1
00:00:00,000 --> 00:00:19,400
朋友們,我的顯卡已經兩天沒核眼了。事情是這樣子的,minimax發布了一款開源的Sota級別的視頻生成模型,minimax H3。

2
00:00:19,400 --> 00:00:29,400
開源前夜,我拿了一條過去的舊AI視頻,丟給H3去重做。原本我只是想跑一條試試,結果效果一出來我就坐不住了。

3
00:00:29,400 --> 00:00:31,900
然后我把过去全部的都冲跑了一遍

4
00:00:31,900 --> 00:00:33,540
做完天都快亮了

5
00:00:33,540 --> 00:00:35,760
我当时脑子里就只有一个想法

6
00:00:35,760 --> 00:00:36,860
后悔

7
00:00:36,860 --> 00:00:40,400
我之前付给C-Dance钱能退吗

8
00:00:40,400 --> 00:00:42,640
那个时候权重还没放出来

9
00:00:42,640 --> 00:00:43,860
于是我就苦等了

10
00:00:43,860 --> 00:00:44,860
等了好几天

11
00:00:44,860 --> 00:00:46,160
他还跳票了好几次

12
00:00:46,160 --> 00:00:48,340
最后他终于来了

13
00:00:48,340 --> 00:00:51,220
我第一时间就把这个权重塞进了我的显卡

14
00:00:51,220 --> 00:00:53,260
从那一刻起到现在

15
00:00:53,260 --> 00:00:54,920
我的GPU就没有停过

16
00:00:54,920 --> 00:00:56,700
所以先下结论

17
00:00:56,700 --> 00:00:59,100
minimax H3就是开放权重

18
00:00:59,100 --> 00:01:01,100
视频模型的deep sick时刻

19
00:01:01,100 --> 00:01:03,120
今天这期视频有点长

20
00:01:03,120 --> 00:01:04,080
建议先点个收藏

21
00:01:04,080 --> 00:01:05,240
我讲三件事

22
00:01:05,240 --> 00:01:06,080
它能干什么

23
00:01:06,080 --> 00:01:07,060
它改变了什么

24
00:01:07,060 --> 00:01:08,400
还有最重要的

25
00:01:08,400 --> 00:01:09,380
你应该做什么

26
00:01:09,380 --> 00:01:10,880
之前看我视频

27
00:01:10,880 --> 00:01:12,680
跟风买了pro6000的朋友

28
00:01:12,680 --> 00:01:14,440
请现在把笑哭了

29
00:01:14,440 --> 00:01:15,400
打在公屏上

30
00:01:15,400 --> 00:01:16,560
那这个模型

31
00:01:16,560 --> 00:01:17,940
参数我就不念了

32
00:01:17,940 --> 00:01:18,840
先说一个

33
00:01:18,840 --> 00:01:20,500
它和以前所有视频模型

34
00:01:20,500 --> 00:01:21,420
不太一样的地方

35
00:01:21,420 --> 00:01:22,800
以前视频模型

36
00:01:22,800 --> 00:01:24,440
你得按它的规矩来

37
00:01:24,440 --> 00:01:26,320
文声视频一个入口

38
00:01:26,320 --> 00:01:27,640
图声视频一个入口

39
00:01:27,640 --> 00:01:28,620
你想搞点复杂

40
00:01:28,620 --> 00:01:30,160
你就得接各种工作流

41
00:01:30,160 --> 00:01:30,440
对吧

42
00:01:30,440 --> 00:01:32,180
每个模型只干一件事

43
00:01:32,180 --> 00:01:33,240
但H3不一样

44
00:01:33,240 --> 00:01:34,700
它文字图片视频

45
00:01:34,700 --> 00:01:35,240
声音

46
00:01:35,240 --> 00:01:36,280
你随便组合

47
00:01:36,280 --> 00:01:37,060
你扔给它

48
00:01:37,060 --> 00:01:38,720
然后用大白话告诉它

49
00:01:38,720 --> 00:01:39,260
你要什么

50
00:01:39,260 --> 00:01:39,960
官方说法

51
00:01:39,960 --> 00:01:42,460
这叫统一的多模态上下文

52
00:01:42,460 --> 00:01:43,380
说人话

53
00:01:43,380 --> 00:01:46,180
它就是个会看会听会拍的创作者

54
00:01:46,180 --> 00:01:47,280
它还有个名字

55
00:01:47,280 --> 00:01:49,840
叫卖相更通用的多模态智能

56
00:01:49,840 --> 00:01:52,580
它做的片子全部都自带声音

57
00:01:52,580 --> 00:01:54,960
对白音效BGM一次生成

58
00:01:54,960 --> 00:01:56,080
口型也能对上

59
00:01:56,080 --> 00:01:58,080
11种语言都能说

60
00:01:58,080 --> 00:02:00,320
你给他一段参考音频

61
00:02:00,320 --> 00:02:02,040
他连音色都能给你换掉

62
00:02:02,040 --> 00:02:05,060
这事能成是因为画面和声音

63
00:02:05,060 --> 00:02:07,980
是在同一个模型里一起长出来的

64
00:02:07,980 --> 00:02:12,100
所以口型呼吸环境天然都是咬合的

65
00:02:12,100 --> 00:02:14,520
虽然说单条最长是15秒

66
00:02:14,520 --> 00:02:15,320
你别嫌短

67
00:02:15,320 --> 00:02:16,500
因为广告啊

68
00:02:16,500 --> 00:02:17,060
demo啊

69
00:02:17,060 --> 00:02:18,320
片头都够用了

70
00:02:18,320 --> 00:02:20,620
先看一个我自己做的demo

71
00:02:20,620 --> 00:02:21,240
讲个故事

72
00:02:21,240 --> 00:02:22,740
我以前是个剧本杀作者

73
00:02:22,740 --> 00:02:25,040
我以前写过一个叫未接来电的本

74
00:02:25,040 --> 00:02:27,420
可惜那个时候没有这种模型啊

75
00:02:27,420 --> 00:02:29,180
物料全部都是手搓的

76
00:02:29,180 --> 00:02:31,760
我把角色卡和主持人手册

77
00:02:31,760 --> 00:02:33,040
都丢给了Image 2

78
00:02:33,040 --> 00:02:35,100
先生成了一套风格图

79
00:02:35,100 --> 00:02:36,480
然后再丢给了H3

80
00:02:36,480 --> 00:02:38,580
等了不到十分钟

81
00:02:38,580 --> 00:02:40,900
一条768P的宣传片就出来了

82
00:02:40,900 --> 00:02:42,260
说实话

83
00:02:42,260 --> 00:02:43,600
看到成片那一刻

84
00:02:43,600 --> 00:02:44,860
我挺感慨的

85
00:02:44,860 --> 00:02:46,040
当年的遗憾

86
00:02:46,040 --> 00:02:47,780
现在十分钟就补上了

87
00:02:47,780 --> 00:02:49,540
随后我就突发奇想

88
00:02:49,540 --> 00:02:51,460
不如让AI来发挥一下创意

89
00:02:51,460 --> 00:02:53,260
不断的Push这个模型的边界

90
00:02:53,260 --> 00:02:56,100
于是我就给他定了一个通宵抽卡任务

91
00:02:56,100 --> 00:02:57,300
看看他能做出来什么

92
00:02:57,300 --> 00:02:59,480
早上起来的时候我被惊到了

93
00:02:59,480 --> 00:03:01,540
37条作品没有一个重复的

94
00:03:01,540 --> 00:03:02,440
也没有废片

95
00:03:02,440 --> 00:03:05,520
各种各样AI的奇思妙想

96
00:03:05,520 --> 00:03:09,340
如果用官方的768P最低价算

97
00:03:09,340 --> 00:03:10,760
两毛三分钱

98
00:03:10,760 --> 00:03:13,400
这一晚上也就消耗了百来块钱

99
00:03:13,400 --> 00:03:16,500
但这么高的成功率和丰富的艺术表达

100
00:03:16,500 --> 00:03:19,200
在过往的开源模型上都从来没见过

101
00:03:19,200 --> 00:03:20,780
我又看了一下官方的demo

102
00:03:20,780 --> 00:03:22,400
那个就更惊艳了

103
00:03:22,400 --> 00:03:24,200
如果你做TVC广告

104
00:03:24,200 --> 00:03:26,020
或者是电商产品游戏demo

105
00:03:26,020 --> 00:03:27,260
或者是产品demo

106
00:03:27,300 --> 00:03:30,820
我很想说H3已经杀死比赛了

107
00:03:30,820 --> 00:03:32,160
在这样的成本下

108
00:03:32,160 --> 00:03:34,140
你很难再回归手措了

109
00:03:34,140 --> 00:03:35,980
还有一个我特别喜欢的能力

110
00:03:35,980 --> 00:03:37,040
视频编辑

111
00:03:37,040 --> 00:03:38,920
有一次我录口播录错了

112
00:03:38,920 --> 00:03:40,340
但我录制时穿着衣服

113
00:03:40,340 --> 00:03:41,460
刚丢进洗衣机

114
00:03:41,460 --> 00:03:43,200
那我就直接给一段素材

115
00:03:43,200 --> 00:03:44,740
再给一段重录的语音

116
00:03:44,740 --> 00:03:45,740
你们看看这段

117
00:03:45,740 --> 00:03:47,540
随时都可以换一身衣服

118
00:03:47,540 --> 00:03:48,700
丰富的艺术表达

119
00:03:48,700 --> 00:03:50,220
加上这种随意的编辑

120
00:03:50,220 --> 00:03:51,960
可以让每一条素材

121
00:03:51,960 --> 00:03:53,860
都可以充分发挥它的价值

122
00:03:53,860 --> 00:03:54,180
OK

123
00:03:54,180 --> 00:03:55,880
爽片咱们放完了

124
00:03:55,880 --> 00:03:57,060
咱们说点硬的

125
00:03:57,060 --> 00:03:59,400
大家肯定想知道本地跑的效果如何

126
00:03:59,400 --> 00:04:02,300
以我这台Pro6000为例吧

127
00:04:02,300 --> 00:04:04,680
我这边测到的极致速度是

128
00:04:04,680 --> 00:04:08,380
1088p的10秒可以24分钟搞定

129
00:04:08,380 --> 00:04:11,560
但是你要知道H3刚出来那个早上

130
00:04:11,560 --> 00:04:13,960
768p的分辨率也要这么长时间

131
00:04:13,960 --> 00:04:17,040
但是没有想到H3的社区如此热情

132
00:04:17,040 --> 00:04:19,480
很快的各种优化方案都出来了

133
00:04:19,480 --> 00:04:20,720
什么Sage Attention

134
00:04:20,720 --> 00:04:21,720
Easy Catch

135
00:04:21,720 --> 00:04:22,280
FBC

136
00:04:22,280 --> 00:04:24,960
我就跟着社区一路换方案

137
00:04:24,960 --> 00:04:26,520
不到48小时

138
00:04:26,520 --> 00:04:28,340
生成速度直接翻了一倍

139
00:04:28,340 --> 00:04:30,020
而且这个优化节奏

140
00:04:30,020 --> 00:04:31,360
我觉得根本不会停啊

141
00:04:31,360 --> 00:04:32,560
社区太狂热了

142
00:04:32,560 --> 00:04:33,180
但是啊

143
00:04:33,180 --> 00:04:35,560
这件事本身比速度更重要

144
00:04:35,560 --> 00:04:36,860
因为权重放出来

145
00:04:36,860 --> 00:04:38,820
全世界最聪明的一群人

146
00:04:38,820 --> 00:04:40,500
就开始帮你免费优化

147
00:04:40,500 --> 00:04:42,180
这在B元模型上

148
00:04:42,180 --> 00:04:43,640
永远都不可能会发生

149
00:04:43,640 --> 00:04:43,960
对吧

150
00:04:43,960 --> 00:04:44,760
更厉害的是

151
00:04:44,760 --> 00:04:46,480
8月3日开元的当天

152
00:04:46,480 --> 00:04:48,040
各种芯片厂商

153
00:04:48,040 --> 00:04:50,120
推理框架都同一天宣布适配

154
00:04:50,120 --> 00:04:52,120
Confi UI当天就能跑

155
00:04:52,120 --> 00:04:54,420
于是卖客的卖提示词的

156
00:04:54,420 --> 00:04:55,280
连夜上架

157
00:04:55,280 --> 00:04:58,520
玩ConfiUI的那些朋友疯狂的持续输出

158
00:04:58,520 --> 00:05:01,720
只有公认的顶级开源

159
00:05:01,720 --> 00:05:03,980
才能一夜之间有这样的生态出现

160
00:05:03,980 --> 00:05:06,880
所以我们得给minimax点一个大大的赞

161
00:05:06,880 --> 00:05:09,060
现在视频生成这个领域

162
00:05:09,060 --> 00:05:11,200
是被避原模型垄断着对吧

163
00:05:11,200 --> 00:05:13,600
像极了ChadGPT时刻

164
00:05:13,600 --> 00:05:15,340
刚刚发生的那段日子

165
00:05:15,340 --> 00:05:17,500
大家又兴奋又焦虑对不对

166
00:05:17,500 --> 00:05:21,000
而minimax他没有选择去成为Anthoropic

167
00:05:21,000 --> 00:05:23,000
他选择了做DeepSeek

168
00:05:23,000 --> 00:05:25,240
把智能评权交给每一个人

169
00:05:25,240 --> 00:05:27,480
把机会评权交给每一个人

170
00:05:27,480 --> 00:05:29,040
所以看到这里的朋友

171
00:05:29,040 --> 00:05:31,080
请把感谢打在公屏上吧

172
00:05:31,080 --> 00:05:31,440
好

173
00:05:31,440 --> 00:05:32,680
接下来我想说的是

174
00:05:32,680 --> 00:05:35,200
这次智能评权会带来多大的机会和变革

175
00:05:35,200 --> 00:05:37,560
这是我们每个人都不能忽视的

176
00:05:37,560 --> 00:05:39,500
大家都想在AI时代赚钱对不对

177
00:05:39,500 --> 00:05:41,820
但是咱不能总盯着提效生产力

178
00:05:41,820 --> 00:05:44,220
因为产品会被海量的生产出来

179
00:05:44,220 --> 00:05:46,860
但传播和货客依旧是最难的

180
00:05:46,860 --> 00:05:48,840
你依然需要获得注意力

181
00:05:48,840 --> 00:05:51,300
而每一次内容形式的升级

182
00:05:51,300 --> 00:05:53,580
都会带来注意力的重新分配

183
00:05:53,580 --> 00:05:58,100
SD2.0,SD2.5解决了AI视频能不能进生产线的问题

184
00:05:58,100 --> 00:06:00,040
但它没解决另一个问题

185
00:06:00,040 --> 00:06:02,620
普通人的试错成本太高了

186
00:06:02,620 --> 00:06:04,140
过去几个月啊

187
00:06:04,140 --> 00:06:06,900
少数负得起抽卡费的人先掌握了这个技巧

188
00:06:06,900 --> 00:06:09,100
他们的内容越来越精彩了

189
00:06:09,100 --> 00:06:11,980
你刷到那些用AI参与叙事的博主

190
00:06:11,980 --> 00:06:13,000
是真的做的好

191
00:06:13,000 --> 00:06:16,020
但是你们看不到的是他们背后的账单

192
00:06:16,020 --> 00:06:19,820
而更多的人是被那个恐怖的成本吓得站在了门外

193
00:06:19,820 --> 00:06:20,820
而H3呢

194
00:06:20,820 --> 00:06:21,740
把试错成本

195
00:06:21,740 --> 00:06:24,240
第一次达到了电费级别

196
00:06:24,240 --> 00:06:25,600
刚才说的社区爆发

197
00:06:25,600 --> 00:06:26,500
就是因为这个原因

198
00:06:26,500 --> 00:06:27,920
我自己就是个样本啊

199
00:06:27,920 --> 00:06:29,740
我的GPU48小时都没停

200
00:06:29,740 --> 00:06:30,620
每做一条

201
00:06:30,620 --> 00:06:32,020
我的提示词就准一分

202
00:06:32,020 --> 00:06:33,580
我的经验就增长一分

203
00:06:33,580 --> 00:06:34,460
而且这一次

204
00:06:34,460 --> 00:06:36,000
经验的传播速度

205
00:06:36,000 --> 00:06:37,420
会比上次快得多

206
00:06:37,420 --> 00:06:38,940
因为一起试错的人

207
00:06:38,940 --> 00:06:40,560
多了好几个数量级

208
00:06:40,560 --> 00:06:41,840
更好的镜头

209
00:06:41,840 --> 00:06:42,740
更好的叙事

210
00:06:42,740 --> 00:06:44,080
会一波一波的冒出来

211
00:06:44,080 --> 00:06:45,600
我不敢说H3的出现

212
00:06:45,600 --> 00:06:47,860
会让每个用AI做视频的人

213
00:06:47,860 --> 00:06:48,800
都会变得更好

214
00:06:48,800 --> 00:06:51,700
但它一定会导致注意力的重新分配

215
00:06:51,700 --> 00:06:54,100
所以为什么我最近睡眠也少

216
00:06:54,100 --> 00:06:55,960
就是因为在用电力和时间

217
00:06:55,960 --> 00:06:57,480
抢跑这波经验差

218
00:06:57,480 --> 00:06:58,820
我的视频制作流程

219
00:06:58,820 --> 00:07:00,100
其实一大半已经被

220
00:07:00,100 --> 00:07:02,660
Hyperframe这样的AI工作流接管了

221
00:07:02,660 --> 00:07:04,380
如果这套H3流程跑动

222
00:07:04,380 --> 00:07:05,900
那剩下的也全都交出去了

223
00:07:05,900 --> 00:07:07,640
所以我永远都是那句

224
00:07:07,640 --> 00:07:09,360
得赶快行动起来啊

225
00:07:09,360 --> 00:07:09,920
朋友们

226
00:07:09,920 --> 00:07:11,340
但是有卡的朋友们

227
00:07:11,340 --> 00:07:12,800
我先泼个冷水啊

228
00:07:12,800 --> 00:07:16,060
本地部署并不是生产级内容的解法

229
00:07:16,060 --> 00:07:18,300
这次minimax有两个重要的模块

230
00:07:18,300 --> 00:07:19,100
没有开源

231
00:07:19,100 --> 00:07:20,400
一个是contextir

232
00:07:20,400 --> 00:07:21,700
就是一个云端的

233
00:07:21,700 --> 00:07:24,260
更智能理解素材的中间服务

234
00:07:24,260 --> 00:07:26,560
它可以返回复杂逻辑

235
00:07:26,560 --> 00:07:28,400
推理生成的结构化提示词

236
00:07:28,400 --> 00:07:31,780
它不是简单把提示词写得更长

237
00:07:31,780 --> 00:07:34,060
而是先判断什么该保留

238
00:07:34,060 --> 00:07:34,960
什么该改变

239
00:07:34,960 --> 00:07:37,740
我做了一个语页小提琴的编辑测试

240
00:07:37,740 --> 00:07:40,960
如果只用简单提示词加原视频

241
00:07:40,960 --> 00:07:43,040
衣服虽然变成了红色语衣

242
00:07:43,040 --> 00:07:44,780
但镜头完全变了

243
00:07:44,780 --> 00:07:48,360
我再用context.ir服务重新生成提示词

244
00:07:48,360 --> 00:07:49,320
看到没有

245
00:07:49,320 --> 00:07:53,300
人物场景和镜头运动都跟原片保持一致了

246
00:07:53,300 --> 00:07:57,500
还有一个是regenerated2k的超分模型没开源

247
00:07:57,882 --> 00:08:07,882
還有一個是regenerated2K的超分模型沒開源。H3是一個原生768P的,而雲端API是通過regenerate2K的超分模型提升到2K分辨率的。它的計算快成本也低。

248
00:08:07,882 --> 00:08:11,102
本地只能说是你能输出1080P

249
00:08:11,102 --> 00:08:13,762
但细节提升是非常有限的

250
00:08:13,762 --> 00:08:15,342
完全达不到云端那个效果

251
00:08:15,342 --> 00:08:17,942
而且我试了其他的超分也完全不可用

252
00:08:17,942 --> 00:08:21,622
而且1088P的时间成本是768P的三倍

253
00:08:21,622 --> 00:08:25,482
所以开源部署的版本是完全没法和官方的2K比质量的

254
00:08:25,482 --> 00:08:26,082
这是事实

255
00:08:26,082 --> 00:08:27,262
所以本地部署的朋友

256
00:08:27,262 --> 00:08:28,902
咱们把定位搞清楚

257
00:08:28,902 --> 00:08:30,782
本地是用来刷经验试错的

258
00:08:30,782 --> 00:08:32,102
不是用来搞生产力的

259
00:08:32,102 --> 00:08:34,102
你早点跑通一条路径

260
00:08:34,102 --> 00:08:37,262
产出高质量的内容和收益挂钩才是最优先的

261
00:08:37,262 --> 00:08:38,562
跑通了这个

262
00:08:38,562 --> 00:08:39,622
Token就是杠杆

263
00:08:39,622 --> 00:08:41,522
跑不通那就是烧Token交学费

264
00:08:41,522 --> 00:08:44,502
我现在策略就是一边拼命的找各种case

265
00:08:44,502 --> 00:08:46,382
验证提炼工作流

266
00:08:46,382 --> 00:08:49,022
另外我会本地出768P

267
00:08:49,022 --> 00:08:50,542
买API超分到2K

268
00:08:50,542 --> 00:08:51,622
按时机的用量走

269
00:08:51,622 --> 00:08:53,202
等我的业务量大了稳定了

270
00:08:53,202 --> 00:08:53,882
消耗的多了

271
00:08:53,882 --> 00:08:55,462
我也会考虑要不要直接买

272
00:08:55,462 --> 00:08:57,322
minimax design的包月包年

273
00:08:57,322 --> 00:09:00,642
直接出包年版简直是个价格屠夫

274
00:09:00,642 --> 00:09:03,542
它输出2K四毛钱一秒

275
00:09:03,542 --> 00:09:06,182
在影视短剧慢剧之外的行业

276
00:09:06,182 --> 00:09:09,282
我可以说SD2.5的价格是毫无竞争力

277
00:09:09,282 --> 00:09:11,282
如果你是原先就有业务

278
00:09:11,282 --> 00:09:13,002
那你就直接走API得了

279
00:09:13,002 --> 00:09:15,082
用Context R的API

280
00:09:15,082 --> 00:09:17,282
可以对复杂的素材做理解

281
00:09:17,282 --> 00:09:18,902
生成复杂的提示词

282
00:09:18,902 --> 00:09:21,102
你甚至可以先用本地抽卡

283
00:09:21,102 --> 00:09:24,102
再调用2K超分API来提升质量

284
00:09:24,102 --> 00:09:25,962
而且API是更灵活的

285
00:09:25,962 --> 00:09:27,642
它能接近你的自动化工作流

286
00:09:27,642 --> 00:09:29,142
总之就是赶快动起来

287
00:09:29,142 --> 00:09:30,582
现在minimax design上

288
00:09:30,582 --> 00:09:32,302
有三次的免费生成

289
00:09:32,302 --> 00:09:33,662
海螺AI也能直接玩

290
00:09:33,662 --> 00:09:36,082
你先得把你第一条片子做出来

291
00:09:36,082 --> 00:09:36,422
对吧

292
00:09:36,422 --> 00:09:38,782
产品demo或者是品牌story

293
00:09:38,782 --> 00:09:41,322
或者是你就是随便想找个提示词玩玩

294
00:09:41,322 --> 00:09:42,202
随便出都可以

295
00:09:42,202 --> 00:09:43,002
但我的建议就是

296
00:09:43,002 --> 00:09:45,022
你得早点开始建立自己的内容体系

297
00:09:45,022 --> 00:09:47,602
尤其是闷头做产品的程序员朋友们

298
00:09:47,602 --> 00:09:49,982
现在是做内容门槛最低的时代了

299
00:09:49,982 --> 00:09:51,982
你得早点尝试建立自己的流量

300
00:09:51,982 --> 00:09:54,222
否则以后就得找博主来买流量

301
00:09:54,222 --> 00:09:55,902
那我们交好学费

302
00:09:55,902 --> 00:09:56,862
跑完试错后

303
00:09:56,862 --> 00:09:58,802
应该进一步的提高内容的质量

304
00:09:58,802 --> 00:10:01,622
因为内容生产的成本大幅的降低

305
00:10:01,622 --> 00:10:03,562
试错成本也大幅降低

306
00:10:03,562 --> 00:10:04,802
所以市场上的内容

307
00:10:04,802 --> 00:10:05,842
数量和质量

308
00:10:05,842 --> 00:10:07,122
都会水涨船高

309
00:10:07,122 --> 00:10:09,322
那我们人类是会审美疲劳的

310
00:10:09,322 --> 00:10:11,262
酷炫效果几天就看腻了

311
00:10:11,262 --> 00:10:11,522
对吧

312
00:10:11,522 --> 00:10:13,062
我们最后还是会回归到

313
00:10:13,062 --> 00:10:14,282
更好的镜头表达

314
00:10:14,282 --> 00:10:15,322
更好的叙事中

315
00:10:15,322 --> 00:10:16,242
所以我认为啊

316
00:10:16,242 --> 00:10:17,522
注意力的重新分配

317
00:10:17,522 --> 00:10:18,662
也是有窗口期的

318
00:10:18,662 --> 00:10:19,602
过了这段时间

319
00:10:19,602 --> 00:10:21,002
还是得内容为王

320
00:10:21,002 --> 00:10:22,402
所以我也在探索

321
00:10:22,402 --> 00:10:24,322
其他的产出创意的方式

322
00:10:24,322 --> 00:10:25,682
比如说更好的工具啊

323
00:10:25,682 --> 00:10:26,922
能帮你快速提升

324
00:10:26,922 --> 00:10:28,242
内容质量的产品啊

325
00:10:28,242 --> 00:10:30,822
能扮演导演和编剧的agent啊

326
00:10:30,822 --> 00:10:32,182
除了minimax design

327
00:10:32,182 --> 00:10:33,542
如果大家有经常

328
00:10:33,542 --> 00:10:36,002
在用的也可以在评论区告诉我

329
00:10:36,002 --> 00:10:37,302
结尾还有一个预判

330
00:10:37,302 --> 00:10:38,802
视频生成这条路

331
00:10:38,802 --> 00:10:40,442
最终还是会回到

332
00:10:40,442 --> 00:10:43,242
Aton的指挥视频模型生成一切的

333
00:10:43,242 --> 00:10:44,902
我绝对相信马斯克说的

334
00:10:44,902 --> 00:10:47,942
我们现在已经处于技术发展的起点时刻了

335
00:10:47,942 --> 00:10:50,682
用技术加速技术肯定是趋势

336
00:10:50,682 --> 00:10:51,882
而minimax呢

337
00:10:51,882 --> 00:10:53,482
有视频生成模型

338
00:10:53,482 --> 00:10:55,562
有minimax design这个数据平台

339
00:10:55,562 --> 00:10:58,702
他们没有理由不把导演编剧的能力

340
00:10:58,702 --> 00:11:00,762
训练到M系列语言模型里的

341
00:11:00,762 --> 00:11:03,802
所以我坚定看好M系列的后续更新

342
00:11:03,802 --> 00:11:05,282
也许几个月半年后

343
00:11:05,282 --> 00:11:07,342
我就再不用去搞提示词了

344
00:11:07,342 --> 00:11:09,202
AI就搞定一切的内容

345
00:11:09,202 --> 00:11:11,662
在那之前谁有更多的注意力

346
00:11:11,662 --> 00:11:13,342
那个时刻一到来

347
00:11:13,342 --> 00:11:15,382
就会更多的被技术加速

348
00:11:15,382 --> 00:11:17,582
这就是指数型世界的特点

349
00:11:17,582 --> 00:11:19,402
所以S3这波机会

350
00:11:19,402 --> 00:11:20,462
你们看懂了吗

351
00:11:20,462 --> 00:11:21,322
好了

352
00:11:21,322 --> 00:11:23,402
以上就是本期的全部内容了

353
00:11:23,402 --> 00:11:24,002
谢谢大家