1
00:00:00,100 --> 00:00:01,500
各位朋友早上好

2
00:00:01,600 --> 00:00:03,933
今天咱们就来看看本周AI界

3
00:00:03,933 --> 00:00:06,200
几件暗流涌动的军备竞赛

4
00:00:06,333 --> 00:00:09,200
从Openai秘密研发的超级新模型

5
00:00:09,300 --> 00:00:12,133
到阿里悄悄上架的Qwen新势力

6
00:00:12,333 --> 00:00:13,833
信息量直接拉满

7
00:00:13,833 --> 00:00:15,900
先看画面这份业内研报

8
00:00:16,133 --> 00:00:19,300
里面指明Gemini3.5Flash表现平平

9
00:00:19,600 --> 00:00:22,033
而Openai正研发规模超越

10
00:00:22,033 --> 00:00:24,000
Estrela的全新预训练模型

11
00:00:24,033 --> 00:00:24,900
与此同时

12
00:00:25,066 --> 00:00:26,933
阿里团队在Almerina

13
00:00:26,933 --> 00:00:29,333
秘密测试代号Kiana的模型

14
00:00:29,600 --> 00:00:32,033
其3D渲染能力让人惊艳

15
00:00:32,033 --> 00:00:34,733
此外图像生成领域有重磅进展

16
00:00:35,033 --> 00:00:37,300
带Openai since day水印的

17
00:00:37,300 --> 00:00:39,100
蒙娜丽莎已现身竞技场

18
00:00:39,100 --> 00:00:41,100
马斯克旗下的Grok image

19
00:00:41,100 --> 00:00:43,300
2.0也已经正式发布

20
00:00:43,500 --> 00:00:46,433
目前在文生图榜单高居第二

21
00:00:46,700 --> 00:00:48,900
综合表现同样不容小觑

22
00:00:49,066 --> 00:00:50,033
遗憾的是

23
00:00:50,333 --> 00:00:52,300
谷歌原本计划推出的

24
00:00:52,300 --> 00:00:56,633
120B超大体量模型Gemma V4 1,200亿

25
00:00:56,833 --> 00:00:59,600
近期爆出该项目已经被取消

26
00:00:59,633 --> 00:01:01,066
看完这轮大盘概览

27
00:01:01,066 --> 00:01:03,633
咱们先来深挖Openai的重磅核弹

28
00:01:04,066 --> 00:01:06,466
知名爆料人Chris在推特指出

29
00:01:06,700 --> 00:01:09,300
GPT5.5绝不是Openai今年

30
00:01:09,300 --> 00:01:11,100
最后一次大规模预训练

31
00:01:11,100 --> 00:01:12,433
爆料里特别强调

32
00:01:12,533 --> 00:01:15,333
虽然GPT6本身已经足够震撼

33
00:01:15,400 --> 00:01:18,666
但Openai在年底前还会掏出代号为

34
00:01:18,933 --> 00:01:20,066
bug的超级模型

35
00:01:20,200 --> 00:01:23,000
预训练规模将直接创下历史新高

36
00:01:23,000 --> 00:01:25,133
推文里甚至极其夸张的声称

37
00:01:25,333 --> 00:01:27,666
bug这款新模型的恐怖性能

38
00:01:27,800 --> 00:01:29,733
绝对会让之前的fable

39
00:01:29,933 --> 00:01:31,666
显得像个原始玩具

40
00:01:32,133 --> 00:01:33,233
不过对于它的发布

41
00:01:33,233 --> 00:01:35,633
时间线与代号对应关系

42
00:01:35,700 --> 00:01:37,033
行业内目前依然

43
00:01:37,033 --> 00:01:38,700
存在不少争论与疑虑

44
00:01:38,700 --> 00:01:39,866
比如知名博主

45
00:01:39,866 --> 00:01:41,833
tokengrammaring就公开反驳

46
00:01:42,266 --> 00:01:45,233
指出SEMI analysis在整理这份报告时

47
00:01:45,300 --> 00:01:47,433
把7月初的dog信息和

48
00:01:47,433 --> 00:01:49,666
后续的Astra给彻底搞混了

49
00:01:49,666 --> 00:01:50,466
在他看来

50
00:01:50,466 --> 00:01:53,066
dog早在Astra之前就已经

51
00:01:53,066 --> 00:01:54,866
在内部报告中出现过

52
00:01:55,200 --> 00:01:57,233
因此它极大概率只是

53
00:01:57,300 --> 00:02:00,300
Asha早期研发阶段的一个内部代号

54
00:02:00,400 --> 00:02:03,000
而不是什么全新的后续迭代产品

55
00:02:03,000 --> 00:02:06,433
毕竟Openai内部取代号向来都很随意

56
00:02:06,600 --> 00:02:08,433
此前就曾用过像

57
00:02:08,433 --> 00:02:10,900
Milthree Milfour这种混乱名称

58
00:02:11,000 --> 00:02:12,933
最后才统一叫Asha

59
00:02:13,233 --> 00:02:16,000
这也让外界很难做出精准判断

60
00:02:16,000 --> 00:02:16,733
所以说

61
00:02:16,733 --> 00:02:19,233
咱们目前依然无法百分百确定

62
00:02:19,466 --> 00:02:21,400
dog到底代表着一个全新

63
00:02:21,400 --> 00:02:23,233
的超大模型预训练流

64
00:02:23,400 --> 00:02:25,400
还是只属于同一批预

65
00:02:25,400 --> 00:02:27,466
训练中的不同阶段检查点

66
00:02:27,700 --> 00:02:29,033
依旧充满神秘感

67
00:02:29,033 --> 00:02:31,100
而如果把视线转向谷歌

68
00:02:31,300 --> 00:02:33,533
研报给出的结论就更残酷了

69
00:02:33,700 --> 00:02:36,100
文章直接断言Gemini 3 Pro

70
00:02:36,100 --> 00:02:38,300
已经是谷歌技术的高峰

71
00:02:38,700 --> 00:02:40,933
近期deep mind高层的变动更

72
00:02:40,933 --> 00:02:42,433
是加剧了内部焦虑

73
00:02:42,433 --> 00:02:43,800
报告甚至预测

74
00:02:43,933 --> 00:02:48,200
接下来的Gemini3.5Pro实力仅相当于Opus4.5

75
00:02:48,333 --> 00:02:50,500
甚至不如GLM5.2

76
00:02:50,600 --> 00:02:53,133
面对Anthropic与Openai的死斗

77
00:02:53,133 --> 00:02:53,933
很难招架

78
00:02:53,933 --> 00:02:56,100
不过其中最为引人注意的

79
00:02:56,133 --> 00:02:57,600
仍然是这份研报

80
00:02:57,600 --> 00:02:59,500
尾部给出的重磅结论

81
00:02:59,800 --> 00:03:01,866
Openai已经彻底克服了

82
00:03:01,866 --> 00:03:03,200
预训练的技术障碍

83
00:03:03,500 --> 00:03:05,000
规模更为庞大的大g

84
00:03:05,100 --> 00:03:06,800
模型正在全力推进

85
00:03:07,100 --> 00:03:10,066
极可能在今年年底前彻底颠覆行业

86
00:03:10,066 --> 00:03:12,300
聊完海外巨头的暗流涌动

87
00:03:12,400 --> 00:03:14,433
咱们把视角转回国内

88
00:03:14,833 --> 00:03:18,433
阿里团队最近在LLM Arena的code

89
00:03:18,433 --> 00:03:21,000
Arena上悄悄上线了一款代号

90
00:03:21,000 --> 00:03:23,833
为piano的全新检查点模型

91
00:03:23,833 --> 00:03:25,933
不少开发者第一时间拿它

92
00:03:25,933 --> 00:03:28,333
跑起了硬核物理弹道测试

93
00:03:28,733 --> 00:03:31,500
将提亚娜与Claude OPPO S5以及

94
00:03:31,500 --> 00:03:34,700
GPT5.6放在同等提示词下

95
00:03:34,733 --> 00:03:38,000
进行3D前端代码生成的大比拼

96
00:03:38,033 --> 00:03:40,600
而从最终实测运行效果来看

97
00:03:40,733 --> 00:03:43,300
提亚娜用threejs生成的

98
00:03:43,300 --> 00:03:45,300
弩炮模拟动画可以说是

99
00:03:45,600 --> 00:03:46,500
极其逼真

100
00:03:46,666 --> 00:03:48,833
不仅整体结构非常完整

101
00:03:49,066 --> 00:03:50,800
拉弦和发射的物理

102
00:03:50,800 --> 00:03:52,633
交互更是丝滑顺畅

103
00:03:52,700 --> 00:03:53,466
在我看来

104
00:03:53,466 --> 00:03:55,000
这款神秘新模型

105
00:03:55,300 --> 00:03:58,400
在前端网页开发与复杂逻辑呈现上

106
00:03:58,466 --> 00:04:01,000
展现出的综合素质完全不

107
00:04:01,000 --> 00:04:03,266
逊色于目前市场上任何

108
00:04:03,266 --> 00:04:05,600
顶尖的第一梯队旗舰模型

109
00:04:05,600 --> 00:04:08,066
如果大家也想亲自体验它的实力

110
00:04:08,400 --> 00:04:10,700
可以去l美Arena的下拉菜单

111
00:04:10,700 --> 00:04:13,100
里选择Beta mode盲测模式

112
00:04:13,466 --> 00:04:15,266
随后进入code Arena

113
00:04:15,333 --> 00:04:16,866
输入你想要的prompt

114
00:04:16,866 --> 00:04:18,900
在提交具体需求之后

115
00:04:19,000 --> 00:04:21,200
系统会有两个完全匿名的

116
00:04:21,200 --> 00:04:23,600
模型同时在后台生成网页

117
00:04:23,900 --> 00:04:25,866
用户需要通过对比双边

118
00:04:25,866 --> 00:04:28,233
生成的结果进行盲测投票

119
00:04:28,400 --> 00:04:30,300
投票结束后就会揭晓

120
00:04:30,300 --> 00:04:32,400
代码到底属于哪两款模型

121
00:04:32,400 --> 00:04:34,266
不过由于抽取是随机的

122
00:04:34,266 --> 00:04:37,100
想碰上Kiana得运气好或多刷几次

123
00:04:37,500 --> 00:04:39,833
下面咱们就来看几个它生成

124
00:04:39,833 --> 00:04:42,233
的代表性高难度实战案例

125
00:04:42,233 --> 00:04:44,866
咱们先来看这个极其惊艳的案例

126
00:04:45,233 --> 00:04:48,200
提示词要求生成一个名为AIRO

127
00:04:48,200 --> 00:04:50,933
one的360度互动耳机展示页面

128
00:04:51,066 --> 00:04:53,433
Tianna生成的UI显得极其精致

129
00:04:53,666 --> 00:04:55,533
黑金高级感直接拉满

130
00:04:55,533 --> 00:04:56,533
更绝的是

131
00:04:56,600 --> 00:04:58,700
这个页面不仅支持用鼠标

132
00:04:58,700 --> 00:05:01,333
全向拖拽旋转和自由缩放

133
00:05:01,700 --> 00:05:03,700
点击耳机上的touch control

134
00:05:03,833 --> 00:05:04,633
DISC按钮

135
00:05:04,933 --> 00:05:07,033
还会及时弹出极其精准

136
00:05:07,033 --> 00:05:09,066
的技术参数标签与提示

137
00:05:09,400 --> 00:05:10,633
完成度非常高

138
00:05:10,633 --> 00:05:13,133
而最让人惊叹的则是页面

139
00:05:13,133 --> 00:05:15,633
底部的动态配色切换功能

140
00:05:16,000 --> 00:05:17,733
当点击不同颜色按钮

141
00:05:17,900 --> 00:05:19,866
3D耳机的表面材质与

142
00:05:19,866 --> 00:05:21,700
光照渲染会实时改变

143
00:05:21,900 --> 00:05:24,433
这种threejs代码实现水平

144
00:05:24,600 --> 00:05:26,900
简直远远超越了大多数对手

145
00:05:26,900 --> 00:05:28,366
看完了单独展示

146
00:05:28,500 --> 00:05:29,666
咱们再来看一组

147
00:05:29,666 --> 00:05:31,533
直观的四方横向对比

148
00:05:32,066 --> 00:05:33,533
在同等提示词下

149
00:05:33,533 --> 00:05:35,166
生成蒸汽朋克飞艇

150
00:05:35,433 --> 00:05:38,300
四个画面分别由Claude OPS 5米x

151
00:05:38,700 --> 00:05:40,633
Tiana Claude Fabo5嗨

152
00:05:40,733 --> 00:05:42,900
以及昆仑3.8Max渲染

153
00:05:42,900 --> 00:05:44,300
大家仔细看右上角

154
00:05:44,300 --> 00:05:46,433
Tiana生成的这艘飞艇

155
00:05:46,833 --> 00:05:48,366
无论是气囊的网格

156
00:05:48,566 --> 00:05:50,133
复杂的金属管道

157
00:05:50,133 --> 00:05:52,666
还是悬挂吊舱的细节丰富度

158
00:05:52,700 --> 00:05:54,533
都明显碾压了左下角

159
00:05:54,533 --> 00:05:56,666
FIB五h生成的简化模型

160
00:05:56,666 --> 00:05:59,333
虽然左上角Opus5在材质与

161
00:05:59,333 --> 00:06:01,500
细节纹理上略微胜出一点点

162
00:06:01,733 --> 00:06:04,266
但Kiara展示出的设计合理性

163
00:06:04,300 --> 00:06:06,900
还有超强的复杂代码构建能力

164
00:06:07,033 --> 00:06:08,700
已经完全能跟顶尖的

165
00:06:08,700 --> 00:06:10,466
Opus模型正面交锋了

166
00:06:10,666 --> 00:06:12,766
结合刚才这一系列惊人表现

167
00:06:12,900 --> 00:06:14,300
业内普遍推测

168
00:06:14,366 --> 00:06:16,266
这个代号QIANA的模型

169
00:06:16,366 --> 00:06:18,733
极可能就是阿里即将发布的

170
00:06:18,866 --> 00:06:22,033
QWEN4.0或Qwen三点九大更新的

171
00:06:22,233 --> 00:06:24,533
前哨检查点太值得期待了

172
00:06:24,533 --> 00:06:26,133
咱们再看这组黄昏

173
00:06:26,133 --> 00:06:28,466
庄园3D场景的对抗

174
00:06:28,566 --> 00:06:30,466
由克亚娜与老牌

175
00:06:30,466 --> 00:06:33,533
葵文3.8Max shy展开直接PK

176
00:06:34,066 --> 00:06:36,633
这次给到的高难度渲染要求是

177
00:06:36,633 --> 00:06:39,433
生成一个极致拟真的黄昏乡村庄园

178
00:06:39,433 --> 00:06:40,933
大家可以清晰的看到

179
00:06:40,966 --> 00:06:43,633
Tianna生成的庄园简直太惊艳了

180
00:06:43,700 --> 00:06:46,133
不仅精细的还原了小巧的

181
00:06:46,133 --> 00:06:48,633
彩灯串与温馨的屋顶光影

182
00:06:48,766 --> 00:06:50,333
周围还布局了层次

183
00:06:50,333 --> 00:06:51,933
丰富的花坛和森林

184
00:06:52,466 --> 00:06:54,300
光影渲染的整体质感

185
00:06:54,333 --> 00:06:56,233
远远超越了下方的老版本

186
00:06:56,233 --> 00:06:58,366
可以说是彻底坐实了它作为

187
00:06:58,366 --> 00:07:00,366
下一代大模型的绝对实力

188
00:07:00,366 --> 00:07:01,833
聊完了代码与3D

189
00:07:01,900 --> 00:07:03,366
咱们再把视线转向

190
00:07:03,366 --> 00:07:05,133
另一个卷到飞起的赛道

191
00:07:05,266 --> 00:07:06,300
图像生成

192
00:07:06,733 --> 00:07:09,166
最近在Lrmina竞技场里

193
00:07:09,166 --> 00:07:11,466
也突然低调上线了一个代号叫

194
00:07:11,466 --> 00:07:13,900
蒙娜丽莎一的全新图像大模型

195
00:07:14,133 --> 00:07:15,633
表现同样非常惹眼

196
00:07:15,633 --> 00:07:17,566
而要说最锤的铁证

197
00:07:17,700 --> 00:07:20,633
就是用Openai官方的验证工具去

198
00:07:20,633 --> 00:07:23,366
检测蒙娜丽莎一跑出来的图片

199
00:07:23,700 --> 00:07:26,266
系统居然直接识别出了Openai

200
00:07:26,266 --> 00:07:29,133
自家专属的seed ID隐形水印

201
00:07:29,366 --> 00:07:31,300
这下可以说是当场破案了

202
00:07:31,300 --> 00:07:33,466
这就意味着Openai在图像

203
00:07:33,466 --> 00:07:35,566
生成领域也要掏出底牌了

204
00:07:35,766 --> 00:07:38,233
无论接下来叫ChatGPT image

205
00:07:38,233 --> 00:07:40,533
2.5还是直接升到3.0

206
00:07:40,700 --> 00:07:42,566
画质与控图能力都将

207
00:07:42,566 --> 00:07:44,733
迎来脱胎换骨般的质飞跃

208
00:07:44,733 --> 00:07:47,633
不过就在各大厂商猛砸算力的时候

209
00:07:47,833 --> 00:07:50,166
谷歌这边却传来了一条坏消息

210
00:07:50,533 --> 00:07:52,833
此前被开源社区寄予厚望的

211
00:07:52,833 --> 00:07:56,066
超大开源模型Gemma V41,200亿

212
00:07:56,133 --> 00:07:59,133
最新爆料显示已被官方彻底砍掉

213
00:07:59,133 --> 00:08:01,333
根据泄露的后台界面显示

214
00:08:01,500 --> 00:08:05,233
这款原本拥有1,230亿参数以及

215
00:08:05,533 --> 00:08:09,100
262k超百万级上下文窗口的开源旗舰

216
00:08:09,100 --> 00:08:12,300
大模型已经被标记为了取消状态

217
00:08:12,466 --> 00:08:14,766
确定不会再面向公众发布了

218
00:08:14,766 --> 00:08:17,700
这也侧面印证了谷歌在开源与闭源

219
00:08:17,700 --> 00:08:20,366
战略之间正在经历剧烈的阵痛调整

220
00:08:20,700 --> 00:08:22,266
他们很可能是想把

221
00:08:22,266 --> 00:08:23,966
全部计算资源集中起来

222
00:08:23,966 --> 00:08:26,433
全力冲刺下一代闭源Gemini

223
00:08:26,700 --> 00:08:27,366
不得不说

224
00:08:27,366 --> 00:08:28,666
真挺令人唏嘘的

225
00:08:28,666 --> 00:08:30,500
而除了Openai与谷歌

226
00:08:30,566 --> 00:08:33,100
马斯克的Grok也不甘示弱

227
00:08:33,233 --> 00:08:34,233
咱们一起来看看

228
00:08:34,233 --> 00:08:36,100
这些精美的生成范例

229
00:08:36,633 --> 00:08:38,033
无论是波普风海报

230
00:08:38,166 --> 00:08:39,500
细腻纺织纹理

231
00:08:39,666 --> 00:08:41,466
还是素描步骤拆解

232
00:08:41,733 --> 00:08:44,166
Grok image 2.0在文字渲染和

233
00:08:44,166 --> 00:08:46,433
审美排版上都极其出色

234
00:08:46,433 --> 00:08:47,266
按照惯例

235
00:08:47,333 --> 00:08:49,633
看完软件层的神仙打架

236
00:08:49,666 --> 00:08:52,333
咱们在结尾再来看一个极其

237
00:08:52,333 --> 00:08:55,100
硬核的具身智能硬件项目

238
00:08:55,366 --> 00:08:58,166
东京大学最新展示了一款名为

239
00:08:58,566 --> 00:09:01,500
Dragon的空中变形飞行机器人

240
00:09:01,566 --> 00:09:03,966
它最令人惊叹的地方就是彻底

241
00:09:03,966 --> 00:09:06,966
打破了传统无人机固定的形态限制

242
00:09:07,266 --> 00:09:09,166
能在空中完成极度

243
00:09:09,166 --> 00:09:11,166
复杂的自由变体运动

244
00:09:11,266 --> 00:09:12,266
大家仔细看

245
00:09:12,466 --> 00:09:14,900
Dirogan机器人在空中飞行时

246
00:09:14,900 --> 00:09:16,933
能精准改变关节构形

247
00:09:17,133 --> 00:09:19,300
甚至能用身体灵活抓取

248
00:09:19,333 --> 00:09:21,333
并包裹移动悬挂物体

249
00:09:21,666 --> 00:09:23,833
在结合了AI控制算法后

250
00:09:23,866 --> 00:09:26,266
它生动展示了具身智能在

251
00:09:26,266 --> 00:09:28,233
复杂空间下的无限潜力

252
00:09:28,300 --> 00:09:30,933
从底层大模型到具身智能

253
00:09:31,066 --> 00:09:33,766
AI的进化爆发正全面到来

254
00:09:34,033 --> 00:09:35,166
感谢你看到这里

255
00:09:35,166 --> 00:09:36,666
这里是大雷早上好

256
00:09:36,666 --> 00:09:37,533
关注频道

257
00:09:37,533 --> 00:09:38,633
咱们下期见