1
00:00:00,000 --> 00:00:26,000
ChatGPT主創人員之一表示他讓他感到心碎。Diogo Ameta曾是OpenAI研究員,參與構建了聊天機器人並共同發明了RLHF,即基於人類反饋的強化學習。這很可能是造就當前AI時代最關鍵的技術。但他卻感到失望用他的話來說,我們空有驚人潛力,卻無法將其轉化為實際用途。

2
00:00:26,000 --> 00:00:34,380
他的观点是,整个行业一直在针对人类语言进行优化,模型在这方面已出色了四年之久,但这并不能带来自动化。

3
00:00:34,800 --> 00:00:41,800
毕竟计算机的语言与人类不同,他在发布文中,反问,模型在对话方面已超越人类多年。

4
00:00:42,600 --> 00:00:50,020
那自动化究竟在哪?他曾以为对话模型能通向API调用,但尽管外界喧嚣不断,显然还缺了关键的一环。

5
00:00:50,020 --> 00:00:52,500
于是两年前他离开了OpenAI

6
00:00:52,500 --> 00:00:54,300
创办了TypeSafe AI

7
00:00:54,300 --> 00:00:57,740
并出任联合创始人兼CEO

8
00:00:57,740 --> 00:00:59,520
他们融资4000万美元

9
00:00:59,520 --> 00:01:00,940
潜心研发两年

10
00:01:00,940 --> 00:01:04,020
并于9月15日周二发布了JEFF

11
00:01:04,020 --> 00:01:06,400
这是他们首个系统一模型一类

12
00:01:06,400 --> 00:01:07,780
全新的前沿模型

13
00:01:07,780 --> 00:01:10,780
能直接为软件提供快速的结构化决策

14
00:01:10,780 --> 00:01:14,960
最奇特的一点是他基于Transformer架构

15
00:01:14,960 --> 00:01:16,840
但他并非大语言模型

16
00:01:16,840 --> 00:01:18,980
他不输出文本而是返回概率

17
00:01:18,980 --> 00:01:21,200
即他们所说的校准决策

18
00:01:21,200 --> 00:01:24,460
Almeta将其称为机器原生人工智能

19
00:01:24,460 --> 00:01:29,500
因为如果人工智能要改变工作流程

20
00:01:29,500 --> 00:01:31,900
人类不能是智能的唯一受众

21
00:01:31,900 --> 00:01:35,260
TypeSafe将其称为前沿智能函数

22
00:01:35,260 --> 00:01:38,440
输入非结构化状态输出类型化的概率决策

23
00:01:38,440 --> 00:01:41,360
输出结果可以是Jason

24
00:01:41,360 --> 00:01:43,400
也可以只是一个简单的字符串

25
00:01:43,400 --> 00:01:45,540
比如我的卡被重复扣费了

26
00:01:45,540 --> 00:01:48,780
随后你可以使用名为Choice Score的原语来提问

27
00:01:48,780 --> 00:01:52,000
以及NEO他们会返回类型化的选择评分

28
00:01:52,000 --> 00:01:55,480
或是带有概率的不耳直答案全部并行评估

29
00:01:55,480 --> 00:01:56,780
以工单分类为例

30
00:01:56,780 --> 00:02:00,500
假设你需要决定有哪个部门处理客户咨询

31
00:02:00,500 --> 00:02:03,040
你会得到账务0.08

32
00:02:03,040 --> 00:02:04,360
技术0.85

33
00:02:04,360 --> 00:02:07,720
销售0.07

34
00:02:07,720 --> 00:02:08,840
智信度0.82

35
00:02:08,840 --> 00:02:10,800
这对人类而言毫无意义

36
00:02:10,800 --> 00:02:12,900
但对代码来说却非常完美

37
00:02:12,900 --> 00:02:15,180
因为可以直接据此进行逻辑分支

38
00:02:15,180 --> 00:02:16,560
这就是名字的由来

39
00:02:16,560 --> 00:02:19,380
类型安全是指程序捕获异常数据的方式

40
00:02:19,380 --> 00:02:22,020
即防止程序遇到预期之外的情况

41
00:02:22,020 --> 00:02:23,340
比如试图用整数

42
00:02:23,340 --> 00:02:25,480
去除以字符串而使用Jeff

43
00:02:25,480 --> 00:02:27,520
你可以预先定义输出格式

44
00:02:27,520 --> 00:02:29,900
它绝不会输出结构以外的内容

45
00:02:29,900 --> 00:02:31,920
省去了解析与验证的麻烦

46
00:02:31,920 --> 00:02:34,300
而大语言模型输出的只是一个字符串

47
00:02:34,300 --> 00:02:35,480
可能是对话回复

48
00:02:35,480 --> 00:02:36,160
可能是代码

49
00:02:36,160 --> 00:02:37,220
可能是幻觉

50
00:02:37,220 --> 00:02:38,320
也可能是拒绝回答

51
00:02:38,320 --> 00:02:40,560
又或是带有风险的结构化数据

52
00:02:40,560 --> 00:02:42,080
稍不留神就会失控

53
00:02:42,080 --> 00:02:44,360
你当然能用代码封装大语言模型

54
00:02:44,360 --> 00:02:46,220
但这正是Jeff的优势所在

55
00:02:46,220 --> 00:02:46,860
Type Safe

56
00:02:46,860 --> 00:02:48,640
声称它不会产生幻觉

57
00:02:48,640 --> 00:02:50,880
在处理系统一任务时

58
00:02:50,880 --> 00:02:53,080
其表现足以媲美大语言模型

59
00:02:53,080 --> 00:02:54,880
且速度快了两个数量级

60
00:02:54,880 --> 00:02:57,240
效率也更高在其底层架构下

61
00:02:57,240 --> 00:03:00,140
采用了一种全新的并行采样器架构

62
00:03:00,140 --> 00:03:03,320
以及一种名为强化学习的训练方法

63
00:03:03,320 --> 00:03:04,920
即基于校准决策

64
00:03:04,920 --> 00:03:06,280
简称RLCD

65
00:03:06,280 --> 00:03:10,080
The Register将RLCD归类为一种架构

66
00:03:10,080 --> 00:03:12,200
但它本质上是一种训练方法

67
00:03:12,200 --> 00:03:14,220
也是他们对RLHF的回应

68
00:03:14,220 --> 00:03:16,600
像GPD 5.6 Terra这类模型

69
00:03:16,600 --> 00:03:18,200
每次只能生成一个Token

70
00:03:18,200 --> 00:03:21,220
每个Token都依赖前一个Jeff

71
00:03:21,220 --> 00:03:23,040
就能一次性生成所有结果

72
00:03:23,040 --> 00:03:25,400
通过单次硬件感知查询实现

73
00:03:25,400 --> 00:03:27,360
它基于结构化程序状态

74
00:03:27,360 --> 00:03:28,660
而非序列消息构建

75
00:03:28,660 --> 00:03:30,700
他们认为放弃字符串

76
00:03:30,700 --> 00:03:32,180
赋予了模型超能力

77
00:03:32,180 --> 00:03:33,640
而且速度快得惊人

78
00:03:33,640 --> 00:03:34,540
这就是TypeSafe

79
00:03:34,540 --> 00:03:36,420
前沿大模型处理耗时

80
00:03:36,420 --> 00:03:38,660
3到329秒端到端

81
00:03:38,660 --> 00:03:40,340
人类用者还行

82
00:03:40,340 --> 00:03:42,260
但在代码自动化场景下

83
00:03:42,260 --> 00:03:43,580
就是瓶颈Jeff只需

84
00:03:43,580 --> 00:03:45,100
70到500毫秒

85
00:03:45,100 --> 00:03:47,200
速度快了40到200倍

86
00:03:47,200 --> 00:03:50,400
在TechCrunch报道的对比演示中

87
00:03:50,400 --> 00:03:55,320
截图显示Jeff响应仅需0.114秒

88
00:03:55,320 --> 00:03:58,540
而Tara则用了8.566秒

89
00:03:58,540 --> 00:03:59,840
更惊人的是价格

90
00:03:59,840 --> 00:04:02,160
大模型每百万输入Token收费

91
00:04:02,160 --> 00:04:03,680
20美分到10美元

92
00:04:03,680 --> 00:04:05,800
输出费用是输入的5倍

93
00:04:05,800 --> 00:04:10,600
而Jeff每百万Token仅需0.042美元

94
00:04:10,600 --> 00:04:13,300
也就是每10亿Token仅需42美元

95
00:04:13,300 --> 00:04:14,760
没错他们按10亿单位计费

96
00:04:14,760 --> 00:04:16,000
输出成本极低

97
00:04:16,000 --> 00:04:16,980
几乎可以忽略不计

98
00:04:16,980 --> 00:04:19,080
Tara模型输入为每百万两美元

99
00:04:19,080 --> 00:04:21,680
输出则为每百万十二美元

100
00:04:21,680 --> 00:04:23,320
数据显示Jeff模型

101
00:04:23,320 --> 00:04:25,880
比Fable 5.1便宜了238倍

102
00:04:25,880 --> 00:04:27,420
再来说说自信度

103
00:04:27,420 --> 00:04:30,900
这恐怕是最被低估的一点大模型

104
00:04:30,900 --> 00:04:31,940
往往过度自信

105
00:04:31,940 --> 00:04:34,440
且对自身确信度的评估前后不一

106
00:04:34,440 --> 00:04:37,700
如果一个模型能在95%的情况下完成任务

107
00:04:37,700 --> 00:04:38,900
但却无法告诉你

108
00:04:38,900 --> 00:04:39,980
它何时会出错

109
00:04:39,980 --> 00:04:41,200
你就无法实现自动化

110
00:04:41,200 --> 00:04:42,500
Jeff的自信度

111
00:04:42,500 --> 00:04:43,420
经过了较准

112
00:04:43,420 --> 00:04:45,400
自信度越高

113
00:04:45,400 --> 00:04:46,560
准确率也就越高

114
00:04:46,560 --> 00:04:48,800
相似的输入会得到相似的答案

115
00:04:48,800 --> 00:04:50,880
Iron DL的首席技术官

116
00:04:50,880 --> 00:04:52,160
Armond Risser表示

117
00:04:52,160 --> 00:04:54,300
这让开源模型得以发挥其效用

118
00:04:54,300 --> 00:04:56,720
他告诉TechCrunch

119
00:04:56,720 --> 00:04:58,060
这实际上是将

120
00:04:58,060 --> 00:05:00,580
幻觉问题交由用户自行把控

121
00:05:00,580 --> 00:05:03,120
自信度5x10如抛硬币直接忽略

122
00:05:03,120 --> 00:05:05,020
9x50则果断执行

123
00:05:05,020 --> 00:05:07,020
面对高调的声称TypeSafe

124
00:05:07,020 --> 00:05:07,800
心知肚明

125
00:05:07,800 --> 00:05:09,500
正如文中那句

126
00:05:09,500 --> 00:05:11,540
非凡的结论需要非凡的证据

127
00:05:11,540 --> 00:05:14,540
下方还附上了有图有真相的证据配文

128
00:05:14,540 --> 00:05:16,020
还带了个图指甲油的表情

129
00:05:16,020 --> 00:05:17,680
他们自称欢迎质疑

130
00:05:17,680 --> 00:05:19,200
因为他们自己就是怀疑论者

131
00:05:19,200 --> 00:05:20,460
值得称赞的是

132
00:05:20,460 --> 00:05:23,020
他们在速度优化上确实下足了功夫

133
00:05:23,020 --> 00:05:24,200
你可以亲自验证

134
00:05:24,200 --> 00:05:27,200
这些模型多运行在西海岸的笔记本上

135
00:05:27,200 --> 00:05:30,020
也就是该服务的所在地定价方面很透明

136
00:05:30,020 --> 00:05:32,900
但他们坦言无法证明其未获补贴

137
00:05:32,900 --> 00:05:35,440
并预计价格未来会进一步下降

138
00:05:35,440 --> 00:05:37,220
且不存在类型错误

139
00:05:37,220 --> 00:05:39,060
虽然一个反例就能将其推翻

140
00:05:39,060 --> 00:05:41,100
但这在数学上是不可能的

141
00:05:41,100 --> 00:05:42,940
他们说在Terra演示中

142
00:05:42,940 --> 00:05:45,460
他们坦言查询过程经过了高度简化

143
00:05:45,460 --> 00:05:47,580
使用了可读的问题件外加一段

144
00:05:47,580 --> 00:05:49,580
有利于Jeff的密集状态描述

145
00:05:49,580 --> 00:05:50,900
从而美化了其表现

146
00:05:50,900 --> 00:05:52,920
他们选用了默认推理的Terra

147
00:05:52,920 --> 00:05:55,260
认为其智力水平最接近Jeff

148
00:05:55,260 --> 00:05:58,360
唯一的意见分歧在于流失可能性的判断

149
00:05:58,360 --> 00:06:01,280
他们称之为真正的魔能两可有趣的是

150
00:06:01,280 --> 00:06:04,460
正是类似的演示说服了他们全力投入

151
00:06:04,460 --> 00:06:08,340
更有利的证据是他们的工作流评估

152
00:06:08,340 --> 00:06:10,420
旨在衡量AI在代码环境下的表现

153
00:06:10,420 --> 00:06:11,800
没有标准答案标签

154
00:06:11,800 --> 00:06:13,460
也不对评估框架进行微调

155
00:06:13,460 --> 00:06:15,900
迎演导致过你和每个模型

156
00:06:15,900 --> 00:06:17,300
都运行相同的工作流

157
00:06:17,300 --> 00:06:20,740
并根据GT6 Extra的平均表现评分

158
00:06:20,740 --> 00:06:22,580
以及Fable 5.1他们说

159
00:06:22,580 --> 00:06:24,340
Jeff 稳居 帕雷托 前沿

160
00:06:24,340 --> 00:06:27,540
在近两个数量级的范围内

161
00:06:27,540 --> 00:06:30,100
且由大语言模型处理所有逻辑

162
00:06:30,100 --> 00:06:32,920
比起遵循工作流思维链的表现

163
00:06:32,920 --> 00:06:33,880
明显差得多

164
00:06:33,880 --> 00:06:36,340
这些调用远比演示案例复杂

165
00:06:36,340 --> 00:06:37,840
更接近实际生产环境

166
00:06:37,840 --> 00:06:40,880
他们将结果发布在评估网站上

167
00:06:40,880 --> 00:06:42,680
他们认为可靠的工作流

168
00:06:42,680 --> 00:06:44,820
是有许多小型的独立

169
00:06:44,820 --> 00:06:46,300
由概率驱动的问题

170
00:06:46,300 --> 00:06:48,240
组成最终导向一个分支

171
00:06:48,240 --> 00:06:48,720
这就是

172
00:06:48,720 --> 00:06:50,720
他们主页上那些数据的来源

173
00:06:50,720 --> 00:06:51,360
也就是100

174
00:06:51,360 --> 00:06:53,300
93.6倍的提速

175
00:06:53,300 --> 00:06:56,240
以及444.6倍的成本缩减

176
00:06:56,240 --> 00:06:58,800
Virtue将其四舍侮入为100

177
00:06:58,800 --> 00:07:00,260
94和400

178
00:07:00,260 --> 00:07:01,460
即45倍

179
00:07:01,460 --> 00:07:03,280
这大概是实际应用中的上限

180
00:07:03,280 --> 00:07:04,980
他们声称这些工作流

181
00:07:04,980 --> 00:07:06,020
并非刻意挑选

182
00:07:06,020 --> 00:07:08,220
且不在训练数据分布之内

183
00:07:08,220 --> 00:07:10,600
但毕竟是他们自家团队构建的

184
00:07:10,600 --> 00:07:12,760
所以难免存在偏差Astra

185
00:07:12,760 --> 00:07:14,280
和Fable的参考案例

186
00:07:14,280 --> 00:07:16,740
可能更偏向OpenAI和Anthlopic

187
00:07:16,740 --> 00:07:19,400
所以他们认为自己可能还低估了性能

188
00:07:19,400 --> 00:07:21,220
以及DeepSeek这些大模型

189
00:07:21,220 --> 00:07:23,960
都是通过TypeSafe的开源系统运行的

190
00:07:23,960 --> 00:07:26,280
即针对Python的大模型封装库

191
00:07:26,280 --> 00:07:29,360
他们称这是获取大模型决策最准确的方式

192
00:07:29,360 --> 00:07:32,060
但处理幻觉时速度较慢成本也更高

193
00:07:32,060 --> 00:07:34,380
他们视类型安全为基本门槛

194
00:07:34,380 --> 00:07:35,560
对智能体来说

195
00:07:35,560 --> 00:07:37,840
工具调用产生幻觉是不可接受的

196
00:07:37,840 --> 00:07:39,780
但对于有延迟要求的系统

197
00:07:39,780 --> 00:07:40,820
这是致命的缺陷

198
00:07:40,820 --> 00:07:42,940
或是深埋在依赖链的底层

199
00:07:42,940 --> 00:07:45,380
即便是最强的模型也难免出错

200
00:07:45,380 --> 00:07:48,820
它们的大模型指标来自OpenRouter

201
00:07:48,820 --> 00:07:51,780
复杂查询会被自动路由到更强的模型

202
00:07:51,780 --> 00:07:53,820
Jeff的零错误率不是测出来的

203
00:07:53,820 --> 00:07:55,480
而是由模式匹配保证的

204
00:07:55,480 --> 00:07:57,180
The Register的反驳很重肯

205
00:07:57,180 --> 00:07:58,880
Jeff并不生成自然语言

206
00:08:14,423 --> 00:08:16,503
以此演示其具备的实时智能

207
00:08:16,503 --> 00:08:18,003
据他们说

208
00:08:18,003 --> 00:08:20,403
工程师曾担心每秒十次查询

209
00:08:20,403 --> 00:08:22,583
以及每小时七美元的成本

210
00:08:22,583 --> 00:08:24,503
团队认为这比预期更低

211
00:08:24,503 --> 00:08:26,503
他读取的是文本

212
00:08:26,503 --> 00:08:28,223
而非图像格式的游戏状态

213
00:08:28,223 --> 00:08:29,163
暂时还不是

214
00:08:29,163 --> 00:08:32,003
他们也承认非AI机器人玩起来会更强

215
00:08:32,003 --> 00:08:33,163
但他们想要的是

216
00:08:33,163 --> 00:08:35,503
能适应多种游戏状态表达的AI

217
00:08:35,503 --> 00:08:37,023
且能听从指令

218
00:08:37,023 --> 00:08:38,963
用他们的话说着酷毙了

219
00:08:38,963 --> 00:08:41,883
演示和黑客松活动即将开启

220
00:08:41,883 --> 00:08:42,743
还有危机竞速

221
00:08:42,743 --> 00:08:47,543
仅靠点击链接在成百上千个页面间穿梭

222
00:08:47,543 --> 00:08:49,583
每一步都有上千种选择

223
00:08:49,583 --> 00:08:51,863
充分展现了其极高的处理效率

224
00:08:51,863 --> 00:08:56,683
面对海量选项不产生幻觉的优势会不断叠加

225
00:08:56,683 --> 00:08:59,143
Jeff支持的基数上限为200

226
00:08:59,143 --> 00:09:00,603
一旦超过这个数值

227
00:09:00,603 --> 00:09:03,263
他会独立评估各个链接再进行选择

228
00:09:03,263 --> 00:09:05,203
因此偶尔会出现减速

229
00:09:05,203 --> 00:09:08,883
毕竟大模型没有推理过程整体提速空间有限

230
00:09:08,883 --> 00:09:12,623
但Astra设为最低档以保证流畅度的情况除外

231
00:09:12,623 --> 00:09:15,763
而且Jeff往往能以更少的步骤完成任务

232
00:09:15,763 --> 00:09:16,963
这是我最喜欢的一点

233
00:09:16,963 --> 00:09:20,563
第二和第三项挑战的起始页面都随机抽到了小黄鸭

234
00:09:20,563 --> 00:09:23,343
直到团队指出来作者才发现这一点

235
00:09:23,343 --> 00:09:26,943
开发者们蜂拥而至导致需求激增

236
00:09:26,943 --> 00:09:30,523
以至于TypeSafe的API一度无法为用户提供服务

237
00:09:30,523 --> 00:09:31,543
据TechCrunch报道

238
00:09:31,543 --> 00:09:33,543
开发者将其视为一种更经济

239
00:09:33,543 --> 00:09:35,923
更稳健的智能集成方案

240
00:09:35,923 --> 00:09:39,323
Adversal公司的工程师Branet Sharma表示

241
00:09:39,323 --> 00:09:42,443
他们此前一直在使用指令安全分类器

242
00:09:42,443 --> 00:09:45,023
基于OpenAI的GPD-4O Mini模型

243
00:09:45,023 --> 00:09:46,183
但换成Jeff后

244
00:09:46,183 --> 00:09:48,563
处理速度提升了5到18倍

245
00:09:48,563 --> 00:09:50,223
且准确度更高

246
00:09:50,223 --> 00:09:53,063
Brio AI的CTO Nikhil Matokar

247
00:09:53,063 --> 00:09:54,703
在处理商务邮件时

248
00:09:54,703 --> 00:09:57,123
对比了Gemini的准确率

249
00:09:57,123 --> 00:09:57,723
略胜一筹

250
00:09:57,723 --> 00:10:00,523
但成本却高出10到20倍

251
00:10:00,523 --> 00:10:01,943
让他感到震惊的是

252
00:10:01,943 --> 00:10:04,443
况且Jeff是唯一能提供

253
00:10:04,443 --> 00:10:06,063
真实概率反馈的模型

254
00:10:06,063 --> 00:10:07,783
他认为这使其成为

255
00:10:07,783 --> 00:10:09,703
自动化工作流的理想之选

256
00:10:09,703 --> 00:10:12,163
AI Gateway的数据表现简直惊人

257
00:10:12,163 --> 00:10:13,663
发布24小时内

258
00:10:13,663 --> 00:10:15,983
Jeff吸引的付费团队数量

259
00:10:15,983 --> 00:10:17,763
是此前任何模型的两倍多

260
00:10:17,763 --> 00:10:20,583
这是Gateway历史上普及最快的模型

261
00:10:20,583 --> 00:10:22,803
12小时内就超越了所有竞品

262
00:10:22,803 --> 00:10:24,583
第18小时就覆盖了

263
00:10:24,583 --> 00:10:25,843
1%的付费团队

264
00:10:25,843 --> 00:10:28,523
第24小时是接近13%

265
00:10:28,523 --> 00:10:31,523
是GPT 5.6系列的两倍

266
00:10:31,523 --> 00:10:34,383
更是Fable 5.1的6倍多

267
00:10:34,383 --> 00:10:36,603
而其他近期发布的模型

268
00:10:36,603 --> 00:10:38,263
均未超过7%

269
00:10:38,263 --> 00:10:39,883
首日数据就足以证明

270
00:10:39,883 --> 00:10:43,743
专用模型投入生产环境的速度有多快

271
00:10:43,743 --> 00:10:47,223
但真正的考验在于持久性谈到具体用途

272
00:10:47,223 --> 00:10:50,363
Perser提到智能体可以自主选择下一个工具

273
00:10:50,363 --> 00:10:55,883
或在子智能体的工作流中决定是继续重视还是询问用户

274
00:10:55,883 --> 00:10:58,823
亦或在行动前先评估风险并决定是否终止

275
00:10:58,823 --> 00:11:01,543
并将无法确定的情况转交人工处理

276
00:11:01,543 --> 00:11:04,683
TypeSafe将其称为智能异服语句或模糊规则

277
00:11:04,683 --> 00:11:07,143
因为手写逻辑往往过于脆弱

278
00:11:07,143 --> 00:11:08,383
他们还建议使用Map

279
00:11:08,383 --> 00:11:11,923
将PB级数据精炼为特征以支持实时应用

280
00:11:11,923 --> 00:11:15,743
毕竟100毫秒的延迟完全在用户体验的可控范围内

281
00:11:15,743 --> 00:11:18,663
从推理过程到越狱尝试全程进行验证

282
00:11:18,663 --> 00:11:21,543
Amata希望人们能用它来监控智能体

283
00:11:21,543 --> 00:11:23,963
由于智能体互间成本高昂

284
00:11:23,963 --> 00:11:26,863
Runiker建议采用实时模型路由

285
00:11:26,863 --> 00:11:31,263
The Register补充到聊天机器人人由LLM架构主导

286
00:11:31,263 --> 00:11:36,343
编程助手代码智能体以及数学证明或演示等可验证任务

287
00:11:36,343 --> 00:11:39,043
而这正如Type Safe戏称的那样

288
00:11:39,043 --> 00:11:40,743
效果时好时坏

289
00:11:40,743 --> 00:11:42,303
至于内部究竟如何

290
00:11:42,303 --> 00:11:43,643
Amita手口如瓶

291
00:11:43,643 --> 00:11:46,483
外界普遍怀疑其底层使用的是

292
00:11:46,483 --> 00:11:47,843
开源权重的大模型

293
00:11:47,843 --> 00:11:50,643
尽管Jeff称其FAQ文档内容并不短

294
00:11:50,643 --> 00:11:54,283
Jeff模型完全采用合成数据进行训练

295
00:11:54,283 --> 00:11:56,263
Amita告诉TechCrunch

296
00:11:56,263 --> 00:11:59,663
资产数据是他人生中最成功的豪赌之一

297
00:11:59,663 --> 00:12:02,263
比发布产品更成功

298
00:12:02,263 --> 00:12:03,803
也优于RLHF

299
00:12:03,803 --> 00:12:07,143
TypeSafe自诩为一家核心的数据研究实验室

300
00:12:07,143 --> 00:12:09,443
公司半数员工几乎主导了

301
00:12:09,443 --> 00:12:12,583
统计学可控合成数据这一细分领域

302
00:12:12,583 --> 00:12:14,383
这已成他毕生乐趣

303
00:12:14,383 --> 00:12:17,563
即便你请求他们也不会用你的数据训练请见谅

304
00:12:17,563 --> 00:12:18,723
想了解细节

305
00:12:18,723 --> 00:12:20,463
得先被他们录用

306
00:12:20,463 --> 00:12:23,283
这就是他们所奉行的苦涩教训

307
00:12:23,283 --> 00:12:26,503
优化任务远比堆砌数据和算力重要

308
00:12:26,503 --> 00:12:29,943
或是RLHF这类针对文本优化的算法

309
00:12:29,943 --> 00:12:32,163
人类评估员偏好虽然适合聊天

310
00:12:32,163 --> 00:12:33,503
却不适用于自动化任务

311
00:12:33,503 --> 00:12:37,543
且RLHF仅适用于程序可校验的任务

312
00:12:37,543 --> 00:12:40,283
导致其现实判断波动极大

313
00:12:40,283 --> 00:12:41,803
智能表现不够稳健

314
00:12:41,803 --> 00:12:43,343
系统1这一名称

315
00:12:43,343 --> 00:12:45,263
源自卡尼曼的思考快与曼

316
00:12:45,263 --> 00:12:46,963
尽管在传统认知中

317
00:12:46,963 --> 00:12:48,803
系统1往往意味着一出错

318
00:12:48,803 --> 00:12:51,163
但他们认为这些模型

319
00:12:51,163 --> 00:12:52,803
其实比其他方案更可靠

320
00:12:52,803 --> 00:12:55,403
他们还完全放弃了公开基准测试

321
00:12:55,403 --> 00:12:56,843
仅提供一次性评估

322
00:12:56,843 --> 00:12:59,483
并推动用户自行构建评估体系

323
00:12:59,483 --> 00:13:02,403
因为系统1任务更易于评估

324
00:13:02,403 --> 00:13:05,603
甚至还有篇题为反基准优化的文章

325
00:13:05,603 --> 00:13:08,003
Jeff 这个名字源自威廉 Tauffell

326
00:13:08,003 --> 00:13:11,403
这位19世纪的经济学家提出了一个悖论

327
00:13:11,403 --> 00:13:13,203
当蒸汽机效率提升

328
00:13:13,203 --> 00:13:16,403
煤炭需求反而不降反升他们压住的是

329
00:13:17,803 --> 00:13:20,003
智能成本每降低一个数量级

330
00:13:20,003 --> 00:13:22,803
就会解锁多出数个数量级的应用场景

331
00:13:24,003 --> 00:13:26,803
阿尔梅达预见智能软件将无处不再呈现出涌现性

332
00:13:26,803 --> 00:13:30,203
且更像早期互联网那样分布化而非如今的超级应用

333
00:13:30,203 --> 00:13:31,803
但The Register 指出

334
00:13:31,803 --> 00:13:33,163
这建立在一个假设之上

335
00:13:33,163 --> 00:13:35,063
即Token的需求会像能源一样广泛

336
00:13:35,063 --> 00:13:37,263
这尚未有定论PO研究中心

337
00:13:37,263 --> 00:13:40,703
发现许多能用上AI的人根本找不到用处

338
00:13:40,703 --> 00:13:44,923
CNBC调查显示另一些人则出于道德顾虑排斥AI

339
00:13:44,923 --> 00:13:47,543
目前Jeff依然是特例

340
00:13:47,543 --> 00:13:50,083
但Renneker预计竞争终会出现

341
00:13:50,083 --> 00:13:52,383
他感叹我们本应早点看透这一点

342
00:13:52,383 --> 00:13:54,343
毕竟大语言模型太便宜了

343
00:13:54,343 --> 00:13:56,983
再加上补贴人们往往懒得去创新

344
00:13:56,983 --> 00:14:00,143
但TypeSafe正在开发更多新模态版本

345
00:14:00,143 --> 00:14:03,203
并开始从后补名单中邀请开发者

346
00:14:03,203 --> 00:14:05,403
询问他们想实现什么样的科幻构想

347
00:14:05,403 --> 00:14:07,163
当TypeCrunch问起TypeSafe

348
00:14:07,163 --> 00:14:08,743
是否属于前沿实验室时

349
00:14:08,743 --> 00:14:11,183
Almeta直言前沿实验室的主打产品

350
00:14:11,183 --> 00:14:12,323
无非是恐惧和炒作

351
00:14:12,323 --> 00:14:14,203
它追求的是真正智能

352
00:14:14,203 --> 00:14:15,923
而非一场无限财富的豪赌

353
00:14:15,923 --> 00:14:17,303
也不是某种宗教

354
00:14:17,303 --> 00:14:19,123
更不是在数据中心里创造神

355
00:14:19,123 --> 00:14:20,523
总之 这是Jeff

356
00:14:20,523 --> 00:14:22,303
欢迎在评论区留言分享

357
00:14:22,303 --> 00:14:23,463
你会如何使用它

358
00:14:23,463 --> 00:14:24,083
感谢观看

359
00:14:24,083 --> 00:14:25,303
我们下期见