caption", "text_original": "但普通用户根本跑不起来", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 67.833, "end": 69.5, "text": "他們需要昂貴GPU", "source": "source_caption", "text_original": "他们需要昂贵GPU", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 69.7, "end": 71.1, "text": "需要雲端推理", "source": "source_caption", "text_original": "需要云端推理", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 71.166, "end": 73.1, "text": "需要訂閱和調用額度", "source": "source_caption", "text_original": "需要订阅和调用额度", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 73.4, "end": 75.466, "text": "Meta選擇30B這個體量", "source": "source_caption", "text_original": "Meta选择30B这个体量", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 75.466, "end": 76.566, "text": "就是在說", "source": "source_caption", "text_original": "就是在说", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 76.7, "end": 79.633, "text": "模型不一定越大越接近未來", "source": "source_caption", "text_original": "模型不一定越大越接近未来", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 79.7, "end": 82.433, "text": "能部署能修改能嵌進", "source": "source_caption", "text_original": "能部署能修改能嵌进", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 82.433, "end": 85.2, "text": "工作流才更接近真實使用", "source": "source_caption", "text_original": "工作流才更接近真实使用", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 85.433, "end": 87.9, "text": "這和LLAMA時代的路線一脈相承", "source": "source_caption", "text_original": "这和LLAMA时代的路线一脉相承", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 88.6, "end": 90.966, "text": "Meta一直在開放模型上押注", "source": "source_caption", "text_original": "Meta一直在开放模型上押注", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 91.166, "end": 93.033, "text": "只是中間經歷過搖擺", "source": "source_caption", "text_original": "只是中间经历过摇摆", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 93.266, "end": 97.7, "text": "當OpenAI Anthropic Google把高端模型", "source": "source_caption", "text_original": "当Openai Anthropic Google把高端模型", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 97.7, "end": 100.0, "text": "能力不斷往閉源體系裡收", "source": "source_caption", "text_original": "能力不断往闭源体系里收", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 100.5, "end": 104.033, "text": "Meta反而重新把開放權重當成差異化", "source": "source_caption", "text_original": "Meta反而重新把开放权重当成差异化", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 104.1, "end": 106.433, "text": "它沒必要在每一項榜單上贏", "source": "source_caption", "text_original": "它没必要在每一项榜单上赢", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 106.666, "end": 108.6, "text": "它要贏的是開發者心智", "source": "source_caption", "text_original": "它要赢的是开发者心智", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 109.1, "end": 111.066, "text": "誰能让更多人拿去改", "source": "source_caption", "text_original": "谁能让更多人拿去改", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 111.166, "end": 113.233, "text": "拿去跑拿去做產品", "source": "source_caption", "text_original": "拿去跑拿去做产品", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 113.466, "end": 116.466, "text": "誰就能在AI底層生態裡留下位置", "source": "source_caption", "text_original": "谁就能在AI底层生态里留下位置", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 116.766, "end": 120.266, "text": "Muse Glimmer還帶有Execute Torch和PTE", "source": "source_caption", "text_original": "musglm还带有execute torch和PTE", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 120.266, "end": 121.9, "text": "這樣的邊緣部署意味", "source": "source_caption", "text_original": "这样的边缘部署意味", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 122.066, "end": 123.566, "text": "这个细节很重要", "source": "source_caption" }, { "start": 124.066, "end": 126.466, "text": "它不是只给服务器看的模型卡", "source": "source_caption" }, { "start": 126.633, "end": 129.1, "text": "而是在暗示手機筆記本", "source": "source_caption", "text_original": "而是在暗示手机笔记本", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 129.1, "end": 131.7, "text": "邊緣設備本地工作站都", "source": "source_caption", "text_original": "边缘设备本地工作站都", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 131.7, "end": 134.2, "text": "可能成为AI agent的运行环境", "source": "source_caption" }, { "start": 134.3, "end": 135.366, "text": "未来AI agent", "source": "source_caption" }, { "start": 135.366, "end": 137.433, "text": "不一定每一步都要回云端", "source": "source_caption" }, { "start": 137.766, "end": 139.666, "text": "他可以在本地看文件讀", "source": "source_caption", "text_original": "他可以在本地看文件读", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 139.666, "end": 142.566, "text": "圖片調工具處理私人數據", "source": "source_caption", "text_original": "图片调工具处理私人数据", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 142.6, "end": 143.866, "text": "再把真正复杂的", "source": "source_caption" }, { "start": 143.866, "end": 145.666, "text": "任务交给云端大模型", "source": "source_caption" }, { "start": 145.8, "end": 149.4, "text": "这种路线最大的价值是隐私和成本", "source": "source_caption" }, { "start": 149.866, "end": 153.2, "text": "用戶把本地文件工作資料聊天", "source": "source_caption", "text_original": "用户把本地文件工作资料聊天", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 153.2, "end": 156.166, "text": "記錄圖片和代碼交給雲端模型", "source": "source_caption", "text_original": "记录图片和代码交给云端模型", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 156.433, "end": 158.0, "text": "永远会有心理门槛", "source": "source_caption" }, { "start": 158.4, "end": 159.466, "text": "企業更敏感", "source": "source_caption", "text_original": "企业企业更敏感", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 159.566, "end": 161.833, "text": "尤其涉及客戶數據合同", "source": "source_caption", "text_original": "尤其涉及客户数据合同", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 162.0, "end": 163.833, "text": "研發資料和內部流程", "source": "source_caption", "text_original": "研发资料和内部流程", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 163.966, "end": 165.966, "text": "一個能在本地跑的模型", "source": "source_caption", "text_original": "一个能在本地跑的模型", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 166.1, "end": 167.833, "text": "就算能力不是最强", "source": "source_caption" }, { "start": 168.0, "end": 169.4, "text": "設備而變得有价值", "source": "source_caption", "text_original": "也会因为数据不出", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 169.4, "end": 170.966, "text": "设备而变得有价值", "source": "source_caption" }, { "start": 171.166, "end": 172.633, "text": "很多真實場景不需要", "source": "source_caption", "text_original": "很多真实场景不需要", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 172.633, "end": 174.266, "text": "世界最聰明的模型", "source": "source_caption", "text_original": "世界最聪明的模型", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 174.8, "end": 176.5, "text": "只需要一個足夠聰明", "source": "source_caption", "text_original": "只需要一个足够聪明", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 176.566, "end": 179.266, "text": "足夠便宜足夠可控的助手", "source": "source_caption", "text_original": "足够便宜足够可控的助手", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 179.366, "end": 181.0, "text": "這也是開源模型真正", "source": "source_caption", "text_original": "这也是开源模型真正", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 181.0, "end": 182.666, "text": "打閉源模型的地方", "source": "source_caption", "text_original": "打闭源模型的地方", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 183.233, "end": 186.033, "text": "閉源模型擅長提供最強通用能力", "source": "source_caption", "text_original": "闭源模型擅长提供最强通用能力", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 186.266, "end": 189.2, "text": "開源模型擅長被改造成具體工具", "source": "source_caption", "text_original": "开源模型擅长被改造成具体工具", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 189.266, "end": 190.866, "text": "小團隊可以微調它", "source": "source_caption", "text_original": "小团队可以微调它", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 191.3, "end": 193.3, "text": "企業可以把它部署到內網", "source": "source_caption", "text_original": "企业可以把它部署到内网", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 193.466, "end": 194.7, "text": "開發者可以把它接", "source": "source_caption", "text_original": "开发者可以把它接", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 194.7, "end": 196.4, "text": "進自己的agent框架", "source": "source_caption", "text_original": "进自己的agent框架", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 196.966, "end": 199.5, "text": "硬體廠商可以把它塞進邊緣設備", "source": "source_caption", "text_original": "硬件厂商可以把它塞进边缘设备", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 199.7, "end": 201.466, "text": "閉源模型賣的是能力", "source": "source_caption", "text_original": "闭源模型卖的是能力", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 201.466, "end": 203.566, "text": "開放模型賣的是可塑性", "source": "source_caption", "text_original": "开放模型卖的是可塑性", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 203.766, "end": 205.433, "text": "兩者不是同一場比賽", "source": "source_caption", "text_original": "两者不是同一场比赛", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 205.6, "end": 207.8, "text": "Meta這次還把數據中心社區", "source": "source_caption", "text_original": "Meta这次还把数据中心社区", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 207.8, "end": 209.766, "text": "基金放在同一個敘事裡", "source": "source_caption", "text_original": "基金放在同一个叙事里", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 209.8, "end": 211.966, "text": "說明它知道AI擴張正在", "source": "source_caption", "text_original": "说明它知道AI扩张正在", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 211.966, "end": 213.566, "text": "遇到新的政治成本", "source": "source_caption" }, { "start": 214.2, "end": 216.4, "text": "數據中心需要電力、土地", "source": "source_caption", "text_original": "数据中心需要电力土地", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 216.4, "end": 218.966, "text": "水、輸電設施和社區許可", "source": "source_caption", "text_original": "水输电设施和社区许可", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 219.033, "end": 221.066, "text": "過去科技公司建機房", "source": "source_caption", "text_original": "过去科技公司建机房", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 221.066, "end": 222.666, "text": "地方政府往往歡迎", "source": "source_caption", "text_original": "地方政府往往欢迎", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 222.666, "end": 224.6, "text": "因為它帶來投資和稅收", "source": "source_caption", "text_original": "因为它带来投资和税收", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 224.7, "end": 226.1, "text": "現在情況變了", "source": "source_caption", "text_original": "现在情况变了", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 226.166, "end": 228.233, "text": "居民開始擔心電費上漲", "source": "source_caption", "text_original": "居民开始担心电费上涨", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 228.666, "end": 231.0, "text": "環境壓力和基礎設施被佔用", "source": "source_caption", "text_original": "环境压力和基础设施被占用", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 231.1, "end": 233.433, "text": "AI越龐大越不可能只用", "source": "source_caption", "text_original": "AI越庞大越不可能只用", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 233.433, "end": 235.2, "text": "技術語言解釋自己", "source": "source_caption", "text_original": "技术语言解释自己", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 235.433, "end": 238.033, "text": "10億美元社區基金表面上是補償", "source": "source_caption", "text_original": "10亿美元社区基金表面上是补偿", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 238.366, "end": 240.866, "text": "背後是AI基建時代的通行費", "source": "source_caption", "text_original": "背后是AI基建时代的通行费", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 241.566, "end": 243.233, "text": "Meta想告訴當地居民", "source": "source_caption", "text_original": "Meta想告诉当地居民", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 243.5, "end": 246.0, "text": "數據中心不是只把電和水抽走", "source": "source_caption", "text_original": "数据中心不是只把电和水抽走", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 246.0, "end": 249.033, "text": "也會把錢、崗位和公共服務帶回來", "source": "source_caption", "text_original": "也会把钱岗位和公共服务带回来", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 249.4, "end": 252.7, "text": "警察、消防員、社區成員被點名", "source": "source_caption", "text_original": "警察消防员社区成员被点名", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 252.766, "end": 254.966, "text": "是一種很現實的政治溝通", "source": "source_caption", "text_original": "是一种很现实的政治沟通", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 255.1, "end": 256.833, "text": "AI公司已經發現", "source": "source_caption", "text_original": "AI公司已经发现", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 256.966, "end": 259.433, "text": "光說技術改變世界沒用了", "source": "source_caption", "text_original": "光说技术改变世界没用了", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 259.7, "end": 262.3, "text": "真正被影響的人要看到具體回報", "source": "source_caption", "text_original": "真正被影响的人要看到具体回报", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 262.566, "end": 265.966, "text": "Muse Glimmer和社區基金看似是兩件事", "source": "source_caption", "text_original": "Musglimmer和社区基金看似是两件事", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 265.966, "end": 267.666, "text": "其實對應同一個問題", "source": "source_caption", "text_original": "其实对应同一个问题", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 267.966, "end": 270.233, "text": "AI權力會不會過度集中", "source": "source_caption", "text_original": "AI权力会不会过度集中", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 270.7, "end": 272.266, "text": "模型集中在雲端", "source": "source_caption", "text_original": "模型集中在云端", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 272.433, "end": 274.466, "text": "算力集中在數據中心", "source": "source_caption", "text_original": "算力集中在数据中心", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 274.566, "end": 276.633, "text": "收益集中在少數公司", "source": "source_caption", "text_original": "收益集中在少数公司", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 276.9, "end": 279.6, "text": "成本卻擴散給社區和消費者", "source": "source_caption", "text_original": "成本却扩散给社区和消费者", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 279.666, "end": 281.3, "text": "這套結構如果不調整", "source": "source_caption", "text_original": "这套结构如果不调整", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 281.3, "end": 282.7, "text": "反彈會越來越大", "source": "source_caption", "text_original": "反弹会越来越大", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 282.866, "end": 286.566, "text": "Meta現在講開放、講本地、講回饋社區", "source": "source_caption", "text_original": "Meta现在讲开放讲本地讲回馈社区", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 286.6, "end": 288.166, "text": "就是在給自己的AI", "source": "source_caption", "text_original": "就是在给自己的AI", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 288.166, "end": 290.033, "text": "擴張換一種合法性", "source": "source_caption", "text_original": "扩张换一种合法性", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 290.266, "end": 292.9, "text": "這不代表Meta突然變成公益組織", "source": "source_caption", "text_original": "这不代表Meta突然变成公益组织", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 293.1, "end": 295.0, "text": "開源也不是純粹善意", "source": "source_caption", "text_original": "开源也不是纯粹善意", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 295.366, "end": 297.466, "text": "Meta的商業邏輯很清楚", "source": "source_caption", "text_original": "Meta的商业逻辑很清楚", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 297.866, "end": 299.8, "text": "它沒有像OpenAI那樣把", "source": "source_caption", "text_original": "它没有像Openai那样把", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 299.8, "end": 302.0, "text": "訂閱模型做成第一入口", "source": "source_caption", "text_original": "订阅模型做成第一入口", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 302.2, "end": 304.866, "text": "也沒有Google這樣的搜索默認位置", "source": "source_caption", "text_original": "也没有Google那样的搜索默认位置", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 305.266, "end": 307.033, "text": "它最擅长的是平台", "source": "source_caption" }, { "start": 307.066, "end": 309.633, "text": "分发广告和开发者网络", "source": "source_caption" }, { "start": 309.8, "end": 311.4, "text": "开放模型能削弱", "source": "source_caption" }, { "start": 311.4, "end": 313.3, "text": "闭源模型的收费能力", "source": "source_caption" }, { "start": 313.466, "end": 315.6, "text": "降低竞争对手的护城河", "source": "source_caption" }, { "start": 315.9, "end": 317.8, "text": "同时让更多应用围绕", "source": "source_caption" }, { "start": 317.8, "end": 319.666, "text": "Meta的模型生态生长", "source": "source_caption" }, { "start": 319.966, "end": 322.066, "text": "如果一个开放模型足够好", "source": "source_caption" }, { "start": 322.066, "end": 324.2, "text": "大量开发者就会围绕它做", "source": "source_caption" }, { "start": 324.2, "end": 327.1, "text": "工具插件微调和部署方案", "source": "source_caption" }, { "start": 327.2, "end": 329.966, "text": "哪怕Meta不直接收每一次调用费", "source": "source_caption" }, { "start": 330.0, "end": 333.166, "text": "它也能影响标准框架和生态方向", "source": "source_caption" }, { "start": 333.4, "end": 336.2, "text": "安卓当年不是靠系统授权费赚钱", "source": "source_caption" }, { "start": 336.566, "end": 339.2, "text": "而是靠开放系统占住移动入口", "source": "source_caption" }, { "start": 339.466, "end": 342.233, "text": "LLAMA和Muse Glimmer也有类似逻辑", "source": "source_caption", "text_original": "LLAMA和Muse glimmer也有类似逻辑", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 342.433, "end": 345.166, "text": "用开放打破别人的封闭利润池", "source": "source_caption" }, { "start": 345.3, "end": 347.8, "text": "再在更大的生态里寻找收益", "source": "source_caption" }, { "start": 348.1, "end": 352.066, "text": "Muse Glimmer的30B体量还有一个现实意义", "source": "source_caption", "text_original": "Muse glimmer的30B体量还有一个现实意义", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 352.266, "end": 354.566, "text": "它让AI从演示回到产品", "source": "source_caption" }, { "start": 354.833, "end": 357.8, "text": "很多AI发布会喜欢展示极限能力", "source": "source_caption" }, { "start": 357.9, "end": 360.7, "text": "仿佛模型能解决所有复杂问题", "source": "source_caption" }, { "start": 360.9, "end": 362.4, "text": "但真正落地时", "source": "source_caption" }, { "start": 362.4, "end": 364.766, "text": "企业关心的是成本延迟", "source": "source_caption" }, { "start": 365.033, "end": 368.066, "text": "部署安全稳定可维护", "source": "source_caption" }, { "start": 368.166, "end": 369.6, "text": "一个30B模型", "source": "source_caption" }, { "start": 369.6, "end": 371.566, "text": "如果能在普通设备或", "source": "source_caption" }, { "start": 371.566, "end": 374.033, "text": "消费级GPU上完成稳定任务", "source": "source_caption" }, { "start": 374.366, "end": 377.2, "text": "它的商业价值可能比一个只能在", "source": "source_caption" }, { "start": 377.2, "end": 380.366, "text": "云端高价运行的巨型模型更直接", "source": "source_caption" }, { "start": 380.433, "end": 381.966, "text": "尤其是agent场景", "source": "source_caption" }, { "start": 382.166, "end": 383.766, "text": "agent不是一次聊天", "source": "source_caption" }, { "start": 383.766, "end": 385.266, "text": "而是一连串动作", "source": "source_caption" }, { "start": 385.5, "end": 387.566, "text": "理解目标拆任务", "source": "source_caption" }, { "start": 387.866, "end": 389.7, "text": "读文件调用工具", "source": "source_caption" }, { "start": 389.833, "end": 390.866, "text": "检查结果", "source": "source_caption" }, { "start": 390.9, "end": 392.2, "text": "失败后重试", "source": "source_caption" }, { "start": 392.3, "end": 395.0, "text": "这里面很多步骤并不需要最强模型", "source": "source_caption" }, { "start": 395.1, "end": 398.166, "text": "文件分类表格检查界面操作", "source": "source_caption" }, { "start": 398.2, "end": 401.866, "text": "图片理解简单代码修改本地搜索", "source": "source_caption" }, { "start": 402.033, "end": 404.0, "text": "都可以由较小模型承担", "source": "source_caption" }, { "start": 404.1, "end": 406.3, "text": "把这些步骤放到本地执行", "source": "source_caption" }, { "start": 406.866, "end": 408.9, "text": "云端只处理最难的推理", "source": "source_caption" }, { "start": 409.2, "end": 411.4, "text": "整个系统成本会明显下降", "source": "source_caption" }, { "start": 411.433, "end": 412.866, "text": "这也解释了为什么", "source": "source_caption" }, { "start": 412.866, "end": 415.466, "text": "本地模型会越来越重要", "source": "source_caption" }, { "start": 415.466, "end": 417.2, "text": "AI的下一阶段不是谁", "source": "source_caption" }, { "start": 417.2, "end": 418.633, "text": "能回答得更漂亮", "source": "source_caption" }, { "start": 419.166, "end": 422.1, "text": "而是谁能更便宜地完成更多动作", "source": "source_caption", "text_original": "而是谁能更便宜的完成更多动作", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 422.3, "end": 424.866, "text": "一个公司每天调用百万次模型", "source": "source_caption" }, { "start": 425.1, "end": 427.8, "text": "如果每一步都走最贵的闭源API", "source": "source_caption" }, { "start": 427.966, "end": 429.433, "text": "成本会很快失控", "source": "source_caption" }, { "start": 429.466, "end": 431.1, "text": "本地模型可以承担大量", "source": "source_caption" }, { "start": 431.1, "end": 433.966, "text": "低风险重复性隐私敏感任务", "source": "source_caption" }, { "start": 434.033, "end": 435.8, "text": "它不必取代旗舰模型", "source": "source_caption" }, { "start": 436.166, "end": 437.666, "text": "只要把旗舰模型从低", "source": "source_caption" }, { "start": 437.666, "end": 439.433, "text": "价值任务里解放出来", "source": "source_caption" }, { "start": 439.433, "end": 440.766, "text": "就已经很有意义", "source": "source_caption" }, { "start": 440.866, "end": 443.8, "text": "对普通创作者和小公司来说", "source": "source_caption" }, { "start": 443.833, "end": 445.8, "text": "这条路线尤其关键", "source": "source_caption" }, { "start": 445.833, "end": 447.5, "text": "大公司可以买算力", "source": "source_caption" }, { "start": 447.633, "end": 448.7, "text": "签云合同", "source": "source_caption" }, { "start": 448.7, "end": 449.866, "text": "雇安全团队", "source": "source_caption" }, { "start": 449.966, "end": 451.633, "text": "小团队没有这个条件", "source": "source_caption" }, { "start": 452.233, "end": 454.033, "text": "如果未来所有AI能力都", "source": "source_caption" }, { "start": 454.033, "end": 456.4, "text": "绑定高价订阅和云端调用", "source": "source_caption" }, { "start": 456.433, "end": 458.566, "text": "小团队会永远站在门外", "source": "source_caption" }, { "start": 458.6, "end": 459.9, "text": "开放权重模型", "source": "source_caption" }, { "start": 459.9, "end": 461.5, "text": "让他们至少有机会在", "source": "source_caption" }, { "start": 461.5, "end": 463.366, "text": "本地搭一个能跑的系统", "source": "source_caption" }, { "start": 463.666, "end": 464.633, "text": "哪怕不完美", "source": "source_caption" }, { "start": 464.633, "end": 466.766, "text": "也能开始改造自己的工作流", "source": "source_caption" }, { "start": 466.9, "end": 469.3, "text": "当然开源模型也有风险", "source": "source_caption" }, { "start": 469.4, "end": 471.6, "text": "开放权重意味着能力会扩散", "source": "source_caption" }, { "start": 471.866, "end": 473.3, "text": "好的用途会扩散", "source": "source_caption" }, { "start": 473.466, "end": 475.033, "text": "坏的用途也会扩散", "source": "source_caption" }, { "start": 475.233, "end": 477.766, "text": "安全过滤、滥用检测", "source": "source_caption", "text_original": "安全过滤滥用检测", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 478.033, "end": 479.8, "text": "责任归属都会变复杂", "source": "source_caption" }, { "start": 479.966, "end": 482.566, "text": "闭源公司常用安全作为封闭理由", "source": "source_caption" }, { "start": 482.566, "end": 484.2, "text": "不能说完全没有道理", "source": "source_caption" }, { "start": 484.4, "end": 487.466, "text": "但如果只因为风险就把能力锁死", "source": "source_caption" }, { "start": 487.9, "end": 490.566, "text": "最后得到的是少数公司", "source": "source_caption" }, { "start": 490.566, "end": 492.3, "text": "控制所有AI基础设施", "source": "source_caption" }, { "start": 492.566, "end": 494.666, "text": "Meta现在押的是另一种答案", "source": "source_caption" }, { "start": 495.3, "end": 496.633, "text": "风险需要治理", "source": "source_caption" }, { "start": 496.7, "end": 499.766, "text": "但不能用风险作为永久垄断的理由", "source": "source_caption" }, { "start": 500.0, "end": 502.066, "text": "Muse Glimmer真正要证明的", "source": "source_caption" }, { "start": 502.066, "end": 505.066, "text": "不是30B能不能击败所有大模型", "source": "source_caption" }, { "start": 505.233, "end": 508.166, "text": "而是开放路线能不能在AI", "source": "source_caption" }, { "start": 508.166, "end": 510.5, "text": "商业化最激烈的时候继续存在", "source": "source_caption" }, { "start": 510.666, "end": 512.033, "text": "闭源路线说", "source": "source_caption" }, { "start": 512.3, "end": 514.1, "text": "最强能力必须集中", "source": "source_caption" }, { "start": 514.4, "end": 517.466, "text": "才能安全可靠可收费", "source": "source_caption" }, { "start": 517.5, "end": 520.1, "text": "开放路线说能力必须分发", "source": "source_caption" }, { "start": 520.266, "end": 522.833, "text": "社会才不会被少数入口控制", "source": "source_caption" }, { "start": 523.0, "end": 524.633, "text": "两条路线都会存在", "source": "source_caption" }, { "start": 524.7, "end": 526.9, "text": "但它们代表完全不同的未来", "source": "source_caption" }, { "start": 527.233, "end": 529.833, "text": "接下来要看的不是一两次榜单成绩", "source": "source_caption" }, { "start": 529.9, "end": 532.3, "text": "而是开发者是否真的使用它", "source": "source_caption" }, { "start": 532.4, "end": 534.6, "text": "有没有人把它部署到本地agent", "source": "source_caption" }, { "start": 534.966, "end": 537.6, "text": "有没有企业拿它做内部流程", "source": "source_caption" }, { "start": 537.766, "end": 540.766, "text": "有没有硬件厂商把它放进设备", "source": "source_caption" }, { "start": 541.166, "end": 542.7, "text": "有没有开源社区围绕", "source": "source_caption" }, { "start": 542.7, "end": 544.4, "text": "它做微调和工具链", "source": "source_caption" }, { "start": 544.666, "end": 546.233, "text": "如果这些事情发生", "source": "source_caption" }, { "start": 546.4, "end": 547.966, "text": "Muse Glimmer的意义就不", "source": "source_caption", "text_original": "muscoglamer的意义就不", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 547.966, "end": 549.6, "text": "只是一个模型发布", "source": "source_caption" }, { "start": 549.666, "end": 552.066, "text": "而是Meta在重新争夺AI", "source": "source_caption" }, { "start": 552.066, "end": 553.766, "text": "基础设施的话语权", "source": "source_caption" }, { "start": 553.766, "end": 556.033, "text": "这场竞争还有一个更大的背景", "source": "source_caption" }, { "start": 556.7, "end": 558.833, "text": "AI正在从模型能力竞争", "source": "source_caption" }, { "start": 558.9, "end": 560.766, "text": "转向分发方式竞争", "source": "source_caption" }, { "start": 561.0, "end": 563.666, "text": "Openai想把模型做成超级应用", "source": "source_caption" }, { "start": 563.866, "end": 566.266, "text": "Google想把AI塞回搜索和安卓", "source": "source_caption" }, { "start": 566.566, "end": 569.033, "text": "苹果想把AI放进硬件和系统", "source": "source_caption" }, { "start": 569.266, "end": 571.1, "text": "Meta则想让开放模型", "source": "source_caption" }, { "start": 571.1, "end": 572.9, "text": "成为开发者默认底座", "source": "source_caption" }, { "start": 573.266, "end": 575.6, "text": "Glimmer只是其中一块拼图", "source": "source_caption", "text_original": "Musglimer只是其中一块拼图", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 575.633, "end": 577.566, "text": "但它指向的方向很清楚", "source": "source_caption" }, { "start": 577.7, "end": 580.366, "text": "让AI不止存在于云端", "source": "source_caption" }, { "start": 580.566, "end": 582.8, "text": "也存在于每台电脑每个", "source": "source_caption" }, { "start": 582.8, "end": 585.0, "text": "应用每个本地工作流里", "source": "source_caption" }, { "start": 585.0, "end": 586.6, "text": "所以这次发布不能只", "source": "source_caption" }, { "start": 586.6, "end": 589.433, "text": "看成Meta又开源了一个模型", "source": "source_caption" }, { "start": 589.8, "end": 591.5, "text": "它更像一次路线宣言", "source": "source_caption" }, { "start": 591.766, "end": 594.366, "text": "AI如果只能在数据中心运行", "source": "source_caption" }, { "start": 594.633, "end": 596.5, "text": "用户就永远是租客", "source": "source_caption" }, { "start": 596.666, "end": 598.9, "text": "AI如果可以在本地运行", "source": "source_caption" }, { "start": 599.233, "end": 601.3, "text": "用户才有机会成为主人", "source": "source_caption" }, { "start": 601.466, "end": 603.6, "text": "Muse Glimmer 30B 的价值就在这里", "source": "source_caption", "text_original": "musglimmer的价值就在这里", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 603.766, "end": 605.566, "text": "它可能不是最强的模型", "source": "source_caption" }, { "start": 605.866, "end": 607.666, "text": "但他在提醒整个行业", "source": "source_caption" }, { "start": 607.866, "end": 610.666, "text": "未来的AI不应该只属于机房", "source": "source_caption" }, { "start": 610.666, "end": 612.266, "text": "也应该属于桌面", "source": "source_caption" }, { "start": 612.3, "end": 615.766, "text": "这里还要看Costello对AI投资的判断", "source": "source_caption", "text_original": "这里还要看Costla对AI投资的判断", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 615.966, "end": 617.966, "text": "他认为Openai和Anthropic", "source": "source_caption" }, { "start": 617.966, "end": 619.5, "text": "不是AI故事的终点", "source": "source_caption" }, { "start": 619.5, "end": 620.666, "text": "而只是开头", "source": "source_caption" }, { "start": 620.8, "end": 622.1, "text": "未来会出现更多", "source": "source_caption" }, { "start": 622.1, "end": 624.766, "text": "千亿美元万亿美元级公司", "source": "source_caption" }, { "start": 625.0, "end": 627.3, "text": "影响科学医疗能源", "source": "source_caption" }, { "start": 627.366, "end": 629.666, "text": "教育制造这些基础领域", "source": "source_caption" }, { "start": 629.7, "end": 632.033, "text": "这个判断放到 Muse Glimmer 30B 上", "source": "source_caption", "text_original": "这个判断放到museglimpse上", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 632.066, "end": 633.9, "text": "就能看见Meta的算盘", "source": "source_caption" }, { "start": 634.0, "end": 637.066, "text": "如果未来有大量垂直AI公司出现", "source": "source_caption" }, { "start": 637.233, "end": 640.166, "text": "他们不一定都愿意从闭源API开始", "source": "source_caption" }, { "start": 640.433, "end": 642.6, "text": "科学研究就是一个典型场景", "source": "source_caption" }, { "start": 642.766, "end": 645.166, "text": "实验室需要处理论文图像", "source": "source_caption" }, { "start": 645.2, "end": 647.866, "text": "代码仪器数据和内部记录", "source": "source_caption" }, { "start": 648.0, "end": 649.5, "text": "很多资料不能随便", "source": "source_caption" }, { "start": 649.5, "end": 651.1, "text": "上传到外部云服务", "source": "source_caption" }, { "start": 651.166, "end": 653.3, "text": "一个本地多模态模型", "source": "source_caption" }, { "start": 653.366, "end": 654.666, "text": "就算不是最强", "source": "source_caption" }, { "start": 654.7, "end": 656.666, "text": "也可以承担文献整理", "source": "source_caption" }, { "start": 657.066, "end": 659.6, "text": "实验记录搜索图表理解", "source": "source_caption" }, { "start": 659.633, "end": 662.233, "text": "工具调用和初步假设生成", "source": "source_caption" }, { "start": 662.433, "end": 663.966, "text": "它不会替代科学家", "source": "source_caption" }, { "start": 664.066, "end": 666.3, "text": "但会把很多低价值搜索", "source": "source_caption" }, { "start": 666.3, "end": 668.4, "text": "和整理动作压低成本", "source": "source_caption" }, { "start": 668.633, "end": 669.633, "text": "Meta很清楚", "source": "source_caption" }, { "start": 670.0, "end": 671.5, "text": "闭源模型越强", "source": "source_caption" }, { "start": 671.666, "end": 674.766, "text": "外界对模型集中化的担心就越重", "source": "source_caption" }, { "start": 674.8, "end": 676.866, "text": "一个公司如果控制模型", "source": "source_caption" }, { "start": 677.3, "end": 678.466, "text": "控制分发", "source": "source_caption" }, { "start": 679.0, "end": 680.033, "text": "控制价格", "source": "source_caption" }, { "start": 680.233, "end": 681.633, "text": "控制安全规则", "source": "source_caption" }, { "start": 681.966, "end": 683.966, "text": "开发者就会变成依附者", "source": "source_caption" }, { "start": 684.3, "end": 686.166, "text": "今天接口价格可以接受", "source": "source_caption" }, { "start": 686.233, "end": 687.766, "text": "明天规则可能变", "source": "source_caption" }, { "start": 687.766, "end": 689.433, "text": "今天调用额度够用", "source": "source_caption" }, { "start": 689.433, "end": 692.166, "text": "明天业务增长后成本可能翻倍", "source": "source_caption" }, { "start": 692.266, "end": 695.066, "text": "开放模型给开发者留了一条退路", "source": "source_caption" }, { "start": 695.633, "end": 698.1, "text": "这条退路本身就是谈判筹码", "source": "source_caption" }, { "start": 698.3, "end": 699.666, "text": "这也是为什么开放权重会压低行业利润率", "source": "source_caption", "text_original": "这也是为什么开放", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 699.666, "end": 701.866, "text": "权重会压低行业利润率", "source": "source_caption" }, { "start": 701.966, "end": 704.2, "text": "只要有足够好的开源替代品", "source": "source_caption" }, { "start": 704.8, "end": 706.433, "text": "闭源公司就很难对中低端任务收太高价格", "source": "source_caption", "text_original": "闭源公司就很难对中", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 706.433, "end": 708.266, "text": "低端任务收太高价格", "source": "source_caption" }, { "start": 708.466, "end": 710.633, "text": "旗舰模型仍然可以卖高价", "source": "source_caption" }, { "start": 710.766, "end": 712.566, "text": "但大量日常任务会被本地模型和开放模型吃掉", "source": "source_caption", "text_original": "但大量日常任务会被", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 712.566, "end": 714.866, "text": "本地模型和开放模型吃掉", "source": "source_caption" }, { "start": 715.033, "end": 717.166, "text": "Meta不靠模型订阅赚钱", "source": "source_caption" }, { "start": 717.566, "end": 719.966, "text": "反而适合推动这种价格下行", "source": "source_caption" }, { "start": 720.066, "end": 721.833, "text": "他要让AI变成空气", "source": "source_caption" }, { "start": 722.0, "end": 723.866, "text": "然后在空气里做平台", "source": "source_caption" }, { "start": 723.9, "end": 726.866, "text": "有人会问30B模型真的够吗", "source": "source_caption" }, { "start": 726.966, "end": 728.433, "text": "这个问题要看任务", "source": "source_caption" }, { "start": 728.7, "end": 730.4, "text": "写一段复杂法律意见", "source": "source_caption" }, { "start": 730.4, "end": 731.3, "text": "可能不够", "source": "source_caption" }, { "start": 731.433, "end": 734.033, "text": "做一个本地客服分类器够", "source": "source_caption" }, { "start": 734.266, "end": 737.433, "text": "通读一本手册并回答设备维修问题", "source": "source_caption" }, { "start": 737.566, "end": 741.266, "text": "可能够整理图片和文字生成初稿", "source": "source_caption" }, { "start": 741.4, "end": 744.566, "text": "可能够控制一组本地工具跑流程，也可能够", "source": "source_caption", "text_original": "可能够控制一组本地工具跑流程不", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 744.866, "end": 745.966, "text": "可能也够", "source": "source_caption" }, { "start": 746.233, "end": 747.8, "text": "AI产品不需要每次都", "source": "source_caption" }, { "start": 747.8, "end": 749.466, "text": "派最强模型上场", "source": "source_caption" }, { "start": 749.666, "end": 751.366, "text": "就像公司不会让总裁", "source": "source_caption" }, { "start": 751.366, "end": 753.033, "text": "去处理每张报销单", "source": "source_caption" }, { "start": 753.3, "end": 755.166, "text": "真正成熟的AI系统", "source": "source_caption" }, { "start": 755.3, "end": 757.233, "text": "很可能是多模型协作", "source": "source_caption" }, { "start": 757.366, "end": 758.866, "text": "小模型负责本地", "source": "source_caption" }, { "start": 758.9, "end": 761.466, "text": "便宜低风险高频动作", "source": "source_caption" }, { "start": 761.5, "end": 763.666, "text": "中型模型负责复杂判断", "source": "source_caption" }, { "start": 764.366, "end": 766.766, "text": "旗舰模型负责少数关键推理", "source": "source_caption" }, { "start": 766.9, "end": 768.766, "text": "Muse Glimmer的位置就在", "source": "source_caption" }, { "start": 768.766, "end": 770.566, "text": "第一层和第二层之间", "source": "source_caption" }, { "start": 770.7, "end": 772.066, "text": "它不抢所有舞台", "source": "source_caption" }, { "start": 772.3, "end": 774.166, "text": "但可以进入很多工作流", "source": "source_caption" }, { "start": 774.2, "end": 776.166, "text": "这个位置一旦铺开", "source": "source_caption" }, { "start": 776.3, "end": 777.866, "text": "使用量会非常大", "source": "source_caption" }, { "start": 777.9, "end": 781.166, "text": "所以Muse Glimmer看起来是技术事件", "source": "source_caption" }, { "start": 781.166, "end": 783.066, "text": "背后其实是分发事件", "source": "source_caption" }, { "start": 783.266, "end": 785.7, "text": "AI行业未来最大的战场", "source": "source_caption" }, { "start": 785.7, "end": 787.966, "text": "不只是模型谁更聪明", "source": "source_caption" }, { "start": 788.033, "end": 790.466, "text": "而是谁能接触更多开发者", "source": "source_caption" }, { "start": 790.8, "end": 791.666, "text": "更多设备", "source": "source_caption" }, { "start": 791.9, "end": 793.1, "text": "更多工作流", "source": "source_caption" }, { "start": 793.233, "end": 794.633, "text": "更多真实业务", "source": "source_caption" }, { "start": 794.7, "end": 797.3, "text": "开放权重提供了一种复制速度", "source": "source_caption" }, { "start": 797.433, "end": 799.3, "text": "一个模型被下载后", "source": "source_caption" }, { "start": 799.3, "end": 801.9, "text": "可以在无数环境里变成不同产品", "source": "source_caption" }, { "start": 802.033, "end": 803.433, "text": "闭源模型再强", "source": "source_caption" }, { "start": 803.666, "end": 805.8, "text": "也必须等用户来到它的入口", "source": "source_caption" }, { "start": 805.966, "end": 808.5, "text": "当然开放模型要赢也不容易", "source": "source_caption" }, { "start": 808.7, "end": 810.233, "text": "开发体验必须好", "source": "source_caption" }, { "start": 810.366, "end": 811.633, "text": "文档必须清楚", "source": "source_caption" }, { "start": 811.8, "end": 813.3, "text": "工具链必须稳定", "source": "source_caption" }, { "start": 813.466, "end": 815.233, "text": "推理速度必须能接受", "source": "source_caption" }, { "start": 815.433, "end": 817.366, "text": "硬件适配不能太痛苦", "source": "source_caption" }, { "start": 817.633, "end": 819.3, "text": "如果一个模型开源了", "source": "source_caption" }, { "start": 819.3, "end": 821.233, "text": "但部署起来像拆炸弹", "source": "source_caption" }, { "start": 821.4, "end": 823.233, "text": "开发者很快就会放弃", "source": "source_caption" }, { "start": 823.466, "end": 827.1, "text": "Meta这次把exec torch等部署线索放进来", "source": "source_caption" }, { "start": 827.233, "end": 829.866, "text": "说明它知道只开源权重还不够", "source": "source_caption" }, { "start": 829.9, "end": 831.966, "text": "真正要争的是落地链路", "source": "source_caption" }, { "start": 832.166, "end": 834.033, "text": "安全问题也不能回避", "source": "source_caption" }, { "start": 834.166, "end": 835.7, "text": "一个能在本地运行", "source": "source_caption" }, { "start": 836.033, "end": 837.2, "text": "能调用工具", "source": "source_caption" }, { "start": 837.566, "end": 839.866, "text": "能处理多模态输入的模型", "source": "source_caption" }, { "start": 840.066, "end": 841.566, "text": "如果被错误使用", "source": "source_caption" }, { "start": 841.6, "end": 843.466, "text": "也会带来自动化滥用", "source": "source_caption" }, { "start": 843.566, "end": 845.2, "text": "开放社区需要评估", "source": "source_caption" }, { "start": 845.4, "end": 848.066, "text": "紅隊使用限制和透明報告", "source": "source_caption", "text_original": "红队使用限制和透明报告", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 848.266, "end": 850.233, "text": "Meta不能一邊強調開放", "source": "source_caption", "text_original": "Meta不能一边强调开放", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 850.6, "end": 852.633, "text": "一邊把風險全丟給社區", "source": "source_caption", "text_original": "一边把风险全丢给社区", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 852.8, "end": 854.433, "text": "開放路線要走得久", "source": "source_caption", "text_original": "开放路线要走得久", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 854.5, "end": 856.366, "text": "必須同時建立開放治理", "source": "source_caption", "text_original": "必须同时建立开放治理", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 856.5, "end": 858.6, "text": "但閉源也不是天然安全", "source": "source_caption", "text_original": "但闭源也不是天然安全", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 858.833, "end": 860.766, "text": "閉源模型的能力集中後", "source": "source_caption", "text_original": "闭源模型的能力集中后", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 860.766, "end": 863.2, "text": "外界更难知道它如何訓練", "source": "source_caption", "text_original": "外界更难知道它如何训练", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 863.2, "end": 864.033, "text": "如何過濾", "source": "source_caption", "text_original": "如何过滤", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 864.1, "end": 865.033, "text": "如何拒絕", "source": "source_caption", "text_original": "如何拒绝", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 865.033, "end": 866.7, "text": "如何偏向某些服務", "source": "source_caption", "text_original": "如何偏向某些服务", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 866.866, "end": 869.566, "text": "安全不是封閉和開放的簡單選擇", "source": "source_caption", "text_original": "安全不是封闭和开放的简单选择", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 870.266, "end": 871.966, "text": "而是透明度、責任和可驗證性之間的平衡", "source": "source_caption", "text_original": "而是透明度责任和", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 871.966, "end": 873.766, "text": "可驗證性之間的平衡", "source": "source_caption", "text_original": "可验证性之间的平衡", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 874.033, "end": 876.6, "text": "Muse Glimmer的價值之一就是讓", "source": "source_caption", "text_original": "Muse glimmer的价值之一就是让", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 876.6, "end": 879.1, "text": "更多研究者能直接測試模型", "source": "source_caption", "text_original": "更多研究者能直接测试模型", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 879.766, "end": 881.966, "text": "而不是只能相信廠商聲明", "source": "source_caption", "text_original": "而不是只能相信厂商声明", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 882.2, "end": 884.033, "text": "Meta設立社區基金", "source": "source_caption", "text_original": "Meta设立社区基金", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 884.366, "end": 885.666, "text": "某種意義上也是在", "source": "source_caption", "text_original": "某种意义上也是在", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 885.666, "end": 888.0, "text": "承認AI基建的外部成本", "source": "source_caption", "text_original": "承认AI基建的外部成本", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 888.166, "end": 890.066, "text": "數據中心建在某個地方", "source": "source_caption", "text_original": "数据中心建在某个地方", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 890.2, "end": 893.166, "text": "當地居民可能面對電價用水", "source": "source_caption", "text_original": "当地居民可能面对电价用水", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 893.166, "end": 895.633, "text": "噪音、土地和稅收分配問題", "source": "source_caption", "text_original": "噪音土地和税收分配问题", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 895.9, "end": 897.066, "text": "科技公司過去常把", "source": "source_caption", "text_original": "科技公司过去常把", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 897.066, "end": 898.9, "text": "這些問題當成地方溝通", "source": "source_caption", "text_original": "这些问题当成地方沟通", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 898.9, "end": 900.966, "text": "現在必須當成核心戰略", "source": "source_caption", "text_original": "现在必须当成核心战略", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 901.1, "end": 902.7, "text": "因為沒有社區許可", "source": "source_caption", "text_original": "因为没有社区许可", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 902.866, "end": 905.433, "text": "再強的模型也需要機房供電", "source": "source_caption", "text_original": "再强的模型也需要机房供电", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 905.5, "end": 907.2, "text": "把這兩條線合起來", "source": "source_caption", "text_original": "把这两条线合起来", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 907.233, "end": 909.4, "text": "Meta的新敘事就完整了", "source": "source_caption", "text_original": "Meta的新叙事就完整了", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 909.633, "end": 912.833, "text": "一邊用開放模型說AI能力要分散", "source": "source_caption", "text_original": "一边用开放模型说AI能力要分散", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 913.066, "end": 915.033, "text": "一邊用社區基金說AI", "source": "source_caption", "text_original": "一边用社区基金说AI", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 915.033, "end": 916.6, "text": "基建收益要回流", "source": "source_caption" }, { "start": 916.633, "end": 918.433, "text": "這不一定全是理想主義", "source": "source_caption", "text_original": "这不一定全是理想主义", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 918.766, "end": 921.466, "text": "但他比單純喊超級智能更聰明", "source": "source_caption", "text_original": "但他比单纯喊超级智能更聪明", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 921.6, "end": 923.666, "text": "他知道AI行業如果繼續只", "source": "source_caption", "text_original": "他知道AI行业如果继续只", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 923.666, "end": 926.033, "text": "展示估值、參數和算力", "source": "source_caption", "text_original": "展示估值参数和算力", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 926.1, "end": 928.0, "text": "公眾反彈會越來越強", "source": "source_caption", "text_original": "公众反弹会越来越强", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 928.166, "end": 931.266, "text": "未來真正有殺傷力的Muse Glimmer應用", "source": "source_caption", "text_original": "未来真正有杀伤力的缪斯Glimmer应用", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 931.766, "end": 934.0, "text": "可能不会出现在发布新闻里", "source": "source_caption" }, { "start": 934.033, "end": 936.766, "text": "而会出现在很小的工作场景里", "source": "source_caption" }, { "start": 936.766, "end": 939.366, "text": "一个设计师用他本地整理素材", "source": "source_caption" }, { "start": 939.833, "end": 942.666, "text": "一个律师事务所用它检索内部文件", "source": "source_caption" }, { "start": 942.8, "end": 945.2, "text": "一个工厂用它读设备图片", "source": "source_caption" }, { "start": 945.4, "end": 947.033, "text": "一个个人开发者用它", "source": "source_caption" }, { "start": 947.033, "end": 948.433, "text": "搭自己的桌面agent", "source": "source_caption" }, { "start": 948.8, "end": 950.2, "text": "每个场景都不震撼", "source": "source_caption" }, { "start": 950.2, "end": 951.8, "text": "但加起来就是生态", "source": "source_caption" }, { "start": 951.866, "end": 954.166, "text": "这也是开放模型最强的地方", "source": "source_caption" }, { "start": 954.6, "end": 956.5, "text": "它的价值不是一次性释放", "source": "source_caption" }, { "start": 956.5, "end": 958.766, "text": "而是被别人不断重新发明", "source": "source_caption" }, { "start": 958.966, "end": 961.1, "text": "闭源模型像一座中央电站", "source": "source_caption" }, { "start": 961.266, "end": 963.966, "text": "开放模型像一批可移动工具箱", "source": "source_caption" }, { "start": 964.233, "end": 965.9, "text": "中央电站功率更大", "source": "source_caption" }, { "start": 965.9, "end": 967.7, "text": "工具箱更容易进入角落", "source": "source_caption" }, { "start": 968.066, "end": 969.9, "text": "AI要真正改变社会", "source": "source_caption" }, { "start": 970.0, "end": 971.8, "text": "不能只靠中央电站", "source": "source_caption" }, { "start": 971.866, "end": 973.2, "text": "也需要工具箱", "source": "source_caption" }, { "start": 973.366, "end": 976.066, "text": "所以Muse Glimmer不是Meta对闭源", "source": "source_caption", "text_original": "所以musglmmer不是Meta对闭源", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 976.066, "end": 977.833, "text": "模型的一次正面冲锋", "source": "source_caption" }, { "start": 977.9, "end": 979.566, "text": "而是一次侧翼包抄", "source": "source_caption" }, { "start": 979.866, "end": 982.466, "text": "它不说自己在所有能力上第一", "source": "source_caption" }, { "start": 982.866, "end": 984.633, "text": "他说自己更容易被拿走", "source": "source_caption" }, { "start": 984.7, "end": 986.033, "text": "更容易本地跑", "source": "source_caption" }, { "start": 986.1, "end": 987.833, "text": "更容易进入普通设备", "source": "source_caption" }, { "start": 987.9, "end": 989.5, "text": "这个策略如果成功", "source": "source_caption" }, { "start": 990.1, "end": 992.266, "text": "Meta会在AI入口战争里拿到", "source": "source_caption" }, { "start": 992.266, "end": 994.066, "text": "一个很难被关闭的位置", "source": "source_caption" }, { "start": 994.266, "end": 996.0, "text": "Execute Torch的存在", "source": "source_caption", "text_original": "execute torch的存在", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 996.4, "end": 999.566, "text": "说明Meta对端侧部署有长期准备", "source": "source_caption" }, { "start": 999.8, "end": 1004.633, "text": "端侧AI不是把云端模型硬塞进手机", "source": "source_caption" }, { "start": 1004.833, "end": 1006.9, "text": "而是重新考虑模型大小", "source": "source_caption" }, { "start": 1006.9, "end": 1011.4, "text": "延迟内存占用电池和隐私。Muse Glimmer", "source": "source_caption", "text_original": "延迟内存占用电池和隐私MUS glimmer", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1012.066, "end": 1014.8, "text": "如果能和端侧框架形成完整链路", "source": "source_caption" }, { "start": 1014.866, "end": 1017.633, "text": "就会成为Meta连接硬件厂商的桥", "source": "source_caption" }, { "start": 1017.833, "end": 1021.666, "text": "以后手机眼镜耳机笔记本桌面设备", "source": "source_caption" }, { "start": 1021.7, "end": 1023.866, "text": "都可能需要一个能本地", "source": "source_caption" }, { "start": 1023.866, "end": 1026.166, "text": "处理多模态任务的模型底座", "source": "source_caption" }, { "start": 1026.266, "end": 1028.166, "text": "这对开发者很有吸引力", "source": "source_caption" }, { "start": 1028.433, "end": 1031.4, "text": "闭源API最舒服的地方是省事", "source": "source_caption" }, { "start": 1031.8, "end": 1033.666, "text": "最痛的地方是不可控", "source": "source_caption" }, { "start": 1033.8, "end": 1037.033, "text": "接口变更价格变更限流地区", "source": "source_caption" }, { "start": 1037.033, "end": 1040.0, "text": "限制内容政策都可能影响产品", "source": "source_caption" }, { "start": 1040.233, "end": 1042.433, "text": "开放模型麻烦一点但可控", "source": "source_caption" }, { "start": 1042.466, "end": 1044.633, "text": "对长期做产品的人来说", "source": "source_caption" }, { "start": 1044.766, "end": 1047.566, "text": "可控往往比短期省事更重要", "source": "source_caption" }, { "start": 1047.7, "end": 1049.9, "text": "Muse Glimmer如果性能足够稳定", "source": "source_caption", "text_original": "musglimer如果性能足够稳定", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1050.0, "end": 1051.7, "text": "就会被放进很多不愿", "source": "source_caption" }, { "start": 1051.7, "end": 1053.833, "text": "完全依赖云端的产品里", "source": "source_caption" }, { "start": 1054.033, "end": 1055.466, "text": "企业采用AI时", "source": "source_caption" }, { "start": 1055.633, "end": 1057.433, "text": "常常不是被能力卡住", "source": "source_caption" }, { "start": 1057.5, "end": 1059.1, "text": "而是被审批卡住", "source": "source_caption" }, { "start": 1059.266, "end": 1060.866, "text": "法务问数据去哪", "source": "source_caption" }, { "start": 1061.0, "end": 1063.066, "text": "安全团队问权限怎么管", "source": "source_caption" }, { "start": 1063.433, "end": 1065.366, "text": "财务问每月调用成本", "source": "source_caption" }, { "start": 1065.566, "end": 1068.033, "text": "业务部门问延迟能不能接受", "source": "source_caption" }, { "start": 1068.2, "end": 1069.5, "text": "开放本地模型可以", "source": "source_caption" }, { "start": 1069.5, "end": 1071.033, "text": "同时回答几个问题", "source": "source_caption" }, { "start": 1071.2, "end": 1072.8, "text": "数据可以留在内网", "source": "source_caption" }, { "start": 1072.866, "end": 1074.1, "text": "成本可以固定", "source": "source_caption" }, { "start": 1074.266, "end": 1075.566, "text": "延迟可以降低", "source": "source_caption" }, { "start": 1075.633, "end": 1077.2, "text": "权限可以自己定义", "source": "source_caption" }, { "start": 1077.4, "end": 1078.633, "text": "它不是完美答案", "source": "source_caption" }, { "start": 1078.633, "end": 1080.233, "text": "但比纯云端更容易", "source": "source_caption" }, { "start": 1080.233, "end": 1081.9, "text": "通过某些企业流程", "source": "source_caption" }, { "start": 1082.3, "end": 1083.6, "text": "Meta还可以借开放", "source": "source_caption" }, { "start": 1083.6, "end": 1085.566, "text": "路线给监管传递信号", "source": "source_caption" }, { "start": 1085.8, "end": 1087.466, "text": "闭源超级模型越强", "source": "source_caption" }, { "start": 1087.466, "end": 1090.466, "text": "监管越担心权力集中和黑箱决策", "source": "source_caption" }, { "start": 1090.766, "end": 1092.866, "text": "开放模型让更多", "source": "source_caption" }, { "start": 1092.866, "end": 1095.266, "text": "研究者能审计能力和风险", "source": "source_caption" }, { "start": 1095.966, "end": 1097.433, "text": "至少表面上更符合", "source": "source_caption" }, { "start": 1097.433, "end": 1099.2, "text": "广泛分发的公共叙事", "source": "source_caption" }, { "start": 1099.366, "end": 1101.6, "text": "对一家掌握社交平台和", "source": "source_caption" }, { "start": 1101.6, "end": 1103.4, "text": "广告系统的公司来说", "source": "source_caption" }, { "start": 1103.466, "end": 1105.066, "text": "这种姿态很重要", "source": "source_caption" }, { "start": 1105.266, "end": 1106.833, "text": "它需要证明自己不是", "source": "source_caption" }, { "start": 1106.833, "end": 1109.0, "text": "把AI变成下一层垄断", "source": "source_caption" }, { "start": 1109.3, "end": 1111.5, "text": "但市场不会因为姿态就买单", "source": "source_caption" }, { "start": 1111.666, "end": 1113.6, "text": "开发者最后只看三件事", "source": "source_caption" }, { "start": 1113.8, "end": 1114.8, "text": "够不够好", "source": "source_caption" }, { "start": 1114.8, "end": 1115.866, "text": "跑不跑得动", "source": "source_caption" }, { "start": 1115.866, "end": 1117.366, "text": "出问题能不能修", "source": "source_caption" }, { "start": 1117.6, "end": 1120.8, "text": "Muse Glimmer需要在真实任务里证明自己", "source": "source_caption", "text_original": "musglimer需要在真实任务里证明自己", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1121.266, "end": 1124.0, "text": "比如本地图片理解、文档问答", "source": "source_caption", "text_original": "比如本地图片理解文档问答", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1124.066, "end": 1127.233, "text": "桌面自动化工具调用失败恢复", "source": "source_caption" }, { "start": 1127.366, "end": 1128.9, "text": "只要这些能力能稳定", "source": "source_caption" }, { "start": 1128.9, "end": 1131.0, "text": "解决一批中等难度问题", "source": "source_caption" }, { "start": 1131.066, "end": 1133.233, "text": "30B就不再是小模型", "source": "source_caption" }, { "start": 1133.266, "end": 1135.833, "text": "而是刚好够用的生产力模型", "source": "source_caption" }, { "start": 1136.066, "end": 1138.633, "text": "AI行业正在出现一种新分工", "source": "source_caption" }, { "start": 1138.766, "end": 1140.9, "text": "最强闭源模型负责天花板", "source": "source_caption" }, { "start": 1141.233, "end": 1143.666, "text": "开放模型负责地板和中间层", "source": "source_caption" }, { "start": 1143.833, "end": 1146.033, "text": "天花板决定行业想象力", "source": "source_caption" }, { "start": 1146.2, "end": 1148.0, "text": "地板决定普及速度", "source": "source_caption" }, { "start": 1148.0, "end": 1149.066, "text": "没有天花板", "source": "source_caption" }, { "start": 1149.633, "end": 1151.066, "text": "AI缺少突破", "source": "source_caption" }, { "start": 1151.266, "end": 1152.166, "text": "没有地板", "source": "source_caption" }, { "start": 1152.4, "end": 1155.233, "text": "AI只会变成少数公司的高价服务", "source": "source_caption" }, { "start": 1155.366, "end": 1158.266, "text": "Muse Glimmer的意义就在地板和中间层", "source": "source_caption", "text_original": "musglimer的意义就在地板和中间层", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1158.766, "end": 1160.666, "text": "他让更多人能踩上去", "source": "source_caption" }, { "start": 1160.833, "end": 1164.666, "text": "扎克伯格讲发明创造而不是自动化", "source": "source_caption" }, { "start": 1164.833, "end": 1167.7, "text": "其实是在修复AI的公共形象", "source": "source_caption" }, { "start": 1167.966, "end": 1169.066, "text": "过去两年", "source": "source_caption" }, { "start": 1169.6, "end": 1173.066, "text": "很多人听到AI就想到裁员替代", "source": "source_caption" }, { "start": 1173.8, "end": 1177.2, "text": "数据中心耗电和大公司集中全力", "source": "source_caption" }, { "start": 1177.4, "end": 1179.166, "text": "Meta要把话题拉回", "source": "source_caption" }, { "start": 1179.166, "end": 1180.966, "text": "创造力和个人赋能", "source": "source_caption" }, { "start": 1181.1, "end": 1182.966, "text": "Muse Glimmer适合这个叙事", "source": "source_caption" }, { "start": 1183.4, "end": 1184.866, "text": "因为它看起来不像一个", "source": "source_caption" }, { "start": 1184.866, "end": 1187.066, "text": "只服务企业裁员的工具", "source": "source_caption" }, { "start": 1187.166, "end": 1188.266, "text": "而像一个可以被", "source": "source_caption" }, { "start": 1188.266, "end": 1190.2, "text": "个人拿走的创造工具", "source": "source_caption" }, { "start": 1190.633, "end": 1193.266, "text": "当然这种叙事也要接受现实检验", "source": "source_caption" }, { "start": 1193.433, "end": 1195.2, "text": "如果开放模型最后主要", "source": "source_caption" }, { "start": 1195.2, "end": 1197.233, "text": "被大公司拿去降低成本", "source": "source_caption" }, { "start": 1197.4, "end": 1199.1, "text": "普通人没有明显受益", "source": "source_caption" }, { "start": 1199.2, "end": 1201.666, "text": "那赋能就会变成漂亮话", "source": "source_caption" }, { "start": 1201.7, "end": 1203.433, "text": "真正能证明Meta的", "source": "source_caption" }, { "start": 1203.433, "end": 1204.966, "text": "是它有没有让个人", "source": "source_caption" }, { "start": 1204.966, "end": 1207.466, "text": "開發者、小企業、學校", "source": "source_caption", "text_original": "开发者小企业学校", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1207.866, "end": 1209.633, "text": "研究者用更低成本", "source": "source_caption" }, { "start": 1209.633, "end": 1211.633, "text": "做出过去做不了的东西", "source": "source_caption" }, { "start": 1211.766, "end": 1214.366, "text": "开源的价值不能停在下载量", "source": "source_caption" }, { "start": 1214.866, "end": 1217.466, "text": "要落到新作品和新产品上", "source": "source_caption" }, { "start": 1217.566, "end": 1218.5, "text": "未来一年", "source": "source_caption" }, { "start": 1218.8, "end": 1220.366, "text": "Muse Glimmer 最值得觀察", "source": "source_caption", "text_original": "musicglimmer最值得观察", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1220.366, "end": 1221.666, "text": "的不是发布热度", "source": "source_caption" }, { "start": 1221.866, "end": 1223.6, "text": "而是二次开发热度", "source": "source_caption" }, { "start": 1223.766, "end": 1225.6, "text": "有没有高质量微调版本", "source": "source_caption" }, { "start": 1226.066, "end": 1229.2, "text": "有沒有中文代碼、醫療、法律", "source": "source_caption", "text_original": "有没有中文代码医疗法律", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1229.233, "end": 1231.666, "text": "教育设计方向的社区版本", "source": "source_caption" }, { "start": 1231.766, "end": 1234.366, "text": "有沒有人把它做成一鍵本地 Agent", "source": "source_caption", "text_original": "有没有人把它做成一键本地agent", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1234.7, "end": 1236.6, "text": "有没有硬件厂商围绕它", "source": "source_caption" }, { "start": 1236.6, "end": 1238.5, "text": "优化驱动和推理速度", "source": "source_caption" }, { "start": 1238.8, "end": 1241.7, "text": "这些才是开放模型真正的生命迹象", "source": "source_caption" }, { "start": 1241.9, "end": 1243.7, "text": "如果这些生态长出来", "source": "source_caption" }, { "start": 1244.0, "end": 1246.6, "text": "Meta就会完成一次很聪明的布局", "source": "source_caption" }, { "start": 1246.8, "end": 1249.433, "text": "它不需要每个用户直接打开Meta产品", "source": "source_caption" }, { "start": 1249.766, "end": 1252.366, "text": "也不需要每次推理都走Meta服务器", "source": "source_caption" }, { "start": 1252.433, "end": 1254.566, "text": "只要大量AI工具底层", "source": "source_caption" }, { "start": 1254.566, "end": 1256.8, "text": "运行的是Meta开放模型", "source": "source_caption" }, { "start": 1257.0, "end": 1259.433, "text": "它就会成为看不见的基础设施", "source": "source_caption" }, { "start": 1259.8, "end": 1261.833, "text": "基础设施的力量往往比", "source": "source_caption" }, { "start": 1261.833, "end": 1263.7, "text": "一个热门应用更持久", "source": "source_caption" }, { "start": 1263.966, "end": 1265.233, "text": "Meta的优势是", "source": "source_caption" }, { "start": 1265.266, "end": 1268.2, "text": "它不急着把模型本身变成利润中心", "source": "source_caption" }, { "start": 1268.5, "end": 1270.6, "text": "它可以允许模型价格下降", "source": "source_caption" }, { "start": 1270.766, "end": 1272.766, "text": "因为它真正赚钱的地方在", "source": "source_caption" }, { "start": 1272.766, "end": 1275.566, "text": "广告社交关系和平台分发", "source": "source_caption" }, { "start": 1275.7, "end": 1279.433, "text": "Openai必须证明订阅和API能覆盖成本", "source": "source_caption" }, { "start": 1279.766, "end": 1282.0, "text": "Anthropic必须证明企业安全", "source": "source_caption" }, { "start": 1282.0, "end": 1284.066, "text": "和高端能力能卖出价格", "source": "source_caption" }, { "start": 1284.366, "end": 1286.233, "text": "Meta则可以用开放模型", "source": "source_caption" }, { "start": 1286.233, "end": 1288.033, "text": "削弱对手收费能力", "source": "source_caption" }, { "start": 1288.166, "end": 1289.4, "text": "这种打法很狠", "source": "source_caption" }, { "start": 1289.466, "end": 1292.366, "text": "如果未来个人设备都能跑中型模型", "source": "source_caption" }, { "start": 1292.766, "end": 1294.766, "text": "AI产品形态也会变", "source": "source_caption" }, { "start": 1294.833, "end": 1296.466, "text": "很多應用不再需要把", "source": "source_caption", "text_original": "很多应用不再需要把", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1296.466, "end": 1298.3, "text": "所有請求傳到伺服器", "source": "source_caption", "text_original": "所有请求传到服务器", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1298.4, "end": 1301.0, "text": "而是把模型嵌進本地功能裡", "source": "source_caption", "text_original": "而是把模型嵌进本地功能里", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1301.033, "end": 1303.366, "text": "寫作軟體可以本地整理草稿", "source": "source_caption", "text_original": "写作软件可以本地整理草稿", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1303.566, "end": 1305.866, "text": "剪輯軟體可以本地識別素材", "source": "source_caption", "text_original": "剪辑软件可以本地识别素材", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1306.166, "end": 1308.366, "text": "瀏覽器可以本地總結網頁", "source": "source_caption", "text_original": "浏览器可以本地总结网页", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1308.433, "end": 1310.766, "text": "開發工具可以本地讀取專案", "source": "source_caption", "text_original": "开发工具可以本地读项目", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1310.966, "end": 1312.233, "text": "雲端仍然存在", "source": "source_caption", "text_original": "云端仍然存在", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1312.466, "end": 1314.833, "text": "但本地會承擔更多基礎動作", "source": "source_caption", "text_original": "但本地会承担更多基础动作", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1314.866, "end": 1317.866, "text": "所以這件事最後要落到一個判斷", "source": "source_caption", "text_original": "所以这件事最后要落到一个判断", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1318.3, "end": 1320.833, "text": "AI的未來不會只有一種入口", "source": "source_caption", "text_original": "AI的未来不会只有一种入口", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1320.966, "end": 1322.966, "text": "雲端旗艦模型會存在", "source": "source_caption", "text_original": "云端旗舰模型会存在", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1323.066, "end": 1325.2, "text": "本地開放模型也會存在", "source": "source_caption", "text_original": "本地开放模型也会存在", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1325.366, "end": 1327.066, "text": "前者負責極限能力", "source": "source_caption", "text_original": "前者负责极限能力", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1327.2, "end": 1329.433, "text": "後者負責普及和控制權", "source": "source_caption", "text_original": "后者负责普及和控制权", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1329.6, "end": 1331.233, "text": "Meta現在壓的是後者", "source": "source_caption", "text_original": "Meta现在压的是后者", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1331.6, "end": 1332.9, "text": "而且押得很聰明", "source": "source_caption", "text_original": "而且押得很聪明", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1333.066, "end": 1334.7, "text": "它不一定贏走所有收入", "source": "source_caption", "text_original": "它不一定赢走所有收入", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1334.833, "end": 1337.166, "text": "但可能贏走很多預設選擇", "source": "source_caption", "text_original": "但可能赢走很多默认选择", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1337.266, "end": 1339.233, "text": "如果把它放進更長週期", "source": "source_caption", "text_original": "如果把它放进更长周期", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1339.566, "end": 1340.5, "text": "Muse Glimmer", "source": "source_caption" }, { "start": 1340.866, "end": 1343.5, "text": "代表的是AI商品化之後的下一步", "source": "source_caption", "text_original": "代表的是AI商品化之后的下一步", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1343.7, "end": 1346.266, "text": "模型能力會越來越像基礎能力", "source": "source_caption", "text_original": "模型能力会越来越像基础能力", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1346.4, "end": 1348.7, "text": "真正差異會轉移到數據", "source": "source_caption", "text_original": "真正差异会转移到数据", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1348.833, "end": 1351.633, "text": "場景部署體驗和生態", "source": "source_caption", "text_original": "场景部署体验和生态", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1351.866, "end": 1354.433, "text": "Meta提前把模型底層打開", "source": "source_caption", "text_original": "Meta提前把模型底座打开", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1354.5, "end": 1357.6, "text": "是在承認模型本身遲早會變便宜", "source": "source_caption", "text_original": "是在承认模型本身迟早会变便宜", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1357.9, "end": 1359.266, "text": "既然便宜不可避免", "source": "source_caption" }, { "start": 1359.266, "end": 1361.0, "text": "不如主動讓它變便宜", "source": "source_caption", "text_original": "不如主动让它变便宜", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1361.1, "end": 1363.366, "text": "再去控制更大的生態入口", "source": "source_caption", "text_original": "再去控制更大的生态入口", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1363.366, "end": 1366.033, "text": "這對閉源公司是一種壓力測試", "source": "source_caption", "text_original": "这对闭源公司是一种压力测试", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1366.266, "end": 1368.7, "text": "閉源模型如果只是略強一點", "source": "source_caption", "text_original": "闭源模型如果只是略强一点", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1368.766, "end": 1369.966, "text": "卻貴很多", "source": "source_caption", "text_original": "却贵很多", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1370.2, "end": 1372.166, "text": "很多企業會改用開放模型", "source": "source_caption", "text_original": "很多企业会改用开放模型", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1372.5, "end": 1374.1, "text": "閉源模型必須在關鍵", "source": "source_caption", "text_original": "闭源模型必须在关键", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1374.1, "end": 1375.9, "text": "任務上強到無法替代", "source": "source_caption", "text_original": "任务上强到无法替代", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1376.166, "end": 1379.466, "text": "或者在產品體驗上省心得足夠明顯", "source": "source_caption", "text_original": "或者在产品体验上省心得足够明显", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1379.7, "end": 1382.266, "text": "否則開源模型每進步一步", "source": "source_caption", "text_original": "否则开源模型每进步一步", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1382.4, "end": 1385.0, "text": "閉源模型的中低端收入就少一塊", "source": "source_caption", "text_original": "闭源模型的中低端收入就少一块", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1385.233, "end": 1387.633, "text": "Muse Glimmer也會推動AI Agent", "source": "source_caption", "text_original": "Muse Glimmer也会推动AI agent", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1388.066, "end": 1389.966, "text": "從玩具走向基礎設施", "source": "source_caption", "text_original": "从玩具走向基础设施", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1390.2, "end": 1393.233, "text": "以前很多agent演示看起来很聪明", "source": "source_caption" }, { "start": 1393.433, "end": 1396.1, "text": "但实际运行成本高延迟长", "source": "source_caption" }, { "start": 1396.166, "end": 1397.4, "text": "隐私风险大", "source": "source_caption" }, { "start": 1397.466, "end": 1399.9, "text": "把一部分能力放到本地之后", "source": "source_caption" }, { "start": 1400.066, "end": 1402.633, "text": "agent可以更安静的处理日常任务", "source": "source_caption" }, { "start": 1402.8, "end": 1404.4, "text": "它不需要每次都上云", "source": "source_caption" }, { "start": 1404.6, "end": 1406.366, "text": "不需要每一步都花钱", "source": "source_caption" }, { "start": 1406.666, "end": 1409.266, "text": "也不需要把所有私人资料交出去", "source": "source_caption" }, { "start": 1409.3, "end": 1410.866, "text": "這就是它的賣點", "source": "source_caption", "text_original": "这就是它的流量点", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1411.3, "end": 1413.233, "text": "Meta不是在送一个模型", "source": "source_caption" }, { "start": 1413.466, "end": 1414.9, "text": "而是在告诉市场", "source": "source_caption" }, { "start": 1415.4, "end": 1418.166, "text": "AI的权力不一定只能往云端集中", "source": "source_caption" }, { "start": 1418.3, "end": 1420.233, "text": "普通设备也可以有智能", "source": "source_caption" }, { "start": 1420.366, "end": 1422.066, "text": "小团队也可以有底座", "source": "source_caption" }, { "start": 1422.4, "end": 1424.066, "text": "开发者也可以有退路", "source": "source_caption" }, { "start": 1424.1, "end": 1425.633, "text": "这个故事比单纯", "source": "source_caption" }, { "start": 1425.633, "end": 1427.6, "text": "参数发布更有传播性", "source": "source_caption" }, { "start": 1427.666, "end": 1428.7, "text": "因为它集中了", "source": "source_caption" }, { "start": 1428.7, "end": 1430.9, "text": "很多人对AI垄断的焦虑", "source": "source_caption" }, { "start": 1431.0, "end": 1434.033, "text": "所以缪斯glimmer值得被单独拿出来讲", "source": "source_caption" }, { "start": 1434.233, "end": 1434.9, "text": "不是因为它", "source": "source_caption" }, { "start": 1434.9, "end": 1437.4, "text": "一夜之间改写大模型格局", "source": "source_caption" }, { "start": 1437.5, "end": 1438.8, "text": "而是因为它把一个被", "source": "source_caption" }, { "start": 1438.8, "end": 1440.633, "text": "忽略的问题摆到台面", "source": "source_caption" }, { "start": 1440.866, "end": 1443.266, "text": "AI普及到底靠更大的模型", "source": "source_caption" }, { "start": 1443.433, "end": 1445.9, "text": "还是靠更容易被部署的模型", "source": "source_caption" }, { "start": 1446.233, "end": 1448.166, "text": "亿元巨头会继续冲天花板", "source": "source_caption" }, { "start": 1448.166, "end": 1450.466, "text": "但真正改变日常工作的往往", "source": "source_caption" }, { "start": 1450.466, "end": 1452.833, "text": "是那些能被普通人装进电脑", "source": "source_caption" }, { "start": 1452.9, "end": 1453.866, "text": "切入流程", "source": "source_caption", "text_original": "接近流程", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "start": 1453.9, "end": 1455.3, "text": "反复修改的东西", "source": "source_caption" } ], "transcript_selection": { "schema": "videonote.transcript_selection.v1", "status": "SELECTED", "requested_language": "zh", "duration": 1455.937625, "reference_path": "/Users/sagawa/AI_WORK/video_notes/url/20260812_085947__扎克伯格瘋了_Muse_Glimmer_30B_殺瘋雲端大廠_AI_霸權宣告終結/reference_subtitle.vtt", "reference_kind": "manual", "reference_language": "zh", "selected_source": "source_with_whisper_gap_repair", "decision_reason": "trusted_source_selected_after_timestamp_alignment", "whisper_rejected": false, "source_profile": { "segment_count": 643, "multilingual": false, "first_start": 0.166, "final_end": 1455.8, "duration_coverage": 0.999905, "maximum_gap_seconds": 0.734, "inverted_count": 0, "language": { "requested": "zh", "matched": true, "latin": 739, "han": 6060, "kana": 0, "hangul": 0 }, "hallucination_region_count": 0, "hallucination_regions": [], "usable_full": true, "usable_partial": true }, "whisper_profile": { "segment_count": 594, "multilingual": false, "first_start": 0.0, "final_end": 1455.625, "duration_coverage": 0.999785, "maximum_gap_seconds": 0.52, "inverted_count": 0, "language": { "requested": "zh", "matched": true, "latin": 776, "han": 6055, "kana": 0, "hangul": 0 }, "hallucination_region_count": 0, "hallucination_regions": [], "usable_full": true, "usable_partial": true }, "comparison": { "window_seconds": 30.0, "compared_window_count": 49, "median_score": 0.9609, "divergent_window_count": 0, "divergent_ratio": 0.0, "divergent_windows": [] }, "timestamp_alignment": { "applied": true, "estimated_offset_seconds": -0.5, "matched_segment_count": 506, "match_scores": [ 0.8333, 0.6184, 0.7, 0.7273, 0.7333, 0.8278, 0.75, 0.9317, 0.6419, 0.6906, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.7512, 0.8421, 0.9499, 0.8435, 1.0, 0.7991, 0.6667, 1.0, 1.0, 1.0, 0.7364, 0.6371, 1.0, 0.6963, 1.0, 1.0, 0.6983, 0.6284, 0.9651, 0.753, 0.6875, 0.6087, 0.72, 1.0, 0.6667, 0.6667, 1.0, 0.8917, 0.6897, 0.6429, 1.0, 1.0, 0.6667, 0.6667, 0.6923, 0.64, 1.0, 0.8118, 0.7125, 0.8571, 1.0, 1.0, 1.0, 0.6287, 0.6242, 1.0, 0.7778, 1.0, 1.0, 0.9275, 0.8889, 0.7922, 0.8889, 0.8889, 0.9091, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.719, 0.6092, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.9429, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.6833, 0.6833, 0.6493, 1.0, 1.0, 1.0, 0.6848, 0.6439, 1.0, 0.624, 0.7066, 0.6705, 1.0, 0.6554, 0.6743, 1.0, 0.6667, 0.6667, 1.0, 1.0, 1.0, 0.8462, 0.9506, 0.9231, 0.9306, 0.7362, 1.0, 1.0, 1.0, 1.0, 0.6897, 0.6429, 1.0, 0.7312, 0.8928, 0.9009, 1.0, 1.0, 1.0, 0.7273, 0.7273, 1.0, 0.8, 1.0, 0.6, 1.0, 1.0, 1.0, 1.0, 0.6493, 0.6573, 0.7163, 1.0, 1.0, 0.9297, 1.0, 0.8696, 0.6667, 1.0, 0.875, 0.8, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.9506, 0.9565, 1.0, 1.0, 0.9412, 0.7619, 1.0, 0.9167, 0.7143, 0.625, 0.7235, 0.7778, 1.0, 1.0, 0.6404, 0.7047, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.6833, 0.6493, 1.0, 0.7394, 1.0, 1.0, 1.0, 1.0, 0.9299, 1.0, 0.6554, 0.9086, 1.0, 0.8961, 0.9072, 0.7778, 0.7875, 0.7675, 1.0, 1.0, 1.0, 1.0, 1.0, 0.9197, 1.0, 0.8889, 1.0, 1.0, 0.9204, 1.0, 1.0, 1.0, 0.7778, 0.8322, 0.8, 0.75, 0.866, 1.0, 1.0, 0.8687, 1.0, 0.9, 0.9, 0.6667, 0.6667, 1.0, 1.0, 1.0, 0.75, 0.7778, 1.0, 0.6983, 0.6512, 0.7625, 0.8333, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.6154, 0.7143, 1.0, 0.6667, 1.0, 0.6368, 0.6268, 1.0, 1.0, 0.884, 1.0, 1.0, 1.0, 0.8182, 0.9524, 0.9677, 1.0, 0.9655, 0.6923, 0.64, 1.0, 1.0, 1.0, 0.6154, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.7143, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.9091, 1.0, 1.0, 1.0, 1.0, 1.0, 0.8889, 1.0, 1.0, 0.9368, 0.8462, 0.9166, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.9, 1.0, 1.0, 1.0, 1.0, 0.8, 0.8, 1.0, 1.0, 1.0, 1.0, 1.0, 0.64, 0.6923, 0.8825, 0.8889, 1.0, 1.0, 0.6006, 0.7267, 1.0, 0.6928, 0.6539, 0.6154, 0.7143, 1.0, 1.0, 0.9167, 1.0, 1.0, 0.7067, 0.9072, 0.8954, 1.0, 1.0, 0.8533, 0.7143, 1.0, 0.6185, 0.6923, 0.6923, 0.8182, 0.7273, 0.7273, 0.8462, 0.875, 0.9, 0.7681, 0.875, 1.0, 0.7854, 0.8571, 0.875, 1.0, 1.0, 1.0, 1.0, 1.0, 0.9186, 0.9091, 1.0, 1.0, 1.0, 1.0, 0.7677, 1.0, 1.0, 1.0, 0.6235, 0.7072, 1.0, 0.8488, 1.0, 0.8, 0.7, 0.8421, 1.0, 1.0, 0.9086, 0.875, 0.9091, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.6667, 0.6667, 1.0, 1.0, 1.0, 1.0, 1.0, 0.6667, 0.6667, 0.6613, 0.6699, 0.7826, 0.7273, 0.7939, 0.64, 0.6923, 0.6897, 0.6429, 1.0, 0.6512, 0.6816, 1.0, 1.0, 1.0, 1.0, 1.0, 0.9316, 0.8, 0.6332, 0.6983, 1.0, 1.0, 1.0, 0.9091, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.927, 0.8889, 0.8, 1.0, 1.0, 0.9182, 0.647, 0.687, 0.6435, 1.0, 0.6667, 0.6667, 0.6184, 0.7305, 1.0, 0.6667, 0.6667, 1.0, 0.8, 0.8318, 0.8356, 1.0, 1.0, 1.0, 1.0, 1.0, 0.6625, 0.875, 1.0, 0.6591, 0.674, 0.9326, 0.6512, 1.0, 1.0, 1.0, 1.0, 1.0, 0.6396, 0.6922, 1.0, 0.6816, 0.6512, 1.0, 1.0, 1.0, 0.9524, 0.9565, 1.0, 0.7356, 0.9556, 0.8, 1.0, 1.0, 1.0, 0.6667, 0.6667, 1.0, 0.9167, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.8571, 0.9, 0.9091, 1.0, 1.0, 1.0, 1.0, 0.7143, 1.0, 1.0, 1.0, 1.0, 1.0, 0.8462, 0.8333, 1.0, 1.0, 0.6667, 1.0, 1.0, 0.7143, 1.0, 1.0, 1.0, 0.8421, 1.0, 1.0, 0.9348, 0.8889, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.6087, 0.72, 0.7799, 0.8683, 0.8148, 1.0, 0.9167, 0.8318, 0.624, 1.0, 0.7143 ] }, "filled_source_profile": { "segment_count": 631, "multilingual": false, "first_start": 0.0, "final_end": 1455.3, "duration_coverage": 0.999562, "maximum_gap_seconds": 0.734, "inverted_count": 0, "language": { "requested": "zh", "matched": true, "latin": 755, "han": 6062, "kana": 0, "hangul": 0 }, "hallucination_region_count": 0, "hallucination_regions": [], "usable_full": true, "usable_partial": true }, "whisper_duration_guard": { "media_duration": 1455.938, "rejected_segment_count": 0, "rejected_segments": [] } }, "presegmentation_correction": { "applied": true, "correction_count": 209 } }