1
00:00:00,000 --> 00:00:07,000
來,各位家人們,最近幾天,整個硅谷都被一款叫做Jill的模型刷屏了。

2
00:00:07,000 --> 00:00:14,000
然後呢,發布前三天已經被Wassel,Cloudflare,LongChang等主流平台迅速給集成了。

3
00:00:14,000 --> 00:00:16,000
這個模型有什麼特點呢?

4
00:00:16,000 --> 00:00:23,000
這款模型,它相比普通的模型推理速度,最高可以提升200倍,最高可以高到200倍。

5
00:00:23,000 --> 00:00:27,000
然後呢,這個成本可以降低,最高可以降到400倍。

6
00:00:27,000 --> 00:00:28,880
是一个非常非常恐怖的数字啊

7
00:00:28,880 --> 00:00:31,220
而且是可以组合的前沿之冷

8
00:00:31,220 --> 00:00:34,280
那么这种牛扮的东西到底是一个什么东西呢

9
00:00:34,280 --> 00:00:38,500
它会不会把open AI把这个Athrope给颠覆掉呢

10
00:00:38,500 --> 00:00:40,440
别张奇哈我们来认识一下

11
00:00:40,440 --> 00:00:41,380
J5这个模型

12
00:00:41,380 --> 00:00:44,260
那么J5这个模型呢一句话来干口啊

13
00:00:44,260 --> 00:00:46,020
J5呢是一个决策模型

14
00:00:46,020 --> 00:00:47,520
决策模型有什么特点呢

15
00:00:47,520 --> 00:00:48,800
它不生成文本

16
00:00:48,800 --> 00:00:50,360
不生成图片

17
00:00:50,360 --> 00:00:53,080
只做结果化判断输出概率

18
00:00:53,080 --> 00:00:54,500
也就是它只输出概率啊

19
00:00:54,500 --> 00:00:57,840
专门给agent的程序做高频快判断

20
00:00:57,840 --> 00:00:59,340
它是一个什么东西啊

21
00:00:59,340 --> 00:01:00,960
它其实是一个决策模型

22
00:01:00,960 --> 00:01:02,300
或者是一个判断模型

23
00:01:02,300 --> 00:01:05,500
它其实跟GPT跟Cloud

24
00:01:05,500 --> 00:01:07,620
跟GP9是有本质的区别的

25
00:01:07,620 --> 00:01:08,760
GPT这种模型呢

26
00:01:08,760 --> 00:01:10,180
其实是一个生成式模型

27
00:01:10,180 --> 00:01:11,220
生成文本

28
00:01:11,220 --> 00:01:12,180
生成图片

29
00:01:12,180 --> 00:01:15,100
但是GP呢只做决策

30
00:01:15,100 --> 00:01:16,960
它们是有本质的区别的

31
00:01:16,960 --> 00:01:18,340
所以它们之间并不是一个

32
00:01:18,340 --> 00:01:20,040
谁替代是这么一种关系

33
00:01:20,040 --> 00:01:21,820
还是一种互补的关系

34
00:01:21,820 --> 00:01:23,380
那么有种人讲

35
00:01:23,380 --> 00:01:25,540
现在是生成式的天下

36
00:01:25,540 --> 00:01:27,320
你这个东西有什么用呢

37
00:01:27,320 --> 00:01:28,960
这个生成式模型

38
00:01:28,960 --> 00:01:30,140
GPT

39
00:01:30,140 --> 00:01:31,140
Cloat

40
00:01:31,140 --> 00:01:33,100
真来也可以做决策对吧

41
00:01:33,100 --> 00:01:36,460
你这个Jill有什么特点呢

42
00:01:36,460 --> 00:01:36,820
对不对

43
00:01:36,820 --> 00:01:40,220
凭什么这么招扬

44
00:01:40,220 --> 00:01:40,600
是吧

45
00:01:40,600 --> 00:01:41,300
也别着急

46
00:01:41,300 --> 00:01:43,100
他其实有非常非常多的场景

47
00:01:43,100 --> 00:01:44,540
我们稍后会来进行介绍

48
00:01:44,540 --> 00:01:46,140
那我们再来看一下

49
00:01:46,140 --> 00:01:47,780
Jill这个模型到底是怎么来的

50
00:01:47,780 --> 00:01:49,080
这个模型呢

51
00:01:49,080 --> 00:01:50,060
其实是由

52
00:01:50,060 --> 00:01:52,720
前欧巴亚的研究员叫Digo

53
00:01:52,720 --> 00:01:56,200
他这个创办的TypeSafe AI公司

54
00:01:56,200 --> 00:01:58,120
在前几天发布的

55
00:01:58,120 --> 00:01:59,180
在9月15号发布的

56
00:01:59,180 --> 00:02:00,020
也前几天发布的

57
00:02:00,020 --> 00:02:01,400
那么这个创始人

58
00:02:01,400 --> 00:02:02,820
是非常非常厉害的一个人物

59
00:02:02,820 --> 00:02:05,160
他参与了ChairGPT的创造

60
00:02:05,160 --> 00:02:07,560
是一个非常非常有经验的

61
00:02:07,560 --> 00:02:08,900
这么一个大佬

62
00:02:08,900 --> 00:02:10,060
然后他自己呢

63
00:02:10,060 --> 00:02:12,940
就是说在创造完ChairGPT之后

64
00:02:12,940 --> 00:02:14,460
他一直在问自己一个问题

65
00:02:14,460 --> 00:02:16,640
就是说这些超人的聊天模型

66
00:02:16,640 --> 00:02:18,380
为什么没有带来这个AGI

67
00:02:18,380 --> 00:02:19,960
那过去的两年

68
00:02:19,960 --> 00:02:20,920
我一直在眼神

69
00:02:20,920 --> 00:02:21,940
抓大家做两件事情

70
00:02:21,940 --> 00:02:24,240
就是一种新的模型训练方法

71
00:02:24,240 --> 00:02:26,480
就是它提出了一种新的训练方法

72
00:02:26,480 --> 00:02:27,980
叫RLCD

73
00:02:27,980 --> 00:02:31,220
然后呢和一种新内心的情形配安模型

74
00:02:31,220 --> 00:02:32,320
就是JIL

75
00:02:32,320 --> 00:02:33,320
这个东西

76
00:02:33,320 --> 00:02:34,460
那

77
00:02:34,460 --> 00:02:36,480
好了那么重点再来介绍这个JIL

78
00:02:36,480 --> 00:02:39,500
那么JIL和这个一般的模型有什么区别呢

79
00:02:39,500 --> 00:02:40,940
那么一般的模型啊

80
00:02:40,940 --> 00:02:42,460
像这个GPT是什么呢

81
00:02:42,460 --> 00:02:43,800
竹子生成Token

82
00:02:43,800 --> 00:02:44,940
是自回归

83
00:02:44,940 --> 00:02:46,760
那么它是呢

84
00:02:46,760 --> 00:02:48,360
高延迟高成本

85
00:02:48,360 --> 00:02:49,900
就是你问它一个问题

86
00:02:49,900 --> 00:02:52,080
他呢就基于你这个问题

87
00:02:52,080 --> 00:02:53,220
这个问题就是一个提示词

88
00:02:53,220 --> 00:02:54,060
基于这个问题

89
00:02:54,060 --> 00:02:56,340
他一个一个的生成这个token

90
00:02:56,340 --> 00:02:57,980
把这个token拼在一起

91
00:02:57,980 --> 00:03:00,080
他其实就是一个完整的这个回答

92
00:03:00,080 --> 00:03:01,840
这么这样子来的

93
00:03:01,840 --> 00:03:02,680
可以生成文本

94
00:03:02,680 --> 00:03:03,340
可以生成逐变

95
00:03:03,340 --> 00:03:05,000
但只有这个模型呢

96
00:03:05,000 --> 00:03:06,500
他是一次效情计算

97
00:03:06,500 --> 00:03:08,560
直接输出判断概率

98
00:03:08,560 --> 00:03:10,340
他的特点是什么呢

99
00:03:10,340 --> 00:03:11,880
延迟极低成本极低

100
00:03:11,880 --> 00:03:12,440
记住啊

101
00:03:12,440 --> 00:03:13,600
延迟极低

102
00:03:13,600 --> 00:03:14,680
也就是说啊

103
00:03:14,680 --> 00:03:16,320
你给一个问题

104
00:03:16,320 --> 00:03:19,240
他基本上在500毫秒之内

105
00:03:19,240 --> 00:03:21,220
可以得到这个结果啊

106
00:03:21,220 --> 00:03:23,500
500毫秒之内其实是非常非常快的

107
00:03:23,500 --> 00:03:25,180
非常非常快啊

108
00:03:25,180 --> 00:03:26,080
那个说实话

109
00:03:26,080 --> 00:03:27,100
你像使用GPT

110
00:03:27,100 --> 00:03:28,460
那基本上几秒啊

111
00:03:28,460 --> 00:03:29,760
十几秒才得到这个结果

112
00:03:29,760 --> 00:03:30,060
对不对

113
00:03:30,060 --> 00:03:31,240
它的延迟极低

114
00:03:31,240 --> 00:03:33,200
500毫秒之内

115
00:03:33,200 --> 00:03:34,100
就可以得到这个结果

116
00:03:34,100 --> 00:03:35,240
成本极低啊

117
00:03:35,240 --> 00:03:37,240
我们后面会介绍的成本

118
00:03:37,240 --> 00:03:39,740
然后再来看一下啊

119
00:03:39,740 --> 00:03:41,020
这个东西怎么使用啊

120
00:03:41,020 --> 00:03:43,080
就是说你怎么去使用这个东西呢

121
00:03:43,080 --> 00:03:45,820
使用这个Jill这个模型呢

122
00:03:45,820 --> 00:03:47,140
使用Jill这个模型呢

123
00:03:47,140 --> 00:03:47,820
通常啊

124
00:03:47,820 --> 00:03:50,760
其实它是开放这个API给大家去使用

125
00:03:50,760 --> 00:03:52,920
就大家通常都是调这个API的方式啊

126
00:03:52,920 --> 00:03:54,000
去使用这个模型

127
00:03:54,000 --> 00:03:55,680
那么它呢

128
00:03:55,680 --> 00:03:57,800
主要提供两个东西

129
00:03:57,800 --> 00:03:58,720
第一个就是state

130
00:03:58,720 --> 00:04:01,800
这个什么东西呢

131
00:04:01,800 --> 00:04:03,360
大家可以把它看作是一个context

132
00:04:03,360 --> 00:04:04,060
一个上下文

133
00:04:04,060 --> 00:04:06,520
然后第二个就是问题

134
00:04:06,520 --> 00:04:09,240
就是开发者预先定义好的派对的任务

135
00:04:09,240 --> 00:04:10,100
主要是两个东西

136
00:04:10,100 --> 00:04:12,080
然后呢交给这个这个模型

137
00:04:12,080 --> 00:04:13,980
这个模型然后再发挥一个结果

138
00:04:13,980 --> 00:04:15,980
那发挥的结果会是个什么东西呢

139
00:04:15,980 --> 00:04:17,300
这个结果啊

140
00:04:17,300 --> 00:04:20,720
跟一般大模型还是有些区别的

141
00:04:20,720 --> 00:04:21,380
一般大模型呢

142
00:04:21,380 --> 00:04:23,600
其实是输出一个啊

143
00:04:23,600 --> 00:04:25,040
会输出一道文本

144
00:04:25,040 --> 00:04:25,460
对不对

145
00:04:25,460 --> 00:04:26,400
那它这里呢

146
00:04:26,400 --> 00:04:28,960
其实是输出一种强烈性的结构化JSON

147
00:04:28,960 --> 00:04:31,340
是一种强烈性的结构化JSON

148
00:04:31,340 --> 00:04:32,740
是一种JSON的文本

149
00:04:32,740 --> 00:04:33,740
啊

150
00:04:33,740 --> 00:04:34,580
JSON文本

151
00:04:34,580 --> 00:04:36,440
它不包含这个自然语言文本

152
00:04:36,440 --> 00:04:38,000
是一种纯JSON的文本

153
00:04:38,000 --> 00:04:40,220
大概是变形结构预测

154
00:04:40,220 --> 00:04:42,040
模型一次性返回所有概率

155
00:04:42,040 --> 00:04:44,720
也就是说你丢给它几个问题啊

156
00:04:44,720 --> 00:04:46,640
它是一次性给你返回所有的概率

157
00:04:46,640 --> 00:04:48,480
它是变形预测的

158
00:04:48,480 --> 00:04:50,940
第三个就是程序可以直接解析

159
00:04:50,940 --> 00:04:52,740
才是反过一个接测文本嘛

160
00:04:52,740 --> 00:04:53,940
是可以解析的

161
00:04:53,940 --> 00:04:55,420
所以我们总结来的话

162
00:04:55,420 --> 00:04:57,440
总结来这个这个这个这个模型

163
00:04:57,440 --> 00:05:00,300
它的输入就是state加预定的question

164
00:05:00,300 --> 00:05:02,460
然后呢输出一个回答对象

165
00:05:02,460 --> 00:05:03,480
是这样子

166
00:05:03,480 --> 00:05:06,260
好那么它的输出呢有三种类型啊

167
00:05:06,260 --> 00:05:07,780
第一种类型呢是no啊

168
00:05:07,780 --> 00:05:09,500
只有三种类型啊

169
00:05:09,500 --> 00:05:11,060
第一个是no啊

170
00:05:11,060 --> 00:05:12,200
no是一个概率类型

171
00:05:12,200 --> 00:05:13,980
这个类型有什么特点呢

172
00:05:13,980 --> 00:05:15,560
也就是不完类型判断

173
00:05:15,560 --> 00:05:17,760
它是返回0到1的伏点数

174
00:05:17,760 --> 00:05:19,960
数值代表成立概率

175
00:05:19,960 --> 00:05:23,220
无需单独设置置信度

176
00:05:23,220 --> 00:05:24,960
这段数字即为置信度

177
00:05:24,960 --> 00:05:26,180
好 我们来看一下

178
00:05:26,180 --> 00:05:28,520
就是它返回一个什么样的东西

179
00:05:28,520 --> 00:05:31,480
这里是它返回的一个结构的一个实例

180
00:05:31,480 --> 00:05:33,280
因为我先说过

181
00:05:33,280 --> 00:05:35,060
它其实是它的使用

182
00:05:35,060 --> 00:05:36,580
它使用的方式

183
00:05:36,580 --> 00:05:38,880
其实跟contract GPT

184
00:05:38,880 --> 00:05:40,180
跟GPT还是有些区别

185
00:05:40,180 --> 00:05:40,600
对吧

186
00:05:40,600 --> 00:05:43,840
通常是以结构的方式对外使用

187
00:05:43,840 --> 00:05:45,840
你直接给他一个

188
00:05:45,840 --> 00:05:46,840
你直接呢

189
00:05:46,840 --> 00:05:50,840
通过这个机构的方式提供这个task需要的内容

190
00:05:50,840 --> 00:05:52,840
state和这个question

191
00:05:52,840 --> 00:05:55,840
然后他返回的是一个json的这样一个sultrime

192
00:05:55,840 --> 00:05:56,840
这个sultrime里面呢

193
00:05:56,840 --> 00:05:57,840
有两个重要的东西

194
00:05:57,840 --> 00:05:59,840
第一个就是类型

195
00:05:59,840 --> 00:06:00,840
类型就是no

196
00:06:00,840 --> 00:06:02,840
它是概率类型

197
00:06:02,840 --> 00:06:05,840
然后这个呢就是它的这个值

198
00:06:05,840 --> 00:06:06,840
0.95

199
00:06:06,840 --> 00:06:07,840
我们前面说过

200
00:06:07,840 --> 00:06:10,840
它这里是0到1的这么一个伏点数

201
00:06:10,840 --> 00:06:11,840
对不对

202
00:06:11,840 --> 00:06:14,280
那么这个0.9.5其实就是表示一个概率

203
00:06:14,280 --> 00:06:16,040
大家可以反而看作是一个概率

204
00:06:16,640 --> 00:06:18,800
那么这个时候同学就会问了

205
00:06:18,800 --> 00:06:19,880
这个这里有什么用呢

206
00:06:20,080 --> 00:06:20,600
对吧

207
00:06:20,600 --> 00:06:22,240
嗯反正0.95是吧

208
00:06:22,440 --> 00:06:24,000
其实场景是非常多的

209
00:06:24,000 --> 00:06:24,680
比如说

210
00:06:25,880 --> 00:06:27,640
检测用户退款对不对

211
00:06:27,840 --> 00:06:29,000
检测用户退款的话

212
00:06:29,000 --> 00:06:29,800
其实呢

213
00:06:30,000 --> 00:06:30,600
这个东西啊

214
00:06:30,600 --> 00:06:32,680
当然你用XenGPT其实也可以回答

215
00:06:32,680 --> 00:06:33,280
是不是

216
00:06:33,400 --> 00:06:35,280
你丢给一个啊

217
00:06:35,280 --> 00:06:37,800
就是用户的这个用户的信息啊

218
00:06:38,200 --> 00:06:39,680
直接给到这个GPT

219
00:06:39,680 --> 00:06:41,000
他也可以回答

220
00:06:41,000 --> 00:06:43,000
说这个用户到底会不会退款

221
00:06:43,000 --> 00:06:43,600
对不对

222
00:06:43,900 --> 00:06:45,000
但是呢其实呢

223
00:06:45,300 --> 00:06:46,700
那个东西有点什么

224
00:06:46,800 --> 00:06:48,500
有点这个杀鸡又牛肉的感觉

225
00:06:48,500 --> 00:06:50,200
其实这种需求呢

226
00:06:50,200 --> 00:06:52,200
其实有更简单的方式去实现

227
00:06:52,200 --> 00:06:55,100
就是所谓的Jell这种方式对吧

228
00:06:55,400 --> 00:06:57,500
那么检测用户退款请求

229
00:06:57,500 --> 00:06:59,600
这用户呢可能是一个信息发过来

230
00:06:59,800 --> 00:07:02,100
那么给到Jell这个模型

231
00:07:02,200 --> 00:07:03,100
那么Jell呢

232
00:07:03,500 --> 00:07:06,100
直接就回答一个概率就可以了

233
00:07:06,400 --> 00:07:07,200
比如说

234
00:07:07,600 --> 00:07:08,800
用户呢不会退款

235
00:07:08,800 --> 00:07:10,000
他就可能就返回给你

236
00:07:10,000 --> 00:07:10,300
对不对

237
00:07:10,300 --> 00:07:10,980
如果会退款

238
00:07:10,980 --> 00:07:12,100
那么就返回这个

239
00:07:12,100 --> 00:07:14,380
比如说0.9或者1就可以了

240
00:07:14,380 --> 00:07:14,620
对不对

241
00:07:14,620 --> 00:07:18,600
那么这样子其实非常方便的解决这个问题

242
00:07:18,600 --> 00:07:20,580
那么它这个

243
00:07:20,580 --> 00:07:22,980
它的成本会远远的低于这个

244
00:07:22,980 --> 00:07:24,280
用GPT来解决这个问题

245
00:07:24,280 --> 00:07:27,820
而且这个速度也会快很多

246
00:07:27,820 --> 00:07:30,460
它的这个延迟的时间是500毫秒

247
00:07:30,460 --> 00:07:33,440
它那个GPT那是几秒钟

248
00:07:33,440 --> 00:07:33,780
对不对

249
00:07:33,780 --> 00:07:35,920
所以它用它这种方式来解决

250
00:07:35,920 --> 00:07:36,520
这个问题的话

251
00:07:36,520 --> 00:07:38,440
其实是成本更低

252
00:07:38,440 --> 00:07:40,000
然后那速度呢更好

253
00:07:40,000 --> 00:07:41,340
是不是速度更快

254
00:07:41,340 --> 00:07:42,080
效率更高

255
00:07:42,942 --> 00:07:46,942
然後我們再來看第二個類型,就是Close類型。

256
00:07:46,942 --> 00:07:56,942
這個什麼意思呢?也就是說,我們在去提供上下文的時候,其實是會帶好幾個選項過去的。

257
00:07:56,942 --> 00:07:58,222
然后呢

258
00:07:58,222 --> 00:08:00,542
这个这个模型说到这个

259
00:08:00,542 --> 00:08:03,722
这个上下和和这个问题之后呢

260
00:08:03,722 --> 00:08:06,182
他会选择其中的一个来作为回答

261
00:08:06,182 --> 00:08:07,942
就是做选择题嘛

262
00:08:07,942 --> 00:08:09,082
就ABC第四个选项

263
00:08:09,082 --> 00:08:10,382
他选一个告诉你

264
00:08:10,382 --> 00:08:11,602
好那么其实这个呢

265
00:08:11,602 --> 00:08:12,682
就是他换回的结果

266
00:08:12,682 --> 00:08:14,882
假设他这个提供的选项

267
00:08:14,882 --> 00:08:16,942
有三个选项啊

268
00:08:16,942 --> 00:08:17,782
一个是binim

269
00:08:17,782 --> 00:08:20,302
这个是结算

270
00:08:20,302 --> 00:08:21,522
然后这个是技术

271
00:08:21,522 --> 00:08:22,502
然后是销售

272
00:08:22,502 --> 00:08:24,202
就是三百三个部门啊

273
00:08:24,202 --> 00:08:25,182
对吧

274
00:08:25,182 --> 00:08:26,762
然后这里面他会选一个

275
00:08:26,762 --> 00:08:29,002
他选择这个billion这个部门

276
00:08:29,002 --> 00:08:30,302
就他选一个

277
00:08:30,302 --> 00:08:31,122
你给他三个选项

278
00:08:31,122 --> 00:08:31,882
他选一个出来

279
00:08:31,882 --> 00:08:33,682
那这有时候这个东西

280
00:08:33,682 --> 00:08:36,702
哪里有这个场景多不多呢

281
00:08:36,702 --> 00:08:37,622
这个很多嘛

282
00:08:37,622 --> 00:08:38,082
对不对

283
00:08:38,082 --> 00:08:39,662
很多重要的场景

284
00:08:39,662 --> 00:08:41,402
比如说这个三件商品

285
00:08:41,402 --> 00:08:42,102
让他选一个

286
00:08:42,102 --> 00:08:42,502
是不是

287
00:08:42,502 --> 00:08:44,442
那其实这种场景是非常多的

288
00:08:44,442 --> 00:08:44,602
是不是

289
00:08:44,602 --> 00:08:46,422
OK我们来看一下

290
00:08:46,422 --> 00:08:47,002
这个关键

291
00:08:47,002 --> 00:08:49,242
只有关键的这个字段

292
00:08:49,242 --> 00:08:50,902
这个是truth

293
00:08:50,902 --> 00:08:52,562
就是概率最高的选项名称

294
00:08:52,562 --> 00:08:53,262
就这个

295
00:08:53,262 --> 00:08:55,042
就概率最高的

296
00:08:55,042 --> 00:08:57,022
他选一个出来

297
00:08:57,022 --> 00:08:58,142
就是ABCD四个选项

298
00:08:58,142 --> 00:08:58,822
他选一个出来

299
00:08:58,822 --> 00:09:01,402
他说不一定是ABCD四个选项

300
00:09:01,402 --> 00:09:03,322
他最多可以255个选项

301
00:09:03,322 --> 00:09:04,542
选项可以很多

302
00:09:04,542 --> 00:09:05,722
你把它列出来

303
00:09:05,722 --> 00:09:07,202
他会选一个出来

304
00:09:07,202 --> 00:09:08,922
然后这个呢

305
00:09:08,922 --> 00:09:10,942
就是概率的总和

306
00:09:10,942 --> 00:09:14,982
就是所有或选项的完整的概率分布

307
00:09:14,982 --> 00:09:17,082
那如果说是255个选项呢

308
00:09:17,082 --> 00:09:20,462
那么就会列出255个选项

309
00:09:20,462 --> 00:09:21,342
每一个选项的概率

310
00:09:21,342 --> 00:09:24,162
那么这个trose这里呢

311
00:09:24,162 --> 00:09:25,462
其实概念最高的那一个

312
00:09:25,462 --> 00:09:27,522
那么这个是confidence

313
00:09:27,522 --> 00:09:28,482
自信度啊

314
00:09:28,482 --> 00:09:30,982
模型对本次分类结果的自信度

315
00:09:30,982 --> 00:09:32,642
那么数值越高呢

316
00:09:32,642 --> 00:09:34,162
就分类呢越确定

317
00:09:34,162 --> 00:09:36,202
好那我们看一下这个自信度是0.8

318
00:09:36,202 --> 00:09:37,442
对吧这是比较高的

319
00:09:37,442 --> 00:09:39,802
是不是当然我们在进行

320
00:09:39,802 --> 00:09:41,122
决策的时候呢

321
00:09:41,122 --> 00:09:43,322
你可以跟这个自信度来进行

322
00:09:43,322 --> 00:09:44,542
评估哈

323
00:09:44,542 --> 00:09:46,722
就是如果自信度高于0.8

324
00:09:46,722 --> 00:09:48,242
或者说高于0.7

325
00:09:48,242 --> 00:09:50,242
我们就可以选择这样的一个结果

326
00:09:50,242 --> 00:09:52,262
就是所谓的决策模型的

327
00:09:52,262 --> 00:09:53,662
价值在这里面

328
00:09:54,162 --> 00:09:56,162
好让我们再看第三个啊

329
00:09:56,162 --> 00:09:57,162
就是score啊

330
00:09:57,162 --> 00:09:58,522
打分啊

331
00:09:58,522 --> 00:09:59,342
打个分数

332
00:09:59,342 --> 00:10:00,782
啊

333
00:10:00,782 --> 00:10:02,362
那么它的特点就是连续打分

334
00:10:02,362 --> 00:10:04,102
各级别概率分布啊

335
00:10:04,102 --> 00:10:05,162
confidence 知性度

336
00:10:05,162 --> 00:10:07,262
我们来看一下这个结果啊

337
00:10:07,262 --> 00:10:07,462
那么

338
00:10:07,462 --> 00:10:10,022
通常它会去

339
00:10:10,022 --> 00:10:12,002
比如说你给它三个选项啊

340
00:10:12,002 --> 00:10:12,622
你给它一个问

341
00:10:12,622 --> 00:10:14,142
就是说你给给问

342
00:10:14,142 --> 00:10:15,622
给上下文给问题

343
00:10:15,622 --> 00:10:16,842
然后提供选项

344
00:10:16,842 --> 00:10:18,602
好那么它反而结果里面呢

345
00:10:18,602 --> 00:10:20,382
它就会对三个选项呢

346
00:10:20,382 --> 00:10:21,682
每个选项打一个分数

347
00:10:21,682 --> 00:10:22,602
我们来看一下啊

348
00:10:22,602 --> 00:10:23,322
好

349
00:10:23,322 --> 00:10:24,262
它的分数是

350
00:10:24,262 --> 00:10:26,462
非诊数打分值

351
00:10:26,462 --> 00:10:27,322
这个是I.86

352
00:10:27,322 --> 00:10:27,722
对不对

353
00:10:27,722 --> 00:10:29,022
知情度是0.94

354
00:10:29,022 --> 00:10:32,322
然后它提供了三个选项

355
00:10:32,322 --> 00:10:33,622
这三个选项

356
00:10:33,622 --> 00:10:34,202
评鉴

357
00:10:34,202 --> 00:10:35,322
轻微不满

358
00:10:35,322 --> 00:10:35,822
生气

359
00:10:35,822 --> 00:10:37,502
极度愤怒

360
00:10:37,502 --> 00:10:37,902
对吧

361
00:10:37,902 --> 00:10:39,222
好上面就是分数

362
00:10:39,222 --> 00:10:41,182
这个第一个评鉴

363
00:10:41,182 --> 00:10:41,962
就是0.0分

364
00:10:41,962 --> 00:10:42,622
就是不评鉴

365
00:10:42,622 --> 00:10:43,242
对吧

366
00:10:43,242 --> 00:10:43,502
OK

367
00:10:43,502 --> 00:10:45,202
这里是0.0爱

368
00:10:45,202 --> 00:10:46,622
这个轻微不满

369
00:10:46,622 --> 00:10:48,342
对应的分数是0.0爱

370
00:10:48,342 --> 00:10:49,962
然后这个是0.1

371
00:10:49,962 --> 00:10:50,822
好

372
00:10:50,822 --> 00:10:51,502
这个生气

373
00:10:51,502 --> 00:10:52,682
然后极度愤怒

374
00:10:52,682 --> 00:10:53,742
是年点88

375
00:10:53,742 --> 00:10:55,042
那么这个分数最高了

376
00:10:55,042 --> 00:10:55,282
对不对

377
00:10:55,282 --> 00:10:57,922
那如果说你是基于这个分数来进行

378
00:10:57,922 --> 00:10:58,362
啊

379
00:10:58,362 --> 00:10:58,982
决策的话

380
00:10:58,982 --> 00:11:00,462
那么你可能会选这个是不是

381
00:11:00,462 --> 00:11:01,162
啊

382
00:11:01,162 --> 00:11:02,482
这个其实就是对于

383
00:11:02,482 --> 00:11:03,462
多个选项

384
00:11:03,462 --> 00:11:04,622
进行打一个分数

385
00:11:04,622 --> 00:11:05,282
啊

386
00:11:05,282 --> 00:11:06,582
这个其实也是长期很多

387
00:11:06,582 --> 00:11:06,942
啊

388
00:11:06,942 --> 00:11:08,282
这个是

389
00:11:08,282 --> 00:11:08,862
呃

390
00:11:08,862 --> 00:11:10,902
我们的Jill这个模型啊

391
00:11:10,902 --> 00:11:11,302
它的

392
00:11:11,302 --> 00:11:14,262
返回结果的三个可能性

393
00:11:14,262 --> 00:11:15,362
OK啊

394
00:11:15,362 --> 00:11:16,102
三种结果

395
00:11:16,102 --> 00:11:17,342
好

396
00:11:17,342 --> 00:11:18,362
那么先怎么总结一下

397
00:11:18,362 --> 00:11:20,902
Jill输出的关键的特性啊

398
00:11:20,902 --> 00:11:22,082
第二点就是

399
00:11:22,082 --> 00:11:23,002
年文本幻觉

400
00:11:23,002 --> 00:11:25,162
它输出结构是强制约定的

401
00:11:25,162 --> 00:11:26,342
宁格式错误

402
00:11:26,342 --> 00:11:27,302
不乱增自断

403
00:11:27,302 --> 00:11:29,842
然后是拼尿变形计算的

404
00:11:29,842 --> 00:11:31,382
一次请就变形处理所有

405
00:11:31,382 --> 00:11:34,742
然后校准级概念

406
00:11:34,742 --> 00:11:36,302
直接预计分支

407
00:11:36,302 --> 00:11:38,262
那么低分支可以人工复制

408
00:11:38,262 --> 00:11:40,222
它会有一个知性度给你

409
00:11:40,222 --> 00:11:41,442
然后输出你的费用

410
00:11:41,442 --> 00:11:44,262
它这个模型输出是不收费的

411
00:11:44,262 --> 00:11:45,962
你用它这个模型输出是不收费的

412
00:11:45,962 --> 00:11:48,482
而且输入的成本也是非常低

413
00:11:48,482 --> 00:11:50,902
它是我们使用

414
00:11:50,902 --> 00:11:54,842
GPT的成本大概就是100万Token

415
00:11:54,842 --> 00:11:55,202
对吧

416
00:11:55,202 --> 00:11:55,842
就一兆

417
00:11:55,842 --> 00:11:58,922
大概就是10美元

418
00:11:58,922 --> 00:12:00,002
10美元

419
00:12:00,002 --> 00:12:01,142
那么他这个

420
00:12:01,142 --> 00:12:03,922
大概就是100万

421
00:12:03,922 --> 00:12:12,342
大概就是0.0.4万美元

422
00:12:12,342 --> 00:12:14,062
就是他这个10美元

423
00:12:14,062 --> 00:12:15,282
他这个年纪年4万美元

424
00:12:15,282 --> 00:12:17,142
所以他的成本低很多

425
00:12:17,142 --> 00:12:19,522
那么也就是说

426
00:12:19,522 --> 00:12:21,682
我们在实际上面进行

427
00:12:21,682 --> 00:12:23,842
在做应用的时候呢

428
00:12:23,842 --> 00:12:25,742
那么其实就给了我们很多的选择

429
00:12:25,742 --> 00:12:28,542
有些场景其实并不完全需要

430
00:12:28,542 --> 00:12:31,242
由GPT或者说

431
00:12:31,242 --> 00:12:33,122
Close这种模型来进行处理

432
00:12:33,122 --> 00:12:34,922
我们完全可以把它交给什么

433
00:12:34,922 --> 00:12:37,082
交给我们的JP这种模型来处理

434
00:12:37,082 --> 00:12:37,542
对不对

435
00:12:37,542 --> 00:12:39,502
所以他们这其实是这样写的关系

436
00:12:39,502 --> 00:12:42,802
并不是说替代的关系

437
00:12:42,802 --> 00:12:43,562
并不是这样子

438
00:12:43,562 --> 00:12:46,642
好那么接下来我们给大家看一个案例

439
00:12:46,642 --> 00:12:47,842
就是到底它是怎么用的

440
00:12:47,842 --> 00:12:48,302
这个东西

441
00:12:48,302 --> 00:12:49,382
我们来看一下

442
00:12:49,382 --> 00:12:51,042
官网上面提供了一个案例

443
00:12:51,042 --> 00:12:52,702
我们直接来看到官网就可以了

444
00:12:52,702 --> 00:12:52,862
好吧

445
00:12:52,862 --> 00:12:54,382
就是汉缩入门

446
00:12:54,382 --> 00:12:56,882
大家可以到这个网站

447
00:12:56,882 --> 00:12:59,042
去看一下它的官网上面提供了一个案例

448
00:12:59,042 --> 00:13:00,362
那么当然首先呢

449
00:13:00,362 --> 00:13:02,382
我们需要去申请

450
00:13:02,382 --> 00:13:03,022
申请

451
00:13:03,022 --> 00:13:04,142
如果说你要使用的话

452
00:13:04,142 --> 00:13:07,862
你需要到它的控制台去申请一个API-KIT

453
00:13:07,862 --> 00:13:10,182
申请还是比较容易的

454
00:13:10,182 --> 00:13:10,822
OK

455
00:13:10,822 --> 00:13:13,862
然后得到一个接口之后呢

456
00:13:13,862 --> 00:13:16,262
你就可以来进行使用了

457
00:13:16,262 --> 00:13:17,142
好

458
00:13:17,142 --> 00:13:17,802
那么这里呢

459
00:13:17,802 --> 00:13:22,562
是CURL这个方式来去发行一个请求

460
00:13:22,802 --> 00:13:24,802
你需要提供的是一个API-K

461
00:13:25,102 --> 00:13:26,102
这是它的接口

462
00:13:26,902 --> 00:13:28,402
然后你提供哪些数据呢

463
00:13:28,402 --> 00:13:29,502
这个数据是State

464
00:13:29,602 --> 00:13:31,602
大家可以把State看作是上线文

465
00:13:32,002 --> 00:13:32,602
OK

466
00:13:32,802 --> 00:13:34,302
这个是它的上线文

467
00:13:34,302 --> 00:13:35,102
是在英文是吧

468
00:13:35,102 --> 00:13:36,602
我们把它翻译一下

469
00:13:36,602 --> 00:13:37,502
什么意思

470
00:13:39,002 --> 00:13:39,502
OK

471
00:13:41,402 --> 00:13:41,802
好

472
00:13:41,802 --> 00:13:43,102
这个上线文是这样子

473
00:13:43,102 --> 00:13:43,502
你好

474
00:13:43,502 --> 00:13:45,902
我已经尝试您借我的strip账户三天了

475
00:13:45,902 --> 00:13:47,102
但其实一直失败

476
00:13:47,102 --> 00:13:48,742
我的销售正在锦上

477
00:13:49,102 --> 00:13:51,142
请尽快帮我帮助我

478
00:13:51,142 --> 00:13:51,662
对吧

479
00:13:51,662 --> 00:13:52,062
这个东西

480
00:13:52,062 --> 00:13:53,742
这个内容应该是一个

481
00:13:54,302 --> 00:13:56,142
是一个street的客户是吧

482
00:13:56,142 --> 00:13:57,942
在在咨询对吧

483
00:13:57,942 --> 00:13:59,462
在向客户进行咨询啊

484
00:13:59,462 --> 00:14:00,462
或者说储藏

485
00:14:00,742 --> 00:14:01,702
储藏这么一个信息啊

486
00:14:01,702 --> 00:14:02,262
ok

487
00:14:02,262 --> 00:14:02,422
好

488
00:14:02,422 --> 00:14:03,342
现在我们来看一下啊

489
00:14:03,862 --> 00:14:04,462
这个模型呢

490
00:14:04,462 --> 00:14:06,302
是Jill latest的这个模型

491
00:14:06,302 --> 00:14:07,342
然后问题是什么呢

492
00:14:08,342 --> 00:14:09,542
问题是这样子啊

493
00:14:09,862 --> 00:14:10,862
在下面啊

494
00:14:11,462 --> 00:14:12,502
这个是那个类型

495
00:14:13,022 --> 00:14:16,142
就是你希望用什么类型来回复你这个问题啊

496
00:14:16,222 --> 00:14:17,062
是no 是吧

497
00:14:17,102 --> 00:14:18,442
也就是给个概率就可以了

498
00:14:18,442 --> 00:14:19,822
我们看一下的问题是什么

499
00:14:19,822 --> 00:14:22,402
这个信息的紧迫度

500
00:14:22,402 --> 00:14:25,922
就是表达一下信息的紧迫度

501
00:14:25,922 --> 00:14:29,082
这个内容的紧迫度是什么样子

502
00:14:29,082 --> 00:14:29,382
对不对

503
00:14:29,382 --> 00:14:30,522
那么然后这个时候

504
00:14:30,522 --> 00:14:33,582
这个模型回复一个概率

505
00:14:33,582 --> 00:14:38,062
年到一这样的一个分值过来

506
00:14:38,062 --> 00:14:38,862
一个概率过来

507
00:14:38,862 --> 00:14:40,922
如果说这个值是一的话

508
00:14:40,922 --> 00:14:41,562
那么说明什么

509
00:14:41,562 --> 00:14:43,122
说明这个紧迫度是很高的

510
00:14:43,122 --> 00:14:43,402
对不对

511
00:14:43,402 --> 00:14:44,502
如果是0.95的话

512
00:14:44,502 --> 00:14:45,902
这个紧迫度也是很高的

513
00:14:45,902 --> 00:14:46,902
好

514
00:14:46,902 --> 00:14:47,702
然后我们再来看一下

515
00:14:47,702 --> 00:14:49,242
那其他的这种

516
00:14:49,242 --> 00:14:51,462
这个案例啊

517
00:14:51,462 --> 00:14:52,342
如果说我们是

518
00:14:52,342 --> 00:14:54,142
写代码掉结果

519
00:14:54,142 --> 00:14:55,902
那么我们自己在代码里面

520
00:14:55,902 --> 00:14:57,102
就要这样来写啊

521
00:14:57,102 --> 00:14:57,362
好

522
00:14:57,362 --> 00:14:58,042
那么同样的

523
00:14:58,042 --> 00:14:58,722
它的这个

524
00:14:58,722 --> 00:15:00,502
上下文还是这个啊

525
00:15:00,502 --> 00:15:01,662
还是这个来举例啊

526
00:15:01,662 --> 00:15:02,902
我们看一下它的模型

527
00:15:02,902 --> 00:15:04,002
是Jolatist

528
00:15:04,002 --> 00:15:05,122
好

529
00:15:05,122 --> 00:15:06,422
那么这个下面是问题啊

530
00:15:06,422 --> 00:15:07,942
它其实是什么呢

531
00:15:07,942 --> 00:15:08,502
呃

532
00:15:08,502 --> 00:15:10,522
它是它什么

533
00:15:10,522 --> 00:15:12,662
它其实是一个一个上下文

534
00:15:12,662 --> 00:15:14,202
它其实是问了三个问题啊

535
00:15:14,202 --> 00:15:14,782
问了三个问题

536
00:15:14,782 --> 00:15:17,262
其实用三种类型来问好

537
00:15:17,262 --> 00:15:18,402
这个是choose

538
00:15:18,402 --> 00:15:19,842
choose ok

539
00:15:19,842 --> 00:15:21,042
这个呢就是它的问题

540
00:15:21,042 --> 00:15:22,102
哪一个团队呢

541
00:15:22,102 --> 00:15:24,202
来处理这样的一个事情

542
00:15:24,202 --> 00:15:24,542
对吧

543
00:15:24,542 --> 00:15:25,662
提供了三个选项

544
00:15:25,662 --> 00:15:27,362
这个是结算这个团队

545
00:15:27,362 --> 00:15:28,262
这个是技术这个团队

546
00:15:28,262 --> 00:15:29,202
这个销售这个团队

547
00:15:29,202 --> 00:15:30,882
这提供了三个选项

548
00:15:31,597 --> 00:15:34,597
因為我們說過,truth嘛,對不對?

549
00:15:34,597 --> 00:15:37,597
你是要大模型給你做選擇,所以你提供這個選項給他。

550
00:15:37,597 --> 00:15:40,597
像ABCD,做選擇題,是這樣子。

551
00:15:40,597 --> 00:15:45,597
好,我們看第二個,就是分數,score,使用這個類型來進行回答。

552
00:15:45,597 --> 00:15:47,797
回答好那么他是什么呢

553
00:15:47,797 --> 00:15:51,477
呃这个现在这个客户的这个什么啊

554
00:15:51,477 --> 00:15:54,777
失望度呢怎么样对吗啊啊

555
00:15:54,777 --> 00:15:55,697
什么怎么样啊OK

556
00:15:55,697 --> 00:15:57,697
他同样的提供了三个什么呢

557
00:15:57,697 --> 00:15:59,377
三种可能啊

558
00:15:59,377 --> 00:16:00,537
第一个呢就是冷静

559
00:16:00,537 --> 00:16:01,917
第二个呢是失望

560
00:16:01,917 --> 00:16:03,937
而第三个是非常生气

561
00:16:03,937 --> 00:16:05,277
就是你要根据什么

562
00:16:05,277 --> 00:16:09,977
根据这个上下文来去对于这三个描述进行打分

563
00:16:09,977 --> 00:16:11,137
打分

564
00:16:11,137 --> 00:16:13,637
OK好我们再看第三个啊

565
00:16:13,637 --> 00:16:15,217
第三个就是什么呢

566
00:16:15,597 --> 00:16:19,197
那Noah这种类型其实就是

567
00:16:19,197 --> 00:16:20,997
概率对吧是概率这个

568
00:16:20,997 --> 00:16:24,197
那么也就是说你的问题一定是跟概率相关的是吧

569
00:16:24,197 --> 00:16:26,197
这个信息它到底

570
00:16:26,197 --> 00:16:29,597
也表示了它多么紧迫对吧

571
00:16:29,597 --> 00:16:31,797
ok 那么这个时候你回答一个概率就可以了

572
00:16:31,797 --> 00:16:33,397
好 我们再来看一下它的回答

573
00:16:33,397 --> 00:16:36,397
因为你这里是三个类型的问题对吧

574
00:16:36,397 --> 00:16:36,997
三个类型的问题

575
00:16:36,997 --> 00:16:38,597
所以它的回答也是三个类型的

576
00:16:38,597 --> 00:16:40,197
我们看choose

577
00:16:40,197 --> 00:16:41,397
choose就是

578
00:16:41,397 --> 00:16:43,597
它首先要做出一个选择对吧

579
00:16:43,597 --> 00:16:45,937
ABCD要选一个选一个选一个答案

580
00:16:45,937 --> 00:16:46,737
对不对啊

581
00:16:46,737 --> 00:16:48,337
但是结果是技术部门是吧

582
00:16:48,337 --> 00:16:50,097
技术部门来处理这个问题啊

583
00:16:50,097 --> 00:16:51,757
之前都是0.78啊

584
00:16:51,757 --> 00:16:54,157
然后当然他也会给三个选项呢

585
00:16:54,157 --> 00:16:55,457
每一个打一个分数啊

586
00:16:56,657 --> 00:16:59,497
我们看第二个第二个问题

587
00:17:00,397 --> 00:17:01,657
第二个问题是打分对吧

588
00:17:02,237 --> 00:17:03,237
打分啊打分的话

589
00:17:03,237 --> 00:17:05,997
他会对于三个情绪啊

590
00:17:05,997 --> 00:17:08,257
但见到打分那很明显啊

591
00:17:08,257 --> 00:17:09,457
这个呢呃

592
00:17:09,457 --> 00:17:12,037
第二个情绪他的这个分值是会高的

593
00:17:12,037 --> 00:17:14,557
也就是他当前的用户所表达这个情绪

594
00:17:14,557 --> 00:17:15,517
应该是这个是吧

595
00:17:16,597 --> 00:17:17,117
啊

596
00:17:17,157 --> 00:17:17,797
是这个

597
00:17:18,357 --> 00:17:19,237
这个情绪啊

598
00:17:20,357 --> 00:17:21,557
好我们看最后这个啊

599
00:17:21,997 --> 00:17:23,477
最后这个是要表

600
00:17:23,477 --> 00:17:25,197
要是输出一个概率是不是

601
00:17:25,197 --> 00:17:26,277
然后直接给了一个概率是不是

602
00:17:26,277 --> 00:17:27,237
这个最简单的啊

603
00:17:27,237 --> 00:17:29,517
一点点说明这个事情的紧迫度啊

604
00:17:29,517 --> 00:17:30,597
是非常非常大的

605
00:17:30,597 --> 00:17:31,157
对不对

606
00:17:31,317 --> 00:17:32,317
这个其实就是

607
00:17:32,597 --> 00:17:33,597
Jayv这个模型啊

608
00:17:33,597 --> 00:17:34,917
使用的这个

609
00:17:35,197 --> 00:17:36,477
这样的一个方式啊

610
00:17:36,477 --> 00:17:37,597
其实就是掉接口

611
00:17:37,597 --> 00:17:38,317
然后呢

612
00:17:38,317 --> 00:17:38,597
啊

613
00:17:38,597 --> 00:17:39,317
接口里面呢

614
00:17:39,317 --> 00:17:41,077
需要描述你的这个问题

615
00:17:41,077 --> 00:17:43,077
其实需要提供一个上下文啊

616
00:17:43,077 --> 00:17:44,677
把你的这个情况给交代啊

617
00:17:44,677 --> 00:17:46,837
这个做用gp

618
00:17:46,837 --> 00:17:47,837
用chat gpt

619
00:17:47,837 --> 00:17:48,837
gpt或者glowd

620
00:17:48,837 --> 00:17:49,317
都是一样的

621
00:17:49,317 --> 00:17:50,637
要提供一个上下文对不对

622
00:17:50,637 --> 00:17:51,637
然后就是问题

623
00:17:51,637 --> 00:17:54,637
如果说是truth或者说score

624
00:17:54,637 --> 00:17:56,517
那么还需要提供这个选项

625
00:17:56,877 --> 00:17:57,637
啊OK

626
00:17:57,637 --> 00:17:58,757
那么结果呢

627
00:17:58,757 --> 00:18:00,277
就对方就会提供一个

628
00:18:00,517 --> 00:18:02,437
啊json的这样的一个结果化的

629
00:18:02,437 --> 00:18:04,277
一个字符传给到我们啊

630
00:18:04,277 --> 00:18:04,917
他这个时候

631
00:18:04,917 --> 00:18:05,637
大家就会问了

632
00:18:05,637 --> 00:18:07,277
这这东西到底有什么用对吧

633
00:18:07,277 --> 00:18:08,637
能用在哪个地方

634
00:18:08,917 --> 00:18:10,917
啊好像没有这个chat gpt

635
00:18:10,917 --> 00:18:11,877
这种这么方便

636
00:18:11,877 --> 00:18:13,537
什么都可以问

637
00:18:13,537 --> 00:18:15,097
然后他多给你一个回复

638
00:18:15,097 --> 00:18:15,337
是不是

639
00:18:15,337 --> 00:18:18,377
但其实这个场景是非常多的

640
00:18:18,377 --> 00:18:19,697
我们来看一下他的场景

641
00:18:19,697 --> 00:18:21,697
其实很多时候

642
00:18:21,697 --> 00:18:23,977
其实并不需要用GPT这种模型

643
00:18:23,977 --> 00:18:25,877
因为他成本太高了

644
00:18:25,877 --> 00:18:27,077
第一个就是成本高

645
00:18:27,077 --> 00:18:27,617
第二个就是什么

646
00:18:27,617 --> 00:18:30,577
就是他的延时太高了

647
00:18:30,577 --> 00:18:32,517
就是处理一个问题要好几秒

648
00:18:32,517 --> 00:18:34,357
他这个几百毫秒就能解决

649
00:18:34,357 --> 00:18:34,617
是不是

650
00:18:34,617 --> 00:18:36,177
所以我们这个时候来看一下

651
00:18:36,177 --> 00:18:37,717
他的场景

652
00:18:37,717 --> 00:18:38,657
第一个场景其实就是

653
00:18:38,657 --> 00:18:40,337
我们用在agent里面

654
00:18:40,337 --> 00:18:44,577
比如说浏览器与这个电脑的使用

655
00:18:44,577 --> 00:18:46,337
比如说什么

656
00:18:46,337 --> 00:18:50,057
识别这个页面状态

657
00:18:50,057 --> 00:18:51,277
判断下一个动作

658
00:18:51,277 --> 00:18:53,317
因为它的这个动作

659
00:18:53,317 --> 00:18:55,097
其实是有限的嘛

660
00:18:55,097 --> 00:18:55,317
对不对

661
00:18:55,317 --> 00:18:57,337
点击输入下拉对不对

662
00:18:57,857 --> 00:18:59,877
就哪怕我玩游戏也是一样的

663
00:18:59,877 --> 00:19:01,737
玩游戏它动作就那么点

664
00:19:01,737 --> 00:19:02,737
选项只有那么多

665
00:19:02,737 --> 00:19:03,037
对不对

666
00:19:03,037 --> 00:19:04,637
你可以把它选项全部列出来

667
00:19:04,637 --> 00:19:04,917
是不是

668
00:19:04,917 --> 00:19:07,577
列出来给它就可以了

669
00:19:07,577 --> 00:19:09,977
也就是说这种操作是有限的操作

670
00:19:09,977 --> 00:19:12,977
其实非常适合用J5这种模型来处理的

671
00:19:12,977 --> 00:19:14,597
还有智能工具路由

672
00:19:14,597 --> 00:19:14,937
对吧

673
00:19:14,937 --> 00:19:16,277
坚准选择工具

674
00:19:16,277 --> 00:19:18,817
搜索数据库邮件询问用户

675
00:19:18,817 --> 00:19:20,897
比如说安全降与分控

676
00:19:20,897 --> 00:19:24,737
就是高危操作前的快速的安全判断

677
00:19:24,737 --> 00:19:26,797
高频游戏

678
00:19:26,797 --> 00:19:27,597
说实话

679
00:19:27,597 --> 00:19:32,757
昨天我还用J5来去玩俄罗斯方块

680
00:19:32,757 --> 00:19:34,277
因为俄罗斯方块的话

681
00:19:34,277 --> 00:19:34,877
其实说实话

682
00:19:34,877 --> 00:19:35,957
它只有那几种选项

683
00:19:35,957 --> 00:19:36,237
对不对

684
00:19:36,237 --> 00:19:37,897
你把选项全部列出来就可以了

685
00:19:37,897 --> 00:19:39,877
用J5进行判断

686
00:19:39,877 --> 00:19:42,277
最高他支持255种选项

687
00:19:42,277 --> 00:19:42,557
是不是

688
00:19:42,557 --> 00:19:45,017
那其实生活中绝大部分

689
00:19:45,017 --> 00:19:46,657
其实实验绝大部分场景的

690
00:19:46,657 --> 00:19:47,517
就有几少数场景

691
00:19:47,517 --> 00:19:49,137
它是一种无限可能的

692
00:19:49,137 --> 00:19:49,537
对不对

693
00:19:49,537 --> 00:19:51,837
绝大部分场景都是有限可能的

694
00:19:51,837 --> 00:19:52,057
是不是

695
00:19:52,057 --> 00:19:54,617
我们这个只要把动作

696
00:19:54,617 --> 00:19:55,437
全部罗列出来

697
00:19:55,437 --> 00:19:57,137
就可以这样交给他去判断

698
00:19:57,137 --> 00:19:59,057
他又快成本又低

699
00:19:59,057 --> 00:19:59,737
是不是

700
00:19:59,737 --> 00:20:00,957
好给我们看一下

701
00:20:00,957 --> 00:20:02,577
比如说客户公单与客户

702
00:20:02,577 --> 00:20:03,777
消息智能处理平台

703
00:20:03,777 --> 00:20:05,797
比如说公单的分类

704
00:20:05,797 --> 00:20:06,077
是吧

705
00:20:06,077 --> 00:20:06,757
是物流问题

706
00:20:06,757 --> 00:20:07,357
账单问题

707
00:20:07,357 --> 00:20:08,377
产品故障还是投诉

708
00:20:08,377 --> 00:20:08,777
对吧

709
00:20:08,777 --> 00:20:10,537
比如说紧急度打分

710
00:20:10,537 --> 00:20:13,097
这个情绪的强烈度

711
00:20:13,097 --> 00:20:13,717
对不对

712
00:20:13,717 --> 00:20:17,257
还有这个比如说像这个Node核心判断

713
00:20:17,257 --> 00:20:18,137
是不是退款

714
00:20:18,137 --> 00:20:19,617
是不是辱骂

715
00:20:19,617 --> 00:20:21,377
是不是需要人工进入

716
00:20:21,377 --> 00:20:21,717
对不对

717
00:20:21,717 --> 00:20:23,497
做的非常非常适合用

718
00:20:23,497 --> 00:20:24,717
我们的Jail这个模型来处理

719
00:20:24,717 --> 00:20:25,797
真的非常非常方便

720
00:20:25,797 --> 00:20:28,617
所以以后做这个智能体

721
00:20:28,617 --> 00:20:29,557
它一定是什么

722
00:20:29,557 --> 00:20:30,637
一定是Jail

723
00:20:30,637 --> 00:20:33,677
一定是Jail和这个GPT或者和Lode

724
00:20:33,677 --> 00:20:35,517
进行配合使用

725
00:20:35,517 --> 00:20:37,257
就现在硅谷已经基本上

726
00:20:37,257 --> 00:20:38,817
是这两个组合在一起使用了

727
00:20:38,817 --> 00:20:39,457
啊

728
00:20:39,457 --> 00:20:39,897
就跟我

729
00:20:39,897 --> 00:20:41,177
根据我所找的信息

730
00:20:41,177 --> 00:20:43,697
就是他们现在已经在调整自己的决策了

731
00:20:43,697 --> 00:20:45,157
都是使用J5加什么

732
00:20:45,157 --> 00:20:47,217
加GPT来做这个智能体

733
00:20:47,217 --> 00:20:49,457
因为GPT其实实在是太贵了

734
00:20:49,457 --> 00:20:50,217
知道吗

735
00:20:50,217 --> 00:20:50,497
啊

736
00:20:50,497 --> 00:20:51,237
好用就好

737
00:20:51,237 --> 00:20:52,017
它实在是贵

738
00:20:52,017 --> 00:20:53,037
而且效率比较低

739
00:20:53,037 --> 00:20:56,177
所以不是所有的场景都适合用GPT键理

740
00:20:56,177 --> 00:20:57,037
啊

741
00:20:57,037 --> 00:20:57,557
其实少

742
00:20:57,557 --> 00:20:58,457
就是大部分场景

743
00:20:58,457 --> 00:21:00,497
其实都是比较适合用J5来做处理的

744
00:21:00,497 --> 00:21:03,177
还比如说这个内容与文档的拼尿删讯

745
00:21:03,177 --> 00:21:04,137
有些分类

746
00:21:04,137 --> 00:21:04,677
是不是

747
00:21:04,677 --> 00:21:05,337
啊

748
00:21:05,337 --> 00:21:05,577
只要

749
00:21:05,577 --> 00:21:07,377
就是只要给一个标记就可以了

750
00:21:07,377 --> 00:21:07,657
对吧

751
00:21:07,657 --> 00:21:08,937
还有评论

752
00:21:08,937 --> 00:21:09,817
消息删选

753
00:21:09,817 --> 00:21:10,777
对不对

754
00:21:10,777 --> 00:21:12,717
说实话

755
00:21:12,717 --> 00:21:14,997
你之前给几个标签给它

756
00:21:14,997 --> 00:21:15,297
是不是

757
00:21:15,297 --> 00:21:16,377
因为这个

758
00:21:16,377 --> 00:21:18,917
这往判断是哪一种类型的

759
00:21:18,917 --> 00:21:20,077
直接就说实话

760
00:21:20,077 --> 00:21:22,037
就很快

761
00:21:22,037 --> 00:21:23,757
就给我们打这个标签了

762
00:21:23,757 --> 00:21:24,857
还能闻到删选

763
00:21:24,857 --> 00:21:25,797
会议记样

764
00:21:25,797 --> 00:21:26,137
是不是

765
00:21:26,137 --> 00:21:28,257
还有实时业务分控

766
00:21:28,257 --> 00:21:28,757
自动化

767
00:21:28,757 --> 00:21:30,797
交易电带分控

768
00:21:30,797 --> 00:21:32,617
服务器与AOT设备

769
00:21:32,617 --> 00:21:33,577
告紧自动化

770
00:21:33,577 --> 00:21:35,237
都适合用

771
00:21:35,237 --> 00:21:37,297
用我们的J5来先处理

772
00:21:37,297 --> 00:21:39,177
比如说模型的路由与大模型

773
00:21:39,177 --> 00:21:39,837
这要评估

774
00:21:39,837 --> 00:21:41,757
模型路由对不对

775
00:21:41,757 --> 00:21:45,277
因为你全部使用GPT或者Close

776
00:21:45,277 --> 00:21:47,177
来解决问题实在是太贵了

777
00:21:47,177 --> 00:21:48,757
公司消耗不起对吧

778
00:21:48,757 --> 00:21:49,717
那么这个时候呢

779
00:21:49,717 --> 00:21:53,197
我就完全可以由J5来做一个模型的调度

780
00:21:53,197 --> 00:21:54,537
来做一个路由

781
00:21:54,537 --> 00:21:56,057
就是说你这个问题

782
00:21:56,057 --> 00:21:58,117
到底适合有哪一个模型来进行处理

783
00:21:58,117 --> 00:21:58,477
对不对

784
00:21:58,477 --> 00:22:01,337
常用的模型可能也就十来个

785
00:22:01,337 --> 00:22:03,437
你把这十来个模型全部累出来

786
00:22:03,437 --> 00:22:06,217
然后你交个J5判断一下就可以了

787
00:22:06,217 --> 00:22:07,797
到底选哪一个模型来处理这个问题

788
00:22:07,797 --> 00:22:08,457
对不对

789
00:22:08,457 --> 00:22:11,337
然后选出来之后

790
00:22:11,337 --> 00:22:13,057
然后我们再交个对应的模型来处理

791
00:22:13,057 --> 00:22:13,357
对不对

792
00:22:13,357 --> 00:22:15,317
比如大模型输出降业预示

793
00:22:15,317 --> 00:22:18,397
判断反复结果是否拼离核心要求

794
00:22:18,397 --> 00:22:20,697
是否存在幻觉分件

795
00:22:20,697 --> 00:22:21,577
做质量预示

796
00:22:21,577 --> 00:22:23,637
比如数据及标注与评测

797
00:22:23,637 --> 00:22:26,097
拼量进行按原判断与分离打分

798
00:22:26,097 --> 00:22:30,577
高效替代大模型的榨纸调用

799
00:22:30,577 --> 00:22:31,757
这只调用

800
00:22:31,757 --> 00:22:33,397
这个其实就是说

801
00:22:33,397 --> 00:22:35,997
它比这个原先用GPT去做决策呢

802
00:22:35,997 --> 00:22:36,957
就要快很多

803
00:22:36,957 --> 00:22:37,717
方便很多

804
00:22:37,717 --> 00:22:40,797
就主打一个高效和年价

805
00:22:40,797 --> 00:22:41,517
对吧

806
00:22:41,517 --> 00:22:43,357
这是我们这种模型的特点

807
00:22:43,357 --> 00:22:44,497
记住啊

808
00:22:44,497 --> 00:22:45,897
它不是要去替代GPT

809
00:22:45,897 --> 00:22:46,877
也不是要替代GPT

810
00:22:46,877 --> 00:22:48,197
也不是替代这Cloud

811
00:22:48,197 --> 00:22:50,817
还是给它们实现一种互补

812
00:22:50,817 --> 00:22:52,137
来完成一个

813
00:22:52,137 --> 00:22:55,157
一个又好用

814
00:22:55,157 --> 00:22:56,397
又省钱的这么一个应用