{ "text": "9月30号\n也就是两周前\nGoogle发布了Gemini\n4系列的第一个模型\n叫Gemini 4 Argon\n官方博客的标题写的是\n我们前沿智能的下一个时代\n署名是Koray Kavukcuoglu\nGoogle DeepMind的高级副总裁\n也是Google的首席AI架构师\n一个新的大版本\n第一批用户却不是开发者\n也不是付费会员\n而是一群做网络安全防御的人\n而且给他们的\n是拿掉了网络安全护栏的版本\n看官方原文这一句\n对于可信的防御方\n和Google内部团队\n我们会发布不带网络安全护栏的\nArgon前面还有一句\nArgon能自主发现\n验证修补关键的软件漏洞\n护栏简单来说\n就是模型遇到\n帮我攻破这个系统\n这类请求时会直接拒绝的限制\n拿掉以后\n它就能放开手去找漏洞\n去验证漏洞能不能被利用\n这批人是通过一个叫\nFairwind的计划拿到的\nGoogle DeepMind说\n这个计划有650多个合作方\n优先给政府和国家级的网络安全机构\n还有医疗\n电信能源\n金融这些关键基础设施\n申请要审核\n要做背景调查\n不允许转卖访问权限\n也不允许拿它写恶意软件\n到了10月13号\n也就是昨天\n第二批用户才开放\n付费的API用户\n还有Gemini App里Pro和\nUltra的订阅用户\n那么它到底强不强\nGoogle放出了一张对比表拿Argon\n和三个对手比\nOpenAI的GPT-6 Astra\nAnthropic的Claude Fable 5.1\n和Claude Opus 5.5\n一共19项测试\nArgon拿了14项第一\n赢得最明显的是两类\n第一类是法律\n金融这种企业里的知识工作\n比如说律所Harvey出的法律Agent\n测试Argon拿了百分之19.6\nGPT-6 Astra只有百分之5.4\n差不多是三倍半\n不过绝对分数还不到百分之20\n十个法律任务\n也就做对两个\n第二类是超长的材料\n在25万到100万token这个长度上\nArgon是百分之84.2\n对手都在百分之72以下\ntoken可以理解成模型读文字的单位\n材料越长\n它和别家拉开的差距就越大\n不过这张表里\n也写着它输了5项像\nTerminal-bench\nOSWorld都是让模型自己在终端里\n在电脑上动手干活的测试\n这几项Argon输给了Claude Opus\n5.5和GPT-6 Astra\nTerminal-bench这一项\nOpus是百分之66.4\nArgon是百分之57.4\n落后9分\n官方把输的几项也老老实实列了出来\n这一点挺难得的\n写代码这块\nGoogle主推的是DeepSWE\n测的是周期很长的软件工程任务\nArgon百分之77.9\nGPT-6 Astra和Opus都是74出头\n领先了但领先不到4分\n第三方的结果\n目前只有Artificial Analysis一家\n他们把10项测试合成一个综合智能分\nArgon拿了53分\n和GPT-6 Astra一样\n可以看到\n图上绿色箭头指的就是Argon\n但排在它前面的\n还有三个\n分别是Opus 5.5是58分\nSonnet 5.5是56分\nFable 5.1也是53分\nArgon真正占优势的地方\n是成本跑完这一整套测试\nArgon花了2407美元\nGPT-6 Astra花了5324美元\n也就是说\n同样的分数\n花的钱不到一半\n接下来给大家看一个容易被忽略的\n变化就是输出的上限\n以前Gemini一次最多输出6万\n4千个token\n这次直接提到了100万\n差不多16倍\n官方的说法是\n模型有了足够的空间\n就能一口气想得更深\n把难题一次做完\n不过要注意的是默认的最大输出\n还是65536\n要专门打开一个叫long decode的长\n解码模式\n才能到100万\n而且这100万是输入和输出共用的\n那Google自己拿它干了些什么\n官方举了三个例子\n第一个一组Argon智能体去分析\nGoogle全公司数据中心的运行数据\n自己找出内存能优化的地方\n自己改掉\n全部上线以后\n释放了超过300 TB的内存\n相当于差不多一万台\n32G内存的电脑加在一起\n第二个优化量子算法里的一段程序\n几分钟时间\n就比已经发表的方案好了百分之40\n第三个把C和\nC++写的老代码迁移成Rust\nRust\n是一种不容易出内存漏洞的编程语言\n其中有一个视频解码器\n迁移完以后\n比之前人写的Rust版本快了2.7倍\n再回到网络安全\n为什么这次要先给防守方看这张\n漏洞修复的排行榜\n叫CWE-bench\nArgon是百分之68\n并列第一\n并列的还有两个\nxAI的Grok 4.7\n和GPT-6 Astra\n也就是说\n在找漏洞这件事上\n几家顶级模型已经差不多站在\n同一个水平了\n这样的话\n关键就不是有没有模型能找到漏洞\n而是谁先用上\nGoogle的思路是\n让防守的人先拿到最强的版本\n赶在攻击者用上同级别的模型之前\n把漏洞先补上\n官方举了个例子\n安全公司Wiz用Argon做扫描\n在全球医院都在用的一款医疗软件里\n找到了一个会泄露个人敏感信息的\n严重漏洞\n之前的前沿模型都没发现\n这里有一个背景\n9月19号\n就在Argon发布前11天\n半岛电视台跟进了华尔街日报的报道\nGoogle的Gemini在安全测试里\n入侵了三家公司\n然后停了下来\n原因是测试环境配置出了错\n模型能连上真实的互联网\n一次是猜对了密码\n另外两次\n用的是网上公开泄露的账号凭据\nGoogle说\n每次都在完成之前停住了\n这不算模型失控\n但这件事至少说明\n这一代模型的攻防能力\n已经强到会在真实世界里出事故了\n我觉得这也是Google这次这么\n谨慎的原因\n官方博客里有一句话\n要安全地放出这种级别的能力\n必须分阶段来\n最后说说普通人怎么用要花多少钱\nGemini App这边\nPro和Ultra的订阅用户\n从昨天开始就能用上\nGoogle说它回答问题更直接\n少了车轱辘话\n另外Gemini的Mac版\n现在可以直接切到写代码的界面\n聊天也可以一键切换成任务模式\n你人不在电脑前\n它也能接着干\nAPI这边要注意\n免费层用不了\n价格是每一百万个token\n输入2美元\n输出10美元\n但这是限时价\n12月1号开始翻倍\n变成4美元和20美元\n就算翻倍以后\n也只有GPT-6 Astra的百分之40\n综合来说\nArgon不是全面第一\n让模型自己操作电脑\n在终端里干活\n它还落后于Claude和GPT\n它真正强的是三件事读很长的材料\n处理法律\n金融这种企业文档\n还有找漏洞\n再加上价格不到OpenAI的一半\n如果你的工作是整理大量文档\n做长篇的分析\n它值得一试想自己验证的话\n有个很简单的办法\n找一份你手头最长的资料\n几百页的合同也好\n一整年的财报也好\n整份丢给它\n问一个需要前后对照才能回答的问题\n再拿同一个问题\n去问你现在在用的模型\n哪个答得对\n马上就清楚了", "language": "zh", "duration": 492.0, "segments": [ { "id": 0, "start": 0.0, "end": 1.12, "text": "9月30号", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 1, "start": 1.31, "end": 2.56, "text": "也就是两周前", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 2, "start": 2.59, "end": 4.0, "text": "Google发布了Gemini", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 3, "start": 4.11, "end": 5.84, "text": "4系列的第一个模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 4, "start": 5.87, "end": 7.52, "text": "叫Gemini 4 Argon", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 5, "start": 7.95, "end": 9.68, "text": "官方博客的标题写的是", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 6, "start": 10.03, "end": 12.4, "text": "我们前沿智能的下一个时代", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 7, "start": 12.91, "end": 14.96, "text": "署名是Koray Kavukcuoglu", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 8, "start": 15.15, "end": 17.28, "text": "Google DeepMind的高级副总裁", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 9, "start": 17.47, "end": 19.76, "text": "也是Google的首席AI架构师", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 10, "start": 20.487, "end": 21.76, "text": "一个新的大版本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 11, "start": 21.95, "end": 24.0, "text": "第一批用户却不是开发者", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 12, "start": 24.11, "end": 25.44, "text": "也不是付费会员", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 13, "start": 25.71, "end": 28.08, "text": "而是一群做网络安全防御的人", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 14, "start": 28.51, "end": 29.6, "text": "而且给他们的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 15, "start": 29.67, "end": 31.92, "text": "是拿掉了网络安全护栏的版本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 16, "start": 33.321, "end": 34.731, "text": "看官方原文这一句", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 17, "start": 35.321, "end": 36.651, "text": "对于可信的防御方", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 18, "start": 36.921, "end": 38.091, "text": "和Google内部团队", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 19, "start": 38.441, "end": 40.891, "text": "我们会发布不带网络安全护栏的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 20, "start": 40.921, "end": 42.971, "text": "Argon前面还有一句", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 21, "start": 43.241, "end": 45.051, "text": "Argon能自主发现", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 22, "start": 45.161, "end": 47.451, "text": "验证修补关键的软件漏洞", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 23, "start": 48.041, "end": 49.291, "text": "护栏简单来说", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 24, "start": 49.321, "end": 50.411, "text": "就是模型遇到", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 25, "start": 50.601, "end": 51.771, "text": "帮我攻破这个系统", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 26, "start": 51.961, "end": 54.731, "text": "这类请求时会直接拒绝的限制", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 27, "start": 55.241, "end": 55.931, "text": "拿掉以后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 28, "start": 55.961, "end": 57.931, "text": "它就能放开手去找漏洞", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 29, "start": 58.041, "end": 59.931, "text": "去验证漏洞能不能被利用", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 30, "start": 60.659, "end": 62.144, "text": "这批人是通过一个叫", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 31, "start": 62.254, "end": 63.905, "text": "Fairwind的计划拿到的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 32, "start": 64.414, "end": 65.664, "text": "Google DeepMind说", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 33, "start": 65.735, "end": 68.065, "text": "这个计划有650多个合作方", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 34, "start": 68.414, "end": 71.505, "text": "优先给政府和国家级的网络安全机构", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 35, "start": 71.535, "end": 72.545, "text": "还有医疗", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 36, "start": 72.815, "end": 73.744, "text": "电信能源", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 37, "start": 73.774, "end": 75.664, "text": "金融这些关键基础设施", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 38, "start": 76.254, "end": 77.425, "text": "申请要审核", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 39, "start": 77.535, "end": 78.705, "text": "要做背景调查", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 40, "start": 78.735, "end": 80.625, "text": "不允许转卖访问权限", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 41, "start": 80.655, "end": 82.625, "text": "也不允许拿它写恶意软件", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 42, "start": 83.3, "end": 84.625, "text": "到了10月13号", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 43, "start": 84.774, "end": 85.425, "text": "也就是昨天", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 44, "start": 85.694, "end": 87.104, "text": "第二批用户才开放", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 45, "start": 87.455, "end": 88.865, "text": "付费的API用户", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 46, "start": 88.895, "end": 90.785, "text": "还有Gemini App里Pro和", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 47, "start": 90.895, "end": 92.145, "text": "Ultra的订阅用户", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 48, "start": 92.779, "end": 94.562, "text": "那么它到底强不强", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 49, "start": 95.101, "end": 97.791, "text": "Google放出了一张对比表拿Argon", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 50, "start": 97.901, "end": 99.071, "text": "和三个对手比", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 51, "start": 99.261, "end": 101.631, "text": "OpenAI的GPT-6 Astra", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 52, "start": 101.702, "end": 103.911, "text": "Anthropic的Claude Fable 5.1", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 53, "start": 104.062, "end": 106.031, "text": "和Claude Opus 5.5", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 54, "start": 106.142, "end": 107.711, "text": "一共19项测试", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 55, "start": 107.741, "end": 109.631, "text": "Argon拿了14项第一", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 56, "start": 110.274, "end": 111.966, "text": "赢得最明显的是两类", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 57, "start": 112.316, "end": 113.646, "text": "第一类是法律", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 58, "start": 113.836, "end": 116.366, "text": "金融这种企业里的知识工作", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 59, "start": 116.396, "end": 118.926, "text": "比如说律所Harvey出的法律Agent", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 60, "start": 119.036, "end": 121.726, "text": "测试Argon拿了百分之19.6", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 61, "start": 122.077, "end": 124.926, "text": "GPT-6 Astra只有百分之5.4", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 62, "start": 124.956, "end": 126.207, "text": "差不多是三倍半", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 63, "start": 126.636, "end": 129.487, "text": "不过绝对分数还不到百分之20", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 64, "start": 129.517, "end": 130.526, "text": "十个法律任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 65, "start": 130.636, "end": 131.647, "text": "也就做对两个", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 66, "start": 132.047, "end": 133.876, "text": "第二类是超长的材料", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 67, "start": 134.066, "end": 137.076, "text": "在25万到100万token这个长度上", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 68, "start": 137.106, "end": 139.076, "text": "Argon是百分之84.2", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 69, "start": 139.266, "end": 141.396, "text": "对手都在百分之72以下", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 70, "start": 141.666, "end": 144.436, "text": "token可以理解成模型读文字的单位", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 71, "start": 144.706, "end": 145.636, "text": "材料越长", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 72, "start": 145.666, "end": 147.876, "text": "它和别家拉开的差距就越大", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 73, "start": 148.366, "end": 149.456, "text": "不过这张表里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 74, "start": 149.486, "end": 151.375, "text": "也写着它输了5项像", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 75, "start": 151.406, "end": 152.255, "text": "Terminal-bench", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 76, "start": 152.445, "end": 155.375, "text": "OSWorld都是让模型自己在终端里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 77, "start": 155.406, "end": 157.536, "text": "在电脑上动手干活的测试", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 78, "start": 157.806, "end": 160.176, "text": "这几项Argon输给了Claude Opus", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 79, "start": 160.205, "end": 162.096, "text": "5.5和GPT-6 Astra", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 80, "start": 162.366, "end": 163.616, "text": "Terminal-bench这一项", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 81, "start": 163.806, "end": 165.775, "text": "Opus是百分之66.4", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 82, "start": 166.046, "end": 168.335, "text": "Argon是百分之57.4", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 83, "start": 168.366, "end": 169.135, "text": "落后9分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 84, "start": 169.486, "end": 172.416, "text": "官方把输的几项也老老实实列了出来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 85, "start": 172.526, "end": 173.695, "text": "这一点挺难得的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 86, "start": 174.401, "end": 175.521, "text": "写代码这块", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 87, "start": 175.551, "end": 177.681, "text": "Google主推的是DeepSWE", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 88, "start": 177.711, "end": 180.401, "text": "测的是周期很长的软件工程任务", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 89, "start": 180.591, "end": 182.481, "text": "Argon百分之77.9", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 90, "start": 182.591, "end": 185.761, "text": "GPT-6 Astra和Opus都是74出头", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 91, "start": 185.951, "end": 188.001, "text": "领先了但领先不到4分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 92, "start": 188.469, "end": 189.559, "text": "第三方的结果", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 93, "start": 189.749, "end": 192.359, "text": "目前只有Artificial Analysis一家", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 94, "start": 192.789, "end": 195.959, "text": "他们把10项测试合成一个综合智能分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 95, "start": 196.15, "end": 198.2, "text": "Argon拿了53分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 96, "start": 198.229, "end": 199.88, "text": "和GPT-6 Astra一样", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 97, "start": 200.389, "end": 201.079, "text": "可以看到", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 98, "start": 201.109, "end": 203.959, "text": "图上绿色箭头指的就是Argon", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 99, "start": 204.029, "end": 205.159, "text": "但排在它前面的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 100, "start": 205.189, "end": 206.2, "text": "还有三个", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 101, "start": 206.229, "end": 208.999, "text": "分别是Opus 5.5是58分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 102, "start": 209.43, "end": 211.399, "text": "Sonnet 5.5是56分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 103, "start": 211.749, "end": 213.799, "text": "Fable 5.1也是53分", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 104, "start": 214.186, "end": 216.04, "text": "Argon真正占优势的地方", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 105, "start": 216.07, "end": 218.519, "text": "是成本跑完这一整套测试", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 106, "start": 218.55, "end": 220.839, "text": "Argon花了2407美元", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 107, "start": 221.029, "end": 224.12, "text": "GPT-6 Astra花了5324美元", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 108, "start": 224.39, "end": 225.0, "text": "也就是说", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 109, "start": 225.19, "end": 226.12, "text": "同样的分数", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 110, "start": 226.15, "end": 227.399, "text": "花的钱不到一半", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 111, "start": 228.616, "end": 231.176, "text": "接下来给大家看一个容易被忽略的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 112, "start": 231.176, "end": 233.386, "text": "变化就是输出的上限", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 113, "start": 234.099, "end": 237.058, "text": "以前Gemini一次最多输出6万", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 114, "start": 237.247, "end": 238.178, "text": "4千个token", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 115, "start": 238.207, "end": 240.338, "text": "这次直接提到了100万", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 116, "start": 240.367, "end": 241.457, "text": "差不多16倍", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 117, "start": 242.047, "end": 243.297, "text": "官方的说法是", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 118, "start": 243.328, "end": 244.977, "text": "模型有了足够的空间", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 119, "start": 245.087, "end": 247.058, "text": "就能一口气想得更深", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 120, "start": 247.087, "end": 248.417, "text": "把难题一次做完", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 121, "start": 248.969, "end": 251.929, "text": "不过要注意的是默认的最大输出", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 122, "start": 251.929, "end": 254.059, "text": "还是65536", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 123, "start": 254.409, "end": 257.368, "text": "要专门打开一个叫long decode的长", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 124, "start": 257.368, "end": 258.298, "text": "解码模式", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 125, "start": 258.409, "end": 259.259, "text": "才能到100万", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 126, "start": 259.689, "end": 262.699, "text": "而且这100万是输入和输出共用的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 127, "start": 263.384, "end": 265.602, "text": "那Google自己拿它干了些什么", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 128, "start": 266.032, "end": 267.443, "text": "官方举了三个例子", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 129, "start": 268.057, "end": 270.986, "text": "第一个一组Argon智能体去分析", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 130, "start": 271.016, "end": 273.707, "text": "Google全公司数据中心的运行数据", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 131, "start": 274.136, "end": 276.347, "text": "自己找出内存能优化的地方", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 132, "start": 276.456, "end": 277.147, "text": "自己改掉", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 133, "start": 277.737, "end": 278.827, "text": "全部上线以后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 134, "start": 278.857, "end": 281.147, "text": "释放了超过300 TB的内存", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 135, "start": 281.496, "end": 283.227, "text": "相当于差不多一万台", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 136, "start": 283.537, "end": 285.626, "text": "32G内存的电脑加在一起", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 137, "start": 286.285, "end": 289.707, "text": "第二个优化量子算法里的一段程序", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 138, "start": 289.737, "end": 290.667, "text": "几分钟时间", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 139, "start": 290.696, "end": 293.467, "text": "就比已经发表的方案好了百分之40", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 140, "start": 294.117, "end": 295.767, "text": "第三个把C和", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 141, "start": 295.876, "end": 298.647, "text": "C++写的老代码迁移成Rust", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 142, "start": 299.076, "end": 299.527, "text": "Rust", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 143, "start": 299.636, "end": 302.327, "text": "是一种不容易出内存漏洞的编程语言", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 144, "start": 302.677, "end": 304.727, "text": "其中有一个视频解码器", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 145, "start": 304.756, "end": 305.686, "text": "迁移完以后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 146, "start": 305.717, "end": 308.887, "text": "比之前人写的Rust版本快了2.7倍", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 147, "start": 309.997, "end": 311.29, "text": "再回到网络安全", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 148, "start": 311.641, "end": 314.6, "text": "为什么这次要先给防守方看这张", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 149, "start": 314.6, "end": 316.09, "text": "漏洞修复的排行榜", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 150, "start": 316.12, "end": 317.691, "text": "叫CWE-bench", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 151, "start": 318.201, "end": 320.01, "text": "Argon是百分之68", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 152, "start": 320.081, "end": 320.73, "text": "并列第一", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 153, "start": 321.16, "end": 322.17, "text": "并列的还有两个", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 154, "start": 322.6, "end": 324.49, "text": "xAI的Grok 4.7", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 155, "start": 324.68, "end": 326.17, "text": "和GPT-6 Astra", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 156, "start": 326.6, "end": 327.29, "text": "也就是说", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 157, "start": 327.4, "end": 329.05, "text": "在找漏洞这件事上", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 158, "start": 329.201, "end": 331.4, "text": "几家顶级模型已经差不多站在", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 159, "start": 331.4, "end": 332.33, "text": "同一个水平了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 160, "start": 332.786, "end": 333.515, "text": "这样的话", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 161, "start": 333.706, "end": 336.315, "text": "关键就不是有没有模型能找到漏洞", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 162, "start": 336.345, "end": 337.756, "text": "而是谁先用上", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 163, "start": 338.265, "end": 339.435, "text": "Google的思路是", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 164, "start": 339.465, "end": 342.075, "text": "让防守的人先拿到最强的版本", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 165, "start": 342.105, "end": 344.876, "text": "赶在攻击者用上同级别的模型之前", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 166, "start": 344.985, "end": 346.075, "text": "把漏洞先补上", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 167, "start": 346.505, "end": 347.595, "text": "官方举了个例子", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 168, "start": 347.945, "end": 350.475, "text": "安全公司Wiz用Argon做扫描", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 169, "start": 350.665, "end": 353.515, "text": "在全球医院都在用的一款医疗软件里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 170, "start": 353.545, "end": 355.945, "text": "找到了一个会泄露个人敏感信息的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 171, "start": 355.945, "end": 356.715, "text": "严重漏洞", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 172, "start": 356.985, "end": 358.955, "text": "之前的前沿模型都没发现", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 173, "start": 359.517, "end": 360.607, "text": "这里有一个背景", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 174, "start": 361.117, "end": 362.207, "text": "9月19号", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 175, "start": 362.237, "end": 364.127, "text": "就在Argon发布前11天", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 176, "start": 364.557, "end": 367.167, "text": "半岛电视台跟进了华尔街日报的报道", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 177, "start": 367.757, "end": 370.077, "text": "Google的Gemini在安全测试里", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 178, "start": 370.077, "end": 371.327, "text": "入侵了三家公司", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 179, "start": 371.597, "end": 372.607, "text": "然后停了下来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 180, "start": 373.117, "end": 375.647, "text": "原因是测试环境配置出了错", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 181, "start": 375.757, "end": 377.807, "text": "模型能连上真实的互联网", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 182, "start": 378.317, "end": 380.047, "text": "一次是猜对了密码", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 183, "start": 380.077, "end": 380.847, "text": "另外两次", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 184, "start": 380.957, "end": 383.567, "text": "用的是网上公开泄露的账号凭据", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 185, "start": 384.157, "end": 384.807, "text": "Google说", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 186, "start": 384.957, "end": 387.167, "text": "每次都在完成之前停住了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 187, "start": 387.197, "end": 388.527, "text": "这不算模型失控", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 188, "start": 388.857, "end": 390.664, "text": "但这件事至少说明", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 189, "start": 390.734, "end": 392.265, "text": "这一代模型的攻防能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 190, "start": 392.415, "end": 394.985, "text": "已经强到会在真实世界里出事故了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 191, "start": 395.415, "end": 397.415, "text": "我觉得这也是Google这次这么", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 192, "start": 397.415, "end": 398.265, "text": "谨慎的原因", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 193, "start": 398.695, "end": 400.104, "text": "官方博客里有一句话", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 194, "start": 400.455, "end": 402.745, "text": "要安全地放出这种级别的能力", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 195, "start": 402.775, "end": 404.025, "text": "必须分阶段来", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 196, "start": 404.732, "end": 407.512, "text": "最后说说普通人怎么用要花多少钱", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 197, "start": 407.941, "end": 409.111, "text": "Gemini App这边", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 198, "start": 409.461, "end": 411.352, "text": "Pro和Ultra的订阅用户", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 199, "start": 411.542, "end": 413.431, "text": "从昨天开始就能用上", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 200, "start": 413.701, "end": 416.071, "text": "Google说它回答问题更直接", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 201, "start": 416.102, "end": 417.191, "text": "少了车轱辘话", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 202, "start": 417.701, "end": 419.431, "text": "另外Gemini的Mac版", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 203, "start": 419.461, "end": 421.912, "text": "现在可以直接切到写代码的界面", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 204, "start": 422.341, "end": 424.871, "text": "聊天也可以一键切换成任务模式", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 205, "start": 425.221, "end": 426.472, "text": "你人不在电脑前", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 206, "start": 426.581, "end": 427.511, "text": "它也能接着干", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 207, "start": 428.093, "end": 429.352, "text": "API这边要注意", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 208, "start": 429.781, "end": 430.871, "text": "免费层用不了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 209, "start": 431.302, "end": 433.271, "text": "价格是每一百万个token", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 210, "start": 433.302, "end": 434.391, "text": "输入2美元", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 211, "start": 434.422, "end": 435.431, "text": "输出10美元", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 212, "start": 435.941, "end": 437.232, "text": "但这是限时价", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 213, "start": 437.461, "end": 439.111, "text": "12月1号开始翻倍", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 214, "start": 439.141, "end": 441.032, "text": "变成4美元和20美元", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 215, "start": 441.542, "end": 442.472, "text": "就算翻倍以后", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 216, "start": 442.662, "end": 445.431, "text": "也只有GPT-6 Astra的百分之40", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 217, "start": 447.198, "end": 448.049, "text": "综合来说", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 218, "start": 448.478, "end": 450.128, "text": "Argon不是全面第一", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 219, "start": 450.559, "end": 452.369, "text": "让模型自己操作电脑", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 220, "start": 452.398, "end": 453.648, "text": "在终端里干活", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 221, "start": 453.679, "end": 455.729, "text": "它还落后于Claude和GPT", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 222, "start": 456.319, "end": 459.808, "text": "它真正强的是三件事读很长的材料", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 223, "start": 459.838, "end": 460.688, "text": "处理法律", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 224, "start": 460.718, "end": 462.528, "text": "金融这种企业文档", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 225, "start": 462.638, "end": 463.409, "text": "还有找漏洞", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 226, "start": 463.918, "end": 466.208, "text": "再加上价格不到OpenAI的一半", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 227, "start": 466.638, "end": 468.929, "text": "如果你的工作是整理大量文档", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 228, "start": 468.958, "end": 470.208, "text": "做长篇的分析", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 229, "start": 470.239, "end": 472.944, "text": "它值得一试想自己验证的话", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 230, "start": 472.974, "end": 474.544, "text": "有个很简单的办法", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 231, "start": 474.574, "end": 476.864, "text": "找一份你手头最长的资料", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 232, "start": 476.894, "end": 478.304, "text": "几百页的合同也好", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 233, "start": 478.334, "end": 479.984, "text": "一整年的财报也好", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 234, "start": 480.094, "end": 481.264, "text": "整份丢给它", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 235, "start": 481.374, "end": 484.544, "text": "问一个需要前后对照才能回答的问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 236, "start": 484.574, "end": 485.664, "text": "再拿同一个问题", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 237, "start": 485.774, "end": 487.744, "text": "去问你现在在用的模型", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 238, "start": 487.774, "end": 488.704, "text": "哪个答得对", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 239, "start": 488.734, "end": 489.584, "text": "马上就清楚了", "source": "video_source_subtitle", "authority": "video_source_subtitle" } ], "transcript_selection": { "schema": "videonote.transcript_selection.v1", "status": "SELECTED", "requested_language": "zh", "duration": 492.0, "reference_path": "/Users/sagawa/AI_WORK/video_notes/url/20261014_105326__谷歌重磅推出_Gemini_4_Argon_14_项测试夺冠_硬刚_GPT-6_价格不到一半/reference_subtitle.vtt", "reference_kind": "manual", "reference_language": "zh-Hans", "selected_source": "uploaded_master_subtitle", "source_caption_kind": "manual", "source_caption_processing": "verbatim_direct", "youtube_overlap_clamps": [], "general_high_confidence_repairs": [], "decision_reason": "explicit_adopt_manual_no_whisper", "source_strategy": "ADOPT_MANUAL", "segmentation_policy": "PRESERVE_SOURCE_CUES", "whisper_rejected": false, "source_profile": { "segment_count": 240, "duration_coverage": 0.872591 } } }