{
  "text": "谷歌还没有发布Gemini4\n但一份写着Gemini4的成绩单已经\n提前把Astra和fable按在了排行榜下面\n更奇怪的是\n真正引爆这场泄露的\n不是某份官方文件\n也不是Google高管的一句暗示\n而是Arena里一个挂着Gemini\n3.8Flash名字的匿名模型\n它顶着一张已经公开的旧身份证\n却干出了公开版本\n根本不像会干的事\n问题也就来了\n这到底是Gemini4不小心露出了底牌\n还是整个AI社区共同\n完成了一次大型误判\n先看最反常的一幕\n有人要求这个模型用原始\nSVG代码画出一台Playstation5\n正常的Flash模型强调速度\n十几秒交出一组大致像样\n的几何图形就算完成任务\n但这个模型没有马上回答\n它沉默了将近10分钟\n随后一次性吐出数千行代码\n把主机的曲线阴影光区\n细节和空间关系都做了出来\n他不像在即时续写\n更像先在后台规划执行检查\n再把成品端上来\n随后测试案例越来越夸张\n像素风宝塔宝马M4鹈鹕骑\n自行车带昼夜切换和车灯\n效果的三维场景空客H145\n直升机完整赛车游戏可交互飞行\n模拟全都开始在开发者社区流传\n这里真正值得注意的\n不是某一张图有多漂亮\n而是它能够把空间坐标\n视觉层级交互逻辑和长\n代码放在同一次生成里完成\n很多模型能写页面\n也能画简单SVG\n但一旦同时加入三维\n动画物理和复杂指令结构\n很容易散掉\n这个匿名模型最引人怀疑的地方\n正是它在复杂任务\n里表现的不像Flash\nArena出现匿名检查点\n本来不是什么新闻\n这个平台的价值就在于盲测\n用户不知道两边是谁\n只看答案投票\n实验室也能避开品牌光环\n收集真实偏好Google\nOpenai和Anthropic都可能把\n候选模型放进去测试\n真正反常的是这次\n模型没有使用陌生代号\n而是套用了已经\n发布的Gemini3.8Flash名称\n社区看到的不是\n一个身份空白的选手\n而是一个能力与\n身份证明显对不上的选手\n这给了泄露故事\n一个非常诱人的解释\nGoogle把Gemini4检查点藏在Gemini3.8Flash名下\n借Arena做公开压力测试\n这个解释很顺\n也符合大模型公司过去\n测试未发布模型的习惯\n但顺不等于已经证实\nArena没有公布能够对应\n到Gemini4的后台标示\nGoogle也没有承认这个\n测试席位属于Gemini4\n开发者看到的是输出差异\n再从输出差异倒推身份\n它是一条合理线索\n却不是完整证据链\n如果故事只停在这里\n最多只能叫疑似内部模型曝光\n真正让它冲上热搜的\n是随后流传出来的一张跑分表\n这张表宣称\nGemini4Pro在Deepseek1.1上达到88.7%\n超过Astra的86.9%\n在GDP Vol AAVR上达到2,064亿楼\nTerminalbench 2.1达到95.3%\nOsworld 2.0达到86.8%\n四项分别考察长程软件工程真实\n知识工作终端操作和计算机使用\n换句话说\n他宣称自己不是只会做漂亮网页\n而是在编程智能体知识\n工作和电脑操作上同时领先\n如果这些数字是真的\n冲击确实很大\nastral和fable代表的不是上一代模型\n而是Openai与Anthropic刚刚摆上\n桌的前沿产品Gemini4若能在\n同一轮测试中完成横扫\n意味着Google不是追上\n而是试图跨过整个竞争梯队\n更狠的是价格传闻\n每百万输入TOKEN2.25美元\n输出11.25美元\n这个价格高于Gemini3.8Flash\n却只有Astra和fable前沿\n档位的一小部分\n能力领先再叠加价格下压\n打击的不只是排行榜\n而是对手的商业模型\n但这里必须踩一脚刹车\n流传的跑分表没有公开完整提示词\n没有样本数量\n没有重复运行结果\n没有评测环境\n也没有原始日志\n社区能够看到一张设计完整的表\n却无法重做这场考试\n大模型跑分对提示词工具\n权限思考预算代理框架和\n采样设置都极其敏感\n即使模型相同\n给他更多测试时间\n更强工具\n或者更高TOKEN预算\n成绩也会明显变化\n没有评测细节\n精确到小数点后的数字反而\n容易制造一种并不存在的确定感\n还有一个容易被忽略的问题\nArena里的视觉演示与那张跑分表\n之间没有被公开证据连起来\n做出精致PS5的模型是否\n就是表里88.7%的模型\n所谓后台规格是否来自同一检查点\n几个泄露账号引用的\n是不是同一个最初来源\n目前都没有闭环\n互联网上常见的做法是把三个\n来源不同的碎片拼成一台完整机器\n一组用户演示证明能力\n一张匿名表证明跑分\n一张终端截图证明规格\n听起来很震撼\n拆开以后\n每一块的证据等级完全不同\n规格泄露尤其如此\n传闻称Gemini4拥有1,000万TOKEN\n输入上限和25.6万TOKEN输出上限\n还带有永久跨会话记忆\n直接联网\n恶意代码隔离环境\n甚至原生机器人电机控制\n任何一项单独拿出来\n都足以成为产品发布会的主角\n可是目前支撑这些说法的\n主要是一张自称来自Google\ndeep mind内部终端的截图\n截图的字段写的很专业\n但专业格式无法证明来源\n没有经过认证的请求记录\n没有公开模型卡\n也没有第二个独立团队复现\n因为Gemini4本身并不是虚构项目\n2026年7月\n桑达尔皮查伊在Alphabet第二\n季度财报电话会上明确表示\nGoogle已经启动公司迄今\n最具野心的预训练任务\n目标就是Gemini4\n这个事实把整场猜测从凭空\n想象推到了项目真实存在\n但泄露身份不明的位置\nGoogle也确实有压力\nGemini3.8Flash已经公开\n定位是低成本高速度的工作模型\n官方定价是每百万输入TOKEN0.75美元\n输出3.75美元\n它在编程和智能体\n任务上比前代更强\n还会在复杂任务上\n执行更长的推理循环\n但Flash终究是效率党\nGoogle需要一个真正站在\n前沿最高位的Pro模型\n去回答Astra和fable带来的竞争\n这也是匿名模型花10分钟\n生成SVG为什么引起怀疑\n它可能是Gemini4\n也可能只是Gemini3.8Flash的高思考配置\n或者一个接入了更复杂\n代理框架的内部变体\n长时间沉默只能说明系统\n愿意投入更多测试时计算\n不能单独证明底层模型换代\n今天的模型能力已经\n不再只由模型权重决定\n外面的脚手架工具验证器执行\n环境和思考预算同样会改变结果\n这一点非常关键\n行业过去喜欢用参数规模解释能力\n后来用训练算力解释\n现在又开始用推理阶段投入解释\n一个模型先写方案\n再调用工具执行\n再检查结果\n再修正错误\n最终输出当然会比\n一次性生成更完整\n但这不等于模型在\n几分钟内学会了新能力\n它可能只是把更多计算\n花在了同一个问题上\n随后被卷进故事的RSI\n让这场泄露又多了一层科幻色彩\nRSI指递归自我改进\n最激进的理解是\nAI能够改进自己\n再由改进后的版本继续改进\n最终让能力增长速度越来越快\n传闻声称\nGemini4之所以表现突然跳升\n甚至可能提前完成预训练\n是因为deep mind在训练\n中闭合了RSI循环\n这个说法听起来能解释所有异常\n但公开证据并不支持这么大的结论\nGoogle参与的dream RSI\n研究确实已经公开\n它解决的问题是智能体在长程\n探索中如何少走重复的弯路\n传统智能体面对算法设计\n数学优化或者GPU内核工程时\n会尝试大量方案\n每次都从头在线执行\n成本非常高\n而且固定的搜索策略\n不会随着经验自动调整\ndream RSI的办法是把过去的\n探索过程保存成一棵历史树\n再把这棵树变成\n一个低成本回放环境\n智能体可以在回放环境\n里尝试不同的探索策略\n先走哪条分支\n同时开几条路线\n什么时候停止\n失败后改去哪边\n因为旧分支的结果已经存在\n它不必每次重新调用\n昂贵模型和评估器\n筛出更好的策略后\n再回到真实任务中探索\n产生新记录\n继续扩充回放环境\n如此循环\n改进的是探索政策\n而不是底层大模型的全部权重\n这和大众想象中的AI重写自己\n第二天智力翻倍\n差别很大\ndream RSI论文明确说明\n底层编码智能体评估器和\n执行接口可以保持固定\n主要变化发生在负责\n分配探索计算的策略层\n它依然重要\n因为智能体真正昂贵的部分\n往往不是提出一个答案\n而是在巨大搜索空间里\n找到值得继续投入的路线\ndream RSI在部分任务中\n把调用量显著压低\n并在算法数学和GPU内核任务上\n改善了质量与成本的平衡\n这也解释了为什么RSI\n值得被放进Gemini4的讨论\n却不能拿来证明Gemini4已经自我进化\nGoogle的方向很清楚\n未来能力提升不止\n来自一次更大的预训练\n也来自模型如何使用历史\n安排搜索\n反复验证和改进工具链\n它更像给研究团队配上一套不会\n忘记试错记录的实验管理系统\n而不是让模型突然拥有\n不受控制的自我复制能力\n真正值得竞争对手警惕的\n恰恰是这套组合\nGoogle同时握着模型TPU\n数据中心搜索浏览器Android\nWorkspace和大规模用户反馈\n单看某个基准\n领先2分未必能决定胜负\n如果同一家公司能够把训练推理\n工具历史回放和产品分发接在一起\n能力改善就可以更快进入真实使用\n再从真实使用回流\n到下一轮系统优化\nOpenai和Anthropic当然也有自己的优势\nAstra的高强度推理和代理执行能力\nfable面向复杂工作流的可靠性\n都不是一张匿名表就能抹掉的\n前沿模型的真实\n竞争也不是百米赛跑\n而是五项全能\n能力延迟成本\n稳定性和生态缺一不可\n一个模型花10分钟做出惊艳网页\n在展示里很有冲击力\n放进企业流程后\n用户还要问它每次是否稳定\n失败如何恢复\n工具权限如何审计\n以及账单是否能承受\n这正是价格传闻如果\n成真会非常危险的原因\n前沿模型公司目前面对同一个矛盾\n训练和推理投入持续增加\n客户却希望TOKEN越来越便宜\n谁能用自研芯片和\n系统优化把单位智能成本压下去\n谁就能把高能力下放到更多任务里\nGoogle拥有TPU和完整云基础设施\n因此具备主动发起价格战的条件\n但所谓2.25美元和\n11.25美元仍然只是传闻\n不能提前写进Google价目表\n长上下文传闻也需要放回成本里看\n1,000万TOKEN听起来像把\n一个大型代码仓库\n几年邮件和整套公司\n文件一次塞进模型\n但能够接收不等于能够有效使用\n上下文越长\n首TOKEN等待注意力计算\n缓存容量和账单压力越高\n真正有价值的不是输入框能塞多大\n而是模型能否在材料\n中稳定找到关键约束\n不丢失中间信息\n不被互相冲突的指令带偏\n没有检索准确率延迟和价格\n单独公布一个窗口上限\n商业意义很有限\n25.6万TOKEN\n输出也是一样\n模型理论上可以一次生成整套项目\n用户实际需要的却不是更多代码\n而是能运行能测试能维护的代码\n输出越长\n局部错误累积的概率越高\n除非系统会持续编译\n执行回归测试和修复\n否则一次吐出几十万TOKEN\n更像产量展示\n不是软件工程能力\n匿名模型那些花10分钟\n完成的案例之所以重要\n正在于它似乎不只是增加输出量\n而是把执行和校验藏进了生成过程\n可惜目前只有成品\n没有完整轨迹\n基准测试同样存在版本选择问题\nTerminalbench Osworld和Deepseek都在快速更新\n新版本会修补数据污染\n提高任务难度\n也会改变工具环境\n把模型在某个旧版本上\n的成绩拿来横向比较\n必须确保所有对手使用同一套任务\n同一种脚手架和相近思考预算\n若Gemini测试时允许10分钟内部执行\nAstra和fable只用默认设置\n所谓领先很可能混入系统配置优势\n反过来如果三者都获得最大预算\nGoogle仍然领先\n那才是更可靠的能力信号\n还有数据污染风险\n公开基准一旦成为营销指标\n题目相似任务和解题\n模式就可能进入训练材料\n分数升高\n有时代表推理能力增强\n有时只是模型更熟悉考试\n真正能说明问题的\n是发布时间晚于训练截止\n任务过程可审计\n失败案例也完整公开的新测试\n匿名表最缺的恰恰是这些细节\n它给出了结果\n却没有给出让外界相信结果的路径\n不过Google并不需要在每个\n榜单上第一才足以改变竞争\n它拥有搜索和Chrome\n可以让模型直接接触实时网页\n拥有Android\n可以让代理进入手机\n拥有Workspace\n可以把代理放进\n企业文档邮件和会议\n拥有云与TPU\n可以控制推理成本\nAstra和fable卖的是模型能力\nGoogle有机会卖一整套默认\n存在于用户工作流里的能力\n只要Gemini4达到同一水平\n分发优势就会开始放大\n这也是机器人控制\n传闻为什么格外吸睛\nGoogle deep mind一直把数字智能与\n具身智能放在同一张路线图上\n如果一个统一多模态\n模型能看懂环境规划动作\n再把指令翻译成电机控制\nGemini就会从聊天窗口走向物理世界\n但支持机器人协议和能够可靠控制\n机器人之间隔着感知误差\n安全约束实时性与硬件差异\n一个后台字段无法\n填平这些工程距离\n从竞争节奏看\nGoogle还有一个理由把未\n发布检查点放进公开盲测\n内部基准容易让团队\n围绕固定指标优化\n真实用户却会提出\n研究者想不到的任务\nArena里的SVG游戏和网页挑战\n正好检验模型面对开放\n需求时能不能自己拆任务\n即使测试模型不是Gemini4\n这种测试方式也可能成为\nGoogle后训练的重要部分\n先让真实用户暴露失败\n再针对失败改进系统\n泄露背后真正的变量\n因此可能不是某个神秘参数\n而是产品开发方式\n过去模型先在实验室完成\n再发布给用户\n现在候选模型可能在匿名\n环境中接受公开压力测试\n社区一边猜身份\n一边免费提供最刁钻的测试集\n公司获得了数据\n用户获得了提前体验\n媒体获得了故事\n唯一缺少的是每条\n结论应有的证据标签\n还有一个判断不能回避\n泄露文化本身已经成为\n前沿模型营销的一部分\n公司不必正式宣布\n匿名测试模糊代号和\n社区截图就能制造关注\n开发者主动寻找神秘模型\n媒体主动拼接线索\n竞争对手也被迫回应\n一个尚未发布的产品\n哪怕最后证明只是高思考版Flash\nGoogle仍然获得了大量免费\n的能力展示和市场预热\n可这套玩法也有风险\n一旦传闻把预期抬到1,000万上下文\n25.6万输出\n全基准第一和极低价格\n正式产品只要少一项\n就会被说成缩水\n泄露不是只有竞争对手承受压力\nGoogle也会被社区制造\n的规格表反向绑架\n尤其当匿名跑分不可复现时\n用户对正式版本的评价可能\n不再基于它实际进步了多少\n而是看它有没有兑现\n一张来历不明的清单\n这件事当前最准确的结论并\n不是Gemini4已经碾压Astra和fable\n而是三层判断\n第一层已经确认\nGemini4项目真实存在\nGoogle正在进行规模空前的预训练\nGemini3.8Flash已经发布\nGoogle也在公开研究更长程的代理\n循环和低成本自我改进策略\n第二层有观察支撑\nArena里确实有人遇到能力\n与标签不匹配的模型\n复杂SVG三维网页和游戏案例\n说明某个测试系统\n投入了更高的推理预算\n它可能属于Google的内部检查点\n也可能是现有模型的特殊配置\n第三层仍然缺证据\n模型就是Gemini4Pro\n四项跑分全面领先\n价格只有对手的约1/4\n支持1,000万上下文\n永久记忆和机器人控制\n以及RSI让它完成了底层自我进化\n把第三层当成新闻\n事实会让文章很炸\n把三层分开才能让判断\n经得住正式发布后的核对\n但这并不代表整场泄露没有价值\n恰恰相反\n它提前暴露了\n下一代模型竞争的方向\n第一测试时计算会继续增加\n模型不再追求每个问题都秒回\n而是按任务价值决定思考多久\n第二前端与三维生成正在从能写\n代码进入能形成完整作品第三\n智能体开始保存探索历史\n减少重复试错第四\n价格会成为能力之外最锋利的武器\n如果Gemini4最终接近流传表现\nAstra和fable真正面对的就\n不是一次跑分失利\n而是Google把前沿能力压进自有\n基础设施后发动的成本战争\n如果正式产品与传闻差距很大\n这场风波也说明了另一件事\n市场已经迫不及待的\n需要一个能把智能体\n长上下文\n视觉代码和自我改进串起来的模型\n需求先于产品出现\n才会让每张截图都像\n未来泄露出来的证据所以\n最值得等待的\n不是Google什么时候揭掉Arena里的伪装\n而是正式模型卡能否回答四个问题\n这些成绩能否复现\n长任务能否稳定\n单位成本能否成立\n所谓RSI到底改进了哪一层\n只有这四件事同时站住\nGemini 4才算真正跨过Astra和fable\n否则今天被碾压的\n可能不是两个竞争模型\n而只是公众对匿名截图\n最后一点必要的怀疑\n事实核验Gemini 4正在训练属实\nGoogle称其为公司最具\n野心的预训练任务属实\nGemini 3.8Flash已经公开属实\ndream RSI论文与相关实验存在属实\nArena中的异常模型就是Gemini4\n尚未证实\n88.7% 2,064error\n95.3%和86.8%等泄露成绩\n尚未由Google或完整评测记录证实\n1,000万TOKEN输入\n25.6万TOKEN输出\n永久记忆与机器人控制尚未证实\nGemini是依靠RSI实现\n底层模型自我进化\n没有公开证据",
  "language": "zh",
  "duration": 1322.0,
  "segments": [
    {
      "id": 0,
      "start": 0.2,
      "end": 2.266,
      "text": "谷歌还没有发布Gemini4",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 1,
      "start": 2.6,
      "end": 5.766,
      "text": "但一份写着Gemini4的成绩单已经",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 2,
      "start": 5.766,
      "end": 9.2,
      "text": "提前把Astra和fable按在了排行榜下面",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 3,
      "start": 9.333,
      "end": 10.366,
      "text": "更奇怪的是",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 4,
      "start": 10.366,
      "end": 12.2,
      "text": "真正引爆这场泄露的",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 5,
      "start": 12.2,
      "end": 13.866,
      "text": "不是某份官方文件",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 6,
      "start": 13.9,
      "end": 15.966,
      "text": "也不是Google高管的一句暗示",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 7,
      "start": 16.533,
      "end": 18.8,
      "text": "而是Arena里一个挂着Gemini",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 8,
      "start": 18.8,
      "end": 21.333,
      "text": "3.8Flash名字的匿名模型",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 9,
      "start": 21.5,
      "end": 23.9,
      "text": "它顶着一张已经公开的旧身份证",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 10,
      "start": 23.933,
      "end": 25.666,
      "text": "却干出了公开版本",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 11,
      "start": 25.766,
      "end": 27.333,
      "text": "根本不像会干的事",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 12,
      "start": 27.566,
      "end": 28.9,
      "text": "问题也就来了",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 13,
      "start": 29.066,
      "end": 32.133,
      "text": "这到底是Gemini4不小心露出了底牌",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 14,
      "start": 32.666,
      "end": 34.466,
      "text": "还是整个AI社区共同",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 15,
      "start": 34.466,
      "end": 36.2,
      "text": "完成了一次大型误判",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 16,
      "start": 36.466,
      "end": 38.1,
      "text": "先看最反常的一幕",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 17,
      "start": 38.2,
      "end": 40.1,
      "text": "有人要求这个模型用原始",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 18,
      "start": 40.1,
      "end": 42.8,
      "text": "SVG代码画出一台Playstation5",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 19,
      "start": 43.066,
      "end": 45.4,
      "text": "正常的Flash模型强调速度",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 20,
      "start": 45.4,
      "end": 47.7,
      "text": "十几秒交出一组大致像样",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 21,
      "start": 47.7,
      "end": 49.9,
      "text": "的几何图形就算完成任务",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 22,
      "start": 50.166,
      "end": 52.066,
      "text": "但这个模型没有马上回答",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 23,
      "start": 52.133,
      "end": 54.1,
      "text": "它沉默了将近10分钟",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 24,
      "start": 54.366,
      "end": 57.133,
      "text": "随后一次性吐出数千行代码",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 25,
      "start": 57.166,
      "end": 59.966,
      "text": "把主机的曲线阴影光区",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 26,
      "start": 60.166,
      "end": 62.766,
      "text": "细节和空间关系都做了出来",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 27,
      "start": 63.1,
      "end": 64.8,
      "text": "他不像在即时续写",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 28,
      "start": 64.933,
      "end": 68.133,
      "text": "更像先在后台规划执行检查",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 29,
      "start": 68.266,
      "end": 69.966,
      "text": "再把成品端上来",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 30,
      "start": 70.066,
      "end": 72.6,
      "text": "随后测试案例越来越夸张",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 31,
      "start": 73.0,
      "end": 76.3,
      "text": "像素风宝塔宝马M4鹈鹕骑",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 32,
      "start": 76.3,
      "end": 78.966,
      "text": "自行车带昼夜切换和车灯",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 33,
      "start": 78.966,
      "end": 81.966,
      "text": "效果的三维场景空客H145",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 34,
      "start": 81.966,
      "end": 85.533,
      "text": "直升机完整赛车游戏可交互飞行",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 35,
      "start": 85.533,
      "end": 88.766,
      "text": "模拟全都开始在开发者社区流传",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 36,
      "start": 88.933,
      "end": 90.6,
      "text": "这里真正值得注意的",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 37,
      "start": 90.6,
      "end": 92.6,
      "text": "不是某一张图有多漂亮",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 38,
      "start": 92.966,
      "end": 95.133,
      "text": "而是它能够把空间坐标",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 39,
      "start": 95.266,
      "end": 98.1,
      "text": "视觉层级交互逻辑和长",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 40,
      "start": 98.1,
      "end": 100.466,
      "text": "代码放在同一次生成里完成",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 41,
      "start": 100.533,
      "end": 102.166,
      "text": "很多模型能写页面",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 42,
      "start": 102.166,
      "end": 103.966,
      "text": "也能画简单SVG",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 43,
      "start": 104.1,
      "end": 106.066,
      "text": "但一旦同时加入三维",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 44,
      "start": 106.166,
      "end": 108.566,
      "text": "动画物理和复杂指令结构",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 45,
      "start": 108.566,
      "end": 109.6,
      "text": "很容易散掉",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 46,
      "start": 109.6,
      "end": 112.3,
      "text": "这个匿名模型最引人怀疑的地方",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 47,
      "start": 112.733,
      "end": 113.9,
      "text": "正是它在复杂任务",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 48,
      "start": 113.9,
      "end": 115.533,
      "text": "里表现的不像Flash",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 49,
      "start": 115.866,
      "end": 117.966,
      "text": "Arena出现匿名检查点",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 50,
      "start": 117.966,
      "end": 119.466,
      "text": "本来不是什么新闻",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 51,
      "start": 119.766,
      "end": 122.133,
      "text": "这个平台的价值就在于盲测",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 52,
      "start": 122.7,
      "end": 124.466,
      "text": "用户不知道两边是谁",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 53,
      "start": 124.466,
      "end": 125.933,
      "text": "只看答案投票",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 54,
      "start": 126.4,
      "end": 128.7,
      "text": "实验室也能避开品牌光环",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 55,
      "start": 128.7,
      "end": 131.0,
      "text": "收集真实偏好Google",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 56,
      "start": 131.733,
      "end": 134.166,
      "text": "Openai和Anthropic都可能把",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 57,
      "start": 134.166,
      "end": 136.0,
      "text": "候选模型放进去测试",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 58,
      "start": 136.166,
      "end": 137.866,
      "text": "真正反常的是这次",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 59,
      "start": 137.866,
      "end": 139.6,
      "text": "模型没有使用陌生代号",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 60,
      "start": 139.9,
      "end": 140.933,
      "text": "而是套用了已经",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 61,
      "start": 140.933,
      "end": 143.4,
      "text": "发布的Gemini3.8Flash名称",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 62,
      "start": 143.666,
      "end": 144.866,
      "text": "社区看到的不是",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 63,
      "start": 144.866,
      "end": 146.6,
      "text": "一个身份空白的选手",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 64,
      "start": 146.733,
      "end": 148.0,
      "text": "而是一个能力与",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 65,
      "start": 148.0,
      "end": 150.1,
      "text": "身份证明显对不上的选手",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 66,
      "start": 150.133,
      "end": 151.4,
      "text": "这给了泄露故事",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 67,
      "start": 151.4,
      "end": 153.066,
      "text": "一个非常诱人的解释",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 68,
      "start": 153.5,
      "end": 158.0,
      "text": "Google把Gemini4检查点藏在Gemini3.8Flash名下",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 69,
      "start": 158.166,
      "end": 160.466,
      "text": "借Arena做公开压力测试",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 70,
      "start": 160.666,
      "end": 161.866,
      "text": "这个解释很顺",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 71,
      "start": 161.866,
      "end": 163.933,
      "text": "也符合大模型公司过去",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 72,
      "start": 163.933,
      "end": 166.0,
      "text": "测试未发布模型的习惯",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 73,
      "start": 166.1,
      "end": 168.466,
      "text": "但顺不等于已经证实",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 74,
      "start": 168.7,
      "end": 170.4,
      "text": "Arena没有公布能够对应",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 75,
      "start": 170.4,
      "end": 172.4,
      "text": "到Gemini4的后台标示",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 76,
      "start": 172.966,
      "end": 174.333,
      "text": "Google也没有承认这个",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 77,
      "start": 174.333,
      "end": 176.333,
      "text": "测试席位属于Gemini4",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 78,
      "start": 176.666,
      "end": 178.566,
      "text": "开发者看到的是输出差异",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 79,
      "start": 178.766,
      "end": 180.966,
      "text": "再从输出差异倒推身份",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 80,
      "start": 181.333,
      "end": 182.9,
      "text": "它是一条合理线索",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 81,
      "start": 183.0,
      "end": 184.733,
      "text": "却不是完整证据链",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 82,
      "start": 184.866,
      "end": 186.666,
      "text": "如果故事只停在这里",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 83,
      "start": 186.8,
      "end": 189.333,
      "text": "最多只能叫疑似内部模型曝光",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 84,
      "start": 189.533,
      "end": 191.333,
      "text": "真正让它冲上热搜的",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 85,
      "start": 191.333,
      "end": 193.9,
      "text": "是随后流传出来的一张跑分表",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 86,
      "start": 193.9,
      "end": 195.1,
      "text": "这张表宣称",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 87,
      "start": 195.333,
      "end": 200.7,
      "text": "Gemini4Pro在Deepseek1.1上达到88.7%",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 88,
      "start": 201.066,
      "end": 203.533,
      "text": "超过Astra的86.9%",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 89,
      "start": 203.6,
      "end": 208.466,
      "text": "在GDP Vol AAVR上达到2,064亿楼",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 90,
      "start": 208.933,
      "end": 212.566,
      "text": "Terminalbench 2.1达到95.3%",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 91,
      "start": 212.933,
      "end": 216.1,
      "text": "Osworld 2.0达到86.8%",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 92,
      "start": 216.2,
      "end": 219.666,
      "text": "四项分别考察长程软件工程真实",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 93,
      "start": 219.666,
      "end": 223.0,
      "text": "知识工作终端操作和计算机使用",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 94,
      "start": 223.4,
      "end": 224.266,
      "text": "换句话说",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 95,
      "start": 224.266,
      "end": 226.8,
      "text": "他宣称自己不是只会做漂亮网页",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 96,
      "start": 226.933,
      "end": 229.533,
      "text": "而是在编程智能体知识",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 97,
      "start": 229.533,
      "end": 232.2,
      "text": "工作和电脑操作上同时领先",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 98,
      "start": 232.333,
      "end": 234.133,
      "text": "如果这些数字是真的",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 99,
      "start": 234.4,
      "end": 235.766,
      "text": "冲击确实很大",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 100,
      "start": 235.966,
      "end": 238.966,
      "text": "astral和fable代表的不是上一代模型",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 101,
      "start": 239.466,
      "end": 242.366,
      "text": "而是Openai与Anthropic刚刚摆上",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 102,
      "start": 242.366,
      "end": 245.2,
      "text": "桌的前沿产品Gemini4若能在",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 103,
      "start": 245.2,
      "end": 247.166,
      "text": "同一轮测试中完成横扫",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 104,
      "start": 247.366,
      "end": 249.166,
      "text": "意味着Google不是追上",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 105,
      "start": 249.6,
      "end": 252.1,
      "text": "而是试图跨过整个竞争梯队",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 106,
      "start": 252.166,
      "end": 253.733,
      "text": "更狠的是价格传闻",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 107,
      "start": 253.933,
      "end": 256.6,
      "text": "每百万输入TOKEN2.25美元",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 108,
      "start": 256.8,
      "end": 258.566,
      "text": "输出11.25美元",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 109,
      "start": 258.666,
      "end": 261.0,
      "text": "这个价格高于Gemini3.8Flash",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 110,
      "start": 261.166,
      "end": 263.366,
      "text": "却只有Astra和fable前沿",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 111,
      "start": 263.366,
      "end": 264.566,
      "text": "档位的一小部分",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 112,
      "start": 264.933,
      "end": 266.966,
      "text": "能力领先再叠加价格下压",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 113,
      "start": 266.966,
      "end": 268.6,
      "text": "打击的不只是排行榜",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 114,
      "start": 268.6,
      "end": 270.366,
      "text": "而是对手的商业模型",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 115,
      "start": 270.466,
      "end": 272.3,
      "text": "但这里必须踩一脚刹车",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 116,
      "start": 272.733,
      "end": 275.366,
      "text": "流传的跑分表没有公开完整提示词",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 117,
      "start": 275.6,
      "end": 276.933,
      "text": "没有样本数量",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 118,
      "start": 277.0,
      "end": 278.566,
      "text": "没有重复运行结果",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 119,
      "start": 278.6,
      "end": 279.8,
      "text": "没有评测环境",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 120,
      "start": 280.2,
      "end": 281.533,
      "text": "也没有原始日志",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 121,
      "start": 281.733,
      "end": 284.3,
      "text": "社区能够看到一张设计完整的表",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 122,
      "start": 284.3,
      "end": 286.3,
      "text": "却无法重做这场考试",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 123,
      "start": 286.4,
      "end": 288.966,
      "text": "大模型跑分对提示词工具",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 124,
      "start": 288.966,
      "end": 291.8,
      "text": "权限思考预算代理框架和",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 125,
      "start": 291.8,
      "end": 293.733,
      "text": "采样设置都极其敏感",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 126,
      "start": 293.8,
      "end": 295.166,
      "text": "即使模型相同",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 127,
      "start": 295.4,
      "end": 297.066,
      "text": "给他更多测试时间",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 128,
      "start": 297.166,
      "end": 298.266,
      "text": "更强工具",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 129,
      "start": 298.333,
      "end": 300.1,
      "text": "或者更高TOKEN预算",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 130,
      "start": 300.3,
      "end": 301.966,
      "text": "成绩也会明显变化",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 131,
      "start": 302.133,
      "end": 303.533,
      "text": "没有评测细节",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 132,
      "start": 304.133,
      "end": 306.466,
      "text": "精确到小数点后的数字反而",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 133,
      "start": 306.466,
      "end": 309.0,
      "text": "容易制造一种并不存在的确定感",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 134,
      "start": 309.066,
      "end": 311.133,
      "text": "还有一个容易被忽略的问题",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 135,
      "start": 311.4,
      "end": 314.3,
      "text": "Arena里的视觉演示与那张跑分表",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 136,
      "start": 314.3,
      "end": 316.766,
      "text": "之间没有被公开证据连起来",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 137,
      "start": 316.9,
      "end": 319.6,
      "text": "做出精致PS5的模型是否",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 138,
      "start": 319.6,
      "end": 322.4,
      "text": "就是表里88.7%的模型",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 139,
      "start": 322.666,
      "end": 325.933,
      "text": "所谓后台规格是否来自同一检查点",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 140,
      "start": 325.966,
      "end": 327.7,
      "text": "几个泄露账号引用的",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 141,
      "start": 327.7,
      "end": 329.5,
      "text": "是不是同一个最初来源",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 142,
      "start": 330.166,
      "end": 331.466,
      "text": "目前都没有闭环",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 143,
      "start": 331.766,
      "end": 334.533,
      "text": "互联网上常见的做法是把三个",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 144,
      "start": 334.533,
      "end": 337.6,
      "text": "来源不同的碎片拼成一台完整机器",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 145,
      "start": 337.666,
      "end": 339.766,
      "text": "一组用户演示证明能力",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 146,
      "start": 339.8,
      "end": 341.9,
      "text": "一张匿名表证明跑分",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 147,
      "start": 341.966,
      "end": 344.3,
      "text": "一张终端截图证明规格",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 148,
      "start": 344.6,
      "end": 345.733,
      "text": "听起来很震撼",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 149,
      "start": 346.266,
      "end": 347.1,
      "text": "拆开以后",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 150,
      "start": 347.1,
      "end": 349.266,
      "text": "每一块的证据等级完全不同",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 151,
      "start": 349.4,
      "end": 351.6,
      "text": "规格泄露尤其如此",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 152,
      "start": 351.666,
      "end": 354.5,
      "text": "传闻称Gemini4拥有1,000万TOKEN",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 153,
      "start": 354.5,
      "end": 357.9,
      "text": "输入上限和25.6万TOKEN输出上限",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 154,
      "start": 358.166,
      "end": 360.133,
      "text": "还带有永久跨会话记忆",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 155,
      "start": 360.166,
      "end": 361.166,
      "text": "直接联网",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 156,
      "start": 361.2,
      "end": 362.766,
      "text": "恶意代码隔离环境",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 157,
      "start": 362.866,
      "end": 365.333,
      "text": "甚至原生机器人电机控制",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 158,
      "start": 365.466,
      "end": 367.266,
      "text": "任何一项单独拿出来",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 159,
      "start": 367.366,
      "end": 369.966,
      "text": "都足以成为产品发布会的主角",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 160,
      "start": 370.166,
      "end": 372.366,
      "text": "可是目前支撑这些说法的",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 161,
      "start": 372.7,
      "end": 374.8,
      "text": "主要是一张自称来自Google",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 162,
      "start": 374.8,
      "end": 376.866,
      "text": "deep mind内部终端的截图",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 163,
      "start": 377.1,
      "end": 378.9,
      "text": "截图的字段写的很专业",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 164,
      "start": 378.9,
      "end": 381.333,
      "text": "但专业格式无法证明来源",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 165,
      "start": 381.5,
      "end": 383.666,
      "text": "没有经过认证的请求记录",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 166,
      "start": 383.733,
      "end": 385.266,
      "text": "没有公开模型卡",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 167,
      "start": 385.333,
      "end": 387.533,
      "text": "也没有第二个独立团队复现",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 168,
      "start": 397.533,
      "end": 400.2,
      "text": "因为Gemini4本身并不是虚构项目",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 169,
      "start": 400.4,
      "end": 402.0,
      "text": "2026年7月",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 170,
      "start": 402.2,
      "end": 404.533,
      "text": "桑达尔皮查伊在Alphabet第二",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 171,
      "start": 404.533,
      "end": 406.966,
      "text": "季度财报电话会上明确表示",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 172,
      "start": 407.366,
      "end": 409.3,
      "text": "Google已经启动公司迄今",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 173,
      "start": 409.3,
      "end": 411.333,
      "text": "最具野心的预训练任务",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 174,
      "start": 411.4,
      "end": 413.166,
      "text": "目标就是Gemini4",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 175,
      "start": 413.2,
      "end": 415.866,
      "text": "这个事实把整场猜测从凭空",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 176,
      "start": 415.866,
      "end": 419.0,
      "text": "想象推到了项目真实存在",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 177,
      "start": 419.1,
      "end": 421.166,
      "text": "但泄露身份不明的位置",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 178,
      "start": 421.366,
      "end": 422.866,
      "text": "Google也确实有压力",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 179,
      "start": 423.333,
      "end": 425.666,
      "text": "Gemini3.8Flash已经公开",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 180,
      "start": 425.7,
      "end": 428.7,
      "text": "定位是低成本高速度的工作模型",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 181,
      "start": 428.966,
      "end": 432.6,
      "text": "官方定价是每百万输入TOKEN0.75美元",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 182,
      "start": 432.8,
      "end": 434.5,
      "text": "输出3.75美元",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 183,
      "start": 434.8,
      "end": 436.366,
      "text": "它在编程和智能体",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 184,
      "start": 436.366,
      "end": 438.1,
      "text": "任务上比前代更强",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 185,
      "start": 438.1,
      "end": 439.566,
      "text": "还会在复杂任务上",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 186,
      "start": 439.566,
      "end": 441.4,
      "text": "执行更长的推理循环",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 187,
      "start": 441.5,
      "end": 443.666,
      "text": "但Flash终究是效率党",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 188,
      "start": 443.9,
      "end": 445.866,
      "text": "Google需要一个真正站在",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 189,
      "start": 445.866,
      "end": 448.133,
      "text": "前沿最高位的Pro模型",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 190,
      "start": 448.266,
      "end": 451.3,
      "text": "去回答Astra和fable带来的竞争",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 191,
      "start": 451.666,
      "end": 453.366,
      "text": "这也是匿名模型花10分钟",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 192,
      "start": 453.366,
      "end": 455.733,
      "text": "生成SVG为什么引起怀疑",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 193,
      "start": 456.066,
      "end": 457.566,
      "text": "它可能是Gemini4",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 194,
      "start": 457.566,
      "end": 461.066,
      "text": "也可能只是Gemini3.8Flash的高思考配置",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 195,
      "start": 461.3,
      "end": 462.866,
      "text": "或者一个接入了更复杂",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 196,
      "start": 462.866,
      "end": 464.566,
      "text": "代理框架的内部变体",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 197,
      "start": 464.933,
      "end": 466.766,
      "text": "长时间沉默只能说明系统",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 198,
      "start": 466.766,
      "end": 469.066,
      "text": "愿意投入更多测试时计算",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 199,
      "start": 469.133,
      "end": 471.6,
      "text": "不能单独证明底层模型换代",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 200,
      "start": 472.566,
      "end": 474.1,
      "text": "今天的模型能力已经",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 201,
      "start": 474.1,
      "end": 476.2,
      "text": "不再只由模型权重决定",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 202,
      "start": 476.466,
      "end": 480.666,
      "text": "外面的脚手架工具验证器执行",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 203,
      "start": 480.666,
      "end": 484.066,
      "text": "环境和思考预算同样会改变结果",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 204,
      "start": 484.1,
      "end": 485.566,
      "text": "这一点非常关键",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 205,
      "start": 486.1,
      "end": 489.366,
      "text": "行业过去喜欢用参数规模解释能力",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 206,
      "start": 489.4,
      "end": 491.5,
      "text": "后来用训练算力解释",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 207,
      "start": 491.9,
      "end": 494.5,
      "text": "现在又开始用推理阶段投入解释",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 208,
      "start": 494.666,
      "end": 496.366,
      "text": "一个模型先写方案",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 209,
      "start": 496.366,
      "end": 498.0,
      "text": "再调用工具执行",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 210,
      "start": 498.0,
      "end": 499.2,
      "text": "再检查结果",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 211,
      "start": 499.2,
      "end": 500.366,
      "text": "再修正错误",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 212,
      "start": 500.366,
      "end": 501.8,
      "text": "最终输出当然会比",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 213,
      "start": 501.8,
      "end": 503.5,
      "text": "一次性生成更完整",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 214,
      "start": 503.933,
      "end": 505.266,
      "text": "但这不等于模型在",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 215,
      "start": 505.266,
      "end": 507.266,
      "text": "几分钟内学会了新能力",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 216,
      "start": 507.366,
      "end": 509.1,
      "text": "它可能只是把更多计算",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 217,
      "start": 509.1,
      "end": 510.9,
      "text": "花在了同一个问题上",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 218,
      "start": 510.933,
      "end": 512.866,
      "text": "随后被卷进故事的RSI",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 219,
      "start": 512.966,
      "end": 515.766,
      "text": "让这场泄露又多了一层科幻色彩",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 220,
      "start": 516.066,
      "end": 518.3,
      "text": "RSI指递归自我改进",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 221,
      "start": 519.066,
      "end": 520.566,
      "text": "最激进的理解是",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 222,
      "start": 520.566,
      "end": 522.1,
      "text": "AI能够改进自己",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 223,
      "start": 522.1,
      "end": 524.666,
      "text": "再由改进后的版本继续改进",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 224,
      "start": 524.966,
      "end": 527.8,
      "text": "最终让能力增长速度越来越快",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 225,
      "start": 527.966,
      "end": 528.933,
      "text": "传闻声称",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 226,
      "start": 528.933,
      "end": 531.333,
      "text": "Gemini4之所以表现突然跳升",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 227,
      "start": 531.6,
      "end": 533.866,
      "text": "甚至可能提前完成预训练",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 228,
      "start": 534.266,
      "end": 535.866,
      "text": "是因为deep mind在训练",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 229,
      "start": 535.866,
      "end": 537.8,
      "text": "中闭合了RSI循环",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 230,
      "start": 538.066,
      "end": 540.933,
      "text": "这个说法听起来能解释所有异常",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 231,
      "start": 541.3,
      "end": 544.266,
      "text": "但公开证据并不支持这么大的结论",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 232,
      "start": 544.5,
      "end": 546.5,
      "text": "Google参与的dream RSI",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 233,
      "start": 546.5,
      "end": 548.4,
      "text": "研究确实已经公开",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 234,
      "start": 548.6,
      "end": 551.466,
      "text": "它解决的问题是智能体在长程",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 235,
      "start": 551.466,
      "end": 554.1,
      "text": "探索中如何少走重复的弯路",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 236,
      "start": 554.566,
      "end": 556.8,
      "text": "传统智能体面对算法设计",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 237,
      "start": 556.966,
      "end": 559.9,
      "text": "数学优化或者GPU内核工程时",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 238,
      "start": 559.933,
      "end": 561.566,
      "text": "会尝试大量方案",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 239,
      "start": 561.733,
      "end": 563.7,
      "text": "每次都从头在线执行",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 240,
      "start": 564.266,
      "end": 565.4,
      "text": "成本非常高",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 241,
      "start": 565.6,
      "end": 567.366,
      "text": "而且固定的搜索策略",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 242,
      "start": 567.366,
      "end": 569.366,
      "text": "不会随着经验自动调整",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 243,
      "start": 569.6,
      "end": 572.1,
      "text": "dream RSI的办法是把过去的",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 244,
      "start": 572.1,
      "end": 574.6,
      "text": "探索过程保存成一棵历史树",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 245,
      "start": 574.733,
      "end": 576.066,
      "text": "再把这棵树变成",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 246,
      "start": 576.066,
      "end": 577.8,
      "text": "一个低成本回放环境",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 247,
      "start": 578.133,
      "end": 579.766,
      "text": "智能体可以在回放环境",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 248,
      "start": 579.766,
      "end": 581.9,
      "text": "里尝试不同的探索策略",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 249,
      "start": 582.1,
      "end": 583.6,
      "text": "先走哪条分支",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 250,
      "start": 583.9,
      "end": 585.6,
      "text": "同时开几条路线",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 251,
      "start": 585.7,
      "end": 586.933,
      "text": "什么时候停止",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 252,
      "start": 587.1,
      "end": 588.8,
      "text": "失败后改去哪边",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 253,
      "start": 588.933,
      "end": 591.3,
      "text": "因为旧分支的结果已经存在",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 254,
      "start": 591.466,
      "end": 593.333,
      "text": "它不必每次重新调用",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 255,
      "start": 593.333,
      "end": 594.966,
      "text": "昂贵模型和评估器",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 256,
      "start": 595.266,
      "end": 596.933,
      "text": "筛出更好的策略后",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 257,
      "start": 597.2,
      "end": 599.133,
      "text": "再回到真实任务中探索",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 258,
      "start": 599.333,
      "end": 600.566,
      "text": "产生新记录",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 259,
      "start": 600.933,
      "end": 602.866,
      "text": "继续扩充回放环境",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 260,
      "start": 603.1,
      "end": 604.166,
      "text": "如此循环",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 261,
      "start": 604.333,
      "end": 606.1,
      "text": "改进的是探索政策",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 262,
      "start": 606.3,
      "end": 608.8,
      "text": "而不是底层大模型的全部权重",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 263,
      "start": 608.933,
      "end": 611.8,
      "text": "这和大众想象中的AI重写自己",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 264,
      "start": 612.2,
      "end": 613.533,
      "text": "第二天智力翻倍",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 265,
      "start": 613.533,
      "end": 614.566,
      "text": "差别很大",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 266,
      "start": 614.8,
      "end": 617.166,
      "text": "dream RSI论文明确说明",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 267,
      "start": 617.3,
      "end": 620.2,
      "text": "底层编码智能体评估器和",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 268,
      "start": 620.2,
      "end": 622.166,
      "text": "执行接口可以保持固定",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 269,
      "start": 622.3,
      "end": 624.133,
      "text": "主要变化发生在负责",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 270,
      "start": 624.133,
      "end": 626.466,
      "text": "分配探索计算的策略层",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 271,
      "start": 626.733,
      "end": 627.8,
      "text": "它依然重要",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 272,
      "start": 628.333,
      "end": 630.533,
      "text": "因为智能体真正昂贵的部分",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 273,
      "start": 630.666,
      "end": 632.5,
      "text": "往往不是提出一个答案",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 274,
      "start": 632.866,
      "end": 634.8,
      "text": "而是在巨大搜索空间里",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 275,
      "start": 634.8,
      "end": 636.933,
      "text": "找到值得继续投入的路线",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 276,
      "start": 637.3,
      "end": 639.2,
      "text": "dream RSI在部分任务中",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 277,
      "start": 639.2,
      "end": 641.333,
      "text": "把调用量显著压低",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 278,
      "start": 641.4,
      "end": 645.366,
      "text": "并在算法数学和GPU内核任务上",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 279,
      "start": 645.366,
      "end": 647.8,
      "text": "改善了质量与成本的平衡",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 280,
      "start": 647.966,
      "end": 650.166,
      "text": "这也解释了为什么RSI",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 281,
      "start": 650.2,
      "end": 652.466,
      "text": "值得被放进Gemini4的讨论",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 282,
      "start": 652.766,
      "end": 656.133,
      "text": "却不能拿来证明Gemini4已经自我进化",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 283,
      "start": 656.333,
      "end": 657.8,
      "text": "Google的方向很清楚",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 284,
      "start": 658.066,
      "end": 659.566,
      "text": "未来能力提升不止",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 285,
      "start": 659.566,
      "end": 661.366,
      "text": "来自一次更大的预训练",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 286,
      "start": 661.566,
      "end": 663.933,
      "text": "也来自模型如何使用历史",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 287,
      "start": 664.366,
      "end": 665.7,
      "text": "安排搜索",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 288,
      "start": 665.733,
      "end": 667.933,
      "text": "反复验证和改进工具链",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 289,
      "start": 668.1,
      "end": 670.366,
      "text": "它更像给研究团队配上一套不会",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 290,
      "start": 670.366,
      "end": 673.066,
      "text": "忘记试错记录的实验管理系统",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 291,
      "start": 673.2,
      "end": 675.133,
      "text": "而不是让模型突然拥有",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 292,
      "start": 675.133,
      "end": 677.333,
      "text": "不受控制的自我复制能力",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 293,
      "start": 677.5,
      "end": 679.866,
      "text": "真正值得竞争对手警惕的",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 294,
      "start": 680.166,
      "end": 681.566,
      "text": "恰恰是这套组合",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 295,
      "start": 681.766,
      "end": 684.3,
      "text": "Google同时握着模型TPU",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 296,
      "start": 684.4,
      "end": 688.133,
      "text": "数据中心搜索浏览器Android",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 297,
      "start": 688.333,
      "end": 690.966,
      "text": "Workspace和大规模用户反馈",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 298,
      "start": 691.0,
      "end": 692.3,
      "text": "单看某个基准",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 299,
      "start": 692.4,
      "end": 694.766,
      "text": "领先2分未必能决定胜负",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 300,
      "start": 695.333,
      "end": 698.133,
      "text": "如果同一家公司能够把训练推理",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 301,
      "start": 698.133,
      "end": 701.966,
      "text": "工具历史回放和产品分发接在一起",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 302,
      "start": 702.133,
      "end": 705.1,
      "text": "能力改善就可以更快进入真实使用",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 303,
      "start": 705.2,
      "end": 707.0,
      "text": "再从真实使用回流",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 304,
      "start": 707.0,
      "end": 708.666,
      "text": "到下一轮系统优化",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 305,
      "start": 708.9,
      "end": 712.266,
      "text": "Openai和Anthropic当然也有自己的优势",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 306,
      "start": 712.866,
      "end": 715.866,
      "text": "Astra的高强度推理和代理执行能力",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 307,
      "start": 716.066,
      "end": 718.733,
      "text": "fable面向复杂工作流的可靠性",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 308,
      "start": 718.8,
      "end": 721.466,
      "text": "都不是一张匿名表就能抹掉的",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 309,
      "start": 721.666,
      "end": 722.766,
      "text": "前沿模型的真实",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 310,
      "start": 722.766,
      "end": 724.666,
      "text": "竞争也不是百米赛跑",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 311,
      "start": 724.666,
      "end": 726.1,
      "text": "而是五项全能",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 312,
      "start": 726.133,
      "end": 728.533,
      "text": "能力延迟成本",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 313,
      "start": 728.6,
      "end": 731.066,
      "text": "稳定性和生态缺一不可",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 314,
      "start": 731.066,
      "end": 733.866,
      "text": "一个模型花10分钟做出惊艳网页",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 315,
      "start": 734.0,
      "end": 735.966,
      "text": "在展示里很有冲击力",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 316,
      "start": 736.3,
      "end": 737.766,
      "text": "放进企业流程后",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 317,
      "start": 737.866,
      "end": 740.3,
      "text": "用户还要问它每次是否稳定",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 318,
      "start": 740.366,
      "end": 741.8,
      "text": "失败如何恢复",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 319,
      "start": 742.066,
      "end": 743.866,
      "text": "工具权限如何审计",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 320,
      "start": 743.966,
      "end": 745.933,
      "text": "以及账单是否能承受",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 321,
      "start": 745.966,
      "end": 747.766,
      "text": "这正是价格传闻如果",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 322,
      "start": 747.766,
      "end": 749.9,
      "text": "成真会非常危险的原因",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 323,
      "start": 750.2,
      "end": 752.866,
      "text": "前沿模型公司目前面对同一个矛盾",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 324,
      "start": 753.266,
      "end": 755.733,
      "text": "训练和推理投入持续增加",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 325,
      "start": 756.066,
      "end": 758.533,
      "text": "客户却希望TOKEN越来越便宜",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 326,
      "start": 758.6,
      "end": 760.3,
      "text": "谁能用自研芯片和",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 327,
      "start": 760.3,
      "end": 763.2,
      "text": "系统优化把单位智能成本压下去",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 328,
      "start": 763.766,
      "end": 766.7,
      "text": "谁就能把高能力下放到更多任务里",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 329,
      "start": 766.9,
      "end": 770.0,
      "text": "Google拥有TPU和完整云基础设施",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 330,
      "start": 770.466,
      "end": 773.266,
      "text": "因此具备主动发起价格战的条件",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 331,
      "start": 773.333,
      "end": 775.333,
      "text": "但所谓2.25美元和",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 332,
      "start": 775.333,
      "end": 777.966,
      "text": "11.25美元仍然只是传闻",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 333,
      "start": 778.0,
      "end": 780.3,
      "text": "不能提前写进Google价目表",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 334,
      "start": 780.566,
      "end": 783.966,
      "text": "长上下文传闻也需要放回成本里看",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 335,
      "start": 783.966,
      "end": 785.933,
      "text": "1,000万TOKEN听起来像把",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 336,
      "start": 785.933,
      "end": 787.5,
      "text": "一个大型代码仓库",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 337,
      "start": 787.6,
      "end": 789.6,
      "text": "几年邮件和整套公司",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 338,
      "start": 789.6,
      "end": 791.4,
      "text": "文件一次塞进模型",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 339,
      "start": 791.666,
      "end": 794.5,
      "text": "但能够接收不等于能够有效使用",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 340,
      "start": 794.6,
      "end": 795.866,
      "text": "上下文越长",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 341,
      "start": 795.933,
      "end": 798.333,
      "text": "首TOKEN等待注意力计算",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 342,
      "start": 798.733,
      "end": 801.0,
      "text": "缓存容量和账单压力越高",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 343,
      "start": 801.2,
      "end": 804.2,
      "text": "真正有价值的不是输入框能塞多大",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 344,
      "start": 804.2,
      "end": 806.066,
      "text": "而是模型能否在材料",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 345,
      "start": 806.066,
      "end": 807.8,
      "text": "中稳定找到关键约束",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 346,
      "start": 807.866,
      "end": 809.4,
      "text": "不丢失中间信息",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 347,
      "start": 809.666,
      "end": 811.966,
      "text": "不被互相冲突的指令带偏",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 348,
      "start": 812.1,
      "end": 815.933,
      "text": "没有检索准确率延迟和价格",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 349,
      "start": 816.166,
      "end": 818.133,
      "text": "单独公布一个窗口上限",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 350,
      "start": 818.5,
      "end": 819.966,
      "text": "商业意义很有限",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 351,
      "start": 820.066,
      "end": 821.766,
      "text": "25.6万TOKEN",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 352,
      "start": 822.1,
      "end": 823.333,
      "text": "输出也是一样",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 353,
      "start": 823.333,
      "end": 826.2,
      "text": "模型理论上可以一次生成整套项目",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 354,
      "start": 826.6,
      "end": 829.333,
      "text": "用户实际需要的却不是更多代码",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 355,
      "start": 829.8,
      "end": 833.1,
      "text": "而是能运行能测试能维护的代码",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 356,
      "start": 833.333,
      "end": 834.5,
      "text": "输出越长",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 357,
      "start": 834.866,
      "end": 837.2,
      "text": "局部错误累积的概率越高",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 358,
      "start": 837.366,
      "end": 839.266,
      "text": "除非系统会持续编译",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 359,
      "start": 839.333,
      "end": 841.6,
      "text": "执行回归测试和修复",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 360,
      "start": 841.9,
      "end": 843.9,
      "text": "否则一次吐出几十万TOKEN",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 361,
      "start": 843.9,
      "end": 845.2,
      "text": "更像产量展示",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 362,
      "start": 845.2,
      "end": 846.933,
      "text": "不是软件工程能力",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 363,
      "start": 847.2,
      "end": 849.133,
      "text": "匿名模型那些花10分钟",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 364,
      "start": 849.133,
      "end": 851.3,
      "text": "完成的案例之所以重要",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 365,
      "start": 851.533,
      "end": 854.6,
      "text": "正在于它似乎不只是增加输出量",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 366,
      "start": 854.666,
      "end": 858.2,
      "text": "而是把执行和校验藏进了生成过程",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 367,
      "start": 858.4,
      "end": 860.1,
      "text": "可惜目前只有成品",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 368,
      "start": 860.1,
      "end": 861.3,
      "text": "没有完整轨迹",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 369,
      "start": 861.533,
      "end": 864.133,
      "text": "基准测试同样存在版本选择问题",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 370,
      "start": 864.766,
      "end": 869.066,
      "text": "Terminalbench Osworld和Deepseek都在快速更新",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 371,
      "start": 869.466,
      "end": 871.5,
      "text": "新版本会修补数据污染",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 372,
      "start": 871.566,
      "end": 872.766,
      "text": "提高任务难度",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 373,
      "start": 872.766,
      "end": 874.333,
      "text": "也会改变工具环境",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 374,
      "start": 874.366,
      "end": 876.1,
      "text": "把模型在某个旧版本上",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 375,
      "start": 876.1,
      "end": 878.1,
      "text": "的成绩拿来横向比较",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 376,
      "start": 878.533,
      "end": 881.6,
      "text": "必须确保所有对手使用同一套任务",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 377,
      "start": 881.7,
      "end": 884.7,
      "text": "同一种脚手架和相近思考预算",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 378,
      "start": 884.933,
      "end": 888.266,
      "text": "若Gemini测试时允许10分钟内部执行",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 379,
      "start": 888.4,
      "end": 890.566,
      "text": "Astra和fable只用默认设置",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 380,
      "start": 890.8,
      "end": 893.866,
      "text": "所谓领先很可能混入系统配置优势",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 381,
      "start": 894.266,
      "end": 897.166,
      "text": "反过来如果三者都获得最大预算",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 382,
      "start": 897.3,
      "end": 898.666,
      "text": "Google仍然领先",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 383,
      "start": 898.733,
      "end": 901.0,
      "text": "那才是更可靠的能力信号",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 384,
      "start": 901.133,
      "end": 902.733,
      "text": "还有数据污染风险",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 385,
      "start": 902.8,
      "end": 905.2,
      "text": "公开基准一旦成为营销指标",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 386,
      "start": 905.466,
      "end": 907.666,
      "text": "题目相似任务和解题",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 387,
      "start": 907.666,
      "end": 910.0,
      "text": "模式就可能进入训练材料",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 388,
      "start": 910.366,
      "end": 911.366,
      "text": "分数升高",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 389,
      "start": 911.366,
      "end": 913.5,
      "text": "有时代表推理能力增强",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 390,
      "start": 913.566,
      "end": 915.966,
      "text": "有时只是模型更熟悉考试",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 391,
      "start": 916.133,
      "end": 917.8,
      "text": "真正能说明问题的",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 392,
      "start": 917.8,
      "end": 919.966,
      "text": "是发布时间晚于训练截止",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 393,
      "start": 920.2,
      "end": 921.733,
      "text": "任务过程可审计",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 394,
      "start": 922.0,
      "end": 924.466,
      "text": "失败案例也完整公开的新测试",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 395,
      "start": 924.933,
      "end": 927.733,
      "text": "匿名表最缺的恰恰是这些细节",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 396,
      "start": 927.9,
      "end": 929.2,
      "text": "它给出了结果",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 397,
      "start": 929.466,
      "end": 932.3,
      "text": "却没有给出让外界相信结果的路径",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 398,
      "start": 932.466,
      "end": 934.166,
      "text": "不过Google并不需要在每个",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 399,
      "start": 934.166,
      "end": 937.1,
      "text": "榜单上第一才足以改变竞争",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 400,
      "start": 937.3,
      "end": 938.9,
      "text": "它拥有搜索和Chrome",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 401,
      "start": 939.0,
      "end": 941.533,
      "text": "可以让模型直接接触实时网页",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 402,
      "start": 941.866,
      "end": 942.766,
      "text": "拥有Android",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 403,
      "start": 942.8,
      "end": 944.6,
      "text": "可以让代理进入手机",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 404,
      "start": 944.8,
      "end": 945.866,
      "text": "拥有Workspace",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 405,
      "start": 946.133,
      "end": 947.3,
      "text": "可以把代理放进",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 406,
      "start": 947.3,
      "end": 949.5,
      "text": "企业文档邮件和会议",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 407,
      "start": 949.666,
      "end": 951.2,
      "text": "拥有云与TPU",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 408,
      "start": 951.3,
      "end": 952.8,
      "text": "可以控制推理成本",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 409,
      "start": 952.966,
      "end": 955.566,
      "text": "Astra和fable卖的是模型能力",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 410,
      "start": 955.8,
      "end": 957.866,
      "text": "Google有机会卖一整套默认",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 411,
      "start": 957.866,
      "end": 960.1,
      "text": "存在于用户工作流里的能力",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 412,
      "start": 960.5,
      "end": 962.7,
      "text": "只要Gemini4达到同一水平",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 413,
      "start": 962.866,
      "end": 964.966,
      "text": "分发优势就会开始放大",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 414,
      "start": 965.0,
      "end": 966.466,
      "text": "这也是机器人控制",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 415,
      "start": 966.466,
      "end": 968.166,
      "text": "传闻为什么格外吸睛",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 416,
      "start": 968.866,
      "end": 971.366,
      "text": "Google deep mind一直把数字智能与",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 417,
      "start": 971.366,
      "end": 973.9,
      "text": "具身智能放在同一张路线图上",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 418,
      "start": 973.966,
      "end": 975.5,
      "text": "如果一个统一多模态",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 419,
      "start": 975.5,
      "end": 978.2,
      "text": "模型能看懂环境规划动作",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 420,
      "start": 978.3,
      "end": 980.666,
      "text": "再把指令翻译成电机控制",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 421,
      "start": 980.866,
      "end": 984.133,
      "text": "Gemini就会从聊天窗口走向物理世界",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 422,
      "start": 984.266,
      "end": 987.8,
      "text": "但支持机器人协议和能够可靠控制",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 423,
      "start": 988.4,
      "end": 991.266,
      "text": "机器人之间隔着感知误差",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 424,
      "start": 991.466,
      "end": 994.733,
      "text": "安全约束实时性与硬件差异",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 425,
      "start": 994.866,
      "end": 996.366,
      "text": "一个后台字段无法",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 426,
      "start": 996.366,
      "end": 997.9,
      "text": "填平这些工程距离",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 427,
      "start": 998.0,
      "end": 999.333,
      "text": "从竞争节奏看",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 428,
      "start": 999.4,
      "end": 1000.9,
      "text": "Google还有一个理由把未",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 429,
      "start": 1000.9,
      "end": 1003.166,
      "text": "发布检查点放进公开盲测",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 430,
      "start": 1003.6,
      "end": 1005.366,
      "text": "内部基准容易让团队",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 431,
      "start": 1005.366,
      "end": 1007.1,
      "text": "围绕固定指标优化",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 432,
      "start": 1007.2,
      "end": 1008.9,
      "text": "真实用户却会提出",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 433,
      "start": 1008.9,
      "end": 1010.566,
      "text": "研究者想不到的任务",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 434,
      "start": 1010.8,
      "end": 1014.066,
      "text": "Arena里的SVG游戏和网页挑战",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 435,
      "start": 1014.066,
      "end": 1016.066,
      "text": "正好检验模型面对开放",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 436,
      "start": 1016.066,
      "end": 1018.3,
      "text": "需求时能不能自己拆任务",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 437,
      "start": 1018.7,
      "end": 1021.0,
      "text": "即使测试模型不是Gemini4",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 438,
      "start": 1021.333,
      "end": 1023.533,
      "text": "这种测试方式也可能成为",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 439,
      "start": 1023.533,
      "end": 1025.566,
      "text": "Google后训练的重要部分",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 440,
      "start": 1026.166,
      "end": 1028.4,
      "text": "先让真实用户暴露失败",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 441,
      "start": 1028.733,
      "end": 1030.8,
      "text": "再针对失败改进系统",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 442,
      "start": 1030.933,
      "end": 1032.733,
      "text": "泄露背后真正的变量",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 443,
      "start": 1032.733,
      "end": 1035.0,
      "text": "因此可能不是某个神秘参数",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 444,
      "start": 1035.5,
      "end": 1037.166,
      "text": "而是产品开发方式",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 445,
      "start": 1037.3,
      "end": 1039.4,
      "text": "过去模型先在实验室完成",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 446,
      "start": 1039.4,
      "end": 1040.666,
      "text": "再发布给用户",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 447,
      "start": 1040.866,
      "end": 1042.733,
      "text": "现在候选模型可能在匿名",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 448,
      "start": 1042.733,
      "end": 1044.866,
      "text": "环境中接受公开压力测试",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 449,
      "start": 1045.333,
      "end": 1046.7,
      "text": "社区一边猜身份",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 450,
      "start": 1046.7,
      "end": 1049.166,
      "text": "一边免费提供最刁钻的测试集",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 451,
      "start": 1049.4,
      "end": 1050.766,
      "text": "公司获得了数据",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 452,
      "start": 1050.866,
      "end": 1052.6,
      "text": "用户获得了提前体验",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 453,
      "start": 1052.733,
      "end": 1054.066,
      "text": "媒体获得了故事",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 454,
      "start": 1054.333,
      "end": 1056.366,
      "text": "唯一缺少的是每条",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 455,
      "start": 1056.366,
      "end": 1058.2,
      "text": "结论应有的证据标签",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 456,
      "start": 1058.366,
      "end": 1060.166,
      "text": "还有一个判断不能回避",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 457,
      "start": 1060.866,
      "end": 1062.533,
      "text": "泄露文化本身已经成为",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 458,
      "start": 1062.533,
      "end": 1064.6,
      "text": "前沿模型营销的一部分",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 459,
      "start": 1065.066,
      "end": 1066.933,
      "text": "公司不必正式宣布",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 460,
      "start": 1067.133,
      "end": 1069.333,
      "text": "匿名测试模糊代号和",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 461,
      "start": 1069.333,
      "end": 1071.533,
      "text": "社区截图就能制造关注",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 462,
      "start": 1071.666,
      "end": 1073.766,
      "text": "开发者主动寻找神秘模型",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 463,
      "start": 1074.0,
      "end": 1075.766,
      "text": "媒体主动拼接线索",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 464,
      "start": 1076.166,
      "end": 1077.733,
      "text": "竞争对手也被迫回应",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 465,
      "start": 1077.733,
      "end": 1079.466,
      "text": "一个尚未发布的产品",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 466,
      "start": 1079.6,
      "end": 1082.5,
      "text": "哪怕最后证明只是高思考版Flash",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 467,
      "start": 1082.8,
      "end": 1084.7,
      "text": "Google仍然获得了大量免费",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 468,
      "start": 1084.7,
      "end": 1087.066,
      "text": "的能力展示和市场预热",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 469,
      "start": 1087.133,
      "end": 1088.933,
      "text": "可这套玩法也有风险",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 470,
      "start": 1089.366,
      "end": 1092.266,
      "text": "一旦传闻把预期抬到1,000万上下文",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 471,
      "start": 1092.666,
      "end": 1094.5,
      "text": "25.6万输出",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 472,
      "start": 1094.766,
      "end": 1097.1,
      "text": "全基准第一和极低价格",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 473,
      "start": 1097.533,
      "end": 1099.333,
      "text": "正式产品只要少一项",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 474,
      "start": 1099.733,
      "end": 1101.3,
      "text": "就会被说成缩水",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 475,
      "start": 1101.466,
      "end": 1104.133,
      "text": "泄露不是只有竞争对手承受压力",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 476,
      "start": 1104.566,
      "end": 1106.266,
      "text": "Google也会被社区制造",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 477,
      "start": 1106.266,
      "end": 1108.0,
      "text": "的规格表反向绑架",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 478,
      "start": 1108.4,
      "end": 1110.9,
      "text": "尤其当匿名跑分不可复现时",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 479,
      "start": 1111.0,
      "end": 1113.266,
      "text": "用户对正式版本的评价可能",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 480,
      "start": 1113.266,
      "end": 1115.6,
      "text": "不再基于它实际进步了多少",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 481,
      "start": 1115.7,
      "end": 1117.1,
      "text": "而是看它有没有兑现",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 482,
      "start": 1117.1,
      "end": 1118.9,
      "text": "一张来历不明的清单",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 483,
      "start": 1118.966,
      "end": 1121.266,
      "text": "这件事当前最准确的结论并",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 484,
      "start": 1121.266,
      "end": 1124.5,
      "text": "不是Gemini4已经碾压Astra和fable",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 485,
      "start": 1124.733,
      "end": 1126.2,
      "text": "而是三层判断",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 486,
      "start": 1126.366,
      "end": 1127.866,
      "text": "第一层已经确认",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 487,
      "start": 1128.2,
      "end": 1130.166,
      "text": "Gemini4项目真实存在",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 488,
      "start": 1130.366,
      "end": 1133.2,
      "text": "Google正在进行规模空前的预训练",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 489,
      "start": 1133.5,
      "end": 1135.933,
      "text": "Gemini3.8Flash已经发布",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 490,
      "start": 1136.133,
      "end": 1138.566,
      "text": "Google也在公开研究更长程的代理",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 491,
      "start": 1138.566,
      "end": 1141.1,
      "text": "循环和低成本自我改进策略",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 492,
      "start": 1141.266,
      "end": 1142.966,
      "text": "第二层有观察支撑",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 493,
      "start": 1143.4,
      "end": 1145.5,
      "text": "Arena里确实有人遇到能力",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 494,
      "start": 1145.5,
      "end": 1147.333,
      "text": "与标签不匹配的模型",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 495,
      "start": 1147.6,
      "end": 1150.966,
      "text": "复杂SVG三维网页和游戏案例",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 496,
      "start": 1151.1,
      "end": 1152.666,
      "text": "说明某个测试系统",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 497,
      "start": 1152.666,
      "end": 1154.6,
      "text": "投入了更高的推理预算",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 498,
      "start": 1154.9,
      "end": 1157.3,
      "text": "它可能属于Google的内部检查点",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 499,
      "start": 1157.8,
      "end": 1160.566,
      "text": "也可能是现有模型的特殊配置",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 500,
      "start": 1160.7,
      "end": 1162.7,
      "text": "第三层仍然缺证据",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 501,
      "start": 1162.933,
      "end": 1164.8,
      "text": "模型就是Gemini4Pro",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 502,
      "start": 1164.9,
      "end": 1166.966,
      "text": "四项跑分全面领先",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 503,
      "start": 1167.166,
      "end": 1169.7,
      "text": "价格只有对手的约1/4",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 504,
      "start": 1169.866,
      "end": 1171.6,
      "text": "支持1,000万上下文",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 505,
      "start": 1171.766,
      "end": 1174.066,
      "text": "永久记忆和机器人控制",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 506,
      "start": 1174.4,
      "end": 1177.766,
      "text": "以及RSI让它完成了底层自我进化",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 507,
      "start": 1177.9,
      "end": 1179.5,
      "text": "把第三层当成新闻",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 508,
      "start": 1179.5,
      "end": 1181.466,
      "text": "事实会让文章很炸",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 509,
      "start": 1181.933,
      "end": 1184.4,
      "text": "把三层分开才能让判断",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 510,
      "start": 1184.4,
      "end": 1186.6,
      "text": "经得住正式发布后的核对",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 511,
      "start": 1186.6,
      "end": 1189.3,
      "text": "但这并不代表整场泄露没有价值",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 512,
      "start": 1189.733,
      "end": 1190.7,
      "text": "恰恰相反",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 513,
      "start": 1190.733,
      "end": 1191.8,
      "text": "它提前暴露了",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 514,
      "start": 1191.8,
      "end": 1193.733,
      "text": "下一代模型竞争的方向",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 515,
      "start": 1194.2,
      "end": 1197.133,
      "text": "第一测试时计算会继续增加",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 516,
      "start": 1197.333,
      "end": 1199.766,
      "text": "模型不再追求每个问题都秒回",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 517,
      "start": 1199.866,
      "end": 1202.4,
      "text": "而是按任务价值决定思考多久",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 518,
      "start": 1202.5,
      "end": 1205.733,
      "text": "第二前端与三维生成正在从能写",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 519,
      "start": 1205.733,
      "end": 1208.866,
      "text": "代码进入能形成完整作品第三",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 520,
      "start": 1209.166,
      "end": 1211.5,
      "text": "智能体开始保存探索历史",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 521,
      "start": 1211.6,
      "end": 1213.8,
      "text": "减少重复试错第四",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 522,
      "start": 1213.8,
      "end": 1216.533,
      "text": "价格会成为能力之外最锋利的武器",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 523,
      "start": 1216.9,
      "end": 1219.7,
      "text": "如果Gemini4最终接近流传表现",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 524,
      "start": 1220.0,
      "end": 1222.333,
      "text": "Astra和fable真正面对的就",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 525,
      "start": 1222.333,
      "end": 1223.9,
      "text": "不是一次跑分失利",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 526,
      "start": 1224.1,
      "end": 1226.666,
      "text": "而是Google把前沿能力压进自有",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 527,
      "start": 1226.666,
      "end": 1229.1,
      "text": "基础设施后发动的成本战争",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 528,
      "start": 1229.3,
      "end": 1231.966,
      "text": "如果正式产品与传闻差距很大",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 529,
      "start": 1232.0,
      "end": 1234.3,
      "text": "这场风波也说明了另一件事",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 530,
      "start": 1234.733,
      "end": 1236.133,
      "text": "市场已经迫不及待的",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 531,
      "start": 1236.133,
      "end": 1237.933,
      "text": "需要一个能把智能体",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 532,
      "start": 1238.166,
      "end": 1239.266,
      "text": "长上下文",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 533,
      "start": 1239.366,
      "end": 1242.733,
      "text": "视觉代码和自我改进串起来的模型",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 534,
      "start": 1243.0,
      "end": 1244.9,
      "text": "需求先于产品出现",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 535,
      "start": 1244.966,
      "end": 1246.733,
      "text": "才会让每张截图都像",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 536,
      "start": 1246.733,
      "end": 1249.2,
      "text": "未来泄露出来的证据所以",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 537,
      "start": 1249.666,
      "end": 1250.8,
      "text": "最值得等待的",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 538,
      "start": 1250.8,
      "end": 1253.6,
      "text": "不是Google什么时候揭掉Arena里的伪装",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 539,
      "start": 1253.733,
      "end": 1256.766,
      "text": "而是正式模型卡能否回答四个问题",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 540,
      "start": 1256.966,
      "end": 1258.666,
      "text": "这些成绩能否复现",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 541,
      "start": 1259.1,
      "end": 1260.666,
      "text": "长任务能否稳定",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 542,
      "start": 1260.9,
      "end": 1262.6,
      "text": "单位成本能否成立",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 543,
      "start": 1262.933,
      "end": 1265.366,
      "text": "所谓RSI到底改进了哪一层",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 544,
      "start": 1265.533,
      "end": 1267.466,
      "text": "只有这四件事同时站住",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 545,
      "start": 1267.8,
      "end": 1270.866,
      "text": "Gemini 4才算真正跨过Astra和fable",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 546,
      "start": 1271.1,
      "end": 1273.066,
      "text": "否则今天被碾压的",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 547,
      "start": 1273.066,
      "end": 1274.866,
      "text": "可能不是两个竞争模型",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 548,
      "start": 1275.1,
      "end": 1277.4,
      "text": "而只是公众对匿名截图",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 549,
      "start": 1277.4,
      "end": 1279.1,
      "text": "最后一点必要的怀疑",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 550,
      "start": 1279.4,
      "end": 1282.666,
      "text": "事实核验Gemini 4正在训练属实",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 551,
      "start": 1283.0,
      "end": 1284.8,
      "text": "Google称其为公司最具",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 552,
      "start": 1284.8,
      "end": 1287.0,
      "text": "野心的预训练任务属实",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 553,
      "start": 1287.366,
      "end": 1290.4,
      "text": "Gemini 3.8Flash已经公开属实",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 554,
      "start": 1290.8,
      "end": 1294.333,
      "text": "dream RSI论文与相关实验存在属实",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 555,
      "start": 1294.8,
      "end": 1297.366,
      "text": "Arena中的异常模型就是Gemini4",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 556,
      "start": 1297.566,
      "end": 1298.533,
      "text": "尚未证实",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 557,
      "start": 1298.866,
      "end": 1302.0,
      "text": "88.7% 2,064error",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 558,
      "start": 1302.366,
      "end": 1306.266,
      "text": "95.3%和86.8%等泄露成绩",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 559,
      "start": 1306.733,
      "end": 1309.333,
      "text": "尚未由Google或完整评测记录证实",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 560,
      "start": 1309.566,
      "end": 1311.133,
      "text": "1,000万TOKEN输入",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 561,
      "start": 1311.2,
      "end": 1313.266,
      "text": "25.6万TOKEN输出",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 562,
      "start": 1313.333,
      "end": 1316.466,
      "text": "永久记忆与机器人控制尚未证实",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 563,
      "start": 1316.733,
      "end": 1318.7,
      "text": "Gemini是依靠RSI实现",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 564,
      "start": 1318.7,
      "end": 1320.3,
      "text": "底层模型自我进化",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    },
    {
      "id": 565,
      "start": 1320.3,
      "end": 1321.4,
      "text": "没有公开证据",
      "source": "video_source_subtitle",
      "authority": "video_source_subtitle"
    }
  ],
  "transcript_selection": {
    "schema": "videonote.transcript_selection.v1",
    "status": "SELECTED",
    "requested_language": "zh",
    "duration": 1322.0,
    "reference_path": "/Users/sagawa/AI_WORK/video_notes/url/20260922_000025__OpenAI_徹底完了_Google_秘密武器_Gemini_4_意外洩漏_這性能簡直是外星科技/reference_subtitle.vtt",
    "reference_kind": "manual",
    "reference_language": "zh",
    "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": 566,
      "duration_coverage": 0.919946
    }
  }
}