{ "text": "一个不会聊天的AI模型\n很可能威胁两家大模型巨头？\n一个完全不会写文章、\n不会陪你聊天\n甚至连一句完整对话都懒得生成的 AI\n最近竟然成了矽谷最火、\n大家抢著用的新模型！\n这个模型叫 Jev\n大家熟悉的 ChatGPT 或者\nClaude\n都在比谁更聪明、\n回答更像真人\n但 Jev 完全反过来\n它只做三件事：\n判断真假、\n做出选择、\n打个分数\n听起来是不是弱得有点可笑？\n但它上线 Vercel 平台之后\n短短 24 小时内就被接近 13% 的付\n费团队直接采用\n打破了所有新模型的历史纪录！\n而且价格便宜到夸张\n每 100 万个词元输入只要 0.04 美元\n输出甚至完全免费\n这件事真正可怕的点\n是公司突然看清了一件事：\n过去大笔大笔花在 OpenAI 和\nAnthropic 旗舰模型身上的钱\n很多根本都白花了！\n像客服要不要退款？\n这笔交易是不是诈骗？\n履历能不能过关？\n这些工作根本不需要一篇文采斐然的作文\n只需要一个极速、\n超便宜\n又能直接让程式执行的判断\n如果海量工作开始转向这种极便宜的小模型\n那大模型巨头OpenAI 和\nAnthropic 的收入还能不能暴增？\n市场给它们逼近一两兆美元的估值\n是不是踩在一个正在松动的假设上？\nJev的出现有没有可能影响你手里的辉达、\n博通跟超微？\n今天 Yuki 就带大家把这笔帐彻底算清楚！\n首先我们要搞清楚\nJev 既不是缩小版的 ChatGPT\n也不是什么打折版的 Claude\n开发它的 TypeSafe AI 把它叫\n做「系统一模型」\n这个概念来自心理学家康纳曼：\n指的是那种快速、\n凭直觉、\n几乎不用花时间思考的反应\n平常大家用的语言模型\n比较像是一位很会写报告的顾问\n你问它问题\n它一个字一个字慢慢打出来\n你的系统拿到这段话之后\n还要专门写程式去拆解文字、\n确认格式\n才能决定下一步要干嘛\n但 Jev 走的是完全不同的路\n你先把选项准备好\n直接让它做单选或复选\n比如一张客服工单进来\n系统可以直接问它：\n这位客户是不是在要求退款？\n答案只有「真」或「假」\n这张单子的紧急程度给个分数？\n这个问题该转给技术部还是客服部？\nJev 会直接给出决定和机率\n程式拿到资料就能立刻执行\n根本不用从一大堆文字里猜它到底想表达什么\n你可能会想：\n把回答限制得这么死\nAI 不就失去灵魂了吗？\n这跟传统程式写死逻辑有什么不一样？\n其实完全不同！\n传统程式需要你把每种文字情境都写清楚\n但 Jev 能看懂人类五花八门的用语\n又能直接产出结构化的结果\n普通聊天模型就像每次都要写完整报告的专家\n而 Jev 就像机场的自动分流闸门\n它不会跟你聊人生\n但可以在几百毫秒内\n决定每一位旅客该走哪条通道\n官方资料显示\nJev 完成一次判断只需要 70 到\n500 毫秒\n在官方的测试流程里\n它比传统大模型最高快了约 194 倍\n便宜了约 445 倍！\n也就是说\nJev 追求的不是讲话多漂亮\n而是让软体能用极低的成本、\n极高的频率去打考绩、\n做决定\n这也是它突然爆火的原因\n工程师不是突然觉得它比 GPT 更聪明\n而是终于拿到了一个可以完美塞进软体流程里\n的超级省钱工具\n那是不是代表它又快又便宜\n就能一路横扫市场、\n把其他模型通通淘汰掉呢？\n那可不一定！\n我们可不能直接把它吹成大模型杀手\n在一些中文测试里\nJev 在分类与判断上的准确率\n其实并没有碾压其他便宜模型\n甚至有测试发现\n只要多花一点点钱换成 MiniMax\n正确率就能高出大约 10 个百分点\n听到这里你可能会怀疑：\n既然准确率没有顶尖\n大家为什么还抢著用？\n因为 Jev 真正的卖点在于：\n当企业面对海量的简单判断时\n可以接受稍微低一点点的能力\n去换取极致的速度和极低的成本；\n真的遇到搞不定的复杂问题\n再转交给 Claude、\nGPT 或者人工来处理\n这就是它最值得投资人注意的地方\n它没有证明小模型全面赢过大模型\n却向全世界证明了：\n大量 AI 任务\n根本不需要每次都去呼叫最昂贵的旗舰模型！\n而当企业开始重新分配这些任务\n被动到蛋糕的\n很可能就是 OpenAI 和\nAnthropic 的收费模式了\n讲完工作原理\n我们直接来算企业的财务帐\n假设一家公司每天要处理 100 万张客服工单\n过去的做法\n是把整张工单丢给最强的大模型\n让它去理解意思、\n分类问题、\n评估风险\n最后再写出一段回复\n这就像客人一进餐厅\n你就让行政主厨亲自出来接客、\n点菜、\n洗菜、\n切菜\n最后再亲自端盘子\n效果当然很好\n但帐单也绝对非常惊人\n现在的新做法\n是先用 Jev 这类快速模型在前面分流\n是不是垃圾讯息？\n是不是要退款？\n风险高不高？\n这些基础问题先交给便宜模型处理\n只有遇到真正复杂、\n需要推理或者写长文的情境\n才升级给 GPT 或 Claude\n你可能会好奇：\n反正最后还是要用到 GPT 或 Claude\n这样拆开来做\n真的能省下大钱吗？\n省下的钱不是一点点\n而是会直接砍半再砍半！\n你想想看\n同一张工单依然用了 AI\n甚至使用的次数还变多了\n但真正落到高价旗舰模型手里的工作比例\n却大幅下降了！\n过去处理一个流程可能要花 100 块呼叫\n高阶模型；\n现在大部分前置判断交给便宜模型\n总成本可能被压缩到只剩原本的一小部分\n昂贵的模型\n从原本包办所有杂事的人\n变成了只有遇到难题才出场的特约专家\n对公司来说\n这当然是省钱的大好消息\n但对 OpenAI 和\nAnthropic 来说\n问题可就大了！\n它们砸了几百亿美元去训练下一代模型\n就是赌市场愿意为更高的智慧付高价\n但现在这种分流机制\n直接把这个算盘给拆了\n企业依然需要 GPT 和 Claude\n但不再愿意让它们包办所有杂事\n这会带来三个非常直接的冲击：\n首先\n模型之间的价格战会变得非常残酷\n当简单工作可以随时被便宜模型接走\n旗舰模型就再也无法只靠品牌卖高价\n它必须不断证明：\n多收的钱\n真的能换来不可替代的结果\n更可怕的是\n使用量会跟收入彻底脱钩\nAI 每天处理的任务量又翻倍了\n但这并不代表大模型公司收到的钱也翻倍了\n如果新增的呼叫全发生在低价模型上\n就跟手机行动网路一样\n大家流量用得再多\n电信公司每个月能收的月租费也不会变多\n最后\n大模型的地位可能从「平台」降级成「高阶供\n应商」\n原本所有流量都经过大模型\n规则跟定价由它说了算\n但如果路由系统成了决定「这一次该叫哪个模\n型」的大脑\n真正控制订单分配权的人\n就变成了这个分流层！\n今天用 Jev\n明天只要有更便宜的就随时换掉\n企业对单一模型根本没有忠诚度可言\n说白了\n小模型真正抢走的不是聊天的使用者\n而是 OpenAI 和\nAnthropic 原本算好的、\n那一张张可以靠旗舰模型统一收费的企业大帐单！\n问题是\nOpenAI和Anthropic早在\nJev出现之前\n就已经押注旗舰模型\n投入巨额资金买晶片、\n签算力合约\n如今\nJev代表的轻量模型与智慧路由开始重新分\n配AI任务\n这些既有的重资产投入\n究竟会变成难以回收的成本\n还是继续支撑两家公司的兆美元估值？\n我们来看看市场砸了多少钱在这两家公司身上\nOpenAI 今年完成了一轮规模空前的融资\n承诺资金约 1220 亿美元\n投后估值直接冲到约 8520 亿美元！\n根据公开报导\n它最新的年化收入速度大约是 400 亿美元\nAnthropic 更是惊人\n今年 5 月完成 650 亿美元融资\n估值被推到约 9650 亿美元\n到了 7 月底\n多家媒体报导它的年化收入速度已经超过\n650 亿美元\n最近甚至有消息说这个数字正在冲向\n1000 亿美元！\n而且它正在筹备上市\n市场传出最高可能募资 1000 亿美元\n估值甚至叫价到接近 2兆美元！\n看到这些数字\n你可能会想：\n既然收入成长得这么猛\n卖算力又供不应求\n这种几兆美元的估值不是很合理吗？\n但大家可千万别被这些营收数字给骗了！\n营收成长\n可完全不等于股东最后能拿进口袋的真金白银啊！\n大模型这门生意最让人头痛的地方\n就是它每多服务一个客户\n背后都要砸更多 GPU、\n建更多资料中心、\n吃更多电\n还要花天价去训练下一代模型\n像 Anthropic 今年 2 月的融资里\n微软和辉达投进来的钱\n其实直接跟 Anthropic 向微软购\n买大约 300 亿美元算力的承诺绑在一起\n这个循环大家一定要看清楚：\n云端与晶片巨头投资模型公司\n模型公司再把这些钱拿回头去买晶片和算力\n营收看起来飙得很快\n但算力开销飙得更快！\n而 Jev 的出现\n恰好刺中了这个估值模式最脆弱的神经\n市场之所以敢给出几万亿美元的估值\n背后假设的前提不只是「大家会用 AI」\n而是「大家会持续花高价买旗舰模型的算力」\n那如果未来的架构变成了「便宜模型处理九成判断\n昂贵模型只处理一成难题」\n这套原本算得好好的收费模型\n不就直接塌掉了一半吗？\n这就让大模型公司直接撞上两道难以跨越的财\n务大墙：\n第一道墙\n是你的收入成长速度\n到底能不能跑赢资本开支？\n如果公司营收成长了 50%\n但为了支持这些营收所砸进去的算力成本却成\n长了 70%\n那规模越大\n离真正赚钱反而越远！\n第二道墙\n是呼叫量暴增之后\n每一次呼叫到底还值多少钱？\n如果大家用得越来越多\n但因为竞争跟分流\n单价跟毛利率持续下滑\n那看似漂亮的营收数字\n背后的含金量其实正在快速下降\n所以\n要判断两兆美元估值到底是不是泡沫\n根本不用去猜 AI 会不会被淘汰\n只要紧盯两组对比：\n一边看「收入增速」能不能跑赢「资本投入增速」；\n另一边看「呼叫量增速」能不能带动「毛利率\n改善」\n如果资本开支永远比收入跑得快\n单价又被小模型压著打\n那么估值倍数就势必要被市场重新血洗\n听到这里\n你心里可能会泛起另一个最大的疑虑：\n如果连 OpenAI 和\nAnthropic 的定价权都被削弱了\n那是不是代表辉达、\n博通跟整个算力供应链\n也要跟著一起倒楣了？\n答案可能正好相反！\n不过\n如果你看到小模型更快、\n更便宜\n就直接认为全球需要的晶片和算力会跟著减少\n那可能又猜错了\n因为科技史上曾经反复出现一种看似矛盾的现象：\n一项资源用得越有效率\n最后的总消耗量反而越大\n更巧的是\nJev这个名字本身就在提示答案\n它的命名灵感\n正是来自经济学上非常著名的「杰文斯悖论」\n（Jevons Paradox）\n19世纪\n英国经济学家威廉・史丹利・杰文斯发现\n蒸汽机变得越来越省煤之后\n英国消耗的煤炭总量不但没有下降\n反而快速增长\n原因并不复杂：\n效率提升降低了使用成本\n原本不划算的应用开始大量出现\n新增需求带来的消耗\n最终超过每一次使用节省下来的资源\n今天的AI\n很可能正在重演同一套剧情\n过去呼叫一次高阶模型又慢又贵\n企业只敢在少数高价值的地方省著用\n但如果现在一次判断只要几百毫秒\n便宜到可以忽略不计\nAI 就会从「偶尔请教的专家」\n变成嵌入在所有软体流程里的「基础零件」\n你可能会怀疑：\n大家都在用这种超省算力的小模型\n晶片的需求不就会大幅减少了吗？\n其实事情完全不是你想的那样！\n你想想看\n以前一张客服工单只呼叫一次 AI\n未来可能从判断语言、\n评估意图、\n分析风险到决定部门\n连续呼叫五次甚至十次！\n每一笔付款、\n每一封邮件、\n每一次广告投放\n背后都多了好几次机器自动判断\n单次价格是下降了\n但总呼叫次数却暴增了几十倍甚至上百倍！\n从这个角度来看\n小模型的崛起\n对大模型公司的「软体高估值」确实是巨大的压力\n但对「算力硬体」来说\n反而像是踩下了一脚超大力的油门！\n辉达卖的是运算平台\n博通做的是客制化晶片与高速网路\n超微也在抢推理市场\n它们根本不在乎每次呼叫是谁在收钱\n它们只在乎全球最终到底产生了多少次运算、\n需要多少晶片跟记忆体！\n只要总需求的爆发速度\n快过单次成本下降的速度\n硬体算力的需求就会继续疯狂扩张\n所以\n这场变局并不是简单的「小模型打败大模型」\n而是整个 AI 产业利润的重新分配：\n大模型公司失去了低价值任务的垄断权；\n路由平台拿到了分单的控制权；\n专用小模型承接了海量的日常判断；\n而晶片与云端巨头\n则坐收整体运算量暴增的红利！\n当然\n投资硬体也不能盲目乐观\n如果小模型省算力的速度\n真的快过新应用产生的速度\n硬体需求依然会面临调整\n但总结来说\nJev 不会在一夜之间杀死 OpenAI\n或 Anthropic\n旗舰模型在复杂程式开发、\n科学研究跟长链推理上\n依然有著不可替代的霸主地位\n它真正刺破的\n是兆美元估值背后最脆弱的那句神话：\n以为 AI 用得越多\n钱就理所当然会源源不绝地流进最昂贵的模型\n口袋里\n理解了这个逻辑\n我们在看未来的投资地图时\n方向就会清晰得多\n如果 Anthropic 未来真的以接近\n两兆美元的估值上市\n你最担心的是它的收入成长不够快\n还是算力开销永远降不下来？\n另外\n你觉得像 Jev 这样的小模型\n最终会不会把辉达跟算力链推向下一轮更大的繁荣？\n欢迎在分享你的看法！\n按赞、\n订阅并开启小铃铛\n我是 Yuki\n我们下期见啰！", "language": "zh", "duration": 836.0, "segments": [ { "id": 0, "start": 0.0, "end": 1.766, "text": "一个不会聊天的AI模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 1, "start": 1.766, "end": 3.833, "text": "很可能威胁两家大模型巨头？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 2, "start": 3.833, "end": 5.5, "text": "一个完全不会写文章、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 3, "start": 5.5, "end": 6.866, "text": "不会陪你聊天", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 4, "start": 6.866, "end": 9.6, "text": "甚至连一句完整对话都懒得生成的 AI", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 5, "start": 9.6, "end": 11.666, "text": "最近竟然成了矽谷最火、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 6, "start": 11.666, "end": 13.4, "text": "大家抢著用的新模型！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 7, "start": 13.4, "end": 14.766, "text": "这个模型叫 Jev", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 8, "start": 14.766, "end": 16.7, "text": "大家熟悉的 ChatGPT 或者", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 9, "start": 16.7, "end": 17.1, "text": "Claude", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 10, "start": 17.1, "end": 18.233, "text": "都在比谁更聪明、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 11, "start": 18.233, "end": 19.366, "text": "回答更像真人", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 12, "start": 19.366, "end": 20.966, "text": "但 Jev 完全反过来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 13, "start": 20.966, "end": 22.1, "text": "它只做三件事：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 14, "start": 22.1, "end": 22.9, "text": "判断真假、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 15, "start": 22.9, "end": 23.733, "text": "做出选择、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 16, "start": 23.733, "end": 24.4, "text": "打个分数", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 17, "start": 24.4, "end": 26.3, "text": "听起来是不是弱得有点可笑？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 18, "start": 26.3, "end": 28.5, "text": "但它上线 Vercel 平台之后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 19, "start": 28.5, "end": 32.2, "text": "短短 24 小时内就被接近 13% 的付", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 20, "start": 32.2, "end": 33.6, "text": "费团队直接采用", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 21, "start": 33.6, "end": 36.166, "text": "打破了所有新模型的历史纪录！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 22, "start": 36.166, "end": 38.0, "text": "而且价格便宜到夸张", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 23, "start": 38.0, "end": 41.1, "text": "每 100 万个词元输入只要 0.04 美元", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 24, "start": 41.1, "end": 42.866, "text": "输出甚至完全免费", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 25, "start": 42.866, "end": 44.7, "text": "这件事真正可怕的点", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 26, "start": 44.7, "end": 47.0, "text": "是公司突然看清了一件事：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 27, "start": 47.0, "end": 49.4, "text": "过去大笔大笔花在 OpenAI 和", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 28, "start": 49.4, "end": 51.366, "text": "Anthropic 旗舰模型身上的钱", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 29, "start": 51.366, "end": 53.166, "text": "很多根本都白花了！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 30, "start": 53.3, "end": 54.7, "text": "像客服要不要退款？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 31, "start": 54.7, "end": 56.533, "text": "这笔交易是不是诈骗？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 32, "start": 56.533, "end": 58.3, "text": "履历能不能过关？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 33, "start": 58.833, "end": 61.733, "text": "这些工作根本不需要一篇文采斐然的作文", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 34, "start": 61.733, "end": 63.2, "text": "只需要一个极速、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 35, "start": 63.2, "end": 63.966, "text": "超便宜", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 36, "start": 63.966, "end": 66.1, "text": "又能直接让程式执行的判断", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 37, "start": 66.1, "end": 69.533, "text": "如果海量工作开始转向这种极便宜的小模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 38, "start": 69.533, "end": 71.666, "text": "那大模型巨头OpenAI 和", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 39, "start": 71.666, "end": 73.766, "text": "Anthropic 的收入还能不能暴增？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 40, "start": 73.766, "end": 76.333, "text": "市场给它们逼近一两兆美元的估值", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 41, "start": 76.333, "end": 78.766, "text": "是不是踩在一个正在松动的假设上？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 42, "start": 78.766, "end": 81.966, "text": "Jev的出现有没有可能影响你手里的辉达、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 43, "start": 82.0, "end": 83.266, "text": "博通跟超微？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 44, "start": 83.266, "end": 86.733, "text": "今天 Yuki 就带大家把这笔帐彻底算清楚！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 45, "start": 89.933, "end": 90.933, "text": "首先我们要搞清楚", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 46, "start": 90.933, "end": 93.166, "text": "Jev 既不是缩小版的 ChatGPT", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 47, "start": 93.166, "end": 95.366, "text": "也不是什么打折版的 Claude", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 48, "start": 95.366, "end": 97.366, "text": "开发它的 TypeSafe AI 把它叫", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 49, "start": 97.366, "end": 98.466, "text": "做「系统一模型」", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 50, "start": 98.466, "end": 100.833, "text": "这个概念来自心理学家康纳曼：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 51, "start": 100.833, "end": 102.3, "text": "指的是那种快速、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 52, "start": 102.3, "end": 103.033, "text": "凭直觉、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 53, "start": 103.033, "end": 104.833, "text": "几乎不用花时间思考的反应", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 54, "start": 104.833, "end": 106.633, "text": "平常大家用的语言模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 55, "start": 106.633, "end": 108.933, "text": "比较像是一位很会写报告的顾问", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 56, "start": 108.933, "end": 109.933, "text": "你问它问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 57, "start": 110.0, "end": 111.766, "text": "它一个字一个字慢慢打出来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 58, "start": 111.766, "end": 113.633, "text": "你的系统拿到这段话之后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 59, "start": 113.633, "end": 115.933, "text": "还要专门写程式去拆解文字、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 60, "start": 115.933, "end": 116.733, "text": "确认格式", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 61, "start": 116.733, "end": 118.833, "text": "才能决定下一步要干嘛", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 62, "start": 118.833, "end": 120.966, "text": "但 Jev 走的是完全不同的路", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 63, "start": 120.966, "end": 122.9, "text": "你先把选项准备好", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 64, "start": 122.9, "end": 124.8, "text": "直接让它做单选或复选", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 65, "start": 124.8, "end": 126.733, "text": "比如一张客服工单进来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 66, "start": 126.733, "end": 128.3, "text": "系统可以直接问它：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 67, "start": 128.3, "end": 130.433, "text": "这位客户是不是在要求退款？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 68, "start": 130.433, "end": 132.366, "text": "答案只有「真」或「假」", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 69, "start": 132.766, "end": 134.866, "text": "这张单子的紧急程度给个分数？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 70, "start": 134.866, "end": 138.033, "text": "这个问题该转给技术部还是客服部？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 71, "start": 138.033, "end": 140.233, "text": "Jev 会直接给出决定和机率", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 72, "start": 140.233, "end": 142.733, "text": "程式拿到资料就能立刻执行", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 73, "start": 142.733, "end": 145.966, "text": "根本不用从一大堆文字里猜它到底想表达什么", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 74, "start": 145.966, "end": 146.966, "text": "你可能会想：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 75, "start": 146.966, "end": 148.7, "text": "把回答限制得这么死", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 76, "start": 148.7, "end": 150.533, "text": "AI 不就失去灵魂了吗？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 77, "start": 151.0, "end": 153.466, "text": "这跟传统程式写死逻辑有什么不一样？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 78, "start": 153.466, "end": 154.833, "text": "其实完全不同！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 79, "start": 155.333, "end": 158.166, "text": "传统程式需要你把每种文字情境都写清楚", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 80, "start": 158.166, "end": 160.766, "text": "但 Jev 能看懂人类五花八门的用语", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 81, "start": 160.766, "end": 162.633, "text": "又能直接产出结构化的结果", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 82, "start": 162.633, "end": 166.266, "text": "普通聊天模型就像每次都要写完整报告的专家", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 83, "start": 166.266, "end": 169.066, "text": "而 Jev 就像机场的自动分流闸门", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 84, "start": 169.066, "end": 170.466, "text": "它不会跟你聊人生", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 85, "start": 170.466, "end": 172.166, "text": "但可以在几百毫秒内", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 86, "start": 172.166, "end": 174.466, "text": "决定每一位旅客该走哪条通道", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 87, "start": 174.466, "end": 175.666, "text": "官方资料显示", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 88, "start": 175.666, "end": 177.8, "text": "Jev 完成一次判断只需要 70 到", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 89, "start": 177.833, "end": 178.466, "text": "500 毫秒", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 90, "start": 178.466, "end": 180.1, "text": "在官方的测试流程里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 91, "start": 180.1, "end": 183.166, "text": "它比传统大模型最高快了约 194 倍", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 92, "start": 183.166, "end": 185.233, "text": "便宜了约 445 倍！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 93, "start": 185.233, "end": 185.833, "text": "也就是说", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 94, "start": 185.833, "end": 187.833, "text": "Jev 追求的不是讲话多漂亮", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 95, "start": 187.833, "end": 190.066, "text": "而是让软体能用极低的成本、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 96, "start": 190.066, "end": 191.566, "text": "极高的频率去打考绩、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 97, "start": 191.566, "end": 192.3, "text": "做决定", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 98, "start": 192.3, "end": 194.066, "text": "这也是它突然爆火的原因", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 99, "start": 194.066, "end": 197.166, "text": "工程师不是突然觉得它比 GPT 更聪明", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 100, "start": 197.166, "end": 200.366, "text": "而是终于拿到了一个可以完美塞进软体流程里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 101, "start": 200.366, "end": 201.833, "text": "的超级省钱工具", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 102, "start": 201.833, "end": 204.166, "text": "那是不是代表它又快又便宜", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 103, "start": 204.166, "end": 205.6, "text": "就能一路横扫市场、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 104, "start": 205.6, "end": 207.733, "text": "把其他模型通通淘汰掉呢？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 105, "start": 207.733, "end": 208.766, "text": "那可不一定！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 106, "start": 208.766, "end": 211.166, "text": "我们可不能直接把它吹成大模型杀手", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 107, "start": 211.166, "end": 212.533, "text": "在一些中文测试里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 108, "start": 212.533, "end": 214.8, "text": "Jev 在分类与判断上的准确率", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 109, "start": 214.8, "end": 217.133, "text": "其实并没有碾压其他便宜模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 110, "start": 217.133, "end": 218.533, "text": "甚至有测试发现", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 111, "start": 218.533, "end": 220.933, "text": "只要多花一点点钱换成 MiniMax", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 112, "start": 221.0, "end": 223.266, "text": "正确率就能高出大约 10 个百分点", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 113, "start": 223.266, "end": 224.9, "text": "听到这里你可能会怀疑：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 114, "start": 224.9, "end": 226.466, "text": "既然准确率没有顶尖", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 115, "start": 226.466, "end": 227.8, "text": "大家为什么还抢著用？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 116, "start": 227.8, "end": 229.933, "text": "因为 Jev 真正的卖点在于：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 117, "start": 229.933, "end": 232.7, "text": "当企业面对海量的简单判断时", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 118, "start": 232.7, "end": 235.433, "text": "可以接受稍微低一点点的能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 119, "start": 235.433, "end": 237.833, "text": "去换取极致的速度和极低的成本；", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 120, "start": 237.833, "end": 239.966, "text": "真的遇到搞不定的复杂问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 121, "start": 239.966, "end": 241.466, "text": "再转交给 Claude、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 122, "start": 241.466, "end": 243.166, "text": "GPT 或者人工来处理", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 123, "start": 243.166, "end": 245.566, "text": "这就是它最值得投资人注意的地方", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 124, "start": 245.733, "end": 248.633, "text": "它没有证明小模型全面赢过大模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 125, "start": 248.633, "end": 249.966, "text": "却向全世界证明了：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 126, "start": 249.966, "end": 251.166, "text": "大量 AI 任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 127, "start": 251.166, "end": 254.5, "text": "根本不需要每次都去呼叫最昂贵的旗舰模型！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 128, "start": 254.5, "end": 256.833, "text": "而当企业开始重新分配这些任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 129, "start": 256.833, "end": 258.566, "text": "被动到蛋糕的", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "真正被动到蛋糕的", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 130, "start": 258.566, "end": 260.266, "text": "很可能就是 OpenAI 和", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 131, "start": 260.266, "end": 262.2, "text": "Anthropic 的收费模式了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 132, "start": 266.1, "end": 267.2, "text": "讲完工作原理", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 133, "start": 267.3, "end": 268.933, "text": "我们直接来算企业的财务帐", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 134, "start": 269.333, "end": 272.533, "text": "假设一家公司每天要处理 100 万张客服工单", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 135, "start": 272.766, "end": 273.666, "text": "过去的做法", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 136, "start": 273.666, "end": 276.366, "text": "是把整张工单丢给最强的大模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 137, "start": 276.566, "end": 277.866, "text": "让它去理解意思、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 138, "start": 277.933, "end": 278.566, "text": "分类问题、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 139, "start": 278.9, "end": 279.6, "text": "评估风险", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 140, "start": 279.833, "end": 281.5, "text": "最后再写出一段回复", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 141, "start": 281.866, "end": 283.166, "text": "这就像客人一进餐厅", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 142, "start": 283.266, "end": 285.333, "text": "你就让行政主厨亲自出来接客、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 143, "start": 285.433, "end": 285.833, "text": "点菜、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 144, "start": 285.966, "end": 286.366, "text": "洗菜、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 145, "start": 286.533, "end": 286.9, "text": "切菜", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 146, "start": 287.166, "end": 288.633, "text": "最后再亲自端盘子", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 147, "start": 288.866, "end": 289.733, "text": "效果当然很好", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 148, "start": 289.966, "end": 291.833, "text": "但帐单也绝对非常惊人", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 149, "start": 291.833, "end": 293.233, "text": "现在的新做法", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 150, "start": 293.233, "end": 296.566, "text": "是先用 Jev 这类快速模型在前面分流", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 151, "start": 296.733, "end": 298.3, "text": "是不是垃圾讯息？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 152, "start": 298.533, "end": 299.5, "text": "是不是要退款？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 153, "start": 299.766, "end": 300.633, "text": "风险高不高？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 154, "start": 300.966, "end": 303.266, "text": "这些基础问题先交给便宜模型处理", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 155, "start": 303.733, "end": 304.966, "text": "只有遇到真正复杂、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 156, "start": 304.966, "end": 306.933, "text": "需要推理或者写长文的情境", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 157, "start": 307.1, "end": 308.833, "text": "才升级给 GPT 或 Claude", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 158, "start": 308.966, "end": 309.766, "text": "你可能会好奇：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 159, "start": 310.033, "end": 312.4, "text": "反正最后还是要用到 GPT 或 Claude", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 160, "start": 312.433, "end": 313.266, "text": "这样拆开来做", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 161, "start": 313.266, "end": 314.766, "text": "真的能省下大钱吗？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 162, "start": 315.1, "end": 316.533, "text": "省下的钱不是一点点", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 163, "start": 316.633, "end": 318.533, "text": "而是会直接砍半再砍半！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 164, "start": 318.533, "end": 319.5, "text": "你想想看", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 165, "start": 319.766, "end": 321.466, "text": "同一张工单依然用了 AI", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 166, "start": 321.866, "end": 324.0, "text": "甚至使用的次数还变多了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 167, "start": 324.033, "end": 327.4, "text": "但真正落到高价旗舰模型手里的工作比例", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 168, "start": 327.533, "end": 328.7, "text": "却大幅下降了！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 169, "start": 328.966, "end": 331.233, "text": "过去处理一个流程可能要花 100 块呼叫", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 170, "start": 331.233, "end": 332.033, "text": "高阶模型；", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 171, "start": 332.333, "end": 334.633, "text": "现在大部分前置判断交给便宜模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 172, "start": 334.866, "end": 337.533, "text": "总成本可能被压缩到只剩原本的一小部分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 173, "start": 337.866, "end": 338.766, "text": "昂贵的模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 174, "start": 339.033, "end": 340.766, "text": "从原本包办所有杂事的人", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 175, "start": 341.033, "end": 344.2, "text": "变成了只有遇到难题才出场的特约专家", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 176, "start": 344.566, "end": 345.4, "text": "对公司来说", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 177, "start": 345.6, "end": 347.066, "text": "这当然是省钱的大好消息", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 178, "start": 347.066, "end": 348.366, "text": "但对 OpenAI 和", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 179, "start": 348.366, "end": 349.233, "text": "Anthropic 来说", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 180, "start": 349.433, "end": 350.633, "text": "问题可就大了！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 181, "start": 350.633, "end": 353.566, "text": "它们砸了几百亿美元去训练下一代模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 182, "start": 353.766, "end": 356.466, "text": "就是赌市场愿意为更高的智慧付高价", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 183, "start": 356.833, "end": 358.333, "text": "但现在这种分流机制", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 184, "start": 358.333, "end": 360.2, "text": "直接把这个算盘给拆了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 185, "start": 360.533, "end": 362.433, "text": "企业依然需要 GPT 和 Claude", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 186, "start": 362.433, "end": 364.933, "text": "但不再愿意让它们包办所有杂事", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 187, "start": 365.133, "end": 367.166, "text": "这会带来三个非常直接的冲击：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 188, "start": 367.533, "end": 367.933, "text": "首先", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 189, "start": 368.133, "end": 370.166, "text": "模型之间的价格战会变得非常残酷", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 190, "start": 370.433, "end": 373.066, "text": "当简单工作可以随时被便宜模型接走", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 191, "start": 373.333, "end": 376.033, "text": "旗舰模型就再也无法只靠品牌卖高价", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 192, "start": 376.033, "end": 377.566, "text": "它必须不断证明：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 193, "start": 377.766, "end": 378.4, "text": "多收的钱", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 194, "start": 378.4, "end": 380.6, "text": "真的能换来不可替代的结果", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 195, "start": 380.9, "end": 381.7, "text": "更可怕的是", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 196, "start": 382.066, "end": 384.066, "text": "使用量会跟收入彻底脱钩", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 197, "start": 384.633, "end": 386.733, "text": "AI 每天处理的任务量又翻倍了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 198, "start": 386.933, "end": 390.2, "text": "但这并不代表大模型公司收到的钱也翻倍了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 199, "start": 390.533, "end": 393.333, "text": "如果新增的呼叫全发生在低价模型上", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 200, "start": 393.433, "end": 394.966, "text": "就跟手机行动网路一样", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 201, "start": 395.2, "end": 396.333, "text": "大家流量用得再多", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 202, "start": 396.466, "end": 399.233, "text": "电信公司每个月能收的月租费也不会变多", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 203, "start": 399.6, "end": 400.033, "text": "最后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 204, "start": 400.266, "end": 403.233, "text": "大模型的地位可能从「平台」降级成「高阶供", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 205, "start": 403.233, "end": 403.6, "text": "应商」", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 206, "start": 403.6, "end": 405.766, "text": "原本所有流量都经过大模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 207, "start": 405.9, "end": 407.666, "text": "规则跟定价由它说了算", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 208, "start": 407.866, "end": 410.666, "text": "但如果路由系统成了决定「这一次该叫哪个模", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 209, "start": 410.666, "end": 411.266, "text": "型」的大脑", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 210, "start": 411.566, "end": 413.3, "text": "真正控制订单分配权的人", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 211, "start": 413.3, "end": 414.833, "text": "就变成了这个分流层！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 212, "start": 415.233, "end": 415.966, "text": "今天用 Jev", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 213, "start": 416.066, "end": 418.233, "text": "明天只要有更便宜的就随时换掉", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 214, "start": 418.533, "end": 420.966, "text": "企业对单一模型根本没有忠诚度可言", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 215, "start": 421.233, "end": 421.866, "text": "说白了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 216, "start": 422.033, "end": 424.366, "text": "小模型真正抢走的不是聊天的使用者", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 217, "start": 424.566, "end": 425.533, "text": "而是 OpenAI 和", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 218, "start": 425.533, "end": 426.866, "text": "Anthropic 原本算好的、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 219, "start": 427.166, "end": 430.566, "text": "那一张张可以靠旗舰模型统一收费的企业大帐单！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 220, "start": 430.566, "end": 431.466, "text": "问题是", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 221, "start": 431.7, "end": 433.566, "text": "OpenAI和Anthropic早在", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 222, "start": 433.566, "end": 434.566, "text": "Jev出现之前", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 223, "start": 434.833, "end": 436.566, "text": "就已经押注旗舰模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 224, "start": 436.733, "end": 438.433, "text": "投入巨额资金买晶片、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 225, "start": 438.633, "end": 439.566, "text": "签算力合约", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 226, "start": 440.0, "end": 440.4, "text": "如今", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 227, "start": 440.433, "end": 443.366, "text": "Jev代表的轻量模型与智慧路由开始重新分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 228, "start": 443.366, "end": 444.233, "text": "配AI任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 229, "start": 444.466, "end": 446.3, "text": "这些既有的重资产投入", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 230, "start": 446.466, "end": 448.333, "text": "究竟会变成难以回收的成本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 231, "start": 448.533, "end": 451.466, "text": "还是继续支撑两家公司的兆美元估值？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 232, "start": 455.333, "end": 458.866, "text": "我们来看看市场砸了多少钱在这两家公司身上", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 233, "start": 458.933, "end": 461.633, "text": "OpenAI 今年完成了一轮规模空前的融资", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 234, "start": 461.933, "end": 464.233, "text": "承诺资金约 1220 亿美元", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 235, "start": 464.5, "end": 467.5, "text": "投后估值直接冲到约 8520 亿美元！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 236, "start": 467.833, "end": 468.933, "text": "根据公开报导", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 237, "start": 469.033, "end": 472.233, "text": "它最新的年化收入速度大约是 400 亿美元", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 238, "start": 472.566, "end": 473.833, "text": "Anthropic 更是惊人", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 239, "start": 473.933, "end": 476.333, "text": "今年 5 月完成 650 亿美元融资", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 240, "start": 476.6, "end": 478.966, "text": "估值被推到约 9650 亿美元", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 241, "start": 479.366, "end": 480.1, "text": "到了 7 月底", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 242, "start": 480.233, "end": 483.133, "text": "多家媒体报导它的年化收入速度已经超过", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 243, "start": 483.133, "end": 484.333, "text": "650 亿美元", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 244, "start": 484.6, "end": 487.2, "text": "最近甚至有消息说这个数字正在冲向", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 245, "start": 487.2, "end": 488.133, "text": "1000 亿美元！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 246, "start": 488.5, "end": 490.033, "text": "而且它正在筹备上市", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 247, "start": 490.233, "end": 493.033, "text": "市场传出最高可能募资 1000 亿美元", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 248, "start": 493.366, "end": 495.733, "text": "估值甚至叫价到接近 2兆美元！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 249, "start": 496.1, "end": 497.1, "text": "看到这些数字", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 250, "start": 497.2, "end": 497.866, "text": "你可能会想：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 251, "start": 498.133, "end": 499.633, "text": "既然收入成长得这么猛", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 252, "start": 499.9, "end": 501.233, "text": "卖算力又供不应求", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 253, "start": 501.533, "end": 503.933, "text": "这种几兆美元的估值不是很合理吗？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 254, "start": 504.266, "end": 507.133, "text": "但大家可千万别被这些营收数字给骗了！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 255, "start": 507.433, "end": 508.066, "text": "营收成长", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 256, "start": 508.066, "end": 511.8, "text": "可完全不等于股东最后能拿进口袋的真金白银啊！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 257, "start": 511.833, "end": 514.3, "text": "大模型这门生意最让人头痛的地方", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 258, "start": 514.433, "end": 516.266, "text": "就是它每多服务一个客户", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 259, "start": 516.4, "end": 518.1, "text": "背后都要砸更多 GPU、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 260, "start": 518.333, "end": 519.533, "text": "建更多资料中心、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 261, "start": 519.633, "end": 520.533, "text": "吃更多电", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 262, "start": 520.666, "end": 522.833, "text": "还要花天价去训练下一代模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 263, "start": 523.2, "end": 525.2, "text": "像 Anthropic 今年 2 月的融资里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 264, "start": 525.333, "end": 526.866, "text": "微软和辉达投进来的钱", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 265, "start": 527.033, "end": 529.133, "text": "其实直接跟 Anthropic 向微软购", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 266, "start": 529.133, "end": 532.1, "text": "买大约 300 亿美元算力的承诺绑在一起", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 267, "start": 532.566, "end": 534.433, "text": "这个循环大家一定要看清楚：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 268, "start": 534.733, "end": 536.733, "text": "云端与晶片巨头投资模型公司", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 269, "start": 536.733, "end": 540.1, "text": "模型公司再把这些钱拿回头去买晶片和算力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 270, "start": 540.366, "end": 541.833, "text": "营收看起来飙得很快", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 271, "start": 542.033, "end": 543.8, "text": "但算力开销飙得更快！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 272, "start": 544.233, "end": 545.233, "text": "而 Jev 的出现", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 273, "start": 545.5, "end": 548.6, "text": "恰好刺中了这个估值模式最脆弱的神经", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 274, "start": 549.133, "end": 551.333, "text": "市场之所以敢给出几万亿美元的估值", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 275, "start": 551.566, "end": 554.0, "text": "背后假设的前提不只是「大家会用 AI」", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 276, "start": 554.4, "end": 557.5, "text": "而是「大家会持续花高价买旗舰模型的算力」", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 277, "start": 557.8, "end": 561.2, "text": "那如果未来的架构变成了「便宜模型处理九成判断", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 278, "start": 561.433, "end": 563.233, "text": "昂贵模型只处理一成难题」", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 279, "start": 563.6, "end": 565.433, "text": "这套原本算得好好的收费模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 280, "start": 565.433, "end": 566.933, "text": "不就直接塌掉了一半吗？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 281, "start": 567.033, "end": 570.433, "text": "这就让大模型公司直接撞上两道难以跨越的财", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 282, "start": 570.433, "end": 570.966, "text": "务大墙：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 283, "start": 571.233, "end": 571.966, "text": "第一道墙", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 284, "start": 572.233, "end": 573.6, "text": "是你的收入成长速度", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 285, "start": 573.766, "end": 575.4, "text": "到底能不能跑赢资本开支？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 286, "start": 575.8, "end": 577.7, "text": "如果公司营收成长了 50%", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 287, "start": 578.0, "end": 581.266, "text": "但为了支持这些营收所砸进去的算力成本却成", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 288, "start": 581.266, "end": 582.2, "text": "长了 70%", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 289, "start": 582.7, "end": 583.633, "text": "那规模越大", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 290, "start": 583.833, "end": 585.3, "text": "离真正赚钱反而越远！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 291, "start": 585.6, "end": 586.133, "text": "第二道墙", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 292, "start": 586.133, "end": 587.566, "text": "是呼叫量暴增之后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 293, "start": 587.7, "end": 589.733, "text": "每一次呼叫到底还值多少钱？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 294, "start": 590.0, "end": 591.333, "text": "如果大家用得越来越多", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 295, "start": 591.533, "end": 593.033, "text": "但因为竞争跟分流", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 296, "start": 593.033, "end": 595.266, "text": "单价跟毛利率持续下滑", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 297, "start": 595.4, "end": 597.4, "text": "那看似漂亮的营收数字", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 298, "start": 597.533, "end": 600.466, "text": "背后的含金量其实正在快速下降", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 299, "start": 600.8, "end": 601.166, "text": "所以", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 300, "start": 601.333, "end": 603.766, "text": "要判断两兆美元估值到底是不是泡沫", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 301, "start": 604.0, "end": 606.133, "text": "根本不用去猜 AI 会不会被淘汰", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 302, "start": 606.4, "end": 607.566, "text": "只要紧盯两组对比：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 303, "start": 607.9, "end": 610.733, "text": "一边看「收入增速」能不能跑赢「资本投入增速」；", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 304, "start": 610.933, "end": 613.666, "text": "另一边看「呼叫量增速」能不能带动「毛利率", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 305, "start": 613.666, "end": 614.1, "text": "改善」", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 306, "start": 614.366, "end": 616.566, "text": "如果资本开支永远比收入跑得快", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 307, "start": 616.8, "end": 618.566, "text": "单价又被小模型压著打", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 308, "start": 618.733, "end": 621.733, "text": "那么估值倍数就势必要被市场重新血洗", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 309, "start": 621.733, "end": 622.6, "text": "听到这里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 310, "start": 622.8, "end": 625.133, "text": "你心里可能会泛起另一个最大的疑虑：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 311, "start": 625.4, "end": 626.9, "text": "如果连 OpenAI 和", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 312, "start": 626.9, "end": 629.333, "text": "Anthropic 的定价权都被削弱了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 313, "start": 629.566, "end": 630.8, "text": "那是不是代表辉达、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 314, "start": 630.8, "end": 632.6, "text": "博通跟整个算力供应链", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 315, "start": 632.6, "end": 634.033, "text": "也要跟著一起倒楣了？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 316, "start": 634.5, "end": 635.7, "text": "答案可能正好相反！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 317, "start": 638.966, "end": 639.5, "text": "不过", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 318, "start": 639.6, "end": 641.2, "text": "如果你看到小模型更快、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 319, "start": 641.433, "end": 641.966, "text": "更便宜", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 320, "start": 642.2, "end": 645.6, "text": "就直接认为全球需要的晶片和算力会跟著减少", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 321, "start": 645.733, "end": 647.1, "text": "那可能又猜错了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 322, "start": 647.366, "end": 650.866, "text": "因为科技史上曾经反复出现一种看似矛盾的现象：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 323, "start": 651.1, "end": 652.766, "text": "一项资源用得越有效率", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 324, "start": 652.9, "end": 654.5, "text": "最后的总消耗量反而越大", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 325, "start": 654.833, "end": 655.366, "text": "更巧的是", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 326, "start": 655.6, "end": 657.766, "text": "Jev这个名字本身就在提示答案", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 327, "start": 657.8, "end": 658.833, "text": "它的命名灵感", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 328, "start": 658.833, "end": 662.1, "text": "正是来自经济学上非常著名的「杰文斯悖论」", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 329, "start": 662.1, "end": 663.2, "text": "（Jevons Paradox）", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 330, "start": 663.5, "end": 664.233, "text": "19世纪", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 331, "start": 664.233, "end": 666.866, "text": "英国经济学家威廉・史丹利・杰文斯发现", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 332, "start": 667.133, "end": 669.066, "text": "蒸汽机变得越来越省煤之后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 333, "start": 669.3, "end": 671.6, "text": "英国消耗的煤炭总量不但没有下降", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 334, "start": 671.766, "end": 673.266, "text": "反而快速增长", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 335, "start": 673.566, "end": 674.566, "text": "原因并不复杂：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 336, "start": 674.8, "end": 676.5, "text": "效率提升降低了使用成本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 337, "start": 676.666, "end": 678.966, "text": "原本不划算的应用开始大量出现", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 338, "start": 679.233, "end": 680.6, "text": "新增需求带来的消耗", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 339, "start": 680.6, "end": 683.366, "text": "最终超过每一次使用节省下来的资源", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 340, "start": 683.633, "end": 684.166, "text": "今天的AI", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 341, "start": 684.166, "end": 686.633, "text": "很可能正在重演同一套剧情", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 342, "start": 686.966, "end": 689.233, "text": "过去呼叫一次高阶模型又慢又贵", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 343, "start": 689.233, "end": 692.266, "text": "企业只敢在少数高价值的地方省著用", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 344, "start": 692.466, "end": 695.033, "text": "但如果现在一次判断只要几百毫秒", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 345, "start": 695.2, "end": 696.633, "text": "便宜到可以忽略不计", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 346, "start": 696.8, "end": 698.733, "text": "AI 就会从「偶尔请教的专家」", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 347, "start": 698.733, "end": 702.2, "text": "变成嵌入在所有软体流程里的「基础零件」", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 348, "start": 702.5, "end": 703.333, "text": "你可能会怀疑：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 349, "start": 703.566, "end": 705.9, "text": "大家都在用这种超省算力的小模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 350, "start": 706.166, "end": 707.966, "text": "晶片的需求不就会大幅减少了吗？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 351, "start": 708.3, "end": 710.3, "text": "其实事情完全不是你想的那样！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 352, "start": 710.633, "end": 711.266, "text": "你想想看", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 353, "start": 711.466, "end": 713.6, "text": "以前一张客服工单只呼叫一次 AI", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 354, "start": 713.9, "end": 715.333, "text": "未来可能从判断语言、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 355, "start": 715.466, "end": 716.133, "text": "评估意图、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 356, "start": 716.3, "end": 717.766, "text": "分析风险到决定部门", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 357, "start": 717.766, "end": 720.166, "text": "连续呼叫五次甚至十次！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 358, "start": 720.266, "end": 721.2, "text": "每一笔付款、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 359, "start": 721.366, "end": 722.133, "text": "每一封邮件、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 360, "start": 722.3, "end": 723.5, "text": "每一次广告投放", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 361, "start": 723.7, "end": 726.3, "text": "背后都多了好几次机器自动判断", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 362, "start": 726.633, "end": 728.1, "text": "单次价格是下降了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 363, "start": 728.2, "end": 731.666, "text": "但总呼叫次数却暴增了几十倍甚至上百倍！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 364, "start": 732.0, "end": 733.0, "text": "从这个角度来看", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 365, "start": 733.2, "end": 734.066, "text": "小模型的崛起", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 366, "start": 734.266, "end": 737.333, "text": "对大模型公司的「软体高估值」确实是巨大的压力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 367, "start": 737.666, "end": 738.966, "text": "但对「算力硬体」来说", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 368, "start": 739.166, "end": 741.433, "text": "反而像是踩下了一脚超大力的油门！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 369, "start": 741.566, "end": 743.3, "text": "辉达卖的是运算平台", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 370, "start": 743.433, "end": 745.7, "text": "博通做的是客制化晶片与高速网路", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 371, "start": 745.7, "end": 747.4, "text": "超微也在抢推理市场", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 372, "start": 747.433, "end": 750.233, "text": "它们根本不在乎每次呼叫是谁在收钱", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 373, "start": 750.233, "end": 753.433, "text": "它们只在乎全球最终到底产生了多少次运算、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 374, "start": 753.633, "end": 755.266, "text": "需要多少晶片跟记忆体！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 375, "start": 755.6, "end": 757.2, "text": "只要总需求的爆发速度", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 376, "start": 757.3, "end": 759.466, "text": "快过单次成本下降的速度", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 377, "start": 759.533, "end": 762.266, "text": "硬体算力的需求就会继续疯狂扩张", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 378, "start": 762.6, "end": 762.9, "text": "所以", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 379, "start": 762.9, "end": 766.366, "text": "这场变局并不是简单的「小模型打败大模型」", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 380, "start": 766.6, "end": 768.933, "text": "而是整个 AI 产业利润的重新分配：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 381, "start": 769.3, "end": 772.066, "text": "大模型公司失去了低价值任务的垄断权；", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 382, "start": 772.366, "end": 774.466, "text": "路由平台拿到了分单的控制权；", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 383, "start": 774.466, "end": 777.333, "text": "专用小模型承接了海量的日常判断；", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 384, "start": 777.366, "end": 778.933, "text": "而晶片与云端巨头", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 385, "start": 778.933, "end": 781.466, "text": "则坐收整体运算量暴增的红利！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 386, "start": 781.7, "end": 782.0, "text": "当然", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 387, "start": 782.0, "end": 784.0, "text": "投资硬体也不能盲目乐观", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 388, "start": 784.266, "end": 785.866, "text": "如果小模型省算力的速度", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 389, "start": 785.866, "end": 787.9, "text": "真的快过新应用产生的速度", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 390, "start": 787.966, "end": 789.866, "text": "硬体需求依然会面临调整", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 391, "start": 790.2, "end": 791.066, "text": "但总结来说", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 392, "start": 791.2, "end": 793.233, "text": "Jev 不会在一夜之间杀死 OpenAI", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 393, "start": 793.233, "end": 794.266, "text": "或 Anthropic", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 394, "start": 794.333, "end": 796.166, "text": "旗舰模型在复杂程式开发、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 395, "start": 796.166, "end": 797.866, "text": "科学研究跟长链推理上", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 396, "start": 797.866, "end": 800.1, "text": "依然有著不可替代的霸主地位", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 397, "start": 800.2, "end": 801.5, "text": "它真正刺破的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 398, "start": 801.5, "end": 804.3, "text": "是兆美元估值背后最脆弱的那句神话：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 399, "start": 804.466, "end": 805.8, "text": "以为 AI 用得越多", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 400, "start": 805.9, "end": 809.3, "text": "钱就理所当然会源源不绝地流进最昂贵的模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 401, "start": 809.3, "end": 809.866, "text": "口袋里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 402, "start": 810.133, "end": 811.166, "text": "理解了这个逻辑", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 403, "start": 811.4, "end": 813.1, "text": "我们在看未来的投资地图时", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 404, "start": 813.366, "end": 814.733, "text": "方向就会清晰得多", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 405, "start": 815.066, "end": 816.8, "text": "如果 Anthropic 未来真的以接近", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 406, "start": 816.8, "end": 818.4, "text": "两兆美元的估值上市", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 407, "start": 818.666, "end": 821.366, "text": "你最担心的是它的收入成长不够快", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 408, "start": 821.633, "end": 823.633, "text": "还是算力开销永远降不下来？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 409, "start": 823.9, "end": 824.266, "text": "另外", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 410, "start": 824.5, "end": 826.166, "text": "你觉得像 Jev 这样的小模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 411, "start": 826.5, "end": 830.033, "text": "最终会不会把辉达跟算力链推向下一轮更大的繁荣？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 412, "start": 830.033, "end": 831.666, "text": "欢迎在分享你的看法！", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 413, "start": 831.666, "end": 832.133, "text": "按赞、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 414, "start": 832.133, "end": 833.3, "text": "订阅并开启小铃铛", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 415, "start": 833.466, "end": 833.966, "text": "我是 Yuki", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 416, "start": 834.166, "end": 835.166, "text": "我们下期见啰！", "source": "video_source_subtitle", "authority": "video_source_subtitle" } ], "transcript_selection": { "schema": "videonote.transcript_selection.v1", "status": "SELECTED", "requested_language": "zh", "duration": 836.0, "reference_path": "/Users/sagawa/AI_WORK/video_notes/url/20260922_072451__Jev衝擊ChatGPT與Claude_AI用得越多反而越難賺_Anthropic兩兆估值還撐得住嗎/reference_subtitle.vtt", "reference_kind": "manual", "reference_language": "zh-Hans", "selected_source": "video_source_subtitle", "source_caption_kind": "manual", "source_caption_processing": "verbatim_direct", "youtube_overlap_clamps": [], "general_high_confidence_repairs": [], "decision_reason": "complete_source_subtitle_fast_path_no_whisper", "whisper_rejected": false, "source_profile": { "segment_count": 417, "duration_coverage": 0.916657 } }, "presegmentation_correction": { "applied": true, "correction_count": 1, "deterministic_fix_count": 0 } }