WEBVTT
Kind: captions
Language: zh-Hans

00:00:00.000 --> 00:00:01.120
9月30号

00:00:01.310 --> 00:00:02.560
也就是两周前

00:00:02.590 --> 00:00:04.000
Google发布了Gemini

00:00:04.110 --> 00:00:05.840
4系列的第一个模型

00:00:05.870 --> 00:00:07.520
叫Gemini 4 Argon

00:00:07.950 --> 00:00:09.680
官方博客的标题写的是

00:00:10.030 --> 00:00:12.400
我们前沿智能的下一个时代

00:00:12.910 --> 00:00:14.960
署名是Koray Kavukcuoglu

00:00:15.150 --> 00:00:17.280
Google DeepMind的高级副总裁

00:00:17.470 --> 00:00:19.760
也是Google的首席AI架构师

00:00:20.487 --> 00:00:21.760
一个新的大版本

00:00:21.950 --> 00:00:24.000
第一批用户却不是开发者

00:00:24.110 --> 00:00:25.440
也不是付费会员

00:00:25.710 --> 00:00:28.080
而是一群做网络安全防御的人

00:00:28.510 --> 00:00:29.600
而且给他们的

00:00:29.670 --> 00:00:31.920
是拿掉了网络安全护栏的版本

00:00:33.321 --> 00:00:34.731
看官方原文这一句

00:00:35.321 --> 00:00:36.651
对于可信的防御方

00:00:36.921 --> 00:00:38.091
和Google内部团队

00:00:38.441 --> 00:00:40.891
我们会发布不带网络安全护栏的

00:00:40.921 --> 00:00:42.971
Argon前面还有一句

00:00:43.241 --> 00:00:45.051
Argon能自主发现

00:00:45.161 --> 00:00:47.451
验证修补关键的软件漏洞

00:00:48.041 --> 00:00:49.291
护栏简单来说

00:00:49.321 --> 00:00:50.411
就是模型遇到

00:00:50.601 --> 00:00:51.771
帮我攻破这个系统

00:00:51.961 --> 00:00:54.731
这类请求时会直接拒绝的限制

00:00:55.241 --> 00:00:55.931
拿掉以后

00:00:55.961 --> 00:00:57.931
它就能放开手去找漏洞

00:00:58.041 --> 00:00:59.931
去验证漏洞能不能被利用

00:01:00.659 --> 00:01:02.144
这批人是通过一个叫

00:01:02.254 --> 00:01:03.905
Fairwind的计划拿到的

00:01:04.414 --> 00:01:05.664
Google DeepMind说

00:01:05.735 --> 00:01:08.065
这个计划有650多个合作方

00:01:08.414 --> 00:01:11.505
优先给政府和国家级的网络安全机构

00:01:11.535 --> 00:01:12.545
还有医疗

00:01:12.815 --> 00:01:13.744
电信能源

00:01:13.774 --> 00:01:15.664
金融这些关键基础设施

00:01:16.254 --> 00:01:17.425
申请要审核

00:01:17.535 --> 00:01:18.705
要做背景调查

00:01:18.735 --> 00:01:20.625
不允许转卖访问权限

00:01:20.655 --> 00:01:22.625
也不允许拿它写恶意软件

00:01:23.300 --> 00:01:24.625
到了10月13号

00:01:24.774 --> 00:01:25.425
也就是昨天

00:01:25.694 --> 00:01:27.104
第二批用户才开放

00:01:27.455 --> 00:01:28.865
付费的API用户

00:01:28.895 --> 00:01:30.785
还有Gemini App里Pro和

00:01:30.895 --> 00:01:32.145
Ultra的订阅用户

00:01:32.779 --> 00:01:34.562
那么它到底强不强

00:01:35.101 --> 00:01:37.791
Google放出了一张对比表拿Argon

00:01:37.901 --> 00:01:39.071
和三个对手比

00:01:39.261 --> 00:01:41.631
OpenAI的GPT-6 Astra

00:01:41.702 --> 00:01:43.911
Anthropic的Claude Fable 5.1

00:01:44.062 --> 00:01:46.031
和Claude Opus 5.5

00:01:46.142 --> 00:01:47.711
一共19项测试

00:01:47.741 --> 00:01:49.631
Argon拿了14项第一

00:01:50.274 --> 00:01:51.966
赢得最明显的是两类

00:01:52.316 --> 00:01:53.646
第一类是法律

00:01:53.836 --> 00:01:56.366
金融这种企业里的知识工作

00:01:56.396 --> 00:01:58.926
比如说律所Harvey出的法律Agent

00:01:59.036 --> 00:02:01.726
测试Argon拿了百分之19.6

00:02:02.077 --> 00:02:04.926
GPT-6 Astra只有百分之5.4

00:02:04.956 --> 00:02:06.207
差不多是三倍半

00:02:06.636 --> 00:02:09.487
不过绝对分数还不到百分之20

00:02:09.517 --> 00:02:10.526
十个法律任务

00:02:10.636 --> 00:02:11.647
也就做对两个

00:02:12.047 --> 00:02:13.876
第二类是超长的材料

00:02:14.066 --> 00:02:17.076
在25万到100万token这个长度上

00:02:17.106 --> 00:02:19.076
Argon是百分之84.2

00:02:19.266 --> 00:02:21.396
对手都在百分之72以下

00:02:21.666 --> 00:02:24.436
token可以理解成模型读文字的单位

00:02:24.706 --> 00:02:25.636
材料越长

00:02:25.666 --> 00:02:27.876
它和别家拉开的差距就越大

00:02:28.366 --> 00:02:29.456
不过这张表里

00:02:29.486 --> 00:02:31.375
也写着它输了5项像

00:02:31.406 --> 00:02:32.255
Terminal-bench

00:02:32.445 --> 00:02:35.375
OSWorld都是让模型自己在终端里

00:02:35.406 --> 00:02:37.536
在电脑上动手干活的测试

00:02:37.806 --> 00:02:40.176
这几项Argon输给了Claude Opus

00:02:40.205 --> 00:02:42.096
5.5和GPT-6 Astra

00:02:42.366 --> 00:02:43.616
Terminal-bench这一项

00:02:43.806 --> 00:02:45.775
Opus是百分之66.4

00:02:46.046 --> 00:02:48.335
Argon是百分之57.4

00:02:48.366 --> 00:02:49.135
落后9分

00:02:49.486 --> 00:02:52.416
官方把输的几项也老老实实列了出来

00:02:52.526 --> 00:02:53.695
这一点挺难得的

00:02:54.401 --> 00:02:55.521
写代码这块

00:02:55.551 --> 00:02:57.681
Google主推的是DeepSWE

00:02:57.711 --> 00:03:00.401
测的是周期很长的软件工程任务

00:03:00.591 --> 00:03:02.481
Argon百分之77.9

00:03:02.591 --> 00:03:05.761
GPT-6 Astra和Opus都是74出头

00:03:05.951 --> 00:03:08.001
领先了但领先不到4分

00:03:08.469 --> 00:03:09.559
第三方的结果

00:03:09.749 --> 00:03:12.359
目前只有Artificial Analysis一家

00:03:12.789 --> 00:03:15.959
他们把10项测试合成一个综合智能分

00:03:16.150 --> 00:03:18.200
Argon拿了53分

00:03:18.229 --> 00:03:19.880
和GPT-6 Astra一样

00:03:20.389 --> 00:03:21.079
可以看到

00:03:21.109 --> 00:03:23.959
图上绿色箭头指的就是Argon

00:03:24.029 --> 00:03:25.159
但排在它前面的

00:03:25.189 --> 00:03:26.200
还有三个

00:03:26.229 --> 00:03:28.999
分别是Opus 5.5是58分

00:03:29.430 --> 00:03:31.399
Sonnet 5.5是56分

00:03:31.749 --> 00:03:33.799
Fable 5.1也是53分

00:03:34.186 --> 00:03:36.040
Argon真正占优势的地方

00:03:36.070 --> 00:03:38.519
是成本跑完这一整套测试

00:03:38.550 --> 00:03:40.839
Argon花了2407美元

00:03:41.029 --> 00:03:44.120
GPT-6 Astra花了5324美元

00:03:44.390 --> 00:03:45.000
也就是说

00:03:45.190 --> 00:03:46.120
同样的分数

00:03:46.150 --> 00:03:47.399
花的钱不到一半

00:03:48.616 --> 00:03:51.176
接下来给大家看一个容易被忽略的

00:03:51.176 --> 00:03:53.386
变化就是输出的上限

00:03:54.099 --> 00:03:57.058
以前Gemini一次最多输出6万

00:03:57.247 --> 00:03:58.178
4千个token

00:03:58.207 --> 00:04:00.338
这次直接提到了100万

00:04:00.367 --> 00:04:01.457
差不多16倍

00:04:02.047 --> 00:04:03.297
官方的说法是

00:04:03.328 --> 00:04:04.977
模型有了足够的空间

00:04:05.087 --> 00:04:07.058
就能一口气想得更深

00:04:07.087 --> 00:04:08.417
把难题一次做完

00:04:08.969 --> 00:04:11.929
不过要注意的是默认的最大输出

00:04:11.929 --> 00:04:14.059
还是65536

00:04:14.409 --> 00:04:17.368
要专门打开一个叫long decode的长

00:04:17.368 --> 00:04:18.298
解码模式

00:04:18.409 --> 00:04:19.259
才能到100万

00:04:19.689 --> 00:04:22.699
而且这100万是输入和输出共用的

00:04:23.384 --> 00:04:25.602
那Google自己拿它干了些什么

00:04:26.032 --> 00:04:27.443
官方举了三个例子

00:04:28.057 --> 00:04:30.986
第一个一组Argon智能体去分析

00:04:31.016 --> 00:04:33.707
Google全公司数据中心的运行数据

00:04:34.136 --> 00:04:36.347
自己找出内存能优化的地方

00:04:36.456 --> 00:04:37.147
自己改掉

00:04:37.737 --> 00:04:38.827
全部上线以后

00:04:38.857 --> 00:04:41.147
释放了超过300 TB的内存

00:04:41.496 --> 00:04:43.227
相当于差不多一万台

00:04:43.537 --> 00:04:45.626
32G内存的电脑加在一起

00:04:46.285 --> 00:04:49.707
第二个优化量子算法里的一段程序

00:04:49.737 --> 00:04:50.667
几分钟时间

00:04:50.696 --> 00:04:53.467
就比已经发表的方案好了百分之40

00:04:54.117 --> 00:04:55.767
第三个把C和

00:04:55.876 --> 00:04:58.647
C++写的老代码迁移成Rust

00:04:59.076 --> 00:04:59.527
Rust

00:04:59.636 --> 00:05:02.327
是一种不容易出内存漏洞的编程语言

00:05:02.677 --> 00:05:04.727
其中有一个视频解码器

00:05:04.756 --> 00:05:05.686
迁移完以后

00:05:05.717 --> 00:05:08.887
比之前人写的Rust版本快了2.7倍

00:05:09.997 --> 00:05:11.290
再回到网络安全

00:05:11.641 --> 00:05:14.600
为什么这次要先给防守方看这张

00:05:14.600 --> 00:05:16.090
漏洞修复的排行榜

00:05:16.120 --> 00:05:17.691
叫CWE-bench

00:05:18.201 --> 00:05:20.010
Argon是百分之68

00:05:20.081 --> 00:05:20.730
并列第一

00:05:21.160 --> 00:05:22.170
并列的还有两个

00:05:22.600 --> 00:05:24.490
xAI的Grok 4.7

00:05:24.680 --> 00:05:26.170
和GPT-6 Astra

00:05:26.600 --> 00:05:27.290
也就是说

00:05:27.400 --> 00:05:29.050
在找漏洞这件事上

00:05:29.201 --> 00:05:31.400
几家顶级模型已经差不多站在

00:05:31.400 --> 00:05:32.330
同一个水平了

00:05:32.786 --> 00:05:33.515
这样的话

00:05:33.706 --> 00:05:36.315
关键就不是有没有模型能找到漏洞

00:05:36.345 --> 00:05:37.756
而是谁先用上

00:05:38.265 --> 00:05:39.435
Google的思路是

00:05:39.465 --> 00:05:42.075
让防守的人先拿到最强的版本

00:05:42.105 --> 00:05:44.876
赶在攻击者用上同级别的模型之前

00:05:44.985 --> 00:05:46.075
把漏洞先补上

00:05:46.505 --> 00:05:47.595
官方举了个例子

00:05:47.945 --> 00:05:50.475
安全公司Wiz用Argon做扫描

00:05:50.665 --> 00:05:53.515
在全球医院都在用的一款医疗软件里

00:05:53.545 --> 00:05:55.945
找到了一个会泄露个人敏感信息的

00:05:55.945 --> 00:05:56.715
严重漏洞

00:05:56.985 --> 00:05:58.955
之前的前沿模型都没发现

00:05:59.517 --> 00:06:00.607
这里有一个背景

00:06:01.117 --> 00:06:02.207
9月19号

00:06:02.237 --> 00:06:04.127
就在Argon发布前11天

00:06:04.557 --> 00:06:07.167
半岛电视台跟进了华尔街日报的报道

00:06:07.757 --> 00:06:10.077
Google的Gemini在安全测试里

00:06:10.077 --> 00:06:11.327
入侵了三家公司

00:06:11.597 --> 00:06:12.607
然后停了下来

00:06:13.117 --> 00:06:15.647
原因是测试环境配置出了错

00:06:15.757 --> 00:06:17.807
模型能连上真实的互联网

00:06:18.317 --> 00:06:20.047
一次是猜对了密码

00:06:20.077 --> 00:06:20.847
另外两次

00:06:20.957 --> 00:06:23.567
用的是网上公开泄露的账号凭据

00:06:24.157 --> 00:06:24.807
Google说

00:06:24.957 --> 00:06:27.167
每次都在完成之前停住了

00:06:27.197 --> 00:06:28.527
这不算模型失控

00:06:28.857 --> 00:06:30.664
但这件事至少说明

00:06:30.734 --> 00:06:32.265
这一代模型的攻防能力

00:06:32.415 --> 00:06:34.985
已经强到会在真实世界里出事故了

00:06:35.415 --> 00:06:37.415
我觉得这也是Google这次这么

00:06:37.415 --> 00:06:38.265
谨慎的原因

00:06:38.695 --> 00:06:40.104
官方博客里有一句话

00:06:40.455 --> 00:06:42.745
要安全地放出这种级别的能力

00:06:42.775 --> 00:06:44.025
必须分阶段来

00:06:44.732 --> 00:06:47.512
最后说说普通人怎么用要花多少钱

00:06:47.941 --> 00:06:49.111
Gemini App这边

00:06:49.461 --> 00:06:51.352
Pro和Ultra的订阅用户

00:06:51.542 --> 00:06:53.431
从昨天开始就能用上

00:06:53.701 --> 00:06:56.071
Google说它回答问题更直接

00:06:56.102 --> 00:06:57.191
少了车轱辘话

00:06:57.701 --> 00:06:59.431
另外Gemini的Mac版

00:06:59.461 --> 00:07:01.912
现在可以直接切到写代码的界面

00:07:02.341 --> 00:07:04.871
聊天也可以一键切换成任务模式

00:07:05.221 --> 00:07:06.472
你人不在电脑前

00:07:06.581 --> 00:07:07.511
它也能接着干

00:07:08.093 --> 00:07:09.352
API这边要注意

00:07:09.781 --> 00:07:10.871
免费层用不了

00:07:11.302 --> 00:07:13.271
价格是每一百万个token

00:07:13.302 --> 00:07:14.391
输入2美元

00:07:14.422 --> 00:07:15.431
输出10美元

00:07:15.941 --> 00:07:17.232
但这是限时价

00:07:17.461 --> 00:07:19.111
12月1号开始翻倍

00:07:19.141 --> 00:07:21.032
变成4美元和20美元

00:07:21.542 --> 00:07:22.472
就算翻倍以后

00:07:22.662 --> 00:07:25.431
也只有GPT-6 Astra的百分之40

00:07:27.198 --> 00:07:28.049
综合来说

00:07:28.478 --> 00:07:30.128
Argon不是全面第一

00:07:30.559 --> 00:07:32.369
让模型自己操作电脑

00:07:32.398 --> 00:07:33.648
在终端里干活

00:07:33.679 --> 00:07:35.729
它还落后于Claude和GPT

00:07:36.319 --> 00:07:39.808
它真正强的是三件事读很长的材料

00:07:39.838 --> 00:07:40.688
处理法律

00:07:40.718 --> 00:07:42.528
金融这种企业文档

00:07:42.638 --> 00:07:43.409
还有找漏洞

00:07:43.918 --> 00:07:46.208
再加上价格不到OpenAI的一半

00:07:46.638 --> 00:07:48.929
如果你的工作是整理大量文档

00:07:48.958 --> 00:07:50.208
做长篇的分析

00:07:50.239 --> 00:07:52.944
它值得一试想自己验证的话

00:07:52.974 --> 00:07:54.544
有个很简单的办法

00:07:54.574 --> 00:07:56.864
找一份你手头最长的资料

00:07:56.894 --> 00:07:58.304
几百页的合同也好

00:07:58.334 --> 00:07:59.984
一整年的财报也好

00:08:00.094 --> 00:08:01.264
整份丢给它

00:08:01.374 --> 00:08:04.544
问一个需要前后对照才能回答的问题

00:08:04.574 --> 00:08:05.664
再拿同一个问题

00:08:05.774 --> 00:08:07.744
去问你现在在用的模型

00:08:07.774 --> 00:08:08.704
哪个答得对

00:08:08.734 --> 00:08:09.584
马上就清楚了