WEBVTT
Kind: captions
Language: zh-HK

00:00:00.266 --> 00:00:01.900
谷歌又发布了新的版本

00:00:01.900 --> 00:00:03.566
而这次与上一代Gemini

00:00:03.566 --> 00:00:05.000
更新的时间相差无几

00:00:05.133 --> 00:00:06.400
甚至有些反常

00:00:06.766 --> 00:00:08.666
当人们看到这个内容的时候

00:00:08.766 --> 00:00:10.766
他们的第一反应大概就是

00:00:10.900 --> 00:00:13.300
谷歌又要开始堆版本的事情了

00:00:13.666 --> 00:00:15.466
但是你如果仔细观察Gemini

00:00:15.466 --> 00:00:16.766
这次更新的方向的话

00:00:16.766 --> 00:00:18.933
就会发现事情并没有那么简单

00:00:19.266 --> 00:00:22.000
由于谷歌此次想要解决的问题并非是

00:00:22.000 --> 00:00:24.800
如何再次创造出一个跑分第一的模型

00:00:24.933 --> 00:00:26.766
而是更加实际的一个问题

00:00:27.200 --> 00:00:29.300
人工智能到底要怎么做到真正的

00:00:29.300 --> 00:00:31.666
进入到人们的日常工作当中去呢

00:00:31.733 --> 00:00:34.900
是开发人员每天都可以愉快的使用它

00:00:35.000 --> 00:00:37.766
并且能够让企业长久的支付费用呢

00:00:38.000 --> 00:00:40.400
因此Gemini这次并没有把所有的注意力

00:00:40.400 --> 00:00:43.000
集中在哪个模型更聪明的问题上

00:00:43.066 --> 00:00:46.333
而是在逐步加强代码开发、工具使用

00:00:46.333 --> 00:00:48.800
和复杂的任务处理等方面的工作

00:00:48.933 --> 00:00:52.100
agent场景下稳定性、效率方面的表现

00:00:52.566 --> 00:00:53.366
也就是说

00:00:53.400 --> 00:00:55.700
谷歌想把Gemini从一个你

00:00:55.700 --> 00:00:57.966
偶尔发几个问题给它的AI中

00:00:57.966 --> 00:00:59.866
变成一个能够长久的

00:00:59.900 --> 00:01:01.900
为你服务的主力模型

00:01:02.300 --> 00:01:03.933
这次的变化看上去只是

00:01:03.933 --> 00:01:05.600
对软件的一个版本升级

00:01:05.700 --> 00:01:07.500
但是它所蕴含的意义却

00:01:07.500 --> 00:01:09.466
远远大于参数提升本身

00:01:09.866 --> 00:01:11.700
因为在人工智能代理真正

00:01:11.700 --> 00:01:13.533
的投入到工作中去之后

00:01:13.533 --> 00:01:14.966
一个任务往往要调用

00:01:14.966 --> 00:01:17.000
模型十几遍到几十遍不等

00:01:17.200 --> 00:01:18.566
此时人们比较的已经

00:01:18.566 --> 00:01:20.733
不是哪一个模型最厉害呢

00:01:20.733 --> 00:01:23.466
而会变成哪一个模型可以以较低

00:01:23.466 --> 00:01:26.333
的成本稳定的性能完成工作呢

00:01:26.666 --> 00:01:27.900
这样就会影响到

00:01:27.900 --> 00:01:30.300
人工智能行业的未来竞争模式

00:01:30.700 --> 00:01:32.300
因此今天我们这期的内容

00:01:32.300 --> 00:01:33.933
就不再只是单纯的去聊

00:01:33.933 --> 00:01:36.200
一下Gemini新版到底有多牛了

00:01:36.266 --> 00:01:37.666
而是从这次更新的背

00:01:37.666 --> 00:01:39.666
后来拆开来思考3个问题

00:01:39.900 --> 00:01:41.500
为什么谷歌会忽然如此

00:01:41.500 --> 00:01:43.733
重视一个不一定要追求第一

00:01:43.733 --> 00:01:46.100
但是非常擅长做事的模型呢

00:01:46.366 --> 00:01:48.100
在agent时代的背景下

00:01:48.133 --> 00:01:50.700
人工智能的价格是否可以不再仅仅

00:01:50.700 --> 00:01:53.400
以每百万个TOKEN多少钱来衡量呢

00:01:53.700 --> 00:01:57.466
另外谷歌、Openai、Deepseek等公司

00:01:57.466 --> 00:01:59.900
接下来又将会怎样进行竞争呢

00:02:00.066 --> 00:02:02.566
如果能解决好这三方面的问题的话

00:02:02.600 --> 00:02:04.000
那么你就能够发现

00:02:04.133 --> 00:02:05.966
谷歌此次发布的新模型

00:02:06.000 --> 00:02:07.866
并非只是单纯的创新

00:02:08.200 --> 00:02:10.500
在于它已经提前预判到了人工智能

00:02:10.500 --> 00:02:12.933
下一阶段最大应用场景的可能性

00:02:13.200 --> 00:02:15.700
大家好我叫财经翻译官

00:02:15.966 --> 00:02:17.566
本期我们从Gemini这次

00:02:17.566 --> 00:02:19.266
的新版本发布说起

00:02:19.300 --> 00:02:21.066
把谷歌想要抢走的东西

00:02:21.066 --> 00:02:22.900
一层层的拆开来给你看

00:02:22.966 --> 00:02:24.600
再说一下这次更新中

00:02:24.600 --> 00:02:26.266
最不符合常理的部分

00:02:26.700 --> 00:02:27.733
过去的几年里

00:02:27.733 --> 00:02:29.900
在人工智能行业的发布会上

00:02:29.900 --> 00:02:32.000
大家都在比的是谁能在最

00:02:32.000 --> 00:02:34.400
困难的基准测试中拿到第一名

00:02:34.666 --> 00:02:36.333
或者解出其他模型无法

00:02:36.333 --> 00:02:38.066
解答的数学题和推理题

00:02:38.100 --> 00:02:39.933
好像模型只要足够聪明

00:02:39.933 --> 00:02:41.600
就可以一统江湖了

00:02:41.800 --> 00:02:45.000
但是这次Gemini3.7Flash并没有

00:02:45.000 --> 00:02:47.300
想要争夺最强大脑的称号

00:02:47.700 --> 00:02:50.400
谷歌给自己定的位置是Workhorse model

00:02:50.533 --> 00:02:52.333
即主力干活的模型

00:02:52.700 --> 00:02:55.700
它并不追求在极难的问题上获得胜利

00:02:56.000 --> 00:02:58.366
它的目标就是解决开发人员每天

00:02:58.366 --> 00:03:01.466
都会频繁使用的各种真实的场景问题

00:03:01.533 --> 00:03:02.700
编写业务代码

00:03:02.966 --> 00:03:04.200
读取很长的文件

00:03:04.333 --> 00:03:05.733
调用各种API

00:03:05.766 --> 00:03:07.766
自动浏览网页查找资料

00:03:07.866 --> 00:03:09.966
处理日常工作中的各项事务

00:03:10.400 --> 00:03:11.933
这些事情并不酷炫

00:03:11.933 --> 00:03:13.966
也不能登上科技新闻头条

00:03:14.100 --> 00:03:15.933
但是他们是开发人员

00:03:15.933 --> 00:03:18.333
每天都要做的80%的工作

00:03:18.566 --> 00:03:21.600
定位的变化要比性能提高更加重要

00:03:21.766 --> 00:03:23.300
这说明整个行业对

00:03:23.300 --> 00:03:25.100
考核的标准发生了变化

00:03:25.266 --> 00:03:27.900
以前人们比的是谁的回答最聪明

00:03:27.900 --> 00:03:29.533
而现在要开始比的是

00:03:29.533 --> 00:03:31.300
谁能用更便宜的价格

00:03:31.700 --> 00:03:33.700
在更多的步骤中遇到错误时

00:03:33.700 --> 00:03:34.900
能够自己纠正

00:03:35.000 --> 00:03:36.400
重新规划路线

00:03:36.700 --> 00:03:38.500
就如公司在招聘的时候

00:03:38.500 --> 00:03:39.933
最聪明的大学毕业生

00:03:39.933 --> 00:03:41.566
当然是最好的人选了

00:03:41.600 --> 00:03:43.133
但是真正能够把日常

00:03:43.133 --> 00:03:45.166
工作做的既快又好

00:03:45.300 --> 00:03:48.133
不需要经常有人来收拾烂摊子的人才

00:03:48.133 --> 00:03:51.400
才是团队中最不可或缺的核心力量

00:03:51.800 --> 00:03:53.066
我跟很多做人工智能

00:03:53.066 --> 00:03:54.366
应用的朋友交流过

00:03:54.466 --> 00:03:55.666
他们之前遇到的最大

00:03:55.666 --> 00:03:57.166
难题就是这样一个问题

00:03:57.333 --> 00:03:59.400
很多模型跑分看起来很厉害

00:03:59.533 --> 00:04:02.100
但是真正放到生产环境中使用

00:04:02.100 --> 00:04:04.000
就会出现各种各样的问题

00:04:04.466 --> 00:04:06.366
让其写出一个完整的功能

00:04:06.500 --> 00:04:08.466
它可以帮到你遗漏了依赖包

00:04:08.466 --> 00:04:10.700
让其修改现有的项目代码

00:04:10.766 --> 00:04:12.900
它可以把其他正常的功能搞砸

00:04:12.966 --> 00:04:14.766
对于需求不明确的地方

00:04:14.966 --> 00:04:16.000
它不会提问

00:04:16.066 --> 00:04:18.500
全部依靠自己的猜测来完成工作

00:04:18.533 --> 00:04:20.933
并且所产出的结果也无法使用

00:04:21.133 --> 00:04:23.333
最后开发人员还要花费更多

00:04:23.333 --> 00:04:25.200
的精力来处理这个问题

00:04:25.266 --> 00:04:26.766
从而降低工作效率

00:04:27.166 --> 00:04:29.766
此次Gemini3.7Flash的出现

00:04:29.800 --> 00:04:32.566
显然是为了应对上述问题而设计的

00:04:32.766 --> 00:04:33.766
我们可以看一下一些

00:04:33.766 --> 00:04:35.400
公开的基准测试数据

00:04:35.400 --> 00:04:36.566
就可以知道它所要

00:04:36.566 --> 00:04:38.133
提高的方向非常明确

00:04:38.400 --> 00:04:41.600
按照Frontier Code 1.1代码基准来计算的话

00:04:41.600 --> 00:04:44.933
在3.6版本中得分为34.4%

00:04:44.933 --> 00:04:46.766
而在43.6%

00:04:47.100 --> 00:04:48.766
在更加接近真实的工程

00:04:48.766 --> 00:04:51.900
项目中进行的DEPSE 1.1测试上

00:04:51.900 --> 00:04:55.300
从49%直接上升到65.3%

00:04:55.466 --> 00:04:56.533
WebDevArena

00:04:56.600 --> 00:04:58.300
面向网页开发的比赛

00:04:58.300 --> 00:05:00.800
已经上升到1,588分

00:05:01.100 --> 00:05:02.933
用终端命令行来测试

00:05:03.100 --> 00:05:04.933
TerminalBench 2.1的时候

00:05:04.933 --> 00:05:07.566
得到的成绩是85.8%

00:05:07.866 --> 00:05:10.366
最极端的情况就是Oswald 2.0

00:05:10.366 --> 00:05:13.200
即测试人工智能控制电脑桌面的能力

00:05:13.200 --> 00:05:18.166
该分数由原来的3.8%猛增到现在的47.9%

00:05:18.533 --> 00:05:20.066
之前用的模型在电脑上

00:05:20.066 --> 00:05:22.133
操作的基本都是点错按钮

00:05:22.133 --> 00:05:23.533
开错软件的水平

00:05:23.600 --> 00:05:24.866
现在可以独立完成

00:05:24.866 --> 00:05:26.733
将近一半的桌面端工作

00:05:26.800 --> 00:05:30.266
例如整理文件、使用软件、填写表格等

00:05:30.266 --> 00:05:31.466
对于agent来说

00:05:31.466 --> 00:05:33.166
这是实实在在的进步

00:05:33.533 --> 00:05:35.800
但是谷歌并没有掩饰自己的缺点

00:05:35.800 --> 00:05:38.666
在更难的TerminalBench 3.0测试中

00:05:38.700 --> 00:05:41.300
谷歌只获得了14.9%的成绩

00:05:41.533 --> 00:05:45.566
小于GPT-5.6terror的20.8%的情况下

00:05:45.566 --> 00:05:47.500
在极端复杂工程任务上

00:05:47.500 --> 00:05:49.300
并不是行业的顶尖水平

00:05:49.733 --> 00:05:50.566
但是对于大多数

00:05:50.566 --> 00:05:52.200
开发者日常的需求而言

00:05:52.266 --> 00:05:54.500
它的能力已经足够使用了

00:05:54.600 --> 00:05:57.600
并且更加懂得怎样把事情做好、做对

00:05:57.900 --> 00:05:59.466
针对性的加强了一些在

00:05:59.466 --> 00:06:01.600
实际开发中经常用到的能力

00:06:01.733 --> 00:06:03.766
严格按照字面意思来执行

00:06:03.900 --> 00:06:05.300
并且不会随意添加

00:06:05.300 --> 00:06:07.000
一些没有被要求的功能

00:06:07.300 --> 00:06:08.800
对于需求不明确的

00:06:09.133 --> 00:06:10.666
地方会主动去询问

00:06:10.700 --> 00:06:12.066
并不是自己瞎猜

00:06:12.200 --> 00:06:13.766
当发现路径走不通

00:06:13.766 --> 00:06:15.066
或者错误很多时

00:06:15.066 --> 00:06:16.300
就会及时止损

00:06:16.366 --> 00:06:17.966
不会一条路走到黑

00:06:18.000 --> 00:06:20.266
浪费大量的时间与计算资源

00:06:20.666 --> 00:06:23.666
这些特点在跑分中是无法体现出来的

00:06:23.700 --> 00:06:25.600
但是真正使用的时候

00:06:25.600 --> 00:06:27.733
体验上的差别就会很明显

00:06:28.066 --> 00:06:29.300
讲完了性能之后

00:06:29.300 --> 00:06:31.200
我们就来聊聊这次更新中最为

00:06:31.200 --> 00:06:33.733
核心的一点agent时代经济学

00:06:33.766 --> 00:06:35.900
也就是人工智能模型成本账

00:06:35.900 --> 00:06:37.900
在算法上已经发生了变化

00:06:38.333 --> 00:06:39.933
人们比较模型的价格时

00:06:39.933 --> 00:06:42.100
只关心每百个TOKEN多少钱

00:06:42.133 --> 00:06:43.933
哪个更便宜就用哪个

00:06:44.066 --> 00:06:46.666
在之前的单轮对话中是正确的

00:06:46.666 --> 00:06:48.500
当你提出一个问题的时候

00:06:48.566 --> 00:06:50.100
模型就会回答一次

00:06:50.100 --> 00:06:51.866
并且计算的非常准确

00:06:52.133 --> 00:06:54.366
但是到了agent代理的时代之后

00:06:54.366 --> 00:06:56.500
该算法就完全不起作用了

00:06:56.866 --> 00:06:59.533
agent不是一次回答一个问题就结束了

00:06:59.566 --> 00:07:02.200
而是一直的去完成一系列的任务

00:07:02.400 --> 00:07:04.866
首先要弄清楚目标是什么

00:07:04.866 --> 00:07:08.066
然后调动相应的工具来实现这个目标

00:07:08.366 --> 00:07:09.966
接着读取相关的文件

00:07:10.066 --> 00:07:11.733
编写代码并进行测试

00:07:11.766 --> 00:07:14.566
一旦发现问题就会及时修正和重试

00:07:14.900 --> 00:07:16.466
整个过程中可能会使用

00:07:16.466 --> 00:07:18.500
到十几乃至几十个模型

00:07:18.900 --> 00:07:21.266
此时衡量成本的主要指标不再

00:07:21.266 --> 00:07:23.100
是每百万个TOKEN的价格

00:07:23.133 --> 00:07:25.300
而变成了完成一个完整的

00:07:25.300 --> 00:07:27.000
任务所需要的所有费用

00:07:27.500 --> 00:07:29.666
有些模型的价格看上去很低廉

00:07:29.733 --> 00:07:31.500
但是第一回做不正确的话

00:07:31.500 --> 00:07:33.166
就要反复尝试几次

00:07:33.266 --> 00:07:35.966
期间还要调用很多没有用到过的工具

00:07:36.100 --> 00:07:37.666
最后算总账的时候

00:07:37.900 --> 00:07:40.766
费用反而要比单个模型的价格高一些

00:07:40.900 --> 00:07:42.666
但是只要有一次能够做到

00:07:42.666 --> 00:07:44.700
的事情的模型就值钱多了

00:07:44.933 --> 00:07:47.166
就比如你要请个装修工人

00:07:47.166 --> 00:07:48.200
一个工钱便宜

00:07:48.200 --> 00:07:49.400
但是干活粗糙

00:07:49.533 --> 00:07:52.133
装完了还要拆掉重新安装三次

00:07:52.400 --> 00:07:54.200
最后花费的钱和时间都比

00:07:54.200 --> 00:07:55.700
找一个工钱稍微高一点

00:07:55.700 --> 00:07:57.066
但是只用一次就可以

00:07:57.066 --> 00:07:59.266
完成的工作的人要多得多

00:07:59.533 --> 00:08:02.466
Gemini3.7Flash本次优化的重点是

00:08:02.466 --> 00:08:04.533
提高第一次就做对的概率

00:08:04.600 --> 00:08:06.200
降低不必要的工具使用

00:08:06.200 --> 00:08:08.000
以及无意义的重复尝试

00:08:08.466 --> 00:08:10.266
它不一定是最便宜的型号

00:08:10.366 --> 00:08:12.800
但是只要能减少重试次数

00:08:12.800 --> 00:08:14.266
那么完成同一项工作

00:08:14.266 --> 00:08:15.866
的总成本就会降低

00:08:16.100 --> 00:08:18.733
这才是企业用户真正关心的账单

00:08:18.933 --> 00:08:21.300
毕竟企业在使用人工智能的时候

00:08:21.300 --> 00:08:22.800
是为了降低成本

00:08:22.866 --> 00:08:23.800
提高效率

00:08:23.933 --> 00:08:26.166
并不是为了购买最低价的代币

00:08:26.666 --> 00:08:28.733
按照这样的成本考量来分析

00:08:28.733 --> 00:08:30.700
它所采取的价格策略的话

00:08:30.766 --> 00:08:32.100
就可以很容易的理解

00:08:32.100 --> 00:08:33.933
到谷歌的战略意图了

00:08:34.133 --> 00:08:35.666
官方公布的售价为

00:08:35.666 --> 00:08:38.766
在2026年底之前是推广价

00:08:38.866 --> 00:08:41.700
输入TOKEN每百万0.75美元

00:08:41.733 --> 00:08:44.400
输出TOKEN每百万3.75美元

00:08:44.766 --> 00:08:48.300
自2027年1月1日开始实行正式价格

00:08:48.400 --> 00:08:50.566
在上涨至1.5美元的时候

00:08:50.666 --> 00:08:52.666
输出上涨到7.5美元

00:08:52.666 --> 00:08:53.666
也就是翻倍

00:08:54.000 --> 00:08:55.200
从价格上来看

00:08:55.200 --> 00:08:57.600
它要比很多开源模型要贵一些

00:08:57.600 --> 00:08:59.766
但是结合其性能以及

00:08:59.766 --> 00:09:01.666
任务完成率来考量的话

00:09:01.666 --> 00:09:03.800
其实它的性价比是非常高的

00:09:04.000 --> 00:09:05.466
谷歌给出这样的价格

00:09:05.600 --> 00:09:08.533
并不是依靠该模型来短期内获利

00:09:08.533 --> 00:09:10.700
而是有两重明确的战略目的

00:09:11.100 --> 00:09:12.733
第一项任务就是夺取

00:09:12.733 --> 00:09:14.366
开发者默认的选择权

00:09:14.566 --> 00:09:16.700
现在的AI开发迁移成本其实要

00:09:16.700 --> 00:09:18.566
比人们想象中的要高一些

00:09:18.733 --> 00:09:20.166
当开发者习惯于某

00:09:20.166 --> 00:09:21.666
一个模型回答的方式

00:09:21.733 --> 00:09:24.366
并且把提示词库工具链工

00:09:24.366 --> 00:09:26.400
作流都建立在该模型之上时

00:09:26.400 --> 00:09:28.566
就不会轻易更换其他模型了

00:09:28.866 --> 00:09:31.466
更换一次就需要重新调整和测试

00:09:31.566 --> 00:09:32.533
费用很高

00:09:32.766 --> 00:09:34.666
目前谷歌采用的是接近

00:09:34.666 --> 00:09:36.300
成本的价格进行推广

00:09:36.400 --> 00:09:38.133
目的是为了让开发者以较

00:09:38.133 --> 00:09:40.166
低的成本使用到产品中去

00:09:40.166 --> 00:09:42.133
在开发者习惯之后再逐步

00:09:42.133 --> 00:09:44.500
绑定到Gemini生态系统上来

00:09:44.866 --> 00:09:47.400
即使到了2027年再提价

00:09:47.400 --> 00:09:50.400
如果总的费用仍然低于竞争对手的话

00:09:50.400 --> 00:09:52.866
大多数人是不会选择离开的

00:09:53.100 --> 00:09:54.766
第二个目的就是夺取

00:09:54.766 --> 00:09:57.066
企业的模型路由的主导权

00:09:57.466 --> 00:09:59.000
稍微有点规模的企业并

00:09:59.000 --> 00:10:00.733
不是只会使用一种模型

00:10:00.766 --> 00:10:02.533
而是会进行分层路由

00:10:02.600 --> 00:10:05.600
大约有5%的最复杂的最难的

00:10:05.600 --> 00:10:07.500
任务会被交由最昂贵且能力

00:10:07.500 --> 00:10:09.900
最强的顶级模型来处理

00:10:10.066 --> 00:10:12.666
剩下的高频重复难度较低的

00:10:12.666 --> 00:10:15.000
日程工作都交给性价比高

00:10:15.000 --> 00:10:17.533
速度快、稳定性强的主要模型来完成

00:10:17.966 --> 00:10:19.600
95%的任务中

00:10:19.666 --> 00:10:22.500
调用量是顶级模型的几十上百倍

00:10:22.533 --> 00:10:25.400
这才是人工智能市场的最大基础盘

00:10:25.733 --> 00:10:28.366
Gemini3.7Flash打向的是这个市场

00:10:28.366 --> 00:10:31.700
不会和顶级模型争夺那5%的高光场景

00:10:31.700 --> 00:10:34.900
而是吃掉剩下的95%的高频调用

00:10:35.266 --> 00:10:36.533
看上去单价不高

00:10:36.533 --> 00:10:38.366
但是架不住使用次数多

00:10:38.466 --> 00:10:41.066
细水长流的收入要比依靠顶尖的

00:10:41.066 --> 00:10:43.566
人工智能来赚钱更有实际意义

00:10:43.866 --> 00:10:45.366
除了模型本身之外

00:10:45.400 --> 00:10:48.133
谷歌还拥有其他厂商所不具备的优势

00:10:48.166 --> 00:10:50.333
那就是它自己的生态系统

00:10:50.733 --> 00:10:53.133
而此次同步推出的新品Spark

00:10:53.166 --> 00:10:55.466
全天候代理则是用这个牌

00:10:55.866 --> 00:10:58.533
即使是第三方的人工智能做的最好了

00:10:58.533 --> 00:11:00.766
也免不了有一个绕不过去的问题

00:11:00.900 --> 00:11:03.066
它拿不到你个人上下文的信息

00:11:03.500 --> 00:11:05.366
要让其帮助你处理邮件的话

00:11:05.366 --> 00:11:07.066
就需要你自己把邮件的

00:11:07.066 --> 00:11:09.066
内容导出来再上传上去

00:11:09.066 --> 00:11:10.566
让其为你安排日程

00:11:10.700 --> 00:11:12.800
需要自己把日程表发送给他

00:11:13.000 --> 00:11:14.266
让其修改文档

00:11:14.333 --> 00:11:15.900
上传文件等事情也是

00:11:15.900 --> 00:11:17.266
一次又一次的进行

00:11:17.333 --> 00:11:18.200
十分不便

00:11:18.500 --> 00:11:20.366
但是Spark是由谷歌直接

00:11:20.366 --> 00:11:21.866
开发出来的代理软件

00:11:21.900 --> 00:11:24.533
它可以非常深入的与Gmail  Google

00:11:24.566 --> 00:11:26.733
日历、Google文档等几乎每一个

00:11:26.733 --> 00:11:28.700
职场人使用的软件进行连接

00:11:29.100 --> 00:11:30.466
它可以自动读取你的

00:11:30.466 --> 00:11:32.900
邮件日程文档等信息

00:11:32.933 --> 00:11:34.466
并且不需要你不断的

00:11:34.466 --> 00:11:36.000
进行数据同步操作

00:11:36.066 --> 00:11:38.400
可以直接帮助你回复邮件

00:11:38.400 --> 00:11:41.333
安排会议、修改文档、整理资料等等

00:11:41.333 --> 00:11:42.733
这样的用户体验是

00:11:42.733 --> 00:11:45.300
第三方人工智能很难做到的

00:11:45.500 --> 00:11:47.066
但是权限越高

00:11:47.066 --> 00:11:49.200
所承担的风险也就越大

00:11:49.566 --> 00:11:51.966
如果人工智能可以随意读取和

00:11:51.966 --> 00:11:54.566
修改你的电子邮件文档等信息

00:11:54.600 --> 00:11:56.600
那么一次错误的操作就会

00:11:56.600 --> 00:11:58.400
导致重要的文件被删除

00:11:58.666 --> 00:12:00.800
或者给客户提供错误的内容

00:12:00.800 --> 00:12:02.666
所造成的影响要比普通

00:12:02.666 --> 00:12:04.866
的AI胡言乱语要大得多

00:12:05.266 --> 00:12:07.666
因此谷歌也需要做好三件事情

00:12:07.866 --> 00:12:09.333
极高的操作透明度

00:12:09.333 --> 00:12:10.566
让使用者知道

00:12:10.566 --> 00:12:12.566
人工智能每一步要做什么

00:12:12.600 --> 00:12:15.133
重要的操作要人工二次确认

00:12:15.400 --> 00:12:17.333
例如发送电子邮件删除

00:12:17.333 --> 00:12:19.300
文件之前都要得到你的同意

00:12:19.333 --> 00:12:21.066
还有一整套的审计日志

00:12:21.200 --> 00:12:23.500
所有的操作都可以被倒退回去

00:12:23.900 --> 00:12:26.666
另外它也加入了Antigravity开发环境

00:12:26.666 --> 00:12:29.566
把人工智能的操作放在一个可以回滚

00:12:29.666 --> 00:12:31.400
受到控制的边界之中

00:12:31.733 --> 00:12:33.700
即使人工智能操作出错

00:12:33.700 --> 00:12:35.400
或者是遇到网络不稳定

00:12:35.533 --> 00:12:37.933
验证码等现实生活中的意外情况

00:12:37.933 --> 00:12:39.500
也可以通过一键回滚

00:12:39.500 --> 00:12:41.200
功能回到之前的状况

00:12:41.400 --> 00:12:43.533
并不会把你的本地环境弄垮

00:12:43.966 --> 00:12:45.800
虽然这些细节并不酷炫

00:12:45.933 --> 00:12:48.400
但是他们是agent真正的去面对

00:12:48.400 --> 00:12:50.200
和解决工作中遇到的各种

00:12:50.200 --> 00:12:52.266
问题所必须跨越的一道坎

00:12:52.733 --> 00:12:54.333
把目光放远一些的话

00:12:54.366 --> 00:12:56.100
目前人工智能界的三大

00:12:56.100 --> 00:12:59.000
巨头们所走的道路已经十分明确

00:12:59.166 --> 00:13:02.200
并且这三条道路根本就不是一回事

00:13:02.600 --> 00:13:05.700
谷歌的Gemini3.7Flash主要卖点为

00:13:05.700 --> 00:13:08.300
性价比、工程完成率、工具链绑定

00:13:08.300 --> 00:13:10.533
以及全套生态入口

00:13:10.566 --> 00:13:13.166
目的是成为整个行业的主要工作模式

00:13:13.400 --> 00:13:15.800
并且占领最大的日常使用市场

00:13:16.400 --> 00:13:19.266
OpenAi所走的道路是性能上限

00:13:19.266 --> 00:13:21.466
与极速路线相结合的方式

00:13:21.766 --> 00:13:24.700
它所推出的新一代推理服务Ultrafast

00:13:25.133 --> 00:13:26.600
在最高峰的时候可以达到

00:13:26.600 --> 00:13:29.400
每秒钟750个输出TOKEN的速度

00:13:29.466 --> 00:13:30.900
就是为了消除agent

00:13:30.900 --> 00:13:32.466
连续工作时延的问题

00:13:32.766 --> 00:13:34.800
以前agent做任务的时候要等

00:13:34.800 --> 00:13:36.500
它一步一步的反应出来

00:13:36.700 --> 00:13:37.866
卡顿感很强

00:13:37.933 --> 00:13:39.300
极速推理可以让整个

00:13:39.300 --> 00:13:41.500
过程像真人操作一样流畅

00:13:41.966 --> 00:13:44.800
主要针对的是高端场景以及实时交互

00:13:44.966 --> 00:13:47.066
例如实时语音对话

00:13:47.066 --> 00:13:49.133
实时设备控制、复杂的推理等等

00:13:49.133 --> 00:13:50.933
对于延迟和能力上限

00:13:50.933 --> 00:13:52.400
有较高要求的场景

00:13:52.866 --> 00:13:55.133
而Deepseek走的是差异化的路线

00:13:55.200 --> 00:13:57.366
V4Pro版本反而提高了售价

00:13:57.366 --> 00:13:59.466
并且开始追求更高的毛利额

00:13:59.600 --> 00:14:02.666
之前是以廉价替代品的身份出现的

00:14:02.800 --> 00:14:05.766
现在则要向高附加值的产品发展了

00:14:05.766 --> 00:14:07.200
不再进行价格战

00:14:07.266 --> 00:14:09.366
利用一定的场景来获得收益

00:14:09.800 --> 00:14:11.066
三条不同的道路

00:14:11.066 --> 00:14:12.166
实际上表明了

00:14:12.166 --> 00:14:14.166
人工智能市场的分化趋势

00:14:14.300 --> 00:14:16.566
并且不会形成一家独大的局面

00:14:16.800 --> 00:14:19.866
高端场景、大众市场、性价比场景等

00:14:19.866 --> 00:14:22.400
未来的各个场景都会出现相应的玩家

00:14:22.466 --> 00:14:24.300
而竞争也不再是简单的

00:14:24.300 --> 00:14:26.266
谁跑分高谁赢的问题了

00:14:26.666 --> 00:14:28.966
但是我们也应该清醒地认识到

00:14:29.000 --> 00:14:30.666
谷歌此次的战略部署

00:14:30.700 --> 00:14:32.300
并不是完全没有风险

00:14:32.366 --> 00:14:33.300
在一些方面

00:14:33.300 --> 00:14:34.300
如果没有做好

00:14:34.300 --> 00:14:36.700
可能会对它实现目标产生影响

00:14:37.166 --> 00:14:39.700
第一个问题就是发布的频率过快

00:14:39.900 --> 00:14:42.533
每隔三周进行一次大的版本更新

00:14:42.533 --> 00:14:45.000
对于普通的用户来说是一件好事

00:14:45.133 --> 00:14:47.000
但是对于企业用户而言

00:14:47.133 --> 00:14:50.266
兼容性测试以及维护的成本将会很高

00:14:50.600 --> 00:14:52.866
企业把主要业务都放在API上

00:14:52.900 --> 00:14:54.366
每次模型更新都需要

00:14:54.366 --> 00:14:55.966
重新测试所有的功能

00:14:56.066 --> 00:14:58.200
害怕新的版本改变了输出格式

00:14:58.266 --> 00:14:59.766
或者功能发生变化

00:14:59.933 --> 00:15:01.666
从而影响到线上的业务

00:15:02.133 --> 00:15:03.866
迭代速度过快会使得保守

00:15:03.866 --> 00:15:06.166
企业无法将关键业务流程上线

00:15:06.400 --> 00:15:08.466
第二个问题就是跑分与

00:15:08.466 --> 00:15:09.933
实际业务之间的差异

00:15:10.200 --> 00:15:12.466
公开基准测试是标准的问题

00:15:12.533 --> 00:15:15.366
但是企业的实际业务环境千变万化

00:15:15.500 --> 00:15:18.066
有的企业有着几十年的老系统

00:15:18.133 --> 00:15:20.900
有的企业则存在复杂的内部后端

00:15:21.100 --> 00:15:23.333
还有一些非常奇特的边缘场景

00:15:23.700 --> 00:15:25.333
跑分高并不意味着在这样的

00:15:25.333 --> 00:15:27.600
环境中可以保持稳定的运行状态

00:15:27.766 --> 00:15:30.066
在实际的企业用户的使用过程中

00:15:30.066 --> 00:15:32.900
是否能够持久的为企业用户提供服务

00:15:33.066 --> 00:15:35.066
则要根据实际情况来判断

00:15:35.500 --> 00:15:36.966
第三个问题是优惠

00:15:36.966 --> 00:15:38.766
期满之后的价格翻倍

00:15:38.966 --> 00:15:41.066
目前低价为推广期的价格

00:15:41.066 --> 00:15:44.566
在2,027年正式售价时会直接翻倍

00:15:44.600 --> 00:15:46.700
到那时很多小微企业和个人

00:15:46.700 --> 00:15:49.266
开发者的成本账都要重新算了

00:15:49.366 --> 00:15:50.800
到时候竞品推出

00:15:50.800 --> 00:15:52.466
性价比更高的产品的话

00:15:52.466 --> 00:15:55.333
就会有一部分价格敏感型用户流失

00:15:56.000 --> 00:15:59.500
第四问是关于隐私以及越权的风险

00:15:59.966 --> 00:16:01.933
agent获得了邮箱、文档

00:16:01.933 --> 00:16:03.566
日程等高权限之后

00:16:03.566 --> 00:16:05.733
在一次误读误写中所造成的后果

00:16:05.733 --> 00:16:08.166
要比普通的AI幻觉要严重得多

00:16:08.566 --> 00:16:10.700
普通的幻觉就是说错了某句话

00:16:10.700 --> 00:16:13.000
而高权限agent所出现的幻觉

00:16:13.000 --> 00:16:14.766
有可能会带来实际的损失

00:16:14.933 --> 00:16:16.866
如果这个问题处理不好的话

00:16:16.966 --> 00:16:18.900
企业和个人用户都不敢

00:16:18.900 --> 00:16:20.800
真的把高权限交给别人

00:16:21.133 --> 00:16:22.300
第五个问题是

00:16:22.300 --> 00:16:24.100
谷歌的产品线非常分散

00:16:24.466 --> 00:16:27.766
目前Gemini有Pro Flash Ultra等版本

00:16:27.800 --> 00:16:30.600
并且还推出了Spark代理和Antigravity

00:16:30.866 --> 00:16:33.666
开发环境等一系列名称繁多的产品

00:16:33.966 --> 00:16:35.700
普通的用户根本无法

00:16:35.700 --> 00:16:38.266
分辨出哪些是哪一种产品

00:16:38.366 --> 00:16:40.700
自己选择的话就选哪一个吧

00:16:41.000 --> 00:16:42.266
产品的线太多

00:16:42.266 --> 00:16:44.666
反而会使用户的挑选成本增加

00:16:44.866 --> 00:16:45.866
说到这的时候

00:16:45.866 --> 00:16:47.700
可能会有人联想到之前谷歌

00:16:47.700 --> 00:16:49.766
的核心高管离职创业的事情

00:16:49.900 --> 00:16:50.700
在那个时候

00:16:50.700 --> 00:16:51.933
很多人都认为谷歌的

00:16:51.933 --> 00:16:53.366
人工智能要不行了

00:16:53.500 --> 00:16:55.100
核心的人才走了之后

00:16:55.100 --> 00:16:57.300
后续的迭代也一定跟不上

00:16:57.700 --> 00:16:59.400
结果才过了几个月的时间

00:16:59.533 --> 00:17:01.566
Gemini三周发布一次新版本

00:17:01.666 --> 00:17:03.800
迭代的速度一点都没有减缓过

00:17:03.800 --> 00:17:05.400
反而越来越快了

00:17:05.600 --> 00:17:07.100
这件事情实际上也是

00:17:07.100 --> 00:17:08.333
给我们的一个警示

00:17:08.700 --> 00:17:10.800
目前的人工智能竞争并不是

00:17:10.800 --> 00:17:12.766
依靠少数顶尖的研究人员

00:17:12.766 --> 00:17:14.733
单打独斗就可以做到的了

00:17:15.066 --> 00:17:16.266
现在已经变成了一种

00:17:16.266 --> 00:17:18.100
工业化流水线式的竞争

00:17:18.300 --> 00:17:20.866
依靠的是算力规模、数据飞轮

00:17:20.900 --> 00:17:22.466
自动化评测体系

00:17:22.466 --> 00:17:25.400
产品分发渠道组成的一个完整的系统

00:17:25.766 --> 00:17:27.666
顶尖的人才固然很重要

00:17:27.700 --> 00:17:30.133
但是系统的迭代能力才是

00:17:30.133 --> 00:17:32.000
目前大厂的核心竞争力

00:17:32.466 --> 00:17:34.000
流水线运转的情况下

00:17:34.000 --> 00:17:35.666
个别人员的调动并不会

00:17:35.666 --> 00:17:37.566
改变整个系统迭代的速度

00:17:37.866 --> 00:17:38.933
对普通人来说

00:17:38.933 --> 00:17:41.133
这同样有着非常直接的启示

00:17:41.266 --> 00:17:43.266
以后人类的核心价值观并

00:17:43.266 --> 00:17:45.366
不是跟AI比谁写代码更快

00:17:45.366 --> 00:17:46.766
谁做执行更强

00:17:46.933 --> 00:17:48.966
这些问题模型都会做的很快

00:17:48.966 --> 00:17:49.966
而且很便宜

00:17:50.466 --> 00:17:52.200
人的真正不可替代之处

00:17:52.200 --> 00:17:53.933
在于确定正确的目标

00:17:53.966 --> 00:17:55.300
控制最后结果

00:17:55.300 --> 00:17:56.566
对决策负责

00:17:56.566 --> 00:17:59.166
而模型只起到高效的执行作用

00:17:59.466 --> 00:18:02.800
倒不如先学会如何利用好这个执行层

00:18:02.800 --> 00:18:04.366
来为模型定方向

00:18:04.400 --> 00:18:05.666
查结果的人

00:18:06.000 --> 00:18:08.266
最后谷歌这次能不能成功

00:18:08.266 --> 00:18:10.600
其实不需要看发布会吹的多么好

00:18:10.600 --> 00:18:13.566
只需要关注4个主要的数据就可以了

00:18:13.733 --> 00:18:16.000
第二是开发者选择的标准

00:18:16.166 --> 00:18:17.666
它是否真的会成为大家

00:18:17.666 --> 00:18:19.733
生产环境中默认使用的模型

00:18:19.766 --> 00:18:21.966
而不会只是用来测试和玩耍

00:18:22.000 --> 00:18:24.200
并且编写一些简单的脚本而已

00:18:24.666 --> 00:18:26.100
第二步就是对整个

00:18:26.100 --> 00:18:27.700
项目的成本进行计算

00:18:27.766 --> 00:18:30.500
包括了重试、人工修复等费用之后

00:18:30.500 --> 00:18:32.800
该项目真正完成的任务总成本

00:18:32.800 --> 00:18:35.166
到底是不是真的比竞争对手低呢

00:18:35.533 --> 00:18:37.366
第三是用户的粘性

00:18:37.900 --> 00:18:40.166
Spark这样的代理产品能否使

00:18:40.166 --> 00:18:42.333
用户形成高频使用的好习惯

00:18:42.366 --> 00:18:43.933
而不能只是新鲜感

00:18:43.933 --> 00:18:45.100
一下就忘了

00:18:45.166 --> 00:18:47.000
第四是生态治理方面

00:18:47.133 --> 00:18:48.733
谷歌能否把杂乱无章的

00:18:48.733 --> 00:18:50.466
产品线整理的井然有序

00:18:50.600 --> 00:18:52.533
并且能够有效的降低用户的

00:18:52.533 --> 00:18:54.733
决策难度以及使用成本呢

00:18:55.133 --> 00:18:56.266
因此回头来看

00:18:56.333 --> 00:18:59.333
谷歌这次又发布了Gemini新版本

00:18:59.400 --> 00:19:01.166
真正值得我们关注的是

00:19:01.166 --> 00:19:02.766
它多出了多少个参数

00:19:02.800 --> 00:19:05.366
并不是哪一个排行榜上超越了谁

00:19:05.666 --> 00:19:06.700
更主要的是

00:19:06.766 --> 00:19:09.300
谷歌正把Gemini的功能由原来的

00:19:09.300 --> 00:19:11.866
回答问题的AI逐渐转变为现在

00:19:11.866 --> 00:19:14.566
的可以帮你做事情的工作型AI

00:19:14.933 --> 00:19:16.500
而其背后所代表的是

00:19:16.500 --> 00:19:18.600
人工智能行业的一个重要变革

00:19:18.800 --> 00:19:21.200
以前人们比拼的是模型有多聪明

00:19:21.333 --> 00:19:24.300
接下来要拼的是模型是否能完成任务

00:19:24.700 --> 00:19:27.333
谁能使得agent更加稳定的使用工具

00:19:27.366 --> 00:19:29.066
并且减少错误的发生

00:19:29.066 --> 00:19:31.166
也不再反复尝试同样的工作

00:19:31.166 --> 00:19:33.966
而又能把每次任务的成本降到最低

00:19:34.166 --> 00:19:35.700
那么他就有可能真正的

00:19:35.700 --> 00:19:37.666
进入到企业的工作当中去

00:19:38.066 --> 00:19:39.566
因此Gemini这次更新

00:19:39.566 --> 00:19:42.166
我认为更值得关注的是3个方面

00:19:42.466 --> 00:19:44.133
开发人员是否真的会把

00:19:44.133 --> 00:19:46.366
它们放到生产环境中去使用

00:19:46.533 --> 00:19:49.133
企业在计算了全部的成本之后

00:19:49.366 --> 00:19:51.366
是否还会愿意长久的使用

00:19:51.366 --> 00:19:53.866
还有谷歌是否能够利用Gmail

00:19:53.966 --> 00:19:56.700
Docs  Calendar等整个生态系统

00:19:56.766 --> 00:19:58.800
并且把Gemini变成用户每天

00:19:58.800 --> 00:20:00.300
都要用到的应用程序

00:20:00.766 --> 00:20:02.166
如果上述的事情最后

00:20:02.166 --> 00:20:03.600
都能够运行起来的话

00:20:03.600 --> 00:20:05.966
那么谷歌此次不断更新Gemini

00:20:05.966 --> 00:20:07.666
并且频繁的进行更新

00:20:07.700 --> 00:20:09.933
就不只是普通的版本升级了

00:20:10.166 --> 00:20:11.666
在某种程度上来说

00:20:11.700 --> 00:20:13.600
它已经预示着人工智能

00:20:13.600 --> 00:20:15.466
代理时代到来的可能性

00:20:15.600 --> 00:20:18.333
但是现在下结论为时尚早

00:20:18.500 --> 00:20:20.466
人工智能行业的变化很快

00:20:20.666 --> 00:20:23.000
今天处于领先地位的企业并不

00:20:23.000 --> 00:20:25.600
代表明天仍然会保持领先地位

00:20:26.100 --> 00:20:27.600
值得我们长期关注的是

00:20:27.600 --> 00:20:29.500
这些模型是否可以从发布会

00:20:29.500 --> 00:20:31.966
中发布的参数、跑分等信息中

00:20:31.966 --> 00:20:34.000
真正走入到普通人的工作流程

00:20:34.000 --> 00:20:36.166
以及企业的实际业务当中去

00:20:36.333 --> 00:20:38.100
这也是我一直想要做的

00:20:38.100 --> 00:20:40.266
在于告诉人们谁发布过什么

00:20:40.300 --> 00:20:41.566
并不是说只把新闻

00:20:41.566 --> 00:20:43.300
本身讲清楚就可以了

00:20:43.466 --> 00:20:45.600
而要把新闻背后所涉及的产业

00:20:45.600 --> 00:20:47.966
逻辑也一并拆开来给大家看

00:20:48.300 --> 00:20:50.733
如果你想继续学习人工智能

00:20:50.733 --> 00:20:53.500
科技巨头以及美股市场的深层次原理

00:20:53.500 --> 00:20:54.800
请关注本频道

00:20:55.133 --> 00:20:56.666
最后还想问一下你

00:20:56.700 --> 00:20:59.600
你觉得Gemini将来最大的优势是什么

00:20:59.600 --> 00:21:01.266
是模型本身的实力

00:21:01.266 --> 00:21:03.966
还是背后谷歌整个生态系统的力量

00:21:04.366 --> 00:21:07.000
欢迎大家在评论区发表自己的观点

00:21:07.366 --> 00:21:08.933
本期内容仅为产业

00:21:08.933 --> 00:21:10.700
信息分享与观点交流

00:21:10.700 --> 00:21:13.600
不构成任何投资建议或交易依据

00:21:14.066 --> 00:21:15.566
AI行业变化非常快

00:21:15.600 --> 00:21:17.600
任何判断都存在不确定性

00:21:17.600 --> 00:21:18.966
请大家独立思考

00:21:19.066 --> 00:21:20.666
理性看待市场变化

00:21:20.666 --> 00:21:21.833
我们下一期见

00:00:00.266 --> 00:00:01.900
Google has released a new version

00:00:01.900 --> 00:00:03.566
And this time with the previous generation Gemini

00:00:03.566 --> 00:00:05.000
The update time is almost the same.

00:00:05.133 --> 00:00:06.400
Even a little abnormal

00:00:06.766 --> 00:00:08.666
When people see this content

00:00:08.766 --> 00:00:10.766
Their first reaction was probably

00:00:10.900 --> 00:00:13.300
Google is about to start the heap version thing again

00:00:13.666 --> 00:00:15.466
But if you look closely at Gemini,

00:00:15.466 --> 00:00:16.766
The direction of this update

00:00:16.766 --> 00:00:18.933
You will find that things are not so simple

00:00:19.266 --> 00:00:22.000
Since the problem Google wants to solve this time is not

00:00:22.000 --> 00:00:24.800
How to create a benchmark-first model again

00:00:24.933 --> 00:00:26.766
It's a more practical question.

00:00:27.200 --> 00:00:29.300
How does artificial intelligence really work?

00:00:29.300 --> 00:00:31.666
Get into people's daily work?

00:00:31.733 --> 00:00:34.900
Yes, developers can happily use it every day.

00:00:35.000 --> 00:00:37.766
And can the company pay for it for a long time?

00:00:38.000 --> 00:00:40.400
So Gemini didn't pay all the attention this time

00:00:40.400 --> 00:00:43.000
Focusing on the question of which model is smarter

00:00:43.066 --> 00:00:46.333
Instead, it is gradually strengthening code development and tool use

00:00:46.333 --> 00:00:48.800
And complex task processing, etc

00:00:48.933 --> 00:00:52.100
Performance of stability and efficiency in agent scenarios

00:00:52.566 --> 00:00:53.366
That is to say

00:00:53.400 --> 00:00:55.700
Google wants to replace Gemini with you

00:00:55.700 --> 00:00:57.966
Occasionally send a few questions to its AI

00:00:57.966 --> 00:00:59.866
Become a long-lasting

00:00:59.900 --> 00:01:01.900
The main model that serves you

00:01:02.300 --> 00:01:03.933
This change seems to be just

00:01:03.933 --> 00:01:05.600
Upgrade a version of the software

00:01:05.700 --> 00:01:07.500
But the significance it contains

00:01:07.500 --> 00:01:09.466
Far greater than the parameter promotion itself

00:01:09.866 --> 00:01:11.700
Because in artificial intelligence agents really

00:01:11.700 --> 00:01:13.533
After putting in the work

00:01:13.533 --> 00:01:14.966
A task is often called

00:01:14.966 --> 00:01:17.000
Models range from a dozen to dozens of times

00:01:17.200 --> 00:01:18.566
At this point, people are already comparing

00:01:18.566 --> 00:01:20.733
Which model is not the most powerful?

00:01:20.733 --> 00:01:23.466
And which model can be turned into a lower

00:01:23.466 --> 00:01:26.333
What about cost-stable performance to get the job done?

00:01:26.666 --> 00:01:27.900
This will affect

00:01:27.900 --> 00:01:30.300
The future competitive model of the artificial intelligence industry

00:01:30.700 --> 00:01:32.300
Therefore, the content of today's issue

00:01:32.300 --> 00:01:33.933
No longer just talk

00:01:33.933 --> 00:01:36.200
How awesome is the new Gemini version?

00:01:36.266 --> 00:01:37.666
But from the back of this update

00:01:37.666 --> 00:01:39.666
Then I took it apart and thought about 3 questions

00:01:39.900 --> 00:01:41.500
Why did Google suddenly do this?

00:01:41.500 --> 00:01:43.733
Valuing one doesn't have to be the first

00:01:43.733 --> 00:01:46.100
But what about models that are really good at doing things?

00:01:46.366 --> 00:01:48.100
In the context of the agent era

00:01:48.133 --> 00:01:50.700
Can the price of artificial intelligence be no longer just

00:01:50.700 --> 00:01:53.400
How much is it per million tokens?

00:01:53.700 --> 00:01:57.466
In addition, companies such as Google, Openai, and Deepseek

00:01:57.466 --> 00:01:59.900
How are we going to compete?

00:02:00.066 --> 00:02:02.566
If you can solve these three problems

00:02:02.600 --> 00:02:04.000
Then you can find out

00:02:04.133 --> 00:02:05.966
Google's new model released this time

00:02:06.000 --> 00:02:07.866
It's not just about innovation.

00:02:08.200 --> 00:02:10.500
In anticipating artificial intelligence

00:02:10.500 --> 00:02:12.933
Possibilities for the next phase of maximum use cases

00:02:13.200 --> 00:02:15.700
Hello everyone, my name is Financial Translator.

00:02:15.966 --> 00:02:17.566
In this issue, we will start with Gemini this time.

00:02:17.566 --> 00:02:19.266
The new version is released

00:02:19.300 --> 00:02:21.066
Take what Google wants to take away

00:02:21.066 --> 00:02:22.900
Take it apart layer by layer to show you

00:02:22.966 --> 00:02:24.600
Again, in this update

00:02:24.600 --> 00:02:26.266
The most irrational part

00:02:26.700 --> 00:02:27.733
Over the past few years

00:02:27.733 --> 00:02:29.900
At the artificial intelligence industry conference

00:02:29.900 --> 00:02:32.000
Everyone is competing to see who can be the best

00:02:32.000 --> 00:02:34.400
First place in a difficult benchmark

00:02:34.666 --> 00:02:36.333
Or solve that other models cannot

00:02:36.333 --> 00:02:38.066
Solved math and reasoning problems

00:02:38.100 --> 00:02:39.933
It seems that as long as the model is smart enough

00:02:39.933 --> 00:02:41.600
You can rule the rivers and lakes

00:02:41.800 --> 00:02:45.000
But this time Gemini3.7Flash did not

00:02:45.000 --> 00:02:47.300
Want to compete for the title of the strongest brain

00:02:47.700 --> 00:02:50.400
Google's position for itself is the Workhorse model

00:02:50.533 --> 00:02:52.333
It's the workhorse model.

00:02:52.700 --> 00:02:55.700
It's not about winning the hardest problems.

00:02:56.000 --> 00:02:58.366
Its goal is to solve the problem for developers every day

00:02:58.366 --> 00:03:01.466
All kinds of real-world scenarios that are frequently used

00:03:01.533 --> 00:03:02.700
Writing business code

00:03:02.966 --> 00:03:04.200
Read long files

00:03:04.333 --> 00:03:05.733
Call various APIs

00:03:05.766 --> 00:03:07.766
Automatically browse a web to find information

00:03:07.866 --> 00:03:09.966
Handle various tasks in daily work

00:03:10.400 --> 00:03:11.933
These things are not cool

00:03:11.933 --> 00:03:13.966
It doesn't make the tech headlines.

00:03:14.100 --> 00:03:15.933
But they are developers

00:03:15.933 --> 00:03:18.333
80% of the work you do every day

00:03:18.566 --> 00:03:21.600
Changes in positioning are more important than performance improvements

00:03:21.766 --> 00:03:23.300
This shows that the entire industry is

00:03:23.300 --> 00:03:25.100
The standards of assessment have changed

00:03:25.266 --> 00:03:27.900
People used to compare who had the smartest answers

00:03:27.900 --> 00:03:29.533
And now it's time to start comparing

00:03:29.533 --> 00:03:31.300
Who can afford a cheaper price?

00:03:31.700 --> 00:03:33.700
When you encounter an error in more steps

00:03:33.700 --> 00:03:34.900
Able to correct itself

00:03:35.000 --> 00:03:36.400
Re-route

00:03:36.700 --> 00:03:38.500
Just like when a company is hiring

00:03:38.500 --> 00:03:39.933
The brightest college graduates

00:03:39.933 --> 00:03:41.566
Of course the best candidate

00:03:41.600 --> 00:03:43.133
But it can really bring everyday

00:03:43.133 --> 00:03:45.166
The work is done quickly and well

00:03:45.300 --> 00:03:48.133
People who don't always need someone to clean up the mess

00:03:48.133 --> 00:03:51.400
It is the most indispensable core strength in the team

00:03:51.800 --> 00:03:53.066
I work with a lot of artificial intelligence

00:03:53.066 --> 00:03:54.366
Friends of the app have communicated

00:03:54.466 --> 00:03:55.666
The largest they have encountered before

00:03:55.666 --> 00:03:57.166
The hard problem is one such question

00:03:57.333 --> 00:03:59.400
A lot of the model scores look pretty good.

00:03:59.533 --> 00:04:02.100
But really put it in the production environment

00:04:02.100 --> 00:04:04.000
All kinds of problems arise

00:04:04.466 --> 00:04:06.366
Let it write a complete function

00:04:06.500 --> 00:04:08.466
It can help you leave out the dependent packages.

00:04:08.466 --> 00:04:10.700
Let it modify the existing project code

00:04:10.766 --> 00:04:12.900
It can mess up other normal functions

00:04:12.966 --> 00:04:14.766
Where requirements are not clear

00:04:14.966 --> 00:04:16.000
It doesn't ask questions.

00:04:16.066 --> 00:04:18.500
It's all about guesswork.

00:04:18.533 --> 00:04:20.933
And the output is also unusable.

00:04:21.133 --> 00:04:23.333
In the end, developers have to spend more

00:04:23.333 --> 00:04:25.200
The energy to deal with this problem

00:04:25.266 --> 00:04:26.766
Thereby reducing work efficiency

00:04:27.166 --> 00:04:29.766
The appearance of Gemini3.7 Flash

00:04:29.800 --> 00:04:32.566
Obviously designed to deal with the above problems

00:04:32.766 --> 00:04:33.766
We can take a look at some

00:04:33.766 --> 00:04:35.400
Open benchmark data

00:04:35.400 --> 00:04:36.566
It knows what it wants.

00:04:36.566 --> 00:04:38.133
The direction of improvement is very clear

00:04:38.400 --> 00:04:41.600
Calculated according to the Frontier Code 1.1 code baseline

00:04:41.600 --> 00:04:44.933
It scored 34.4% in version 3.6.

00:04:44.933 --> 00:04:46.766
And at 43.6%

00:04:47.100 --> 00:04:48.766
Closer to real engineering

00:04:48.766 --> 00:04:51.900
On the DEPSE 1.1 test conducted in the project

00:04:51.900 --> 00:04:55.300
From 49% to 65.3% directly

00:04:55.466 --> 00:04:56.533
WebDevArena

00:04:56.600 --> 00:04:58.300
Web development-oriented competitions

00:04:58.300 --> 00:05:00.800
Has risen to 1,588 points

00:05:01.100 --> 00:05:02.933
Test with the end point command line

00:05:03.100 --> 00:05:04.933
When TerminalBench 2.1

00:05:04.933 --> 00:05:07.566
The result is 85.8%.

00:05:07.866 --> 00:05:10.366
The most extreme case is Oswald 2.0.

00:05:10.366 --> 00:05:13.200
That is to test the ability of artificial intelligence to control computer desktops

00:05:13.200 --> 00:05:18.166
The score has skyrocketed from 3.8% to 47.9% now.

00:05:18.533 --> 00:05:20.066
The model I used before is on the computer

00:05:20.066 --> 00:05:22.133
The operation is basically clicking the wrong button.

00:05:22.133 --> 00:05:23.533
The level of opening the wrong software

00:05:23.600 --> 00:05:24.866
It can now be done independently.

00:05:24.866 --> 00:05:26.733
Nearly half of all desktop tasks

00:05:26.800 --> 00:05:30.266
Such as organizing documents, using software, filling out forms, etc

00:05:30.266 --> 00:05:31.466
For the agent

00:05:31.466 --> 00:05:33.166
That's real progress.

00:05:33.533 --> 00:05:35.800
But Google doesn't hide its shortcomings

00:05:35.800 --> 00:05:38.666
In the more difficult TerminalBench 3.0 test

00:05:38.700 --> 00:05:41.300
Google got 14.9 percent.

00:05:41.533 --> 00:05:45.566
Less than 20.8% of GPT-5.6 terror

00:05:45.566 --> 00:05:47.500
On extremely complex engineering tasks

00:05:47.500 --> 00:05:49.300
It is not at the top level of the industry

00:05:49.733 --> 00:05:50.566
However, for the majority

00:05:50.566 --> 00:05:52.200
For the daily needs of developers

00:05:52.266 --> 00:05:54.500
It's powerful enough to use

00:05:54.600 --> 00:05:57.600
And better understand how to do things well and correctly

00:05:57.900 --> 00:05:59.466
Targeted strengthening of some

00:05:59.466 --> 00:06:01.600
Capabilities that are often used in practical development

00:06:01.733 --> 00:06:03.766
Strictly follow the literal meaning

00:06:03.900 --> 00:06:05.300
And will not be added at will.

00:06:05.300 --> 00:06:07.000
Some features that were not requested

00:06:07.300 --> 00:06:08.800
Unclear requirements

00:06:09.133 --> 00:06:10.666
The place will take the initiative to ask.

00:06:10.700 --> 00:06:12.066
I'm not guessing

00:06:12.200 --> 00:06:13.766
When you find that the path is blocked

00:06:13.766 --> 00:06:15.066
Or a lot of mistakes

00:06:15.066 --> 00:06:16.300
It will stop the loss in time.

00:06:16.366 --> 00:06:17.966
It won't go dark one way

00:06:18.000 --> 00:06:20.266
Waste a lot of time and computing resources

00:06:20.666 --> 00:06:23.666
These characteristics cannot be reflected in the running points

00:06:23.700 --> 00:06:25.600
But when it is actually used

00:06:25.600 --> 00:06:27.733
The difference in experience will be obvious

00:06:28.066 --> 00:06:29.300
After talking about the performance

00:06:29.300 --> 00:06:31.200
Let's talk about the latest updates in this update

00:06:31.200 --> 00:06:33.733
The core point of agent-age economics

00:06:33.766 --> 00:06:35.900
That is, the artificial intelligence model cost account

00:06:35.900 --> 00:06:37.900
The algorithms have changed.

00:06:38.333 --> 00:06:39.933
When people compare the prices of models

00:06:39.933 --> 00:06:42.100
Only care how much per hundred TOKEN

00:06:42.133 --> 00:06:43.933
Use whichever is cheaper

00:06:44.066 --> 00:06:46.666
It was correct in the previous one-round conversation.

00:06:46.666 --> 00:06:48.500
When you ask a question

00:06:48.566 --> 00:06:50.100
The model will answer once

00:06:50.100 --> 00:06:51.866
And the calculations are very accurate

00:06:52.133 --> 00:06:54.366
But in the age of agents

00:06:54.366 --> 00:06:56.500
It doesn't work at all.

00:06:56.866 --> 00:06:59.533
Agents don't answer one question at a time

00:06:59.566 --> 00:07:02.200
It's always a series of tasks.

00:07:02.400 --> 00:07:04.866
The first thing to do is to figure out what the goal is

00:07:04.866 --> 00:07:08.066
Then mobilize the appropriate tools to achieve this goal

00:07:08.366 --> 00:07:09.966
Then read the relevant files

00:07:10.066 --> 00:07:11.733
Write code and test it

00:07:11.766 --> 00:07:14.566
Once a problem is found, it will be corrected and retried in time.

00:07:14.900 --> 00:07:16.466
It may be used throughout the process

00:07:16.466 --> 00:07:18.500
To a dozen or even dozens of models

00:07:18.900 --> 00:07:21.266
At this point, the main metric for measuring costs is no longer available

00:07:21.266 --> 00:07:23.100
Is the price per million TOKEN

00:07:23.133 --> 00:07:25.300
And it becomes a complete

00:07:25.300 --> 00:07:27.000
All expenses required for the task

00:07:27.500 --> 00:07:29.666
Some of the models look really cheap.

00:07:29.733 --> 00:07:31.500
But if you do it incorrectly the first time.

00:07:31.500 --> 00:07:33.166
You have to try it over and over again.

00:07:33.266 --> 00:07:35.966
During this period, many unused tools will also be called

00:07:36.100 --> 00:07:37.666
At the end of the general ledger

00:07:37.900 --> 00:07:40.766
The cost is actually higher than the price of a single model

00:07:40.900 --> 00:07:42.666
But as long as you can do it once

00:07:42.666 --> 00:07:44.700
The model of the thing is much more valuable

00:07:44.933 --> 00:07:47.166
Let's say you want to hire a decorator.

00:07:47.166 --> 00:07:48.200
A cheap wage

00:07:48.200 --> 00:07:49.400
But the work is rough

00:07:49.533 --> 00:07:52.133
After installing, you have to remove it and reinstall it three times.

00:07:52.400 --> 00:07:54.200
In the end, it costs more money and time than

00:07:54.200 --> 00:07:55.700
Find a salary that is slightly higher

00:07:55.700 --> 00:07:57.066
But you can only use it once.

00:07:57.066 --> 00:07:59.266
A lot more people get the work done.

00:07:59.533 --> 00:08:02.466
The focus of this optimization of Gemini3.7Flash is

00:08:02.466 --> 00:08:04.533
Increase the probability of getting it right the first time

00:08:04.600 --> 00:08:06.200
Reduce the use of unnecessary tools

00:08:06.200 --> 00:08:08.000
And pointless repetition

00:08:08.466 --> 00:08:10.266
It doesn't have to be the cheapest model

00:08:10.366 --> 00:08:12.800
But as long as the number of retries can be reduced

00:08:12.800 --> 00:08:14.266
Then complete the same task

00:08:14.266 --> 00:08:15.866
The total cost will decrease

00:08:16.100 --> 00:08:18.733
This is the bill that business users really care about

00:08:18.933 --> 00:08:21.300
After all, when companies use artificial intelligence

00:08:21.300 --> 00:08:22.800
It's to cut costs.

00:08:22.866 --> 00:08:23.800
Increase efficiency

00:08:23.933 --> 00:08:26.166
It's not about buying the lowest-priced token.

00:08:26.666 --> 00:08:28.733
Analyze according to such cost considerations

00:08:28.733 --> 00:08:30.700
The price strategy it adopts

00:08:30.766 --> 00:08:32.100
It is easy to understand

00:08:32.100 --> 00:08:33.933
To Google's strategic intentions

00:08:34.133 --> 00:08:35.666
The official selling price is

00:08:35.666 --> 00:08:38.766
The promotion price is before the end of 2026.

00:08:38.866 --> 00:08:41.700
Enter TOKEN $0.75 per million

00:08:41.733 --> 00:08:44.400
Output TOKEN $3.75 per million

00:08:44.766 --> 00:08:48.300
The official price will be implemented from January 1, 2027

00:08:48.400 --> 00:08:50.566
When it rose to $1.50

00:08:50.666 --> 00:08:52.666
Output rose to $7.50

00:08:52.666 --> 00:08:53.666
That's doubled.

00:08:54.000 --> 00:08:55.200
From the price point of view

00:08:55.200 --> 00:08:57.600
It's a bit more expensive than many open source models.

00:08:57.600 --> 00:08:59.766
However, considering its performance and

00:08:59.766 --> 00:09:01.666
Considering the task completion rate

00:09:01.666 --> 00:09:03.800
In fact, its cost performance is very high.

00:09:04.000 --> 00:09:05.466
Google offers such a price

00:09:05.600 --> 00:09:08.533
It's not about short-term profits.

00:09:08.533 --> 00:09:10.700
It has two distinct strategic objectives.

00:09:11.100 --> 00:09:12.733
The first task is to seize

00:09:12.733 --> 00:09:14.366
Developer default options

00:09:14.566 --> 00:09:16.700
The current cost of AI development migration is actually

00:09:16.700 --> 00:09:18.566
A little bit higher than people think.

00:09:18.733 --> 00:09:20.166
When developers become accustomed to a certain

00:09:20.166 --> 00:09:21.666
The way a model answers

00:09:21.733 --> 00:09:24.366
And put the prompt thesaurus tool chain worker

00:09:24.366 --> 00:09:26.400
When the workflow is based on the model

00:09:26.400 --> 00:09:28.566
It will not be easy to replace other models.

00:09:28.866 --> 00:09:31.466
Once replaced, it needs to be readjusted and tested.

00:09:31.566 --> 00:09:32.533
The cost is high

00:09:32.766 --> 00:09:34.666
Google is currently using something close to

00:09:34.666 --> 00:09:36.300
The price of the cost to promote

00:09:36.400 --> 00:09:38.133
The purpose is to enable developers to compare

00:09:38.133 --> 00:09:40.166
Low cost to use in the product

00:09:40.166 --> 00:09:42.133
After developers get used to it, gradually

00:09:42.133 --> 00:09:44.500
Bind to the Gemini ecosystem

00:09:44.866 --> 00:09:47.400
Even if the price increases again in 2027

00:09:47.400 --> 00:09:50.400
If the total cost is still lower than the competition

00:09:50.400 --> 00:09:52.866
Most people are not going to leave.

00:09:53.100 --> 00:09:54.766
The second objective is to seize

00:09:54.766 --> 00:09:57.066
The dominance of enterprise model routing

00:09:57.466 --> 00:09:59.000
A slightly larger company does not

00:09:59.000 --> 00:10:00.733
Not only one model will be used

00:10:00.766 --> 00:10:02.533
Instead, it will perform hierarchical routing

00:10:02.600 --> 00:10:05.600
It's about five percent of the most complex, the most difficult.

00:10:05.600 --> 00:10:07.500
The task will be entrusted to the most expensive and capable

00:10:07.500 --> 00:10:09.900
The most powerful top-level model to handle

00:10:10.066 --> 00:10:12.666
The rest of the high-frequency repetitions are less difficult

00:10:12.666 --> 00:10:15.000
The schedule work is handed over to cost-effective

00:10:15.000 --> 00:10:17.533
Fast and stable main model to complete

00:10:17.966 --> 00:10:19.600
95% of the tasks

00:10:19.666 --> 00:10:22.500
The call volume is dozens or hundreds of times that of the top model

00:10:22.533 --> 00:10:25.400
This is the biggest foundation for the artificial intelligence market

00:10:25.733 --> 00:10:28.366
Gemini3.7 Flash is targeting this market

00:10:28.366 --> 00:10:31.700
Won't compete with top models for that 5% highlight scene

00:10:31.700 --> 00:10:34.900
Instead, eat the remaining 95% of high-frequency calls

00:10:35.266 --> 00:10:36.533
The unit price doesn't seem to be high.

00:10:36.533 --> 00:10:38.366
But I can't stand it, and I use it a lot.

00:10:38.466 --> 00:10:41.066
A steady stream of income is better than relying on the top

00:10:41.066 --> 00:10:43.566
Artificial intelligence to make money is more practical

00:10:43.866 --> 00:10:45.366
Except for the model itself

00:10:45.400 --> 00:10:48.133
Google also has an advantage over other vendors

00:10:48.166 --> 00:10:50.333
It's its own ecosystem.

00:10:50.733 --> 00:10:53.133
And the new product Spark launched simultaneously this time

00:10:53.166 --> 00:10:55.466
All-weather agents use this card

00:10:55.866 --> 00:10:58.533
Even third-party artificial intelligence does the best

00:10:58.533 --> 00:11:00.766
There is inevitably a problem that cannot be bypassed

00:11:00.900 --> 00:11:03.066
It can't get your context.

00:11:03.500 --> 00:11:05.366
If you want it to help you deal with emails

00:11:05.366 --> 00:11:07.066
You need to send the email yourself.

00:11:07.066 --> 00:11:09.066
Export the content and upload it again.

00:11:09.066 --> 00:11:10.566
Let it organize your schedule for you

00:11:10.700 --> 00:11:12.800
You need to send him the schedule yourself.

00:11:13.000 --> 00:11:14.266
Let it modify the document

00:11:14.333 --> 00:11:15.900
Uploading files and other things

00:11:15.900 --> 00:11:17.266
Do it again and again

00:11:17.333 --> 00:11:18.200
Very inconvenient

00:11:18.500 --> 00:11:20.366
But Spark is directly operated by Google

00:11:20.366 --> 00:11:21.866
Developed proxy software

00:11:21.900 --> 00:11:24.533
It can be very deep with Gmail Google.

00:11:24.566 --> 00:11:26.733
Calendars, Google Docs, and almost every

00:11:26.733 --> 00:11:28.700
Workplace users use software to connect

00:11:29.100 --> 00:11:30.466
It can automatically read your

00:11:30.466 --> 00:11:32.900
Email schedule documents and other information

00:11:32.933 --> 00:11:34.466
And you don't need to keep

00:11:34.466 --> 00:11:36.000
Perform data synchronization operations

00:11:36.066 --> 00:11:38.400
I can directly help you reply to emails.

00:11:38.400 --> 00:11:41.333
Schedule meetings, modify documents, organize materials, etc

00:11:41.333 --> 00:11:42.733
This user experience is

00:11:42.733 --> 00:11:45.300
Third-party artificial intelligence is difficult to do

00:11:45.500 --> 00:11:47.066
But the higher the authority

00:11:47.066 --> 00:11:49.200
The greater the risk

00:11:49.566 --> 00:11:51.966
If artificial intelligence can read and

00:11:51.966 --> 00:11:54.566
Modify your email documents and other information

00:11:54.600 --> 00:11:56.600
Then a wrong operation will

00:11:56.600 --> 00:11:58.400
Causing important files to be deleted

00:11:58.666 --> 00:12:00.800
Or provide customers with the wrong content

00:12:00.800 --> 00:12:02.666
The impact is more than ordinary

00:12:02.666 --> 00:12:04.866
The AI gibberish is much bigger

00:12:05.266 --> 00:12:07.666
So Google also needs to do three things well

00:12:07.866 --> 00:12:09.333
Extremely high operational transparency

00:12:09.333 --> 00:12:10.566
Let users know

00:12:10.566 --> 00:12:12.566
What to do every step of artificial intelligence

00:12:12.600 --> 00:12:15.133
Important operations need to be confirmed manually.

00:12:15.400 --> 00:12:17.333
For example, sending an email to delete

00:12:17.333 --> 00:12:19.300
Your consent is required before the document

00:12:19.333 --> 00:12:21.066
There is also a complete set of audit logs

00:12:21.200 --> 00:12:23.500
All operations can be reversed

00:12:23.900 --> 00:12:26.666
It also adds Antigravity development environment

00:12:26.666 --> 00:12:29.566
Put the operation of artificial intelligence in a rollback

00:12:29.666 --> 00:12:31.400
Within the boundaries of control

00:12:31.733 --> 00:12:33.700
Even if the artificial intelligence operation goes wrong

00:12:33.700 --> 00:12:35.400
Or encounter network instability

00:12:35.533 --> 00:12:37.933
CAPTCHAs and other unexpected situations in real life

00:12:37.933 --> 00:12:39.500
It can also be rolled back with one click.

00:12:39.500 --> 00:12:41.200
The function returns to the previous situation

00:12:41.400 --> 00:12:43.533
It's not going to bring down your local environment.

00:12:43.966 --> 00:12:45.800
Although these details are not cool

00:12:45.933 --> 00:12:48.400
But they're real agents to face

00:12:48.400 --> 00:12:50.200
And solve various problems encountered in the work

00:12:50.200 --> 00:12:52.266
A hurdle that the problem must cross

00:12:52.733 --> 00:12:54.333
If you look a little farther

00:12:54.366 --> 00:12:56.100
The current three major artificial intelligence

00:12:56.100 --> 00:12:59.000
The path taken by the giants is very clear

00:12:59.166 --> 00:13:02.200
And they're not the same thing.

00:13:02.600 --> 00:13:05.700
The main selling points of Google's Gemini3.7 Flash are

00:13:05.700 --> 00:13:08.300
Cost performance, project completion rate, tool chain binding

00:13:08.300 --> 00:13:10.533
And a full set of ecological entrances

00:13:10.566 --> 00:13:13.166
The goal is to become the main mode of operation for the entire industry

00:13:13.400 --> 00:13:15.800
And capture the largest market for daily use

00:13:16.400 --> 00:13:19.266
The path taken by OpenAi is the upper limit of performance

00:13:19.266 --> 00:13:21.466
In combination with the extreme speed route

00:13:21.766 --> 00:13:24.700
It launched a new generation of inference service, Ultrafast.

00:13:25.133 --> 00:13:26.600
It can be reached at the peak.

00:13:26.600 --> 00:13:29.400
750 output TOKEN speed per second

00:13:29.466 --> 00:13:30.900
To eliminate the agent

00:13:30.900 --> 00:13:32.466
The problem of continuous working delay

00:13:32.766 --> 00:13:34.800
In the past, when agents did tasks, they had to wait

00:13:34.800 --> 00:13:36.500
It's taking things one step at a time.

00:13:36.700 --> 00:13:37.866
The card feels very strong

00:13:37.933 --> 00:13:39.300
Extremely fast reasoning allows the entire

00:13:39.300 --> 00:13:41.500
The process is as smooth as a human operation

00:13:41.966 --> 00:13:44.800
Mainly targeting high-end scenarios and real-time interaction

00:13:44.966 --> 00:13:47.066
Like real-time voice conversations.

00:13:47.066 --> 00:13:49.133
Real-time device control, complex reasoning, and more

00:13:49.133 --> 00:13:50.933
For latency and capacity limits

00:13:50.933 --> 00:13:52.400
Scenarios with higher requirements

00:13:52.866 --> 00:13:55.133
And Deepseek takes a differentiated route

00:13:55.200 --> 00:13:57.366
The V4Pro version has actually increased the price.

00:13:57.366 --> 00:13:59.466
And they began to pursue higher gross margins.

00:13:59.600 --> 00:14:02.666
It came as a cheap alternative.

00:14:02.800 --> 00:14:05.766
Now we need to move towards higher value-added products

00:14:05.766 --> 00:14:07.200
No more price wars

00:14:07.266 --> 00:14:09.366
Make use of certain scenarios to generate revenue

00:14:09.800 --> 00:14:11.066
Three different paths

00:14:11.066 --> 00:14:12.166
In fact, it shows that

00:14:12.166 --> 00:14:14.166
The differentiation trend of artificial intelligence market

00:14:14.300 --> 00:14:16.566
And it's not going to be monolithic.

00:14:16.800 --> 00:14:19.866
High-end scenarios, mass market, cost-effective scenarios, etc

00:14:19.866 --> 00:14:22.400
There will be corresponding players in each scenario in the future

00:14:22.466 --> 00:14:24.300
And competition is no longer simple

00:14:24.300 --> 00:14:26.266
It's a question of who has the highest score and who wins

00:14:26.666 --> 00:14:28.966
But we should also be soberly aware

00:14:29.000 --> 00:14:30.666
Google's strategic deployment this time

00:14:30.700 --> 00:14:32.300
It's not risk-free.

00:14:32.366 --> 00:14:33.300
In some ways

00:14:33.300 --> 00:14:34.300
If not done well

00:14:34.300 --> 00:14:36.700
Could have an impact on how it achieves its goals

00:14:37.166 --> 00:14:39.700
The first problem is that the frequency of releases is too fast

00:14:39.900 --> 00:14:42.533
A major version update every three weeks

00:14:42.533 --> 00:14:45.000
This is a good thing for the average user

00:14:45.133 --> 00:14:47.000
However, for enterprise users

00:14:47.133 --> 00:14:50.266
The cost of compatibility testing and maintenance will be high

00:14:50.600 --> 00:14:52.866
The enterprise focuses its main business on APIs

00:14:52.900 --> 00:14:54.366
Every model update requires

00:14:54.366 --> 00:14:55.966
Re-test all functions

00:14:56.066 --> 00:14:58.200
Afraid that the new version will change the output format

00:14:58.266 --> 00:14:59.766
Or the function changes.

00:14:59.933 --> 00:15:01.666
And that's going to impact business online.

00:15:02.133 --> 00:15:03.866
Iterating too fast can make it conservative

00:15:03.866 --> 00:15:06.166
Businesses cannot bring critical business processes online

00:15:06.400 --> 00:15:08.466
The second issue is the relationship between the benchmark and

00:15:08.466 --> 00:15:09.933
Differences between actual businesses

00:15:10.200 --> 00:15:12.466
Public benchmarks are a matter of standards

00:15:12.533 --> 00:15:15.366
However, the actual business environment for enterprises is ever-changing

00:15:15.500 --> 00:15:18.066
Some companies have decades-old systems

00:15:18.133 --> 00:15:20.900
Some companies have complex internal backends

00:15:21.100 --> 00:15:23.333
There are also some very peculiar fringe scenes

00:15:23.700 --> 00:15:25.333
A high score does not mean that you are in such a situation

00:15:25.333 --> 00:15:27.600
The environment can maintain a stable operating state

00:15:27.766 --> 00:15:30.066
In the actual use process of enterprise users

00:15:30.066 --> 00:15:32.900
Can we provide sustainable services to enterprise users?

00:15:33.066 --> 00:15:35.066
It should be judged according to the actual situation

00:15:35.500 --> 00:15:36.966
The third issue is discounts

00:15:36.966 --> 00:15:38.766
After the expiration date, the price will double

00:15:38.966 --> 00:15:41.066
The current low price is the price during the promotion period.

00:15:41.066 --> 00:15:44.566
It will directly double the official price in 2,027

00:15:44.600 --> 00:15:46.700
By then, many small businesses and individuals

00:15:46.700 --> 00:15:49.266
The developer's cost account has to be recalculated

00:15:49.366 --> 00:15:50.800
At that time, competing products will be launched.

00:15:50.800 --> 00:15:52.466
More cost-effective products

00:15:52.466 --> 00:15:55.333
A portion of price-sensitive users churn

00:15:56.000 --> 00:15:59.500
The fourth question is about privacy and the risk of overreach

00:15:59.966 --> 00:16:01.933
Agent obtained email, document

00:16:01.933 --> 00:16:03.566
After high authority such as schedule

00:16:03.566 --> 00:16:05.733
The consequences of a misreading or writing

00:16:05.733 --> 00:16:08.166
Much more serious than ordinary AI hallucinations

00:16:08.566 --> 00:16:10.700
A common hallucination is saying something wrong.

00:16:10.700 --> 00:16:13.000
And the hallucinations of high-powered agents

00:16:13.000 --> 00:16:14.766
There may be actual losses

00:16:14.933 --> 00:16:16.866
If this problem is not handled well

00:16:16.966 --> 00:16:18.900
Businesses and individual users are hesitant

00:16:18.900 --> 00:16:20.800
Really delegate high authority to others

00:16:21.133 --> 00:16:22.300
The fifth question is

00:16:22.300 --> 00:16:24.100
Google's product line is very fragmented

00:16:24.466 --> 00:16:27.766
Gemini currently has Pro Flash Ultra and other versions

00:16:27.800 --> 00:16:30.600
And also launched Spark proxy and Antigravity

00:16:30.866 --> 00:16:33.666
A series of products with various names such as development environment

00:16:33.966 --> 00:16:35.700
Ordinary users simply cannot

00:16:35.700 --> 00:16:38.266
Identify which products are which

00:16:38.366 --> 00:16:40.700
Choose whichever one you choose.

00:16:41.000 --> 00:16:42.266
There are too many product lines.

00:16:42.266 --> 00:16:44.666
Instead, it will increase the user's selection cost

00:16:44.866 --> 00:16:45.866
Speaking of which

00:16:45.866 --> 00:16:47.700
Some people may think of Google before

00:16:47.700 --> 00:16:49.766
The core executives' departure to start a business

00:16:49.900 --> 00:16:50.700
At that time

00:16:50.700 --> 00:16:51.933
Many people believe that Google's

00:16:51.933 --> 00:16:53.366
Artificial intelligence is dying

00:16:53.500 --> 00:16:55.100
After the core talents leave

00:16:55.100 --> 00:16:57.300
Subsequent iterations will definitely not keep up

00:16:57.700 --> 00:16:59.400
It's only been a few months

00:16:59.533 --> 00:17:01.566
Gemini releases a new version every three weeks

00:17:01.666 --> 00:17:03.800
It's not slowing down at all.

00:17:03.800 --> 00:17:05.400
It's getting faster

00:17:05.600 --> 00:17:07.100
This thing is actually

00:17:07.100 --> 00:17:08.333
A warning for us

00:17:08.700 --> 00:17:10.800
The current artificial intelligence competition is not

00:17:10.800 --> 00:17:12.766
Relying on a few top researchers

00:17:12.766 --> 00:17:14.733
You can do it alone

00:17:15.066 --> 00:17:16.266
It has now become a

00:17:16.266 --> 00:17:18.100
Industrial assembly line competition

00:17:18.300 --> 00:17:20.866
It relies on computing power scale and data flywheel.

00:17:20.900 --> 00:17:22.466
Automated evaluation system

00:17:22.466 --> 00:17:25.400
A complete system composed of product distribution channels

00:17:25.766 --> 00:17:27.666
Top talent is important

00:17:27.700 --> 00:17:30.133
But the iterative ability of the system is

00:17:30.133 --> 00:17:32.000
The current core competitiveness of large factories

00:17:32.466 --> 00:17:34.000
When the assembly line is running

00:17:34.000 --> 00:17:35.666
The transfer of individual personnel will not

00:17:35.666 --> 00:17:37.566
Altering the rate at which the entire system iterates

00:17:37.866 --> 00:17:38.933
For ordinary people

00:17:38.933 --> 00:17:41.133
And this has a very direct implication as well.

00:17:41.266 --> 00:17:43.266
Future human core values and

00:17:43.266 --> 00:17:45.366
It's not about who writes code faster than AI

00:17:45.366 --> 00:17:46.766
Who executes better?

00:17:46.933 --> 00:17:48.966
And these problems the model does very quickly.

00:17:48.966 --> 00:17:49.966
And it's cheap.

00:17:50.466 --> 00:17:52.200
The true irreplaceability of human beings

00:17:52.200 --> 00:17:53.933
It's about setting the right target.

00:17:53.966 --> 00:17:55.300
Control the final outcome

00:17:55.300 --> 00:17:56.566
Be responsible for decision-making

00:17:56.566 --> 00:17:59.166
The model only plays an efficient execution role

00:17:59.466 --> 00:18:02.800
It's better to learn how to make good use of this execution layer first

00:18:02.800 --> 00:18:04.366
To orient the model

00:18:04.400 --> 00:18:05.666
The person who checked the results

00:18:06.000 --> 00:18:08.266
And will Google do it this time?

00:18:08.266 --> 00:18:10.600
In fact, you don't need to see how well the press conference blows

00:18:10.600 --> 00:18:13.566
You only need to focus on 4 main data

00:18:13.733 --> 00:18:16.000
The second is the criteria for developers to choose

00:18:16.166 --> 00:18:17.666
Will it really become a reality for everyone?

00:18:17.666 --> 00:18:19.733
Default model used in the production environment

00:18:19.766 --> 00:18:21.966
It's not just for testing and playing.

00:18:22.000 --> 00:18:24.200
And write some simple scripts

00:18:24.666 --> 00:18:26.100
The second step is to evaluate the entire

00:18:26.100 --> 00:18:27.700
The cost of the project is calculated

00:18:27.766 --> 00:18:30.500
After including retries, manual repairs, etc.

00:18:30.500 --> 00:18:32.800
The total cost of the task actually completed by the project

00:18:32.800 --> 00:18:35.166
Is it really lower than the competition?

00:18:35.533 --> 00:18:37.366
The third is user stickiness

00:18:37.900 --> 00:18:40.166
Can a proxy product like Spark make

00:18:40.166 --> 00:18:42.333
Users form good habits of high-frequency use

00:18:42.366 --> 00:18:43.933
It can't just be novelty.

00:18:43.933 --> 00:18:45.100
I forgot about it.

00:18:45.166 --> 00:18:47.000
The fourth is ecological governance

00:18:47.133 --> 00:18:48.733
Can Google put the messy

00:18:48.733 --> 00:18:50.466
The product line is organized in an orderly manner

00:18:50.600 --> 00:18:52.533
And it can effectively reduce the user's

00:18:52.533 --> 00:18:54.733
What about the difficulty of decision-making and the cost of use?

00:18:55.133 --> 00:18:56.266
So looking back

00:18:56.333 --> 00:18:59.333
Google has released a new version of Gemini this time

00:18:59.400 --> 00:19:01.166
What really deserves our attention is

00:19:01.166 --> 00:19:02.766
How many more parameters does it have?

00:19:02.800 --> 00:19:05.366
It's not about which leaderboard surpasses whom.

00:19:05.666 --> 00:19:06.700
More importantly

00:19:06.766 --> 00:19:09.300
Google is transforming Gemini's functionality from the original

00:19:09.300 --> 00:19:11.866
The AI that answers questions is gradually transforming into the present

00:19:11.866 --> 00:19:14.566
A working AI that can help you do things.

00:19:14.933 --> 00:19:16.500
And what it represents is

00:19:16.500 --> 00:19:18.600
An important change in the artificial intelligence industry

00:19:18.800 --> 00:19:21.200
People used to compete about how smart models were.

00:19:21.333 --> 00:19:24.300
The next thing to fight is whether the model can complete the task

00:19:24.700 --> 00:19:27.333
Who can make agents use tools more stably?

00:19:27.366 --> 00:19:29.066
And reduce the occurrence of errors

00:19:29.066 --> 00:19:31.166
You stop doing the same thing over and over again.

00:19:31.166 --> 00:19:33.966
And minimize the cost of each task

00:19:34.166 --> 00:19:35.700
Then he may be truly

00:19:35.700 --> 00:19:37.666
Enter the work of the company

00:19:38.066 --> 00:19:39.566
Therefore, Gemini this update

00:19:39.566 --> 00:19:42.166
I think there are three areas that deserve more attention

00:19:42.466 --> 00:19:44.133
Will developers really put

00:19:44.133 --> 00:19:46.366
They are used in the production environment

00:19:46.533 --> 00:19:49.133
After the company calculates the full cost

00:19:49.366 --> 00:19:51.366
Will you still be willing to use it for a long time?

00:19:51.366 --> 00:19:53.866
And will Google be able to leverage Gmail?

00:19:53.966 --> 00:19:56.700
The entire ecosystem, including Docs Calendar

00:19:56.766 --> 00:19:58.800
And turn Gemini into a user every day

00:19:58.800 --> 00:20:00.300
Applications that are required

00:20:00.766 --> 00:20:02.166
If the above things finally

00:20:02.166 --> 00:20:03.600
If it can be run

00:20:03.600 --> 00:20:05.966
Then Google keeps updating Gemini this time.

00:20:05.966 --> 00:20:07.666
And updated frequently

00:20:07.700 --> 00:20:09.933
It's not just an ordinary version upgrade.

00:20:10.166 --> 00:20:11.666
To some extent

00:20:11.700 --> 00:20:13.600
It already heralds artificial intelligence

00:20:13.600 --> 00:20:15.466
The possibility of the age of agency

00:20:15.600 --> 00:20:18.333
But it's too early to tell.

00:20:18.500 --> 00:20:20.466
The artificial intelligence industry is changing rapidly

00:20:20.666 --> 00:20:23.000
The leading companies today are not

00:20:23.000 --> 00:20:25.600
Representatives will still maintain their lead tomorrow

00:20:26.100 --> 00:20:27.600
It is worth our long-term attention that

00:20:27.600 --> 00:20:29.500
Are these models available from the press conference?

00:20:29.500 --> 00:20:31.966
In the parameters, scores and other information published in

00:20:31.966 --> 00:20:34.000
Really step into the workflow of ordinary people

00:20:34.000 --> 00:20:36.166
And the actual business of the enterprise

00:20:36.333 --> 00:20:38.100
This is what I've always wanted to do

00:20:38.100 --> 00:20:40.266
It's about telling people who posted what.

00:20:40.300 --> 00:20:41.566
It's not just the news.

00:20:41.566 --> 00:20:43.300
Just make it clear.

00:20:43.466 --> 00:20:45.600
And to put the industry behind the news

00:20:45.600 --> 00:20:47.966
The logic is also taken apart for everyone to see

00:20:48.300 --> 00:20:50.733
If you want to continue learning artificial intelligence

00:20:50.733 --> 00:20:53.500
Tech giants and the deep principles of the US stock market

00:20:53.500 --> 00:20:54.800
Please follow this channel

00:20:55.133 --> 00:20:56.666
Finally, I would like to ask you

00:20:56.700 --> 00:20:59.600
What do you think will be Gemini's greatest advantage in the future?

00:20:59.600 --> 00:21:01.266
It's the power of the model.

00:21:01.266 --> 00:21:03.966
Or is it the power behind Google's entire ecosystem?

00:21:04.366 --> 00:21:07.000
Everyone is welcome to express their opinions in the comment section

00:21:07.366 --> 00:21:08.933
This issue is for industry only

00:21:08.933 --> 00:21:10.700
Information sharing and exchange of ideas

00:21:10.700 --> 00:21:13.600
Does not constitute any investment advice or trading basis

00:21:14.066 --> 00:21:15.566
The AI industry is changing rapidly

00:21:15.600 --> 00:21:17.600
There is uncertainty in any judgment

00:21:17.600 --> 00:21:18.966
Please think for yourself

00:21:19.066 --> 00:21:20.666
A rational view of market changes

00:21:20.666 --> 00:21:21.833
See you next time.