1 00:00:00,000 --> 00:00:01,400 行業拐點來了。 2 00:00:01,400 --> 00:00:02,900 Deepseek V4 Flash。 3 00:00:02,900 --> 00:00:03,900 展現。 4 00:00:03,900 --> 00:00:04,600 價格。 5 00:00:04,600 --> 00:00:05,600 BBN10倍。 6 00:00:05,600 --> 00:00:07,100 中間內存的模型沒得生存了。 7 00:00:07,100 --> 00:00:07,900 Meta沒了。 8 00:00:07,900 --> 00:00:08,600 Macroke。 9 00:00:08,600 --> 00:00:09,200 Germline。 10 00:00:09,200 --> 00:00:09,800 再見。 11 00:00:09,800 --> 00:00:11,700 日常的活真的是列入醫保了一樣。 12 00:00:11,700 --> 00:00:12,500 就是你隨便用。 13 00:00:12,500 --> 00:00:15,300 這一單雜任我們再去找海外這幾個模型搞一下。 14 00:00:15,300 --> 00:00:18,000 但是你單看Deepseek V4絕對不是行業拐點。 15 00:00:18,000 --> 00:00:19,700 大家應該把第二個事情點起來。 16 00:00:19,700 --> 00:00:21,200 就是Google的DeepMet。 17 00:00:21,200 --> 00:00:21,700 震動。 18 00:00:21,700 --> 00:00:23,700 Pasabis都退二線了。 19 00:00:23,700 --> 00:00:24,800 Jeff Dean離職。 20 00:00:24,800 --> 00:00:25,600 這說明什麼? 21 00:00:25,600 --> 00:00:26,100 同志們。 22 00:00:26,100 --> 00:00:27,400 就說明Pigachin。 23 00:00:30,000 --> 00:00:36,820 一个要上天了 24 00:00:36,820 --> 00:00:37,780 一个要入地了 25 00:00:37,780 --> 00:00:38,160 对不对 26 00:00:38,160 --> 00:00:40,020 你看现在Deep Seek V4 27 00:00:40,020 --> 00:00:41,260 这个事情一出来 28 00:00:41,260 --> 00:00:43,360 价格比别人低10倍 29 00:00:43,360 --> 00:00:43,740 对吧 30 00:00:43,740 --> 00:00:44,680 你这么理解吧 31 00:00:44,680 --> 00:00:45,480 就拼多多 32 00:00:45,480 --> 00:00:45,920 对吧 33 00:00:45,920 --> 00:00:47,820 10块钱美金的东西 34 00:00:47,820 --> 00:00:50,600 在美国卖10块钱人民币 35 00:00:50,600 --> 00:00:51,820 对吧 36 00:00:51,820 --> 00:00:53,300 你就这么理解 37 00:00:53,300 --> 00:00:54,880 干出一个拼多多 38 00:00:54,880 --> 00:00:55,860 那你看 39 00:00:55,860 --> 00:00:57,960 你那个高于这个钱的 40 00:00:57,960 --> 00:00:58,900 这个东西没人买 41 00:00:58,900 --> 00:00:59,580 对不对 42 00:00:59,580 --> 00:01:02,520 我感觉一个更贴近的比喻是现在进入了 43 00:01:02,520 --> 00:01:05,460 AI有医保的名册的时代了 44 00:01:05,460 --> 00:01:11,220 完全就是普通的药直接走几彩就完事了 45 00:01:11,220 --> 00:01:13,940 然后贵的你走进口药就这样 46 00:01:13,940 --> 00:01:15,680 是吧 47 00:01:15,680 --> 00:01:19,160 基本中间那层没得生存了 48 00:01:19,160 --> 00:01:20,420 你说这什么Groak 49 00:01:20,420 --> 00:01:22,140 Germ9再见 50 00:01:22,140 --> 00:01:25,240 我觉得最大的问题是说 51 00:01:25,240 --> 00:01:27,660 大家得重新去定位自己的产品 52 00:01:27,660 --> 00:01:30,980 就是说最首先的是那些定价 53 00:01:30,980 --> 00:01:32,920 可能冲着Opus去干的 54 00:01:32,920 --> 00:01:34,340 或者冲着GPT去干的 55 00:01:34,340 --> 00:01:36,580 然后能力只比Kimi 56 00:01:36,580 --> 00:01:39,140 DeepSeek还有GLM高一点 57 00:01:39,140 --> 00:01:41,180 高一点甚至可能只能持平 58 00:01:41,180 --> 00:01:42,920 包括在不限于 59 00:01:42,920 --> 00:01:43,820 来一个 60 00:01:43,820 --> 00:01:48,120 比如说Meta有一个很出名的叫Spark 61 00:01:48,120 --> 00:01:49,220 其实也不出名 62 00:01:49,220 --> 00:01:50,740 可能很多人都不知道它那个模型 63 00:01:50,740 --> 00:01:52,480 Meta都转去卖算力了 64 00:01:52,480 --> 00:01:52,700 对 65 00:01:52,700 --> 00:01:57,540 Meta跌了那么多过亿年薪的人 66 00:01:57,540 --> 00:01:57,940 做的 67 00:01:57,940 --> 00:01:58,380 对吧 68 00:01:58,380 --> 00:01:58,860 我们就说 69 00:01:58,860 --> 00:01:59,520 说不定Meta 70 00:01:59,520 --> 00:02:00,460 以一己之力 71 00:02:00,460 --> 00:02:01,580 在美国把DeepSeek 72 00:02:01,580 --> 00:02:03,400 V4价格再降六倍 73 00:02:03,400 --> 00:02:06,320 有可能他可能倒卖这些 74 00:02:06,320 --> 00:02:07,620 把自己做成中转站 75 00:02:07,620 --> 00:02:08,660 去倒卖这个数据 76 00:02:08,660 --> 00:02:09,700 他可能更赚钱 77 00:02:09,700 --> 00:02:12,180 所以老修第一个结论 78 00:02:12,180 --> 00:02:13,080 Meta没了 79 00:02:13,080 --> 00:02:15,280 Meta被斩杀掉了 80 00:02:15,280 --> 00:02:16,660 我觉得这个斩杀线的意思 81 00:02:16,660 --> 00:02:16,980 就是什么 82 00:02:16,980 --> 00:02:18,120 就是说在这条线之下 83 00:02:18,120 --> 00:02:19,140 已经很难存活了 84 00:02:19,140 --> 00:02:20,100 就跟超市里东西似的 85 00:02:20,100 --> 00:02:20,480 你看这个 86 00:02:20,480 --> 00:02:21,600 有的超市里东西 87 00:02:21,600 --> 00:02:22,700 就是走得特别快 88 00:02:22,700 --> 00:02:23,800 有的就走不动 89 00:02:23,800 --> 00:02:24,880 甚至说周转率 90 00:02:24,880 --> 00:02:26,280 达到几周几个月 91 00:02:26,280 --> 00:02:26,800 这种情况 92 00:02:26,800 --> 00:02:29,020 那些就是展示下下的商品 93 00:02:29,020 --> 00:02:29,460 对吧 94 00:02:29,460 --> 00:02:31,360 但是现在DeepSafe给了这么一个定价 95 00:02:31,360 --> 00:02:33,040 就把这个锚点标在这了 96 00:02:33,040 --> 00:02:33,540 你怎么样 97 00:02:33,540 --> 00:02:36,420 就跟这个当时吕布那个原文设计一样 98 00:02:36,420 --> 00:02:37,500 我这一键设在这了 99 00:02:37,500 --> 00:02:39,040 大家震慑在我的威力之下 100 00:02:39,040 --> 00:02:40,540 你想拼也不是不能拼 101 00:02:40,540 --> 00:02:42,460 你非要说俩换一个也可以换 102 00:02:42,460 --> 00:02:43,840 但是很多人会先掂点点亮 103 00:02:43,840 --> 00:02:44,900 我要不要干这事 104 00:02:44,900 --> 00:02:47,300 这是对市场的一个震撼 105 00:02:47,300 --> 00:02:47,500 真的 106 00:02:47,500 --> 00:02:48,640 就这一键设出去之后 107 00:02:48,640 --> 00:02:50,740 我觉得很多人都得再重新掂点亮 108 00:02:50,740 --> 00:02:52,520 回去想想这公司要怎么搞了 109 00:02:52,520 --> 00:02:56,020 但是回到比较实际的体验 110 00:02:56,020 --> 00:02:57,040 就问一下老修 111 00:02:57,040 --> 00:03:00,740 就实际在用Deep Seek V4 Fresh的体验上 112 00:03:00,740 --> 00:03:03,660 你觉得取代你大部分的模型吗 113 00:03:03,660 --> 00:03:05,140 或者说你觉得体验上怎么样 114 00:03:05,140 --> 00:03:08,540 体验上其实就是我刚才说的那种重复性的 115 00:03:08,540 --> 00:03:09,040 它很好 116 00:03:09,040 --> 00:03:11,180 但是需要推理的 117 00:03:11,180 --> 00:03:14,280 然后感知一些实时结果 118 00:03:14,280 --> 00:03:16,440 并且调整自己的方案的这种的 119 00:03:16,440 --> 00:03:17,560 这种其实还难 120 00:03:17,560 --> 00:03:20,280 这个其实是要留给它的V4 Pro正式版 121 00:03:20,280 --> 00:03:22,320 或者说现在这些场景 122 00:03:22,320 --> 00:03:24,040 还是需要顶级模型搞一把 123 00:03:24,040 --> 00:03:25,940 但是一些快修快补 124 00:03:25,940 --> 00:03:27,600 一些重复性工作 125 00:03:27,600 --> 00:03:29,100 快速的工作 126 00:03:29,100 --> 00:03:30,080 它就速度非常快 127 00:03:30,080 --> 00:03:30,740 做的还不错 128 00:03:30,740 --> 00:03:33,120 所以我是真的真心感受到 129 00:03:33,120 --> 00:03:36,780 日常的活真的是列入医保了一样 130 00:03:36,780 --> 00:03:37,920 你随便用没事 131 00:03:37,920 --> 00:03:39,240 一难杂认 132 00:03:39,240 --> 00:03:41,940 我们再去找海外这几个模型搞一下 133 00:03:41,940 --> 00:03:43,600 所以你看 134 00:03:43,600 --> 00:03:46,620 就变成说跟中国电车一样 135 00:03:46,620 --> 00:03:48,680 我20万块钱 136 00:03:48,680 --> 00:03:50,540 我可以买到非常好的电车 137 00:03:50,540 --> 00:03:51,740 过往三四十万 138 00:03:51,740 --> 00:03:53,580 四五十万这样车的 139 00:03:53,580 --> 00:03:54,840 这种豪华车的水平 140 00:03:54,840 --> 00:03:55,160 是的 141 00:03:55,160 --> 00:03:56,920 然后你就拍现在中国市场上 142 00:03:56,920 --> 00:03:58,740 这四五十万的车绝技 143 00:03:58,740 --> 00:03:59,480 不行了 144 00:03:59,480 --> 00:04:01,020 BBA都卖不出去了 145 00:04:01,020 --> 00:04:02,460 现在都销量掉百分之三十 146 00:04:02,460 --> 00:04:03,280 最多卖到三十万 147 00:04:03,280 --> 00:04:04,820 超过三十万的车 148 00:04:04,820 --> 00:04:06,000 不太行了 149 00:04:06,000 --> 00:04:07,480 很多都下探到二十几了 150 00:04:07,480 --> 00:04:08,400 BBA都不行了 151 00:04:08,400 --> 00:04:09,140 对吧 152 00:04:09,140 --> 00:04:10,060 连二手车都拍不通了 153 00:04:10,060 --> 00:04:12,100 我觉得这个就是一个很典型的斩杀件 154 00:04:12,100 --> 00:04:14,220 你用中国电车现在市场上来理解 155 00:04:14,220 --> 00:04:16,100 然后要么就最高端的 156 00:04:16,100 --> 00:04:18,060 买的最高端的保时捷 157 00:04:18,060 --> 00:04:19,740 蓝博基尼那些东西 158 00:04:19,740 --> 00:04:21,700 我跟你讲 159 00:04:21,700 --> 00:04:22,700 我跟你讲 160 00:04:22,700 --> 00:04:24,500 我给你一个我那个年代的比喻 161 00:04:24,500 --> 00:04:25,460 就两辆车 162 00:04:25,460 --> 00:04:27,500 一辆车是奇瑞QQ 163 00:04:27,500 --> 00:04:28,820 然后0.8的排量 164 00:04:28,820 --> 00:04:29,300 你知道吗 165 00:04:29,300 --> 00:04:29,860 有个坡 166 00:04:29,860 --> 00:04:30,820 就得下来推那种 167 00:04:30,820 --> 00:04:33,580 然后另外一个就50红光 168 00:04:33,580 --> 00:04:34,720 就50红光 169 00:04:34,720 --> 00:04:35,520 对吧 170 00:04:35,520 --> 00:04:37,500 就是拉人拉货都一样 171 00:04:37,500 --> 00:04:38,040 然后呢 172 00:04:38,040 --> 00:04:38,940 就是那个 173 00:04:38,940 --> 00:04:39,800 你就是说 174 00:04:39,800 --> 00:04:42,000 放里面放个沙发就能坐五个人 175 00:04:42,000 --> 00:04:42,880 但是随便那个 176 00:04:42,880 --> 00:04:44,440 弄个什么东西能拉走那种 177 00:04:44,440 --> 00:04:45,900 就是这两辆神车 178 00:04:45,900 --> 00:04:47,360 一个神车是什么意思 179 00:04:47,360 --> 00:04:49,000 就是说很多小老板 180 00:04:49,000 --> 00:04:50,120 就是刚刚那个什么 181 00:04:50,120 --> 00:04:51,360 就是谈个业务什么的 182 00:04:51,360 --> 00:04:52,360 你有个车代步 183 00:04:52,360 --> 00:04:53,520 我也有个QQ 184 00:04:53,520 --> 00:04:54,640 但是你要想谈女朋友 185 00:04:54,640 --> 00:04:55,420 很寒酸 186 00:04:55,420 --> 00:04:56,240 你知道我有一个朋友 187 00:04:56,240 --> 00:04:58,060 就是当年拉着女朋友去东风 188 00:04:58,060 --> 00:04:59,120 又到一坡上 189 00:04:59,120 --> 00:04:59,760 你知道吗 190 00:04:59,760 --> 00:05:00,900 起步上不去了 191 00:05:00,900 --> 00:05:01,320 你知道吗 192 00:05:01,320 --> 00:05:02,500 挂一档都上不去 193 00:05:02,500 --> 00:05:03,200 后来怎么办 194 00:05:03,200 --> 00:05:04,500 真的是女朋友下车 195 00:05:04,500 --> 00:05:05,960 然后帮着她推上去的那车 196 00:05:05,960 --> 00:05:07,660 当时她只给我们讲的时候 197 00:05:07,660 --> 00:05:09,240 就为这个事哭了好久 198 00:05:09,240 --> 00:05:12,360 但是就这辆车你知道吗 199 00:05:12,360 --> 00:05:13,380 有了这款车之后 200 00:05:13,380 --> 00:05:15,540 很多人就能摸车了 201 00:05:15,540 --> 00:05:15,880 你知道吧 202 00:05:15,880 --> 00:05:17,620 应该是说干活拉货 203 00:05:17,620 --> 00:05:19,820 有了兜底保障 204 00:05:19,820 --> 00:05:21,360 但是你出去接客 205 00:05:21,360 --> 00:05:21,700 对吧 206 00:05:21,700 --> 00:05:22,900 你还是开个麦巴赫 207 00:05:22,900 --> 00:05:24,080 你要搞个商务 208 00:05:24,080 --> 00:05:25,480 比如说你上机场接个人 209 00:05:25,480 --> 00:05:26,500 你开个50红光去 210 00:05:26,500 --> 00:05:28,300 我觉得有点不太合适 211 00:05:28,300 --> 00:05:30,060 你租一辆麦巴赫 212 00:05:30,060 --> 00:05:30,700 对不对 213 00:05:30,700 --> 00:05:33,320 讲了很多比喻都有 214 00:05:33,320 --> 00:05:33,740 对吧 215 00:05:33,740 --> 00:05:34,680 我还想了一个 216 00:05:34,680 --> 00:05:36,920 我说小米出来把山寨机干没了 217 00:05:36,920 --> 00:05:37,260 对不对 218 00:05:37,260 --> 00:05:39,240 对对对 219 00:05:39,240 --> 00:05:42,200 但是摩托的事又没山寨的 220 00:05:42,200 --> 00:05:43,160 是吧 221 00:05:43,160 --> 00:05:45,240 我觉得这个真的 222 00:05:45,240 --> 00:05:46,740 就为什么我开篇说 223 00:05:46,740 --> 00:05:48,420 这行业拐点绝对没有问题 224 00:05:48,420 --> 00:05:50,280 但是你单看Deepseek V4 225 00:05:50,280 --> 00:05:51,360 绝对不是行业拐点 226 00:05:51,360 --> 00:05:53,620 你应该结合 227 00:05:53,620 --> 00:05:55,820 Google的人事震动来看 228 00:05:55,820 --> 00:05:56,660 我们伸个围 229 00:05:56,660 --> 00:05:58,920 为什么要把这两个事情 230 00:05:58,920 --> 00:05:59,500 结合起来 231 00:05:59,500 --> 00:06:01,480 我们人已经走了 232 00:06:01,480 --> 00:06:02,460 我们就先不说了 233 00:06:02,460 --> 00:06:02,760 好不好 234 00:06:02,760 --> 00:06:04,220 Jeff Dean一路平安 235 00:06:04,220 --> 00:06:08,380 对 236 00:06:08,380 --> 00:06:09,740 祝你发大财 237 00:06:09,740 --> 00:06:11,580 感谢你为AI 238 00:06:11,580 --> 00:06:13,160 做出的巨大的贡献 239 00:06:13,160 --> 00:06:17,720 永远AI史上有你一笔 240 00:06:17,720 --> 00:06:19,760 传奇我们就不讲了 241 00:06:19,760 --> 00:06:21,220 但是我觉得 242 00:06:21,220 --> 00:06:22,660 上期有个事情 243 00:06:22,660 --> 00:06:23,880 我没有跟大家聊透 244 00:06:23,880 --> 00:06:24,840 就是Gemini 245 00:06:24,840 --> 00:06:27,200 就你说Gemini拉胯 246 00:06:27,200 --> 00:06:27,680 对吧 247 00:06:27,680 --> 00:06:28,940 我说这次斩杀线 248 00:06:28,940 --> 00:06:30,400 我觉得应该差不多 249 00:06:30,400 --> 00:06:32,020 也把Gemini给斩掉了 250 00:06:32,020 --> 00:06:32,400 对 251 00:06:32,400 --> 00:06:33,420 还留了一点点 252 00:06:33,420 --> 00:06:33,900 还有一点 253 00:06:33,900 --> 00:06:35,940 就那个太监淹掉了 254 00:06:35,940 --> 00:06:37,060 还供起来 255 00:06:37,060 --> 00:06:38,680 还留了一点 256 00:06:38,680 --> 00:06:40,300 我们自己还在用Gemini 257 00:06:40,300 --> 00:06:41,480 还留了一点念想 258 00:06:41,480 --> 00:06:41,740 还有 259 00:06:41,740 --> 00:06:42,040 还有 260 00:06:42,040 --> 00:06:42,400 还有 261 00:06:42,400 --> 00:06:42,680 还有留点 262 00:06:42,680 --> 00:06:43,780 留点一点念想 263 00:06:43,780 --> 00:06:44,300 对 264 00:06:44,300 --> 00:06:46,360 但是我要说的是什么 265 00:06:46,360 --> 00:06:48,200 大家现在在笑Gemini 266 00:06:48,200 --> 00:06:48,580 对吧 267 00:06:48,580 --> 00:06:49,820 说大家都 268 00:06:49,820 --> 00:06:54,700 Gemline 2.5 之 然后现在3.6 269 00:06:54,700 --> 00:06:56,860 大家笑他 270 00:06:56,860 --> 00:06:58,840 换了二二子海沃的兄弟 271 00:06:58,840 --> 00:06:59,580 现在变成 272 00:06:59,580 --> 00:07:01,040 能不能好 273 00:07:01,040 --> 00:07:03,700 但是我跟你讲 274 00:07:03,700 --> 00:07:04,640 我不开玩笑的 275 00:07:04,640 --> 00:07:06,860 Google的No Map 276 00:07:06,860 --> 00:07:08,200 Google路线 277 00:07:08,200 --> 00:07:10,360 真的我们大家要好好学习一下 278 00:07:10,360 --> 00:07:12,240 现在所有人是 279 00:07:12,240 --> 00:07:13,900 我猜跟这个事情有关 280 00:07:13,900 --> 00:07:15,400 因为现在Google全量推的 281 00:07:15,400 --> 00:07:17,600 都是3.6 Flash的模型 282 00:07:17,600 --> 00:07:18,000 对吧 283 00:07:18,000 --> 00:07:20,040 你包括Gymline Spark 284 00:07:20,040 --> 00:07:21,080 这些完成任务 285 00:07:21,080 --> 00:07:23,400 全部默认都是跑3.6 Flash 286 00:07:23,400 --> 00:07:25,280 所以我是觉得说 287 00:07:25,280 --> 00:07:26,340 这代表一个趋势 288 00:07:26,340 --> 00:07:28,700 最新的Flash模型 289 00:07:28,700 --> 00:07:30,580 蒸馏自己上一代的Pro模型 290 00:07:30,580 --> 00:07:31,880 我觉得这个路线 291 00:07:31,880 --> 00:07:32,960 它贯彻得非常好 292 00:07:32,960 --> 00:07:33,760 所以我觉得 293 00:07:33,760 --> 00:07:35,480 我们大方向看Google 294 00:07:35,480 --> 00:07:37,160 包括它所有这些布局 295 00:07:37,160 --> 00:07:38,320 什么做机器人 296 00:07:52,731 --> 00:07:54,951 它的云增长足够了 297 00:07:54,951 --> 00:07:56,911 我们这前一段时间刚交了打 298 00:07:56,911 --> 00:07:58,731 它的云增长非常的高 299 00:07:58,731 --> 00:08:03,671 而且其实蒸留自己前一代的旗舰模型 300 00:08:03,671 --> 00:08:05,131 这个也是业内标配 301 00:08:05,131 --> 00:08:07,031 因为实际上这样的做 302 00:08:07,031 --> 00:08:08,831 通常是不需要那么久的 303 00:08:08,831 --> 00:08:11,411 比如说Deep Seek这一次V4 Flash 304 00:08:11,411 --> 00:08:12,991 为什么这么强 305 00:08:12,991 --> 00:08:17,071 它其实就是把原来的Pro的V4 Pro Preview 306 00:08:17,071 --> 00:08:18,011 就是预览版 307 00:08:18,011 --> 00:08:21,671 把预览版里面那些好的一些场景 308 00:08:21,671 --> 00:08:23,091 全部都蒸馏到非常的 309 00:08:23,091 --> 00:08:25,351 然后后训练在调整 310 00:08:25,351 --> 00:08:25,951 靠后训练 311 00:08:25,951 --> 00:08:27,891 所有的核心参数都没变 312 00:08:27,891 --> 00:08:28,191 是吧 313 00:08:28,191 --> 00:08:29,451 核心参数都没有变 314 00:08:29,451 --> 00:08:30,291 全靠后训练 315 00:08:30,291 --> 00:08:31,671 智能水平上升了 316 00:08:31,671 --> 00:08:32,231 是的 317 00:08:32,231 --> 00:08:33,291 成本还掉了 318 00:08:33,291 --> 00:08:34,291 对成本下了 319 00:08:34,291 --> 00:08:38,151 它其实这个就是把之前的旗舰拿来干的 320 00:08:38,151 --> 00:08:40,691 所以这个事情就是说挺标准的 321 00:08:40,691 --> 00:08:41,791 而且它也没有等很长时间 322 00:08:41,791 --> 00:08:43,811 我们从preview应该是几个月前才讲的 323 00:08:43,811 --> 00:08:44,771 三到四个月 324 00:08:44,771 --> 00:08:45,651 三个月左右 325 00:08:45,651 --> 00:08:47,771 那个时候到现在 326 00:08:47,771 --> 00:08:50,151 就已经flash版本就出来了 327 00:08:50,151 --> 00:08:51,591 而且做的效果还挺好 328 00:08:51,591 --> 00:08:53,351 软件硬件都优化款了 329 00:08:53,351 --> 00:08:54,831 我相信谷歌也能做那么快 330 00:08:54,831 --> 00:08:55,851 但是他可能节奏上 331 00:08:55,851 --> 00:08:56,771 不需要跑那么快 332 00:08:56,771 --> 00:08:57,371 对 333 00:08:57,371 --> 00:08:59,811 但是这个就会出现一个问题 334 00:08:59,811 --> 00:09:01,691 你是什么Transformer 335 00:09:01,691 --> 00:09:02,851 什么MOE 336 00:09:02,851 --> 00:09:04,051 什么什么所有这些东西 337 00:09:04,051 --> 00:09:05,211 都是Google提出来的 338 00:09:05,211 --> 00:09:05,691 对吧 339 00:09:05,691 --> 00:09:06,871 至少什么Jeff Dean的 340 00:09:06,871 --> 00:09:07,831 这些人提出来的 341 00:09:07,831 --> 00:09:09,831 但是这个时候他就会想了 342 00:09:09,831 --> 00:09:11,131 那我要干嘛 343 00:09:11,131 --> 00:09:11,571 对吧 344 00:09:11,571 --> 00:09:13,751 你们现在开始拼Flash了 345 00:09:13,751 --> 00:09:15,271 开始卷价格了 346 00:09:15,271 --> 00:09:17,411 开始进入重复工作了 347 00:09:17,411 --> 00:09:17,731 对吧 348 00:09:17,731 --> 00:09:18,891 我干嘛 349 00:09:18,891 --> 00:09:19,851 我干嘛去 350 00:09:19,851 --> 00:09:20,591 对不对 351 00:09:20,591 --> 00:09:25,191 你看我留意到哈萨比斯最近的一个采访 352 00:09:25,191 --> 00:09:28,391 我觉得有一定程度上是真情流露了 353 00:09:28,391 --> 00:09:28,711 对不对 354 00:09:28,711 --> 00:09:32,131 他就说如果能够由我来拍板的话 355 00:09:32,131 --> 00:09:35,811 我愿意多做几个像AlphaFold这样的产品 356 00:09:35,811 --> 00:09:37,111 待在实验室里面 357 00:09:37,111 --> 00:09:39,431 可能就先把癌症给攻克了 358 00:09:39,431 --> 00:09:40,151 对吧 359 00:09:40,151 --> 00:09:42,911 我不会那么快让他让这个AI出来 360 00:09:42,911 --> 00:09:43,971 给你们写邮件 361 00:09:43,971 --> 00:09:45,311 干办公 362 00:09:45,311 --> 00:09:46,471 是吧 363 00:09:46,471 --> 00:09:48,771 在科学家眼睛里面 364 00:09:48,771 --> 00:09:50,331 这些是还好 365 00:09:50,331 --> 00:09:52,871 但是更急的是那些生命科学 366 00:09:52,871 --> 00:09:53,851 其他的一些问题 367 00:09:53,851 --> 00:09:55,511 所以你抓住这句话 368 00:09:55,511 --> 00:09:58,811 你就大概能够明白 369 00:09:58,811 --> 00:10:00,431 整个Google里面的斗争了 370 00:10:00,431 --> 00:10:01,211 进一段广告 371 00:10:01,211 --> 00:10:04,031 最近早晚报也做了非常多的升级 372 00:10:04,031 --> 00:10:05,051 我们已经发了多少期 373 00:10:05,051 --> 00:10:06,151 500多期了 374 00:10:06,151 --> 00:10:06,671 对 375 00:10:06,671 --> 00:10:09,511 然后深度解读可能都干了100多期了 376 00:10:09,511 --> 00:10:10,191 对 377 00:10:10,191 --> 00:10:12,831 非常感谢大家 378 00:10:12,831 --> 00:10:15,531 对于我们早晚报邮件产品的支持 379 00:10:15,531 --> 00:10:18,991 我们现在也是推出了IM端 380 00:10:18,991 --> 00:10:21,771 那在IM端就更方便大家去使用了 381 00:10:21,771 --> 00:10:22,671 首先第一个 382 00:10:22,671 --> 00:10:25,231 你每天可以收到两条语音的播报 383 00:10:25,231 --> 00:10:25,631 对吧 384 00:10:25,631 --> 00:10:27,651 了解今天的早报跟晚报 385 00:10:27,651 --> 00:10:30,151 然后每次你还会收到三条 386 00:10:30,151 --> 00:10:32,011 就是今天的深度言报 387 00:10:32,011 --> 00:10:33,751 也是用语音播客的形式 388 00:10:33,751 --> 00:10:35,551 跟大家来做一个互动 389 00:10:35,551 --> 00:10:37,431 那关于里面你感兴趣的 390 00:10:37,431 --> 00:10:38,091 不懂的环节 391 00:10:38,091 --> 00:10:39,691 你可以直接在你的IM端 392 00:10:39,691 --> 00:10:40,451 跟他聊天了 393 00:10:40,451 --> 00:10:40,851 对 394 00:10:40,851 --> 00:10:42,451 就可以随时随问了 395 00:10:42,451 --> 00:10:43,811 随时这个做笔记 396 00:10:43,811 --> 00:10:45,431 或者帮自己整理言报 397 00:10:45,431 --> 00:10:47,571 那我们过往500多期的 398 00:10:47,571 --> 00:10:49,251 这个新闻的数据 399 00:10:49,251 --> 00:10:51,531 已经是一个非常详实的 400 00:10:51,531 --> 00:10:52,951 一个AI新闻面的 401 00:10:52,951 --> 00:10:53,831 一个数据库了 402 00:10:53,831 --> 00:10:55,451 那加上我们调优的 403 00:10:55,451 --> 00:10:56,631 我们三个人的 404 00:10:56,631 --> 00:10:58,151 这个模型在里面 405 00:10:58,151 --> 00:11:00,231 我觉得可能很多时候 406 00:11:00,231 --> 00:11:02,031 可以一键帮你洞悉到 407 00:11:02,031 --> 00:11:03,231 就是你最关心的 408 00:11:03,231 --> 00:11:04,051 这个AI动态 409 00:11:04,051 --> 00:11:05,231 所以非常推荐 410 00:11:05,231 --> 00:11:07,391 大家去订阅我们的早晚报 411 00:11:07,391 --> 00:11:08,671 然后大家去绑定 412 00:11:08,671 --> 00:11:09,491 自己的IM端 413 00:11:09,491 --> 00:11:10,711 来使用我们一个 414 00:11:10,711 --> 00:11:11,511 可以交互的 415 00:11:11,511 --> 00:11:12,691 互动版的早晚报 416 00:11:12,691 --> 00:11:14,111 运过都说好 417 00:11:14,111 --> 00:11:15,291 也感谢一直跟随 418 00:11:15,291 --> 00:11:16,711 我们的这些VIP的客户们 419 00:11:16,711 --> 00:11:17,491 帮我们做测试 420 00:11:17,491 --> 00:11:18,191 给我们提意见 421 00:11:18,191 --> 00:11:19,331 让我们这产品就来 422 00:11:19,331 --> 00:11:21,871 先换上来土耳其人 423 00:11:21,871 --> 00:11:22,431 对吧 424 00:11:22,431 --> 00:11:24,371 还有负责Gemini 425 00:11:24,371 --> 00:11:25,371 Josh 426 00:11:25,371 --> 00:11:29,091 这几个都是雕花可以的 427 00:11:29,091 --> 00:11:29,791 对不对 428 00:11:29,791 --> 00:11:31,691 开始围着创始人 429 00:11:31,691 --> 00:11:32,651 开始在功课 430 00:11:32,651 --> 00:11:34,571 说现在在功课编程了 431 00:11:34,571 --> 00:11:34,951 对吧 432 00:11:34,951 --> 00:11:36,331 那就是说明说 433 00:11:36,331 --> 00:11:37,391 在哈萨比斯 434 00:11:37,391 --> 00:11:38,671 Jeff Dean这些人的 435 00:11:38,671 --> 00:11:41,391 工作里面 436 00:11:41,391 --> 00:11:42,951 就这些事情优先级不高 437 00:11:42,951 --> 00:11:45,131 不太想做 438 00:11:45,131 --> 00:11:46,551 对 439 00:11:46,551 --> 00:11:49,271 对 我跟你讲 你说不太想做才是真话 440 00:11:49,271 --> 00:11:52,491 就是说当才你说那句话 我也看到你转发的嘛 441 00:11:52,491 --> 00:11:54,551 就是哈斯比斯说 如果是他拍板 442 00:11:54,551 --> 00:11:55,911 然后你就会先把癌症功课 443 00:11:55,911 --> 00:11:57,831 这个事情真的是很崇高的理想 444 00:11:57,831 --> 00:12:00,691 但是在商业社会中 我觉得是没有人会允许他这么做的 445 00:12:00,691 --> 00:12:02,911 对吧 又不是他一家独揽 446 00:12:02,911 --> 00:12:05,491 把所有AI资源和算力都揽在自己手里面 447 00:12:05,491 --> 00:12:08,171 不是这样的 如果你执着于说一定要科学做贡献 448 00:12:08,171 --> 00:12:09,331 我觉得这是很崇高的理想 449 00:12:09,331 --> 00:12:11,631 但在对于一个公司来讲 对于股东来讲 450 00:12:11,631 --> 00:12:14,731 不见得这是最大化的一个效率选择 451 00:12:14,731 --> 00:12:15,991 所以我觉得他 452 00:12:15,991 --> 00:12:18,151 我觉得导演没那么崇高 453 00:12:18,151 --> 00:12:20,871 可能我觉得像哈萨比斯这个level 454 00:12:20,871 --> 00:12:22,651 他也知道这个生意要怎么做 455 00:12:22,651 --> 00:12:24,351 但他可能觉得说 456 00:12:24,351 --> 00:12:25,731 多几个alpha fold 457 00:12:25,731 --> 00:12:27,291 可能是一个更大的生意 458 00:12:27,291 --> 00:12:28,391 他可能看得更长远 459 00:12:28,391 --> 00:12:31,511 但是可能作为我创始人来讲 460 00:12:31,511 --> 00:12:32,371 作为Google来讲 461 00:12:32,371 --> 00:12:33,511 我眼前更重要 462 00:12:33,511 --> 00:12:35,471 我下一季的财报怎么教 463 00:12:35,471 --> 00:12:37,211 我Google语言故事怎么讲 464 00:12:37,211 --> 00:12:37,731 对吧 465 00:12:37,731 --> 00:12:39,231 其实你看这一次 466 00:12:39,231 --> 00:12:41,231 Google这个财报出来 467 00:12:41,231 --> 00:12:42,631 云啊增长啊 468 00:12:42,631 --> 00:12:43,491 什么都特别明显 469 00:12:43,491 --> 00:12:45,491 去年我们就讲过了 470 00:12:45,491 --> 00:12:48,111 这个云肯定是要成为第二曲线的 471 00:12:48,111 --> 00:12:49,431 你云不成为第二曲线 472 00:12:49,431 --> 00:12:51,331 你这搜索份额掉 473 00:12:51,331 --> 00:12:52,471 这事情你源不过去 474 00:12:52,471 --> 00:12:53,151 对不对 475 00:12:53,151 --> 00:12:53,471 没错 476 00:12:53,471 --> 00:12:54,931 那我云成为第二曲线 477 00:12:54,931 --> 00:12:55,511 很简单 478 00:12:55,511 --> 00:12:57,931 我这个算力怎么分配 479 00:12:57,931 --> 00:12:58,151 来 480 00:12:58,151 --> 00:12:59,971 这个哪些给Anthoropic 481 00:12:59,971 --> 00:13:00,551 对吧 482 00:13:00,551 --> 00:13:02,851 这个哪些给Gemini 483 00:13:02,851 --> 00:13:03,811 对不对 484 00:13:03,811 --> 00:13:05,491 我这个做预算 485 00:13:05,491 --> 00:13:07,931 这个批财开预算会的时候 486 00:13:07,931 --> 00:13:09,771 大家来开个预算会 487 00:13:09,771 --> 00:13:11,331 这个B就开始撕了 488 00:13:11,331 --> 00:13:13,211 对 489 00:13:13,211 --> 00:13:15,871 所以我会想到一个八卦的事情 490 00:13:15,871 --> 00:13:16,671 就有点意思 491 00:13:16,671 --> 00:13:17,831 突然窜起来 492 00:13:17,831 --> 00:13:22,011 本来跟着哈萨比斯在研究蛋白质的诺贝尔奖获得者 493 00:13:22,011 --> 00:13:23,051 应该是上个月 494 00:13:23,051 --> 00:13:24,171 跳起anthropic了 495 00:13:24,171 --> 00:13:28,931 然后而且哈萨比斯是anthropic的早期投资者 496 00:13:28,931 --> 00:13:32,131 不违反各种进业条例的 497 00:13:32,131 --> 00:13:33,411 是应该在很小的时候 498 00:13:33,411 --> 00:13:35,211 anthropic很小的时候投的 499 00:13:35,211 --> 00:13:39,231 然后他现在其实现任CEO有事去做chairman 500 00:13:39,231 --> 00:13:40,311 就是做chairman 501 00:13:40,311 --> 00:13:42,091 chairman应该是董事长是吧 502 00:13:42,091 --> 00:13:45,291 我不知道海外到底是哪个位置 503 00:13:45,291 --> 00:13:45,871 主席 504 00:13:45,871 --> 00:13:47,371 还不知道是什么 505 00:13:47,371 --> 00:13:47,951 话事人 506 00:13:47,951 --> 00:13:49,231 然后他过去以后 507 00:13:49,231 --> 00:13:51,311 他是说他重要在看的 508 00:13:51,311 --> 00:13:55,671 也是关于健康蛋白质相关的项目去了 509 00:13:55,671 --> 00:13:57,371 我就觉得说 510 00:13:57,371 --> 00:14:00,771 其实可能这个事情确实他是很想做 511 00:14:00,771 --> 00:14:02,351 我自己也看过他的自传 512 00:14:02,351 --> 00:14:04,011 他的自传其实写过很多次 513 00:14:04,011 --> 00:14:07,531 他是很想追求科学家是首要身份的 514 00:14:07,531 --> 00:14:10,131 但是应该来说这两年来搞Gemini 515 00:14:10,131 --> 00:14:11,591 肯定没办法做科学家了 516 00:14:11,591 --> 00:14:12,891 有点郁郁不得志 517 00:14:12,891 --> 00:14:14,791 我教大家一个方法 518 00:14:14,791 --> 00:14:16,191 就是你在GPT上面 519 00:14:16,191 --> 00:14:18,491 你跟哈萨比斯换个脑 520 00:14:18,491 --> 00:14:22,131 你用他的角度来想一下这个问题 521 00:14:22,131 --> 00:14:23,211 你就会发现 522 00:14:23,211 --> 00:14:25,051 刚才老书补充了一个很重要的信息 523 00:14:25,051 --> 00:14:26,991 妈的你AI coding什么 524 00:14:26,991 --> 00:14:27,791 我都投了 525 00:14:27,791 --> 00:14:29,111 对吧 526 00:14:29,111 --> 00:14:29,791 我都投了 527 00:14:29,791 --> 00:14:31,051 你Google也投了 528 00:14:31,051 --> 00:14:31,871 你不要说我投了 529 00:14:31,871 --> 00:14:32,711 Google也投了 530 00:14:32,711 --> 00:14:33,711 对谷歌也投了 531 00:14:33,711 --> 00:14:35,731 而且有好多的TPU芯片 532 00:14:35,731 --> 00:14:36,851 全部都先给Iso 533 00:14:36,851 --> 00:14:37,931 给去训练下一代的 534 00:14:37,931 --> 00:14:39,791 可能是Cloud6什么的 535 00:14:39,791 --> 00:14:40,871 我搞个毛线coding 536 00:14:40,871 --> 00:14:42,631 我肯定对吧 537 00:14:42,631 --> 00:14:44,751 Fallet多搞几个 538 00:14:44,751 --> 00:14:47,351 你看我们都知道说 539 00:14:47,351 --> 00:14:48,611 最高级的用Cloud 540 00:14:48,611 --> 00:14:50,731 其他干活的用Deepseek 541 00:14:50,731 --> 00:14:52,731 他们就把Gemini做成Deepseek的 542 00:14:52,731 --> 00:14:54,411 美国版平地都够了 543 00:14:54,411 --> 00:14:54,771 对吧 544 00:14:54,771 --> 00:14:57,231 其实Google就是在做这个动作 545 00:14:57,231 --> 00:14:58,611 对因为你在美国 546 00:14:58,611 --> 00:15:00,371 不太可能说大家都用Deepseek 547 00:15:00,371 --> 00:15:03,531 这肯定就有很多天然的问题 548 00:15:03,531 --> 00:15:05,451 就算都是部署在本地的 549 00:15:05,451 --> 00:15:08,151 他也很想用个美国自己的模型 550 00:15:08,151 --> 00:15:10,311 如果Gemini Flash的后面 551 00:15:10,311 --> 00:15:11,371 跟Deep Seek Flash 552 00:15:11,371 --> 00:15:13,311 同一个来我价格稍微贵一点 553 00:15:13,311 --> 00:15:14,971 肯定是选Jim来就这样 554 00:15:14,971 --> 00:15:16,611 小苏当时那个Flash那块 555 00:15:16,611 --> 00:15:17,751 我特别特别同意 556 00:15:17,751 --> 00:15:19,351 我觉得这个战略本身啊 557 00:15:19,351 --> 00:15:20,051 并没有什么错误 558 00:15:20,051 --> 00:15:20,851 就是战略性放弃 559 00:15:20,851 --> 00:15:23,171 就跟咱们看到一些这个拳击片 560 00:15:23,171 --> 00:15:23,831 看到嘛 561 00:15:23,831 --> 00:15:25,031 就什么Rocky什么的 562 00:15:25,031 --> 00:15:26,591 就作为那个一方来讲 563 00:15:26,591 --> 00:15:27,431 他就是先守着 564 00:15:27,431 --> 00:15:28,811 先防守防守 565 00:15:28,811 --> 00:15:29,631 我能扛得住 566 00:15:29,631 --> 00:15:30,831 能扛的几个回合 567 00:15:30,831 --> 00:15:31,431 七八回合 568 00:15:31,431 --> 00:15:32,971 就是你看我后面怎么样 569 00:15:32,971 --> 00:15:34,611 你只要没动了我的根本 570 00:15:34,611 --> 00:15:35,871 我后面几个走公桥 571 00:15:35,871 --> 00:15:36,311 右勾拳 572 00:15:36,311 --> 00:15:37,931 一个那什么右勾饼 573 00:15:37,931 --> 00:15:38,931 直接就能把那对 574 00:15:39,491 --> 00:15:53,251 我觉得就是说,Google作为老牌公司,上市公司,而且有这么好的Google基础,它不会像那些年轻人雪橘方刚一样,然后从天到晚的就是叫着自己的评分多高,然后能把价格打多低,我觉得它先稳住,让自己的生态能够运转起来。 575 00:15:53,251 --> 00:15:54,531 这是核心 576 00:15:54,531 --> 00:15:55,811 其实我告诉你 577 00:15:55,811 --> 00:15:57,871 我把这个事情拆开来 578 00:15:57,871 --> 00:15:58,511 就一句话 579 00:15:58,511 --> 00:15:59,831 就是Google的战略 580 00:15:59,831 --> 00:16:01,671 在年初就已经定好 581 00:16:01,671 --> 00:16:02,811 就是Deep Seek的话 582 00:16:02,811 --> 00:16:04,971 做美国的Deep Seek 583 00:16:04,971 --> 00:16:06,051 它的战略就是这个 584 00:16:06,051 --> 00:16:06,831 对 585 00:16:06,831 --> 00:16:09,351 但一旦这个战略定下来以后 586 00:16:09,351 --> 00:16:11,491 一旦这个战略定下来以后 587 00:16:11,491 --> 00:16:12,831 就要有所牺牲 588 00:16:12,831 --> 00:16:13,271 对吧 589 00:16:13,271 --> 00:16:13,711 老兄 590 00:16:13,711 --> 00:16:15,171 你反对革命是吧 591 00:16:15,171 --> 00:16:17,011 你反对革命是吧 592 00:16:17,011 --> 00:16:18,371 请坐后面的椅子 593 00:16:18,371 --> 00:16:19,331 去Chill坐一下 594 00:16:19,331 --> 00:16:21,691 不换脑子 595 00:16:21,691 --> 00:16:23,231 不换脑子就换脑子 596 00:16:23,231 --> 00:16:24,511 是不是 597 00:16:24,511 --> 00:16:28,311 我觉得Google的战略非常简单 598 00:16:28,311 --> 00:16:31,191 就是我DeepMind的使命完成了 599 00:16:31,191 --> 00:16:31,651 对 600 00:16:31,651 --> 00:16:34,111 我现在整个公司我要DeepSeek化 601 00:16:34,111 --> 00:16:35,951 你们不要跟我讲别的 602 00:16:35,951 --> 00:16:37,211 我就是要DeepSeek化 603 00:16:37,211 --> 00:16:40,331 然后如果你能帮我把编程搞下来 604 00:16:40,331 --> 00:16:41,051 OK可以 605 00:16:41,051 --> 00:16:43,651 如果你DeepMind不能帮我把编程搞下来 606 00:16:43,651 --> 00:16:43,911 再见 607 00:16:43,911 --> 00:16:46,291 你现在也可以倒过来看 608 00:16:46,291 --> 00:16:48,171 DeepSeek的编码的部分 609 00:16:48,171 --> 00:16:50,271 它现在也才刚刚开始做 610 00:16:50,271 --> 00:16:50,491 是 611 00:16:50,491 --> 00:16:55,031 所以我觉得这个是一个战略的考量 612 00:16:55,031 --> 00:16:57,191 然后当然我觉得 613 00:16:57,191 --> 00:16:58,931 我们刚才讲了一个 614 00:16:58,931 --> 00:16:59,771 反正斩杀线 615 00:16:59,771 --> 00:17:01,851 第一刀肯定是把meta给斩的 616 00:17:01,851 --> 00:17:03,951 你说这个事情对Google好还不好 617 00:17:03,951 --> 00:17:06,031 我的结论就是刚才也讲了 618 00:17:06,031 --> 00:17:08,311 Google就是把根留住 619 00:17:08,311 --> 00:17:10,111 被砍了一刀 620 00:17:10,111 --> 00:17:11,391 然后把根留住 621 00:17:11,391 --> 00:17:12,631 把那个供起来 622 00:17:12,631 --> 00:17:15,111 还能接上 623 00:17:15,111 --> 00:17:15,971 还能接上 624 00:17:15,971 --> 00:17:18,971 但是当时我想说一个话题 625 00:17:18,971 --> 00:17:19,671 我当时就是说 626 00:17:19,671 --> 00:17:21,191 我不想让它滑过去 627 00:17:21,191 --> 00:17:22,031 就当他说到 628 00:17:22,031 --> 00:17:22,951 说拼多多嘛 629 00:17:22,951 --> 00:17:23,271 对吧 630 00:17:23,271 --> 00:17:24,611 你说比如说国内的生产力 631 00:17:24,611 --> 00:17:25,391 加上价格 632 00:17:25,391 --> 00:17:26,931 那在美国卖两块钱东西 633 00:17:26,931 --> 00:17:27,991 在国内卖两块人民币 634 00:17:27,991 --> 00:17:29,871 这东西有七倍到八倍的差价 635 00:17:29,871 --> 00:17:31,771 但是其实真正到这边 636 00:17:31,771 --> 00:17:33,011 本地不是这样的 637 00:17:33,011 --> 00:17:34,051 就是我在本地的感受 638 00:17:34,051 --> 00:17:35,351 在多加拿大也有这个什么 639 00:17:35,351 --> 00:17:36,071 TMO什么的 640 00:17:36,071 --> 00:17:37,391 不行 641 00:17:37,391 --> 00:17:38,751 就是一直扩外不开 642 00:17:38,751 --> 00:17:39,651 有两个原因 643 00:17:39,651 --> 00:17:40,331 一个是物流 644 00:17:40,331 --> 00:17:41,331 一个是物流原因 645 00:17:41,331 --> 00:17:41,931 一个是关税 646 00:17:41,931 --> 00:17:44,991 但是这个东西在AI上不存在 647 00:17:44,991 --> 00:17:45,631 是不是不存在 648 00:17:45,631 --> 00:17:46,011 我不知道 649 00:17:46,011 --> 00:17:46,391 对不对 650 00:17:46,391 --> 00:17:47,851 就说起码我现在看见 651 00:17:47,851 --> 00:17:49,371 我去注册这个deep6 652 00:17:49,371 --> 00:17:50,291 API很顺 653 00:17:50,291 --> 00:17:51,711 然后再把DeepSick 654 00:17:51,711 --> 00:17:52,551 接入到Codex 655 00:17:52,551 --> 00:17:54,491 那个Codex里面也很顺 656 00:17:54,491 --> 00:17:55,811 那就不存在 657 00:17:55,811 --> 00:17:58,431 这个像国内商品物流过来 658 00:17:58,431 --> 00:17:59,891 这种国际贸易的这种差距 659 00:17:59,891 --> 00:18:01,011 那这就很可怕 660 00:18:01,011 --> 00:18:03,251 这真的我觉得这才是关键 661 00:18:03,251 --> 00:18:04,151 就是它怎么能够 662 00:18:04,151 --> 00:18:06,091 接入到现在原有的工具里面 663 00:18:06,091 --> 00:18:06,911 让大家用着顺手 664 00:18:06,911 --> 00:18:08,251 用着爱不释手 665 00:18:08,251 --> 00:18:10,611 这个是我觉得太牛逼的地方 666 00:18:10,611 --> 00:18:12,171 但是其实应该是说 667 00:18:12,171 --> 00:18:13,911 按目前这个成本结构 668 00:18:13,911 --> 00:18:15,711 肯定就是你要出海 669 00:18:15,711 --> 00:18:17,431 它还是有一定的成本的 670 00:18:17,431 --> 00:18:18,591 不管算力电力 671 00:18:18,591 --> 00:18:20,151 token这个成本肯定在 672 00:18:20,151 --> 00:18:22,611 但是现在的成本结构 673 00:18:22,611 --> 00:18:25,631 对于用美金技价的购买来讲 674 00:18:25,631 --> 00:18:27,211 已经足够便宜了 675 00:18:27,211 --> 00:18:29,671 而且他不还要马上做 676 00:18:29,671 --> 00:18:30,791 销锋评股 677 00:18:30,791 --> 00:18:31,271 对吧 678 00:18:31,271 --> 00:18:32,991 就是把播风播股价格 679 00:18:32,991 --> 00:18:34,831 对于我海外来讲太舒服了 680 00:18:34,831 --> 00:18:35,451 对不对 681 00:18:35,451 --> 00:18:36,751 我永远都在播股上用 682 00:18:36,751 --> 00:18:38,371 然后对他来讲 683 00:18:38,371 --> 00:18:39,331 他那些顺路易不浪费 684 00:18:39,331 --> 00:18:41,091 还能有一笔很好的收入 685 00:18:41,091 --> 00:18:43,511 对于海外来讲 686 00:18:43,511 --> 00:18:45,251 我认为这个就是说 687 00:18:45,251 --> 00:18:46,071 全是利好 688 00:18:46,071 --> 00:18:46,591 没有利空 689 00:18:46,591 --> 00:18:48,531 这我也让我觉得很恐怖 690 00:18:48,531 --> 00:18:51,131 这个是一个很有意思的事情 691 00:18:51,131 --> 00:18:51,491 老秀 692 00:18:51,491 --> 00:18:54,731 就是为什么三方平台能够把 693 00:18:54,731 --> 00:18:57,171 这个Deep Seek V4 Flash的价格 694 00:18:57,171 --> 00:18:58,791 卖的比官方还便宜六倍 695 00:18:58,791 --> 00:18:59,991 这个是 696 00:18:59,991 --> 00:19:02,011 我觉得这个也很好笑 697 00:19:02,011 --> 00:19:04,031 对这个就是表面看起来 698 00:19:04,031 --> 00:19:06,371 难道第三方有更深的技术吗 699 00:19:06,371 --> 00:19:08,071 其实就不是这么回事 700 00:19:08,071 --> 00:19:11,971 它就像羊毛出在猪身上一样 701 00:19:11,971 --> 00:19:13,751 你这个东西只要的架起来 702 00:19:13,751 --> 00:19:15,051 它有别的方式可以卖 703 00:19:15,051 --> 00:19:15,711 对吧 704 00:19:15,711 --> 00:19:17,911 可以卖你这个过往的中转的数据 705 00:19:17,911 --> 00:19:19,171 然后数据标注 706 00:19:19,171 --> 00:19:20,151 都很优质的 707 00:19:20,151 --> 00:19:21,731 这些东西也都可以做 708 00:19:21,731 --> 00:19:22,331 对 709 00:19:22,331 --> 00:19:25,511 所以我们之所以叫梁胜 710 00:19:25,511 --> 00:19:27,891 真的是梁文峰做的这个东西 711 00:19:27,891 --> 00:19:29,991 自己标备的是六倍定价 712 00:19:29,991 --> 00:19:32,091 实际上它是很克制的 713 00:19:32,091 --> 00:19:33,071 是真的很克制的 714 00:19:33,071 --> 00:19:35,191 其实应该还是有人去搜了一些 715 00:19:35,191 --> 00:19:36,331 闲散的算力来卖 716 00:19:36,331 --> 00:19:36,911 对吧 717 00:19:36,911 --> 00:19:37,471 对 718 00:19:37,471 --> 00:19:39,311 因为这里面实际上 719 00:19:39,311 --> 00:19:41,491 我觉得中间有一个差别 720 00:19:41,491 --> 00:19:43,711 是说你很多人其实用不上 721 00:19:43,711 --> 00:19:47,051 真正的企业级API的那种服务水准 722 00:19:47,051 --> 00:19:48,931 因为SLA是要求很高的 723 00:19:48,931 --> 00:19:50,831 我们说稳定性 724 00:19:50,831 --> 00:19:52,131 我可能就是好多个9 725 00:19:52,131 --> 00:19:52,411 对吧 726 00:19:52,411 --> 00:19:53,151 可能6个9 727 00:19:53,151 --> 00:19:54,831 但是我普通的伞的 728 00:19:54,831 --> 00:19:56,731 可能就90%也行 729 00:19:56,731 --> 00:19:58,411 有的时候有10%时间用不了 730 00:19:58,411 --> 00:19:58,931 也没关系 731 00:19:58,931 --> 00:20:01,591 我可以有很多其他伞的能补上 732 00:20:01,591 --> 00:20:02,371 差不多就行了 733 00:20:02,371 --> 00:20:03,011 对 734 00:20:03,011 --> 00:20:04,071 然后这里面的价格 735 00:20:04,071 --> 00:20:05,191 其实也给很多 736 00:20:05,191 --> 00:20:07,951 其实中间的做这些平台的也能赚 737 00:20:07,951 --> 00:20:08,651 也能赚 738 00:20:08,651 --> 00:20:08,811 对 739 00:20:08,811 --> 00:20:10,711 这些都是一开始都会想不到 740 00:20:10,711 --> 00:20:11,831 都觉得脑子都想不到 741 00:20:11,831 --> 00:20:14,111 因为甚至连一些挖比特币的 742 00:20:14,111 --> 00:20:14,651 干嘛的 743 00:20:14,651 --> 00:20:15,971 原来的现在都能够 744 00:20:15,971 --> 00:20:19,771 其实我比较佩服梁文峰的两点 745 00:20:19,771 --> 00:20:20,751 一点就是说 746 00:20:20,751 --> 00:20:22,751 你看他又重新启动融资了 747 00:20:22,751 --> 00:20:24,391 跟咱上期判断是一样的 748 00:20:24,391 --> 00:20:24,751 对吧 749 00:20:24,751 --> 00:20:26,351 就这回我看他新闻说 750 00:20:26,351 --> 00:20:27,371 中间停了一段时间 751 00:20:27,371 --> 00:20:28,331 现在又重新启动了 752 00:20:28,331 --> 00:20:30,451 一堆人都在那什么 753 00:20:30,451 --> 00:20:31,151 恶手枪性 754 00:20:31,151 --> 00:20:32,651 就是说这回低调低调 755 00:20:32,651 --> 00:20:33,651 一定要投进去 756 00:20:33,651 --> 00:20:34,631 这回又是500亿 757 00:20:34,631 --> 00:20:36,251 加一起的话1000亿 758 00:20:36,251 --> 00:20:37,311 这个大手比 759 00:20:37,311 --> 00:20:39,431 就比有些上市公司的估值还高了 760 00:20:39,431 --> 00:20:40,911 就是AI相关的上市公司 761 00:20:40,911 --> 00:20:41,491 估值还高了 762 00:20:41,491 --> 00:20:42,991 这样一下但愿就有了 763 00:20:42,991 --> 00:20:43,691 更重要的是说 764 00:20:43,691 --> 00:20:45,031 团队的信心就有了 765 00:20:45,031 --> 00:20:47,071 我觉得这点上特别让我配 766 00:20:47,071 --> 00:20:48,991 问题是他做现金流的能力 767 00:20:48,991 --> 00:20:50,151 也已经被证明了 768 00:20:50,151 --> 00:20:54,151 上次我们被全网下架的视频 769 00:20:54,151 --> 00:20:56,011 别急了 770 00:20:56,011 --> 00:20:57,631 这期还是我们得意之作 771 00:20:57,631 --> 00:21:00,751 很多网友说 772 00:21:00,751 --> 00:21:02,631 这个是拆解DeepSeek 773 00:21:02,631 --> 00:21:05,531 跟梁文峰最佳的分析之一了 774 00:21:05,531 --> 00:21:07,611 大家说为啥只能看片段 775 00:21:07,611 --> 00:21:08,631 我们也没办法 776 00:21:08,631 --> 00:21:10,091 我们没有办法 777 00:21:10,091 --> 00:21:11,411 这个过于真实 778 00:21:11,411 --> 00:21:13,031 但是我觉得上一期 779 00:21:13,031 --> 00:21:14,811 我们给大家讲完这个事情 780 00:21:14,811 --> 00:21:17,391 就是说Deep Seek至少证明了 781 00:21:17,391 --> 00:21:19,431 他今天老子脱了衣服 782 00:21:19,431 --> 00:21:20,711 想要卖要赚钱 783 00:21:20,711 --> 00:21:22,311 这个比谁都快 784 00:21:22,311 --> 00:21:24,971 我现在融你钱 785 00:21:24,971 --> 00:21:26,671 是给你融资 786 00:21:26,671 --> 00:21:28,011 给你面子 787 00:21:28,011 --> 00:21:30,531 给你面子 788 00:21:30,531 --> 00:21:32,391 是给稳住团队 789 00:21:32,391 --> 00:21:34,131 你不要以为说 790 00:21:34,131 --> 00:21:37,351 稳住团队给国家一个交代 791 00:21:37,351 --> 00:21:38,011 这个事情 792 00:21:38,011 --> 00:21:39,631 你看你新闻里说的 793 00:21:39,631 --> 00:21:40,331 就是说这一轮 794 00:21:40,331 --> 00:21:41,731 就是相当于可能是一家轮 795 00:21:41,731 --> 00:21:41,991 或者什么 796 00:21:41,991 --> 00:21:42,691 我也不知道哪轮 797 00:21:42,691 --> 00:21:45,631 说很多上一轮落选的基金 798 00:21:45,631 --> 00:21:48,331 都在积极踊跃的争取 799 00:21:48,331 --> 00:21:51,311 有内鬼终止交易 800 00:21:51,311 --> 00:21:53,151 内鬼已经被清除了 801 00:21:53,151 --> 00:21:54,091 太小众了 802 00:21:54,091 --> 00:21:55,191 当年我们融资的时候 803 00:21:55,191 --> 00:21:55,751 真的就是 804 00:21:55,751 --> 00:21:57,351 FA拿了一张表过来 805 00:21:57,351 --> 00:21:58,171 五六十个基金 806 00:21:58,171 --> 00:21:58,751 然后打电话 807 00:21:58,751 --> 00:21:59,311 能感兴趣 808 00:21:59,311 --> 00:22:00,331 能有十分之一的不错 809 00:22:00,331 --> 00:22:02,091 然后最终人家能愿意答应 810 00:22:02,091 --> 00:22:03,111 就有个几个 811 00:22:03,111 --> 00:22:05,511 人家梁文峰那边是落选 812 00:22:05,511 --> 00:22:07,631 有一些一手轮落选的基金 813 00:22:07,631 --> 00:22:10,571 每个安慰奖的 814 00:22:10,571 --> 00:22:11,191 真的 815 00:22:11,191 --> 00:22:12,291 给你点份 816 00:22:12,291 --> 00:22:13,131 给你点给你点 817 00:22:13,131 --> 00:22:14,331 也怪不容易的 818 00:22:14,331 --> 00:22:15,791 然后这是第一个 819 00:22:15,791 --> 00:22:17,091 但第二个我更佩服 820 00:22:17,091 --> 00:22:17,451 你知道吗 821 00:22:17,451 --> 00:22:18,491 第二个事我更佩服 822 00:22:18,491 --> 00:22:20,151 就是说他不去做多摩泰 823 00:22:20,151 --> 00:22:22,171 当然这个上期我们也讲到了 824 00:22:22,171 --> 00:22:23,971 就是你知道中国战略里面 825 00:22:23,971 --> 00:22:25,731 就无论是孙子兵法还是什么 826 00:22:25,731 --> 00:22:26,451 就讲的一个什么叫 827 00:22:26,451 --> 00:22:27,251 围三缺一 828 00:22:27,251 --> 00:22:28,971 就是你不能把三个门 829 00:22:28,971 --> 00:22:30,091 四个门全围上 830 00:22:30,091 --> 00:22:31,451 一个手围成的时候 831 00:22:31,451 --> 00:22:32,231 你得留一个门 832 00:22:32,231 --> 00:22:32,891 让大家能跑 833 00:22:32,891 --> 00:22:34,411 然后你就追着再打 834 00:22:34,411 --> 00:22:36,071 就是我觉得这就是 835 00:22:36,071 --> 00:22:38,511 围三缺一的一个特别典型的战略 836 00:22:38,511 --> 00:22:39,251 就是什么意思 837 00:22:39,251 --> 00:22:41,351 不做多摩泰都赚钱的对吧 838 00:22:41,351 --> 00:22:46,071 我给其他模型 839 00:22:46,071 --> 00:22:47,071 留条活路 840 00:22:47,071 --> 00:22:48,371 让你们做多模态 841 00:22:48,371 --> 00:22:50,251 让我这边永远不可能完成所有任务 842 00:22:50,251 --> 00:22:51,011 然后干嘛 843 00:22:51,011 --> 00:22:52,611 我专心致志去赚我的钱 844 00:22:52,611 --> 00:22:54,711 你们专心致志去赚你们的名 845 00:22:54,711 --> 00:22:56,171 咱们别都名利双收 846 00:22:56,171 --> 00:22:57,591 我永远留下最后一颗铜板 847 00:22:57,591 --> 00:22:59,691 我反正觉得这是梁文峰 848 00:22:59,691 --> 00:23:01,191 非常非常聪明的一个定位 849 00:23:01,191 --> 00:23:03,591 当然能够抗拒这个诱惑 850 00:23:03,591 --> 00:23:05,711 和抵抗这个舆论的压力 851 00:23:05,711 --> 00:23:06,511 或者是大家期待 852 00:23:06,511 --> 00:23:07,531 这太难了 853 00:23:07,531 --> 00:23:07,991 我不知道老兄 854 00:23:07,991 --> 00:23:08,931 小松你们怎么看 855 00:23:08,931 --> 00:23:10,211 反正我认为这是 856 00:23:10,211 --> 00:23:11,411 梁文鲁特别 857 00:23:11,411 --> 00:23:14,091 我们昨天才变过这个问题 858 00:23:14,091 --> 00:23:15,611 就是多摩泰这个事情 859 00:23:15,611 --> 00:23:15,951 是的 860 00:23:15,951 --> 00:23:18,011 比如说你编程 861 00:23:18,011 --> 00:23:19,891 当然老修肯定知道 862 00:23:19,891 --> 00:23:21,251 编程在做前端 863 00:23:21,251 --> 00:23:23,211 哪个地方会用到多摩泰 864 00:23:23,211 --> 00:23:23,611 对吧 865 00:23:23,611 --> 00:23:24,511 要写一张照片 866 00:23:24,511 --> 00:23:25,091 要分析什么 867 00:23:25,091 --> 00:23:26,531 但我普通人 868 00:23:26,531 --> 00:23:27,151 我肯定不知道 869 00:23:27,151 --> 00:23:27,731 是的 870 00:23:27,731 --> 00:23:28,591 我哪知道 871 00:23:28,591 --> 00:23:30,511 我说你网页给我写着写着写着 872 00:23:30,511 --> 00:23:31,771 然后我跟你说 873 00:23:31,771 --> 00:23:33,391 这个黑不是五彩斑斓的黑 874 00:23:33,391 --> 00:23:34,071 你看不到 875 00:23:34,071 --> 00:23:34,771 你瞎了 876 00:23:34,771 --> 00:23:35,491 对不对 877 00:23:35,491 --> 00:23:36,411 你不知道 878 00:23:36,411 --> 00:23:37,011 对 879 00:23:37,011 --> 00:23:38,451 然后他叫了一个第三方 880 00:23:38,451 --> 00:23:38,891 帮你干 881 00:23:38,931 --> 00:23:40,931 然後他叫了一個第三方幫你幹了,你還以為是他幹的很錯? 882 00:23:40,931 --> 00:23:44,931 其實是他本來就看不見,他叫第三方幫他看的,轉述了一下他在幹的。 883 00:23:44,931 --> 00:23:47,931 那這就是一個很大的分歧嘛。 884 00:23:47,931 --> 00:23:53,931 我作為一個魔情賞倉,我是像Deep Seek這樣子,我就做給這些懂行的人。 885 00:23:53,931 --> 00:23:55,231 我就明确告诉你 886 00:23:55,231 --> 00:23:56,331 我就不做多摩泰 887 00:23:56,331 --> 00:23:57,531 然后我足够平 888 00:23:57,531 --> 00:24:00,111 还是说我是谷歌家大业大 889 00:24:00,111 --> 00:24:02,151 我这个哪怕我要做成Deep Seek 890 00:24:02,151 --> 00:24:04,591 我也要做成一个多摩泰的Deep Seek 891 00:24:04,591 --> 00:24:06,031 就全原生多摩泰 892 00:24:06,031 --> 00:24:10,311 但是这个就是路线的区别了 893 00:24:10,311 --> 00:24:11,751 派系的区别了 894 00:24:11,751 --> 00:24:12,311 对不对 895 00:24:12,311 --> 00:24:13,831 完全战略比较 896 00:24:13,831 --> 00:24:16,671 对这个刚才贤哥问的感受 897 00:24:16,671 --> 00:24:18,551 我觉得我得call back一个事情 898 00:24:18,551 --> 00:24:19,231 就是今天 899 00:24:19,231 --> 00:24:22,651 Jeff Dean离开谷歌的事情 900 00:24:22,651 --> 00:24:24,511 就是Jeff Dean我估计你们俩不会太熟 901 00:24:24,511 --> 00:24:25,911 那我要说他几个事情 902 00:24:25,911 --> 00:24:26,631 你们会知道 903 00:24:26,631 --> 00:24:27,971 TPU是他搞的 904 00:24:27,971 --> 00:24:32,251 TPU现在GEMIA的我们叫做什么 905 00:24:32,251 --> 00:24:37,411 就是他的成本是跟人家领先数量级级别的 906 00:24:37,411 --> 00:24:40,151 这个东西是Jeff Dean在10年前 907 00:24:40,151 --> 00:24:41,191 他就预判到了 908 00:24:41,191 --> 00:24:43,431 这个东西必须得自己做TPU 909 00:24:43,431 --> 00:24:44,491 才能做到这个事情 910 00:24:44,491 --> 00:24:47,671 否则谷歌成本会永远被别人牵着走 911 00:24:47,671 --> 00:24:49,191 10年前他就判到了 912 00:24:49,191 --> 00:24:51,271 而且10年前已经开始投入使用了 913 00:24:51,271 --> 00:24:53,771 十年比斯定调了原生多摩泰 914 00:24:53,771 --> 00:24:55,571 对哈萨比斯定调原生多摩泰 915 00:24:55,571 --> 00:24:57,471 那杰弗丁这个人再更早 916 00:24:57,471 --> 00:24:58,731 他是现在Tensors Flow 917 00:24:58,731 --> 00:25:00,451 现在也一直是 918 00:25:00,451 --> 00:25:01,651 就是全球最大的 919 00:25:01,651 --> 00:25:04,211 machine learning的开源平台 920 00:25:04,211 --> 00:25:05,431 Tensors Flow里面 921 00:25:05,431 --> 00:25:07,011 大量的机器学习的东西 922 00:25:07,011 --> 00:25:09,391 都是杰弗丁在里面去推的 923 00:25:09,391 --> 00:25:10,351 里面有很多的工具 924 00:25:10,351 --> 00:25:11,131 都是他做的 925 00:25:11,131 --> 00:25:12,691 就是非常底层 926 00:25:12,691 --> 00:25:13,791 你就想象一下 927 00:25:13,791 --> 00:25:15,551 你现在除了大模型之外的 928 00:25:15,551 --> 00:25:16,611 那些AI的基建 929 00:25:16,611 --> 00:25:18,751 基本都是他搞的那套基建 930 00:25:18,751 --> 00:25:20,151 跑他编译出来 931 00:25:20,151 --> 00:25:21,271 然后他帮你调试 932 00:25:21,271 --> 00:25:22,391 干嘛优化 933 00:25:22,391 --> 00:25:23,651 都是他之前做的 934 00:25:23,651 --> 00:25:25,211 就这两个事情就已经很大了 935 00:25:25,211 --> 00:25:26,031 他是这么一个人 936 00:25:26,031 --> 00:25:27,611 他现在出去做的事情 937 00:25:27,611 --> 00:25:28,791 我的理解 938 00:25:28,791 --> 00:25:31,571 其实是美国版的Deepseek AGI 939 00:25:31,571 --> 00:25:33,091 他想做的是 940 00:25:33,091 --> 00:25:35,771 要再做一个更底层基建的 941 00:25:35,771 --> 00:25:37,771 普惠的一个东西 942 00:25:37,771 --> 00:25:39,751 虽然现在因为太新了 943 00:25:39,751 --> 00:25:41,011 也没办法讲清楚 944 00:25:41,011 --> 00:25:42,251 但是我感觉 945 00:25:42,251 --> 00:25:44,551 他其实跟我们的Deepseek 946 00:25:44,551 --> 00:25:46,011 说的事情是很像的 947 00:25:46,011 --> 00:25:47,971 但是他要迈入更高的 948 00:25:47,971 --> 00:25:49,451 但是他会更底层 949 00:25:49,451 --> 00:25:50,951 然后他讲了有一个判断 950 00:25:50,951 --> 00:25:52,951 我觉得非常值得大家记住 951 00:25:52,951 --> 00:25:56,171 他就是说到现在这个时间节点 952 00:25:56,171 --> 00:25:57,371 大家应该去做 953 00:25:57,371 --> 00:25:59,671 目前你看到的大模型 954 00:25:59,671 --> 00:26:03,171 只能做1%或者是0%的事情 955 00:26:03,171 --> 00:26:04,591 这个才有价值 956 00:26:04,591 --> 00:26:06,451 20%已经能做了的话 957 00:26:06,451 --> 00:26:08,591 都没有足够的创新价值了 958 00:26:08,591 --> 00:26:11,531 他就是要去做1%左右的这样的一个东西 959 00:26:11,531 --> 00:26:15,271 我觉得这个就是做战略的人 960 00:26:15,271 --> 00:26:17,691 站的角度就做出不同的战略 961 00:26:17,691 --> 00:26:18,071 对吧 962 00:26:18,071 --> 00:26:20,391 我们刚才说Google的优势 963 00:26:20,391 --> 00:26:23,051 当时Gymna刚出来的时候 964 00:26:23,051 --> 00:26:25,511 最大优势操场上叫我们原生多摩泰 965 00:26:25,511 --> 00:26:26,231 对吧 966 00:26:26,231 --> 00:26:28,891 这个一定是科学家级别的定义 967 00:26:28,891 --> 00:26:30,751 投入了巨大资源来成本 968 00:26:30,751 --> 00:26:32,091 来做这样一个基建 969 00:26:32,091 --> 00:26:33,651 做这样一个基建 970 00:26:33,651 --> 00:26:36,071 但是你换成梁胜 971 00:26:36,071 --> 00:26:37,851 来做这个产品 972 00:26:37,851 --> 00:26:38,971 他可能是专家 973 00:26:38,971 --> 00:26:39,931 他不是科学家 974 00:26:39,931 --> 00:26:40,571 我专家 975 00:26:40,571 --> 00:26:41,211 我商人 976 00:26:41,211 --> 00:26:42,171 我搞量化了 977 00:26:42,171 --> 00:26:42,631 对吧 978 00:26:42,631 --> 00:26:43,791 我就知道说 979 00:26:43,791 --> 00:26:46,431 我从非多摩泰的 980 00:26:46,431 --> 00:26:48,171 给你搞到价格最低 981 00:26:48,171 --> 00:26:48,591 对吧 982 00:26:48,591 --> 00:26:50,211 这个是路线的区别 983 00:26:50,211 --> 00:26:51,871 就跟我们企业里面做决策 984 00:26:51,871 --> 00:26:52,251 对吧 985 00:26:52,251 --> 00:26:53,831 我们要做一个项目 986 00:26:53,831 --> 00:26:56,711 我拉贤哥来做这个项目的定义 987 00:26:56,711 --> 00:26:58,531 那贤哥肯定是先奔着 988 00:26:58,531 --> 00:27:00,231 先做现金流 989 00:27:00,231 --> 00:27:00,751 先赚钱 990 00:27:00,751 --> 00:27:00,971 对吧 991 00:27:00,971 --> 00:27:03,751 比如说我拉老修来 992 00:27:03,751 --> 00:27:04,611 老修就要说 993 00:27:04,611 --> 00:27:06,351 我这个架构不能这么设计 994 00:27:06,351 --> 00:27:07,991 不然到时候业务跑到哪里 995 00:27:07,991 --> 00:27:08,911 会出什么问题 996 00:27:08,911 --> 00:27:10,931 你得给我一个月时间 997 00:27:10,931 --> 00:27:12,211 架构要搭好 998 00:27:12,211 --> 00:27:13,191 对吧 999 00:27:13,191 --> 00:27:14,731 我们经常会在企业里面 1000 00:27:14,731 --> 00:27:15,711 会遇到这样的问题 1001 00:27:15,711 --> 00:27:20,091 对不同人做的决策的依据是不一样的 1002 00:27:20,091 --> 00:27:20,991 我觉得一年前 1003 00:27:20,991 --> 00:27:22,471 JM2.5 Pro的时候 1004 00:27:22,471 --> 00:27:25,291 证明的路线是对的 1005 00:27:25,291 --> 00:27:25,731 对吧 1006 00:27:25,731 --> 00:27:26,971 那你就做 1007 00:27:26,971 --> 00:27:27,671 你就对 1008 00:27:27,671 --> 00:27:29,031 我敬你是条汉子 1009 00:27:29,031 --> 00:27:30,431 那今年他妈的你不行 1010 00:27:30,431 --> 00:27:32,891 我也不能说什么 1011 00:27:32,891 --> 00:27:33,751 我们就爱过 1012 00:27:33,751 --> 00:27:35,471 对吧 1013 00:27:35,471 --> 00:27:38,131 但是今年Deep Seek这个路线突然又行了 1014 00:27:38,131 --> 00:27:40,211 但是你回到明年去看 1015 00:27:40,211 --> 00:27:42,831 你又可能会觉得今天Deep Seek格局又低 1016 00:27:42,831 --> 00:27:44,291 很有可能 1017 00:27:44,291 --> 00:27:45,011 很正常 1018 00:27:45,011 --> 00:27:45,731 对 1019 00:27:45,731 --> 00:27:48,791 因为这个东西就像在一年以前 1020 00:27:48,791 --> 00:27:50,751 大家问编程到底行不行 1021 00:27:50,751 --> 00:27:51,031 是吧 1022 00:27:51,031 --> 00:27:52,391 还有疑虑那个时候 1023 00:27:52,391 --> 00:27:52,811 对吧 1024 00:27:52,811 --> 00:27:54,331 主要是我们进步了 1025 00:27:54,331 --> 00:27:56,011 后面一进步 1026 00:27:56,011 --> 00:27:57,391 这个事情马上就追上来了 1027 00:27:57,391 --> 00:27:57,711 是