1
00:00:00,000 --> 00:00:04,400
今天,Deep Seek团队给我们带来了一个全新的模型。

2
00:00:05,600 --> 00:00:09,420
也就是Deep Seek V4系列里的Deep Seek V4.1 Flash。

3
00:00:10,480 --> 00:00:13,920
这是一个临时测试版,会在9月10日过期。

4
00:00:14,960 --> 00:00:17,460
而且里面其实有很多东西值得了。

5
00:00:18,460 --> 00:00:22,420
Deep Seek声称,Version 4.1 Flash基于全新的架构。

6
00:00:23,160 --> 00:00:26,280
原生支持多模态,而且速度明显更快。

7
00:00:26,860 --> 00:00:28,840
但价格还是Version 4的水平。

8
00:00:28,840 --> 00:00:32,640
他们发布这个模型的方式也挺离谱的

9
00:00:32,640 --> 00:00:35,560
现在看他们到底会怎么做

10
00:00:35,560 --> 00:00:36,960
真的很有意思

11
00:00:36,960 --> 00:00:39,520
目前还没有官方基准

12
00:00:39,520 --> 00:00:41,340
参数量或者权重

13
00:00:41,340 --> 00:00:44,000
他们基本上就是直接说

14
00:00:44,000 --> 00:00:45,020
这是接口

15
00:00:45,020 --> 00:00:46,440
你有两天时间

16
00:00:46,440 --> 00:00:48,560
试着把它攻破

17
00:00:48,560 --> 00:00:51,980
你可以通过现有的Deep Seek API访问它

18
00:00:51,980 --> 00:00:55,880
使用Deep Seek Version 4.1 Flash的Model ID

19
00:00:55,880 --> 00:00:58,140
它有效期到9月10

20
00:00:58,140 --> 00:01:00,760
目前价格和4Flash一样

21
00:01:00,760 --> 00:01:05,420
不过测试请求每个账号最多只能同时发20个

22
00:01:05,420 --> 00:01:08,620
Deep Seek甚至还会问测试者

23
00:01:08,620 --> 00:01:12,660
这个新架构未来能不能彻底取代4Pro

24
00:01:12,660 --> 00:01:16,380
这也让这次临时发布更有看透

25
00:01:16,380 --> 00:01:20,080
Deep Seek 4.1Flash的速度快得惊人

26
00:01:24,080 --> 00:01:26,860
他们每秒大约能处理400个Token

27
00:01:26,860 --> 00:01:28,160
说实话

28
00:01:28,160 --> 00:01:31,660
这个速度真的太夸张了

29
00:01:31,660 --> 00:01:33,540
它生成回复快的惊人

30
00:01:33,540 --> 00:01:37,100
现在这个模型在速度上

31
00:01:37,100 --> 00:01:39,040
确实是在和Muse Spark E

32
00:01:39,040 --> 00:01:41,320
三正面竞争

33
00:01:41,320 --> 00:01:42,900
这真是相当离谱

34
00:01:42,900 --> 00:01:46,020
而且它比Muse模型便宜得多

35
00:01:46,020 --> 00:01:47,520
但记住

36
00:01:47,520 --> 00:01:48,640
9月10日之前

37
00:01:48,640 --> 00:01:50,280
这个测试板就会消失了

38
00:01:50,280 --> 00:01:51,440
我也知道

39
00:01:51,440 --> 00:01:54,340
最近因为Deep Seed的新价格

40
00:01:54,340 --> 00:01:56,300
不少人心里都不太舒服

41
00:01:56,860 --> 00:01:59,400
不过看起来,这一切很快就会变。

42
00:02:00,200 --> 00:02:05,740
DC也会在9月10日北京时间12,00下调Version 4 Flash的价格。

43
00:02:06,620 --> 00:02:10,500
有了新的Cash,输入成本会从每百万Input Tokens赢。

44
00:02:11,460 --> 00:02:15,180
0075美分降到0.003美分,缓存被命重视。

45
00:02:16,040 --> 00:02:21,420
输入数据的成本也会从每百万Token约22美分降到15美分。

46
00:02:21,420 --> 00:02:28,000
输出数据的成本则会从每百万OPPO tokens67美分降到60美分

47
00:02:28,000 --> 00:02:30,360
至于高峰时段的价格

48
00:02:30,360 --> 00:02:31,660
它们将保持不变

49
00:02:31,660 --> 00:02:34,180
虽然在两倍速下看起来没变

50
00:02:34,180 --> 00:02:35,180
但总体来说

51
00:02:35,180 --> 00:02:38,540
Deep6 Flash正在变得更快也更便宜

52
00:02:38,540 --> 00:02:42,400
如果4.1 Flash真的只是这种架构的预览

53
00:02:42,400 --> 00:02:45,980
那我接下来要展示的几个演示版本

54
00:02:45,980 --> 00:02:48,780
都会让我们在新的Pearl上

55
00:02:48,780 --> 00:02:52,300
连同这条基于新架构的新模型线一起

56
00:02:52,300 --> 00:02:53,600
明显处于劣势

57
00:02:53,600 --> 00:02:55,860
AI领域有很多内容

58
00:02:55,860 --> 00:02:57,800
我不会发到YouTube频道上

59
00:02:57,800 --> 00:03:01,540
但你可以在我免费的newsletter里找到它们

60
00:03:01,540 --> 00:03:03,780
链接就在下方简介里

61
00:03:03,780 --> 00:03:05,420
你可以完全免费订阅

62
00:03:05,420 --> 00:03:06,900
不过朋友们

63
00:03:06,900 --> 00:03:08,920
我拍这条视频的原因是

64
00:03:08,920 --> 00:03:12,240
这个模型在生成各种代码

65
00:03:12,240 --> 00:03:14,440
视力方面真的非常强

66
00:03:14,440 --> 00:03:16,400
这就是Deep6

67
00:03:16,400 --> 00:03:22,480
E-Flash在我们频道之前测试过的几乎所有领域都超过的地方

68
00:03:22,480 --> 00:03:28,140
从3D环境编码到前端输出还有很多其他方面

69
00:03:28,140 --> 00:03:33,500
这次测试里4.1 Flash和另一个神秘模型做了对比

70
00:03:33,500 --> 00:03:37,620
而这个神秘模型很可能是Block 4.7

71
00:03:37,620 --> 00:03:39,240
你可以看到

72
00:03:39,240 --> 00:03:41,520
这个模型有多强

73
00:03:42,160 --> 00:03:45,860
尤其是在3.js环境里搭建七大奇迹的时候

74
00:03:45,860 --> 00:03:47,380
区别在于

75
00:03:47,380 --> 00:03:50,960
4.1 Flash 版本完成这两个测试

76
00:03:50,960 --> 00:03:52,440
大概花了两个小时

77
00:03:52,440 --> 00:03:53,280
花费二

78
00:03:53,280 --> 00:03:54,760
60美元

79
00:03:54,760 --> 00:03:59,320
而 Secret Model 只用了30分钟就完成了两个测试

80
00:03:59,320 --> 00:04:02,080
第二个模型的价格没有公布

81
00:04:02,080 --> 00:04:03,860
不过目前 Version 4

82
00:04:03,860 --> 00:04:06,420
E Flash 最大的问题是

83
00:04:06,420 --> 00:04:11,060
它花太多时间在思考和做一些没必要的测试上

84
00:04:11,060 --> 00:04:13,120
严重拖慢了流程

85
00:04:13,120 --> 00:04:16,320
但最终拿到的结果能力还是挺惊人的

86
00:04:16,320 --> 00:04:19,540
我知道我刚才说过Deep Seek 4

87
00:04:19,540 --> 00:04:22,740
E Flash有点太爱过度思考了

88
00:04:22,740 --> 00:04:23,680
但很多时候

89
00:04:23,680 --> 00:04:26,140
这个模型其实快得离谱

90
00:04:26,140 --> 00:04:27,900
我真的很好

91
00:04:27,900 --> 00:04:30,500
其Deep Seek背后到底用了什么架构

92
00:04:30,500 --> 00:04:33,480
因为从这些测试里你能看到

93
00:04:33,480 --> 00:04:37,920
它在解码时平均每秒能输出大约350个Token

94
00:04:37,920 --> 00:04:40,600
对这种能力级别的推理模型来说

95
00:04:40,600 --> 00:04:42,220
这速度真的非常夸张

96
00:04:42,220 --> 00:04:44,580
挺让人印象深刻的

97
00:04:44,580 --> 00:04:47,700
尤其考虑到这还只是一个临时测试版

98
00:04:47,700 --> 00:04:49,600
在3D建模里

99
00:04:49,600 --> 00:04:50,660
Deep Seek Version 4

100
00:04:50,660 --> 00:04:54,340
E Flash相比前代已经迈出了决定性的一步

101
00:04:54,340 --> 00:04:58,380
它能用3JS生成相当出色的内容

102
00:04:58,380 --> 00:05:01,540
尤其是在配合Deep Seek Toolkit

103
00:05:01,540 --> 00:05:04,780
并把推理能力开到最高时

104
00:05:04,780 --> 00:05:08,800
当被要求创建一个经典的中式园林时

105
00:05:08,800 --> 00:05:12,620
它一气呵成的生成了所有回廊

106
00:05:12,620 --> 00:05:16,020
甚至连池塘都一次完成了

107
00:05:16,020 --> 00:05:18,660
还用Shader反射出所有建筑

108
00:05:18,660 --> 00:05:22,920
细节多的超出你对任何其他模型的预期

109
00:05:22,920 --> 00:05:25,380
这是我完全没想到的

110
00:05:25,380 --> 00:05:28,420
尤其是Deep Seek Version 4这一系列模型

111
00:05:28,420 --> 00:05:29,380
不过4

112
00:05:29,380 --> 00:05:32,200
E Flash模型却一下就做出来了

113
00:05:32,200 --> 00:05:34,500
这不是一张静态图

114
00:05:34,500 --> 00:05:36,680
整个环境都是可交互的

115
00:05:36,680 --> 00:05:39,760
而且直接在浏览器里运行

116
00:05:39,760 --> 00:05:41,640
所以很高兴看到

117
00:05:41,640 --> 00:05:45,040
这个模型已经能生成这种水平了

118
00:05:45,040 --> 00:05:45,840
当然

119
00:05:45,840 --> 00:05:48,440
推理过程还是比较慢

120
00:05:48,440 --> 00:05:51,220
不过程序代码本身生成的很快

121
00:05:51,220 --> 00:05:54,500
大约每秒300到400个Token

122
00:05:54,500 --> 00:05:57,320
接下来是一个由DIT-6

123
00:05:57,320 --> 00:06:00,680
E-Flash生成的火箭发射模拟

124
00:06:00,680 --> 00:06:02,540
我们把它和Fable 5.1

125
00:06:02,540 --> 00:06:06,020
GPT-6 Astra以及Chemic A3做了对比

126
00:06:06,020 --> 00:06:07,140
然后你

127
00:06:07,140 --> 00:06:09,160
你能明显看出

128
00:06:09,160 --> 00:06:13,280
Deep Seek 4.1相比上一版有了很明显的视觉提升

129
00:06:13,280 --> 00:06:16,280
环境和整体画质都挺稳

130
00:06:16,280 --> 00:06:19,180
但物理效果还是有点让人怀疑

131
00:06:19,180 --> 00:06:20,340
当然

132
00:06:20,340 --> 00:06:23,180
它没有Fable和GPT6那么聪明

133
00:06:23,180 --> 00:06:24,760
甚至还不如Keymatt 3

134
00:06:24,760 --> 00:06:26,800
这算是唯一的取点

135
00:06:26,800 --> 00:06:28,240
不过火箭发射后

136
00:06:28,240 --> 00:06:30,780
你会发现它不是按正常轨迹飞

137
00:06:31,240 --> 00:06:33,560
而是在空中相桶一样开始打转

138
00:06:33,560 --> 00:06:35,800
更现实的走向

139
00:06:36,020 --> 00:06:40,320
和其他模型相比

140
00:06:40,320 --> 00:06:42,420
质量上确实还有一点差距

141
00:06:42,420 --> 00:06:46,320
但作为一个Fresh版本

142
00:06:46,320 --> 00:06:47,840
它已经很惊艳了

143
00:06:47,840 --> 00:06:51,940
它在质量不差的同时

144
00:06:51,940 --> 00:06:54,020
速度比Kimiko 3更快

145
00:06:54,020 --> 00:06:55,160
成本也更低

146
00:06:55,160 --> 00:06:58,140
这对开源社区来说

147
00:06:58,140 --> 00:06:59,580
是个很大的优势

148
00:06:59,580 --> 00:07:03,260
这是DipSeq Version 4.1 Flash的

149
00:07:03,260 --> 00:07:04,440
体速生成结果

150
00:07:04,440 --> 00:07:07,500
真的完全超出我的预期

151
00:07:07,500 --> 00:07:10,260
它加了一层很轻的氛围感

152
00:07:10,260 --> 00:07:14,640
这种环境里大多数模型通常都做不到

153
00:07:14,640 --> 00:07:15,720
还加了鸟

154
00:07:15,720 --> 00:07:18,200
能正常工作的瀑布

155
00:07:18,200 --> 00:07:19,980
以及一些其他元素

156
00:07:19,980 --> 00:07:20,920
比如风车

157
00:07:20,920 --> 00:07:23,860
能看到Deep Seek Version 4

158
00:07:23,860 --> 00:07:27,240
E Flash竟然能做出这种东西

159
00:07:27,240 --> 00:07:28,600
真的很不可思议

160
00:07:28,600 --> 00:07:31,860
这里Deep Seek Version 4.1 Flash

161
00:07:31,860 --> 00:07:32,840
已经建起了一个

162
00:07:32,840 --> 00:07:35,780
一个完整的Minecraft克隆版

163
00:07:35,780 --> 00:07:37,400
而且看起来还不错

164
00:07:37,400 --> 00:07:41,520
我真没想到Deep Seek模型能做到这一点

165
00:07:41,520 --> 00:07:44,700
因为我拿来测试的多数模型

166
00:07:44,700 --> 00:07:47,800
碰到这种请求通常都会失败

167
00:07:47,800 --> 00:07:50,880
根本没法像原来的Sandbox那样

168
00:07:50,880 --> 00:07:52,920
在世界里生成任何组件

169
00:07:52,920 --> 00:07:55,260
不过这次我们看到

170
00:07:55,260 --> 00:07:59,640
它成功把不同的纹理和组件都完整写出来了

171
00:07:59,640 --> 00:08:00,620
当然

172
00:08:14,718 --> 00:08:19,598
能看到一个完整的Minecraft克隆版只要这么点钱

173
00:08:19,598 --> 00:08:21,818
还是挺离谱的

174
00:08:21,818 --> 00:08:24,898
要是把同样的请求发给GVT6

175
00:08:24,898 --> 00:08:28,358
甚至Fable 5.1肯定会贵得多

176
00:08:28,358 --> 00:08:31,918
但用这个工具几乎花不了什么钱

177
00:08:31,918 --> 00:08:34,218
还是很值得考虑的

178
00:08:34,218 --> 00:08:36,538
当然如果你有时间

179
00:08:36,538 --> 00:08:41,898
可以用这个模型来做很多基础任务或者复杂任务

180
00:08:41,898 --> 00:08:47,898
然后再切到GVT或Fable模型去处理更快更复杂的工作

181
00:08:47,898 --> 00:08:54,298
我真的很喜欢这个生成效果

182
00:08:57,598 --> 00:09:01,278
因为他们居然能做出这个靠近火山的Packet

183
00:09:01,278 --> 00:09:11,518
把它放到一个岛上还加了背景音效和音乐

184
00:09:11,898 --> 00:09:18,998
简直太惊艳了

185
00:09:18,998 --> 00:09:22,478
模型还能把整个环境都捕捉出来

186
00:09:22,478 --> 00:09:24,738
连周围和地形都考虑进去了

187
00:09:24,738 --> 00:09:26,338
你还能看到不同的

188
00:09:26,338 --> 00:09:30,378
把火山灯光和氛围都放进一个画面里

189
00:09:30,378 --> 00:09:31,478
真的很棒

190
00:09:31,478 --> 00:09:36,218
下面是Deep Seek 4.1 Flash和上一代4 Flash的对比

191
00:09:37,918 --> 00:09:39,358
在这次测试里

192
00:09:39,358 --> 00:09:42,238
模型用了大约11.2万个Token

193
00:09:42,238 --> 00:09:45,898
20步就完成了任务

194
00:09:45,898 --> 00:09:47,798
只花了7分41秒

195
00:09:47,798 --> 00:09:51,998
两个模型都要用BoxoArt做一个Packet

196
00:09:51,998 --> 00:09:58,618
其中速度一度高达每秒427个Token

197
00:09:58,618 --> 00:10:01,098
但最夸张的是

198
00:10:01,098 --> 00:10:04,618
4.1版本整个流程只花了30美分

199
00:10:04,618 --> 00:10:11,478
考虑到它的价格速度和效果

200
00:10:11,478 --> 00:10:12,938
这真的很惊人

201
00:10:12,938 --> 00:10:19,018
这确实是个很厉害的模型

202
00:10:19,018 --> 00:10:21,578
相比你们屏幕上看到的4Flash

203
00:10:21,578 --> 00:10:22,658
它前进了一大步

204
00:10:22,658 --> 00:10:24,518
如果你喜欢这个视频

205
00:10:24,518 --> 00:10:25,518
想支持频道

206
00:10:25,518 --> 00:10:28,078
可以通过下面的Super Thanks打赏

207
00:10:28,078 --> 00:10:28,798
或者

208
00:10:28,798 --> 00:10:31,918
你也可以加入我们的私人Discord服务器

209
00:10:31,918 --> 00:10:35,658
可以免费每月获取各种AI工具的订阅

210
00:10:35,658 --> 00:10:39,118
还有AI圈每天的新闻独家内容

211
00:10:39,118 --> 00:10:41,578
更多好东西也都在这

212
00:10:41,578 --> 00:10:44,198
朋友们最有意思的是

213
00:10:44,198 --> 00:10:45,938
这还只是测试版

214
00:10:45,938 --> 00:10:49,698
他们还会继续完善这个最终版本

215
00:10:49,698 --> 00:10:53,458
所以这次只测试这个模型两天

216
00:10:53,458 --> 00:10:55,458
好更好地了解它的能力

217
00:10:55,458 --> 00:10:59,118
之后他们会继续调整这些基准

218
00:10:59,118 --> 00:11:01,198
让它适用于大多数情况

219
00:11:01,198 --> 00:11:04,618
用户会根据收到的反馈来调整

220
00:11:04,618 --> 00:11:10,278
但这个模型能够在3D环境里生成非常惊艳的场景

221
00:11:10,278 --> 00:11:15,058
而且在其他格式和领域里也表现得相当不错

222
00:11:15,058 --> 00:11:17,398
这一点很让人高兴

223
00:11:17,398 --> 00:11:20,038
另一个很有意思的测试

224
00:11:20,038 --> 00:11:24,318
是让这个模型生成摄像机的3D投影

225
00:11:25,038 --> 00:11:27,458
第一版大概一分钟就生成好了

226
00:11:27,458 --> 00:11:29,618
经过另外两轮交互后

227
00:11:29,618 --> 00:11:32,738
完整结果大约就准备好了

228
00:11:32,738 --> 00:11:34,398
六分钟

229
00:11:34,398 --> 00:11:36,818
对这样价位的模型来说

230
00:11:36,818 --> 00:11:38,898
最终效果真的很惊艳

231
00:11:38,898 --> 00:11:40,638
你可以看到

232
00:11:40,638 --> 00:11:43,238
它能把相机细节看得很清楚

233
00:11:43,238 --> 00:11:44,738
演示三体交互

234
00:11:44,738 --> 00:11:47,818
还把各个部分都展示出来

235
00:11:47,818 --> 00:11:51,038
这样你就能很轻松地想象出

236
00:11:51,038 --> 00:11:52,318
它的所有组件

237
00:11:52,318 --> 00:11:53,538
这里

238
00:11:53,538 --> 00:11:57,098
模型专注于创建一个独立的地下城游戏

239
00:11:57,098 --> 00:12:00,458
这个需求真正有意思的地方在于

240
00:12:00,458 --> 00:12:01,078
4

241
00:12:01,078 --> 00:12:03,738
Eflash版本居然能执行

242
00:12:03,738 --> 00:12:06,078
它一次就成功了

243
00:12:06,078 --> 00:12:07,818
这就是其他模型

244
00:12:07,818 --> 00:12:09,718
通常至少要是两次

245
00:12:09,718 --> 00:12:12,718
而且经常因为路径规划差

246
00:12:12,718 --> 00:12:14,358
卡在墙边的原因

247
00:12:14,358 --> 00:12:17,898
但Deep Seek模型成功在环境里导航

248
00:12:17,898 --> 00:12:20,418
并且第一次就完成了任务

249
00:12:20,418 --> 00:12:21,558
再看画面

250
00:12:21,558 --> 00:12:23,098
图形其实没那么差

251
00:12:23,098 --> 00:12:24,778
你可以看到

252
00:12:24,778 --> 00:12:27,618
它给每个敌人都加了不同的纹理

253
00:12:27,618 --> 00:12:29,958
它甚至还加了不同的

254
00:12:29,958 --> 00:12:32,078
还有各种怪物

255
00:12:32,078 --> 00:12:32,858
背景音效

256
00:12:32,858 --> 00:12:34,578
各类图形和视觉元素

257
00:12:34,578 --> 00:12:35,398
还有动画

258
00:12:35,398 --> 00:12:36,518
所以

259
00:12:36,518 --> 00:12:38,638
这是另一个很好的例子

260
00:12:38,638 --> 00:12:41,478
说明4.1 Flash不只是快的惊人

261
00:12:41,478 --> 00:12:47,398
它在Agentic行为和空间思维方面

262
00:12:47,398 --> 00:12:49,498
似乎真的有了明显进步

263
00:12:49,498 --> 00:12:51,598
最后

264
00:12:51,598 --> 00:12:53,078
Deep Seek Version 4

265
00:12:53,078 --> 00:12:58,878
E-Flash在环境里做出了一个完整的Morial Card风格游戏

266
00:12:58,878 --> 00:13:02,038
真的太夸张了

267
00:13:02,038 --> 00:13:04,798
这个模型生成了所有音效

268
00:13:04,798 --> 00:13:06,758
还有各种环境和地图

269
00:13:06,758 --> 00:13:08,438
效果真的很惊艳

270
00:13:14,318 --> 00:13:17,118
而且它只花了大约1.10美元

271
00:13:17,118 --> 00:13:19,458
这里还有两个很大的优点

272
00:13:19,458 --> 00:13:23,298
你看到的这个模型运行得快得惊人

273
00:13:23,298 --> 00:13:25,078
而且在这次测试里

274
00:13:25,078 --> 00:13:28,838
他还用自己的视觉能力来完成工作

275
00:13:28,838 --> 00:13:30,418
总的来说

276
00:13:30,418 --> 00:13:34,478
我真的很想看看这个模型最后会变成什么样

277
00:13:34,478 --> 00:13:36,978
这对Deep Seed团队来说

278
00:13:36,978 --> 00:13:38,918
绝对是个巨大进步

279
00:13:38,918 --> 00:13:42,278
从新的架构也能明显看出

280
00:13:42,278 --> 00:13:43,998
他们很重视速度

281
00:13:43,998 --> 00:13:46,078
原生视觉能力

282
00:13:46,078 --> 00:13:48,678
以及全方位交付出色表现

283
00:13:49,458 --> 00:13:52,258
所以我真的很期待看它最后会变成什么样

284
00:13:52,258 --> 00:13:53,238
记住

285
00:13:53,238 --> 00:13:55,938
你只有两天时间可以访问这个模型

286
00:13:55,938 --> 00:13:58,818
所以我非常建议你去试试

287
00:13:58,818 --> 00:13:59,358
跑跑看

288
00:13:59,358 --> 00:14:00,938
看看它能为你做什么

289
00:14:00,938 --> 00:14:02,638
不过话说回来

290
00:14:02,638 --> 00:14:03,238
朋友们

291
00:14:03,238 --> 00:14:07,678
我会把今天视频里用到的所有链接都放在下面的简介里

292
00:14:07,678 --> 00:14:09,778
欢迎来SATA Store

293
00:14:09,778 --> 00:14:11,038
看看AI的世界

294
00:14:11,038 --> 00:14:12,518
那里有免费工具

295
00:14:12,518 --> 00:14:14,538
我的个人建议

296
00:14:14,538 --> 00:14:16,438
社区和各种资源

297
00:14:16,438 --> 00:14:20,298
也欢迎在CodingPath和Benchmark平台探索AI世界

298
00:14:20,298 --> 00:14:22,598
顺便看看我们的第二个频道

299
00:14:22,598 --> 00:14:24,738
记得订阅我们的Nothletters

300
00:14:24,738 --> 00:14:26,458
加入Discord

301
00:14:26,458 --> 00:14:27,978
关注我的Widder

302
00:14:27,978 --> 00:14:30,858
最后记得订阅点开通知灵

303
00:14:30,858 --> 00:14:32,018
给这个视频点赞

304
00:14:32,018 --> 00:14:34,178
也请看看我们之前的视频

305
00:14:34,178 --> 00:14:36,978
这样你就能及时了解最新的AI新闻

306
00:14:36,978 --> 00:14:38,138
说到这

307
00:14:38,138 --> 00:14:40,098
朋友们祝你们今天过得愉快

308
00:14:40,098 --> 00:14:41,538
传播正能量

309
00:14:41,538 --> 00:14:42,558
我们很快再见

310
00:14:42,558 --> 00:14:43,518
再见了

311
00:14:43,518 --> 00:14:43,978
Floss