1 00:00:00,000 --> 00:00:28,540 如果你持续关注硅谷AI产业动态,一定听过这场堪称传奇的人才出走事件。手握20亿美元顶级风头,集结一批从OpenAI出走的核心研发人才。秘密打磨接近两年,所有人都默认他们会拿出一款能够横扫行业的顶尖大模型。可最终的结果却让整个科技圈大跌眼镜。在前不久的发布会结束之后,他们推出的全新模型在主流权威基准评测榜单当中甚至无法冲进前五。 2 00:00:28,540 --> 00:00:30,640 不少开源模型的综合表现 3 00:00:30,640 --> 00:00:32,240 都能够实现对它的压制 4 00:00:32,240 --> 00:00:34,500 回望2024年年底那场震动 5 00:00:34,500 --> 00:00:35,680 全球AI行业的 6 00:00:35,680 --> 00:00:37,240 OpenAI高管离职风波 7 00:00:37,240 --> 00:00:38,780 大家才慢慢意识到 8 00:00:38,780 --> 00:00:41,200 这根本不是简单的人员流动 9 00:00:41,200 --> 00:00:42,560 而是一场精心布局 10 00:00:42,560 --> 00:00:44,820 持续多年的人才收割行动 11 00:00:44,820 --> 00:00:45,780 米拉姆拉蒂 12 00:00:45,780 --> 00:00:48,120 也就是曾经短暂执掌OpenAI 13 00:00:48,120 --> 00:00:49,940 全程主导ChatGPT 14 00:00:49,940 --> 00:00:51,740 落地研发的核心关键人物 15 00:00:51,740 --> 00:00:54,460 在2024年9月正式离开老东家 16 00:00:54,460 --> 00:00:56,300 并且不是孤身一人 17 00:00:56,300 --> 00:00:58,320 他踏出OpenAI大门之后 18 00:00:58,320 --> 00:01:00,160 联合创始人约翰·舒尔曼 19 00:01:00,160 --> 00:01:02,380 前研究副总裁巴雷特·佐夫 20 00:01:02,380 --> 00:01:06,160 还有一大批资深核心研究员紧随其后递交辞呈 21 00:01:06,160 --> 00:01:07,800 这群人的集体出走 22 00:01:07,800 --> 00:01:11,120 几乎直接抽走了OpenAI半数核心研发力量 23 00:01:11,120 --> 00:01:13,180 在硅谷的产业生态里面 24 00:01:13,180 --> 00:01:15,500 顶尖人工智能研究员的稀缺程度 25 00:01:15,500 --> 00:01:18,840 甚至超过英伟达H100系列高端算力芯片 26 00:01:18,840 --> 00:01:21,960 米拉带走的远远不止数十名资深工程师 27 00:01:21,960 --> 00:01:26,060 更是一套经过GPT3到GPT4完整项目验证 28 00:01:26,060 --> 00:01:28,760 成熟可行的大模型训练方法论 29 00:01:28,760 --> 00:01:31,220 还有整套工程落地的实战经验 30 00:01:31,220 --> 00:01:33,060 这支团队亲眼见证过 31 00:01:33,060 --> 00:01:35,660 全球规模顶尖算力集群的运行逻辑 32 00:01:35,660 --> 00:01:37,420 也踩遍了训练过程里 33 00:01:37,420 --> 00:01:40,240 各类会引发模型崩溃的隐蔽技术陷阱 34 00:01:40,240 --> 00:01:42,860 因此当他们对外宣布独立创业的时候 35 00:01:42,860 --> 00:01:45,240 整个风险投资圈子陷入狂热 36 00:01:45,240 --> 00:01:47,460 硅谷风投机构最擅长算账 37 00:01:47,460 --> 00:01:50,300 也最容易陷入害怕错失机遇的情绪 38 00:01:50,300 --> 00:01:51,600 安德森霍洛维茨 39 00:01:51,600 --> 00:01:53,900 直接敲定20亿美元现金投入 40 00:01:53,900 --> 00:01:55,320 投资落地的时候 41 00:01:55,320 --> 00:01:57,960 这家连正式名称都没有完全敲定 42 00:01:57,960 --> 00:02:01,540 一行商用产品代码都还没有产出的初创团队 43 00:02:01,540 --> 00:02:03,700 最终定名Thinking Machines Lab 44 00:02:03,700 --> 00:02:06,180 投后估值直接冲到120亿美元 45 00:02:06,180 --> 00:02:10,020 没有任何成熟产品就能拥有120亿美元估值 46 00:02:10,020 --> 00:02:13,760 放在其他行业很容易被贴上庞氏骗局的标签 47 00:02:13,760 --> 00:02:16,140 但是在近几年火热的AI赛道 48 00:02:16,140 --> 00:02:20,360 资本把这称作抢占通往通用人工智能时代最后的传票 49 00:02:20,360 --> 00:02:22,460 风投机构押注的核心目标 50 00:02:22,460 --> 00:02:24,400 就是米拉带领的这支团队 51 00:02:24,400 --> 00:02:25,720 能够打造出一套 52 00:02:25,720 --> 00:02:27,980 打破OpenAI避原模式垄断的 53 00:02:27,980 --> 00:02:28,960 全新产业生态 54 00:02:28,960 --> 00:02:30,500 所有人静静观望 55 00:02:30,500 --> 00:02:32,380 期待着20亿美元资金 56 00:02:32,380 --> 00:02:34,280 两年时间能够孵化出 57 00:02:34,280 --> 00:02:36,640 怎样惊世骇俗的人工智能产品 58 00:02:36,640 --> 00:02:38,440 整个行业秉析宁神 59 00:02:38,440 --> 00:02:40,580 等待新一代顶级大模型的 60 00:02:40,580 --> 00:02:41,580 技术降维打击 61 00:02:41,580 --> 00:02:43,980 时间来到2026年7月中旬 62 00:02:43,980 --> 00:02:45,060 也就是近期 63 00:02:45,060 --> 00:02:46,360 沉寂一年半之久的 64 00:02:46,360 --> 00:02:47,420 Thinking Machines Lab 65 00:02:47,420 --> 00:02:50,140 终于对外公开酝酿许久的重磅成果 66 00:02:50,140 --> 00:02:52,260 开源基础模型代号Inklin 67 00:02:52,260 --> 00:02:54,020 伴随着发布会正式落幕 68 00:02:54,020 --> 00:02:56,760 整个科技行业陷入巨大的错愕 69 00:02:56,760 --> 00:02:59,860 这款投入重金闭关研发两年打造的模型 70 00:02:59,860 --> 00:03:03,640 在各类权威基准测试当中成绩只能用平庸形容 71 00:03:03,640 --> 00:03:05,360 前方有克劳德·菲波5 72 00:03:05,360 --> 00:03:08,880 OpenAI主流避原模型两座顶尖避原模型 73 00:03:08,880 --> 00:03:10,060 稳定占据高位 74 00:03:10,060 --> 00:03:12,420 同期发布的月之暗面KIMI K3 75 00:03:12,420 --> 00:03:14,740 这款拥有2.8万亿参数的模型 76 00:03:14,740 --> 00:03:16,640 同样能够和它正面抗衡 77 00:03:16,640 --> 00:03:18,480 多项指标实现反超 78 00:03:18,480 --> 00:03:20,720 Inkling在开源模型梯队当中 79 00:03:20,720 --> 00:03:23,260 综合水平最多只能算作中油水准 80 00:03:23,260 --> 00:03:25,860 更加尴尬的还有发布时间窗口 81 00:03:25,860 --> 00:03:28,460 Inkling正式上线仅仅一天之后 82 00:03:28,460 --> 00:03:30,440 Kimi K3热度全面爆发 83 00:03:30,440 --> 00:03:32,120 大量开发者涌入调用 84 00:03:32,120 --> 00:03:33,800 直接把算力资源耗尽 85 00:03:33,800 --> 00:03:36,560 几乎抢走了Inkling全部公众关注度 86 00:03:36,560 --> 00:03:38,000 普通观众看到这里 87 00:03:38,000 --> 00:03:40,980 大概率都会认定这是一场彻头彻尾的失败 88 00:03:40,980 --> 00:03:41,960 顶级团队 89 00:03:41,960 --> 00:03:42,780 顶级资金 90 00:03:42,780 --> 00:03:43,620 顶级资源 91 00:03:43,620 --> 00:03:45,740 最后打出一张平平无奇的牌 92 00:03:45,740 --> 00:03:48,100 甚至被后发竞品全面碾压 93 00:03:48,100 --> 00:03:50,700 但如果你只看到表面的跑分输赢 94 00:03:50,700 --> 00:03:54,180 完全看不懂当下AI行业真正的内卷困局 95 00:03:54,180 --> 00:03:57,360 更看不懂硅谷顶级资本真正的下注逻辑 96 00:03:57,360 --> 00:03:59,680 现在整个AI行业已经陷入一种 97 00:03:59,680 --> 00:04:01,340 极度畸形的内卷怪圈 98 00:04:01,340 --> 00:04:03,900 绝大多数团队、媒体、开发者 99 00:04:03,900 --> 00:04:06,380 都在疯狂迷信基准测试分数 100 00:04:06,380 --> 00:04:09,260 把榜单排名当成唯一的评判标准 101 00:04:09,260 --> 00:04:11,340 为了零点几分的微弱提升 102 00:04:11,340 --> 00:04:14,040 无数公司动辄烧数亿美金算力 103 00:04:14,040 --> 00:04:15,420 堆砌海量参数 104 00:04:15,420 --> 00:04:16,800 清洗冗余数据 105 00:04:16,800 --> 00:04:18,220 反复微调权重 106 00:04:18,220 --> 00:04:20,820 最后做出来的模型跑分看似好看 107 00:04:20,820 --> 00:04:22,300 但是商用成本极高 108 00:04:22,300 --> 00:04:23,880 落地兼容性极差 109 00:04:23,880 --> 00:04:25,580 企业部署难度巨大 110 00:04:25,580 --> 00:04:27,400 完全是为了跑分而生 111 00:04:27,400 --> 00:04:29,480 而非为了商用而生的产品 112 00:04:29,480 --> 00:04:31,040 这也是最近一年来 113 00:04:31,040 --> 00:04:33,240 无数顶级AI模型高开低走 114 00:04:33,240 --> 00:04:35,920 热度爆红却无法变现的核心原因 115 00:04:35,920 --> 00:04:38,260 行业早已不再缺聪明的模型 116 00:04:38,260 --> 00:04:39,640 企业真正缺的 117 00:04:39,640 --> 00:04:41,660 是稳定便宜可控能落地 118 00:04:41,660 --> 00:04:44,060 零合规风险的工业化AI基础设施 119 00:04:44,060 --> 00:04:46,960 而这恰恰就是Inkling精准卡位的赛道 120 00:04:46,960 --> 00:04:49,660 大家一定要点赞订阅锁住频道 121 00:04:49,660 --> 00:04:53,480 接下来我逐层拆解这场瞒天过海的AI商业布局 122 00:04:53,480 --> 00:04:54,840 看懂了你就明白 123 00:04:54,840 --> 00:04:56,240 这根本不是翻车 124 00:04:56,240 --> 00:04:59,240 而是近两年最顶级的AI创业降位布局 125 00:04:59,240 --> 00:05:00,880 我们不妨深入拆解 126 00:05:00,880 --> 00:05:05,520 这笔20亿美元资金孵化出来的技术底子究竟具备哪些特点 127 00:05:05,520 --> 00:05:10,240 Inkling属于总参数量达到9750亿的混合专家模型 128 00:05:10,240 --> 00:05:12,460 也就是业内常说的MOE架构 129 00:05:12,460 --> 00:05:14,860 单单看总参数规模接近万亿 130 00:05:14,860 --> 00:05:16,440 听起来极具冲击力 131 00:05:16,440 --> 00:05:18,920 但是熟悉稀疏架构的从业者 132 00:05:18,920 --> 00:05:21,180 都清楚MOE底层运行逻辑 133 00:05:21,180 --> 00:05:23,300 每一次处理词源信息的时候 134 00:05:23,300 --> 00:05:26,500 仅仅只有410亿参数被激活参与运算 135 00:05:26,500 --> 00:05:29,460 模型原生支持100万长度上下文窗口 136 00:05:29,460 --> 00:05:31,760 依托45万亿词源规模的文本 137 00:05:31,760 --> 00:05:33,680 图像 音频混合数据 138 00:05:33,680 --> 00:05:35,520 完成从零开始的预训链 139 00:05:35,520 --> 00:05:38,160 整套体系当中最具备革新意义的 140 00:05:38,160 --> 00:05:40,800 是它重构了多模态信息处理逻辑 141 00:05:40,800 --> 00:05:44,760 目前市面上绝大多数宣称原生多模态的大模型 142 00:05:44,760 --> 00:05:48,820 本质只是在基础语言模型外部附加一套处理组件 143 00:05:48,820 --> 00:05:50,780 遇到图片、音频信息 144 00:05:50,780 --> 00:05:52,180 需要独立编码器 145 00:05:52,180 --> 00:05:56,580 先把视觉、听觉信号转化为模型能够识别的次元格式 146 00:05:56,580 --> 00:05:58,460 再送入主体网络运算 147 00:05:58,460 --> 00:06:00,780 而Inkling跳过中间转化流程 148 00:06:00,780 --> 00:06:04,260 可以直接读取原始像素与音频波形信号 149 00:06:04,260 --> 00:06:07,160 这种底层架构重构的工程难度极高 150 00:06:07,160 --> 00:06:11,320 也直观体现出这支前OpenAI团队扎实的技术积累 151 00:06:11,320 --> 00:06:14,760 对比Meta很多开源模型饱受争议的授权条款 152 00:06:14,760 --> 00:06:16,160 Thinking Machines Lab 153 00:06:16,160 --> 00:06:17,760 这次选择十分开放 154 00:06:17,760 --> 00:06:20,360 直接采用Apache 2.0开源协议 155 00:06:20,360 --> 00:06:21,980 没有附加繁杂限制 156 00:06:21,980 --> 00:06:25,220 完整模型权重直接上传至哈格面部平台 157 00:06:25,220 --> 00:06:27,080 开发者可以自由修改 158 00:06:27,080 --> 00:06:28,480 部署 商用 159 00:06:28,480 --> 00:06:30,500 即便拥有底层架构创新 160 00:06:30,500 --> 00:06:32,640 搭配宽松友好的开源协议 161 00:06:32,640 --> 00:06:35,500 依旧无法掩盖一处无法回避的短板 162 00:06:35,500 --> 00:06:38,040 模型纯粹推理上线并不突出 163 00:06:38,040 --> 00:06:41,060 当下AI行业早已进入白热化竞争阶段 164 00:06:41,060 --> 00:06:45,360 各家机构哪怕基准测试出现小数点级别的分数差距 165 00:06:45,360 --> 00:06:47,980 都会专门撰写论文大肆宣传 166 00:06:47,980 --> 00:06:49,520 投入20亿美元资金 167 00:06:49,520 --> 00:06:52,340 最终产出一款激活参数410亿 168 00:06:52,340 --> 00:06:54,300 综合表现中等的模型 169 00:06:54,300 --> 00:06:57,460 在外人眼里无疑是一笔严重亏损的投资 170 00:06:57,460 --> 00:06:59,740 不少业内研究员研读论文 171 00:06:59,740 --> 00:07:02,540 实测模型之后给出的评价高度统一 172 00:07:02,540 --> 00:07:04,760 评语大多只有很酷两个字 173 00:07:04,760 --> 00:07:06,260 站在投资人视角 174 00:07:06,260 --> 00:07:08,520 花费巨额资金换来一句很酷 175 00:07:08,520 --> 00:07:11,160 和你邀请他人享用高端米其林晚餐 176 00:07:11,160 --> 00:07:12,580 最后只收到一句 177 00:07:12,580 --> 00:07:14,120 你人很好高度相似 178 00:07:14,120 --> 00:07:16,280 属于十分委婉的负面评价 179 00:07:16,280 --> 00:07:18,140 很多人会顺势产生疑问 180 00:07:18,140 --> 00:07:20,640 难道米拉团队已经将狼才进 181 00:07:20,640 --> 00:07:23,040 脱离大厂充足算力扶持之后 182 00:07:23,040 --> 00:07:26,540 这群天才工程师再也无法完成前沿模型研发 183 00:07:26,540 --> 00:07:28,120 答案显然是否定的 184 00:07:28,120 --> 00:07:31,420 这里藏着一个极异被大众忽略的反常识逻辑 185 00:07:31,420 --> 00:07:35,260 也是绝大多数行业观察者出现判断偏差的关键 186 00:07:35,260 --> 00:07:37,420 外界普遍默认研发大模型 187 00:07:37,420 --> 00:07:39,760 必须无休止开展算力军备竞赛 188 00:07:39,760 --> 00:07:41,980 追求更高参数 更强跑分 189 00:07:41,980 --> 00:07:45,060 谁的榜单成绩领先 谁就是最终赢家 190 00:07:45,060 --> 00:07:48,560 可这套竞争思路已经是两年之前的老旧玩法 191 00:07:48,560 --> 00:07:52,160 现如今想要依靠基准测试提升1%的准确率 192 00:07:52,160 --> 00:07:55,980 需要额外投入数亿美元电费 海量算力资源 193 00:07:56,060 --> 00:07:59,760 还要耗费大量人力筛选稀缺高质量训练数据 194 00:08:00,000 --> 00:08:07,480 边际成本已经膨胀到难以承受的地步。米拉在OpenAI长期生耕,比任何人都清楚算力成本核算逻辑。 195 00:08:07,480 --> 00:08:14,960 深刻明白B原大模型无休止堆参数换取微小性能提升,本身就是一条难以持续的死胡同。 196 00:08:14,960 --> 00:08:18,040 如果运营一家估值120亿美元的企业 197 00:08:18,040 --> 00:08:19,800 核心目标绝对不是 198 00:08:19,800 --> 00:08:22,960 打赢一场没有实际商业价值的跑分竞赛 199 00:08:22,960 --> 00:08:26,600 而是搭建一套能持续自我造血的商业体系 200 00:08:26,600 --> 00:08:30,860 所以当大家看见一群掌握顶尖AI底层认知的技术人才 201 00:08:30,860 --> 00:08:34,360 在关键节点推出一款看起来平平无奇的产品 202 00:08:34,360 --> 00:08:36,920 千万不要简单判定他们决策失误 203 00:08:36,920 --> 00:08:40,320 他们正在执行一套普通人难以看懂的长线布局 204 00:08:40,320 --> 00:08:42,440 Inkling跑分表现相对平庸 205 00:08:42,440 --> 00:08:43,960 并不是研发能力不足 206 00:08:43,960 --> 00:08:46,040 而是精准的商业策略规划 207 00:08:46,040 --> 00:08:48,440 Tai Mankey & G. McKean ES Lab 208 00:08:48,440 --> 00:08:50,160 从一开始就没有争夺 209 00:08:50,160 --> 00:08:52,720 全球最强通用大模型的想法 210 00:08:52,720 --> 00:08:54,440 把一款推理能力中等 211 00:08:54,440 --> 00:08:55,660 算力消耗低廉 212 00:08:55,660 --> 00:08:56,820 响应速度出众 213 00:08:56,820 --> 00:08:58,360 并且完全开源的模型 214 00:08:58,360 --> 00:08:59,560 免费对外开放 215 00:08:59,560 --> 00:09:01,980 仅仅是整套战略布局的第一步 216 00:09:01,980 --> 00:09:03,760 20亿美元的资金投入 217 00:09:03,760 --> 00:09:05,060 绝大部分运用在 218 00:09:05,060 --> 00:09:07,280 Inkling反直觉的底层机制设计 219 00:09:07,280 --> 00:09:09,180 以及他们真正用来实现 220 00:09:09,180 --> 00:09:11,240 商业化变现的秘密武器之上 221 00:09:11,240 --> 00:09:12,880 倘若大家单纯认为 222 00:09:12,880 --> 00:09:15,180 这款无法冲进榜单前五的模型 223 00:09:15,180 --> 00:09:16,140 就是故事终点 224 00:09:16,140 --> 00:09:20,080 那就严重低估企业及弊端市场竞争的残酷程度 225 00:09:20,080 --> 00:09:21,780 为了更好落地商业化 226 00:09:21,780 --> 00:09:24,940 Inkling甚至主动削减一部分自由联想能力 227 00:09:24,940 --> 00:09:28,340 只为持续降低单次调用产生的词源开销 228 00:09:28,340 --> 00:09:30,020 模型内部运行逻辑 229 00:09:30,020 --> 00:09:31,640 加上Thinking Machines Lab 230 00:09:31,640 --> 00:09:33,980 后续依靠估值变现的整套方案 231 00:09:33,980 --> 00:09:37,200 才是整个科技行业真正需要警惕的地方 232 00:09:37,200 --> 00:09:40,900 当我们拆解Inkling内部调节推理强度的核心机制 233 00:09:40,900 --> 00:09:43,900 就能意识到未来人工智能赛道的主战场 234 00:09:43,900 --> 00:09:46,340 根本不会集中在评测排行榜 235 00:09:46,340 --> 00:09:48,340 这款模型看似平庸的跑分 236 00:09:48,340 --> 00:09:50,740 是刻意控制之下形成的结果 237 00:09:50,740 --> 00:09:53,920 米拉团队在Inkling内部设置三处反直觉 238 00:09:53,920 --> 00:09:55,920 甚至有些激进的技术设定 239 00:09:55,920 --> 00:09:57,640 每一项都精准击中 240 00:09:57,640 --> 00:10:00,760 日军千万次AI调用企业客户的核心痛点 241 00:10:00,760 --> 00:10:04,340 第一项设定可以称之为智能调节推理力度旋钮 242 00:10:04,340 --> 00:10:07,360 官方命名为思考力度控制系统 243 00:10:07,360 --> 00:10:09,180 这项设计充满巧思 244 00:10:09,180 --> 00:10:11,440 市面上绝大多数主流大模型 245 00:10:11,440 --> 00:10:13,560 处理简单问答1加1等于计 246 00:10:13,560 --> 00:10:16,120 和撰写后端微服务架构代码 247 00:10:16,120 --> 00:10:18,260 后台占用的算力资源差距极小 248 00:10:18,260 --> 00:10:20,260 好比驾驶重型重载卡车 249 00:10:20,260 --> 00:10:23,180 仅仅只为前往楼下便利店采购物品 250 00:10:23,180 --> 00:10:25,980 在企业大规模批量调用场景之下 251 00:10:25,980 --> 00:10:28,620 这种资源浪费很难长期承受 252 00:10:28,620 --> 00:10:30,240 依托这套调节旋钮 253 00:10:30,240 --> 00:10:32,040 运行规则彻底改变 254 00:10:32,040 --> 00:10:33,880 当模型处理发票分类 255 00:10:33,880 --> 00:10:37,160 标准化客服问答这类重复性基础任务 256 00:10:37,160 --> 00:10:38,420 可以下调思考力度 257 00:10:38,420 --> 00:10:40,660 此时Inkling反馈速度极快 258 00:10:40,660 --> 00:10:42,720 调用成本压缩到极低水平 259 00:10:42,720 --> 00:10:45,100 基础答案完全能够满足业务需求 260 00:10:45,100 --> 00:10:47,160 一旦遭遇核心业务逻辑推演 261 00:10:47,160 --> 00:10:48,400 复杂代码修复 262 00:10:48,400 --> 00:10:50,040 高阶数学推理难题 263 00:10:50,040 --> 00:10:51,920 直接拉高思考力度档位 264 00:10:51,920 --> 00:10:55,340 Inkling会调动更多算力进入深度推理模式 265 00:10:55,340 --> 00:10:58,600 综合推理效率能够对标行业第一梯队大模型 266 00:10:58,600 --> 00:11:01,020 同时算力消耗远低于同级产品 267 00:11:01,020 --> 00:11:03,600 在企业商用场景具备极强优势 268 00:11:03,600 --> 00:11:07,500 这意味着企业客户终于可以动态管控算力支出 269 00:11:07,500 --> 00:11:10,780 当后台每日运行数百万次智能体任务 270 00:11:10,780 --> 00:11:12,580 可以按需调整资源投入 271 00:11:12,580 --> 00:11:16,060 省下的每一笔开支都会直接转化为企业净利润 272 00:11:16,060 --> 00:11:18,840 第二项设定更加打破大众固有认知 273 00:11:18,840 --> 00:11:22,460 研发团队刻意训练模型形成诚实输出特性 274 00:11:22,460 --> 00:11:25,780 当下绝大多数大模型普遍存在幻觉问题 275 00:11:25,780 --> 00:11:29,660 无论面对何种提问都会强行生成完整答案 276 00:11:29,660 --> 00:11:31,520 哪怕内容完全不符合事实 277 00:11:31,520 --> 00:11:33,540 依旧会以笃定语气输出 278 00:11:33,540 --> 00:11:34,680 在文学创作 279 00:11:34,680 --> 00:11:36,060 日常闲聊场景 280 00:11:36,060 --> 00:11:37,600 这种自主生成能力 281 00:11:37,600 --> 00:11:39,000 可以称之为创造力 282 00:11:39,000 --> 00:11:40,220 放到金融预测 283 00:11:40,220 --> 00:11:42,160 医疗诊断这类企业场景 284 00:11:42,160 --> 00:11:43,720 毫无依据的编造内容 285 00:11:43,720 --> 00:11:44,960 会带来致命风险 286 00:11:44,960 --> 00:11:45,680 为此 287 00:11:45,680 --> 00:11:47,080 Thinking Machines Lab 288 00:11:47,080 --> 00:11:49,180 引入认知学导向训练方案 289 00:11:49,180 --> 00:11:50,200 简单来讲 290 00:11:50,200 --> 00:11:52,080 利用强化学习奖励模型 291 00:11:52,080 --> 00:11:54,100 主动承认认知边界的行为 292 00:11:54,100 --> 00:11:55,760 团队持续迭代训练 293 00:11:55,760 --> 00:11:57,600 引导Inkling在信息不足 294 00:11:57,600 --> 00:11:59,540 无法得出可靠结论的时候 295 00:11:59,540 --> 00:12:00,980 不要强行猜测 296 00:12:00,980 --> 00:12:03,100 直接告知使用者无法解答 297 00:12:03,100 --> 00:12:05,120 最终训练成果十分突出 298 00:12:05,120 --> 00:12:06,620 Inkling在事件预测 299 00:12:06,620 --> 00:12:08,680 数据分析类客观评测中 300 00:12:08,680 --> 00:12:11,540 准确率超越多款主流通用大模型 301 00:12:11,540 --> 00:12:13,280 达到行业顶尖水准 302 00:12:13,280 --> 00:12:15,040 对于华尔街对冲基金 303 00:12:15,040 --> 00:12:16,960 医疗行业大型机构而言 304 00:12:16,960 --> 00:12:20,220 一台关键时刻敢于坦诚信息不足的模型 305 00:12:20,220 --> 00:12:23,880 远比一位强行输出答案的通用模型更加可靠 306 00:12:23,880 --> 00:12:26,480 第三项设定带着一丝黑色幽默 307 00:12:26,480 --> 00:12:28,460 为极致压缩运行成本 308 00:12:28,460 --> 00:12:31,260 经过3000万轮强化学习迭代之后 309 00:12:31,260 --> 00:12:34,780 Inkling自主演化出高度精简的内部思考语言 310 00:12:34,780 --> 00:12:36,220 我们可以简单对比 311 00:12:36,220 --> 00:12:38,860 差异主流模型处理复杂问题时 312 00:12:38,860 --> 00:12:41,020 内部推理链条会使用规范 313 00:12:41,020 --> 00:12:42,360 完整英文语句 314 00:12:42,360 --> 00:12:44,540 需要消耗大量词源资源 315 00:12:44,540 --> 00:12:47,320 Inkling内部推演过程抛弃完整语法 316 00:12:47,320 --> 00:12:49,000 只保留核心关键词 317 00:12:49,000 --> 00:12:50,800 推理效果几乎没有衰减 318 00:12:50,800 --> 00:12:52,920 但是极大削减词源占用 319 00:12:52,920 --> 00:12:54,320 听起来十分有趣 320 00:12:54,320 --> 00:12:58,160 背后体现的是对推理效率极致压榨的工程思路 321 00:12:58,460 --> 00:13:03,400 读到这里大家就能明白,这从来不是一款用来冲击榜单排名的展示型模型, 322 00:13:03,800 --> 00:13:09,480 而是面向大规模企业商用量身打造,精打细算,稳定可靠的底层基础设施。 323 00:13:09,840 --> 00:13:14,980 即便Inclin拥有诸多精妙设计,仅仅依靠开源模型本身无法创造收入, 324 00:13:15,080 --> 00:13:22,060 能够支撑120亿估值完成变现的核心载体,是配套同步推出的Tinker自动化微调平台。 325 00:13:22,420 --> 00:13:26,740 这才是SY NTNG MacAINES Lab真正的商业核武器。 326 00:13:26,740 --> 00:13:28,060 前面我们提到 327 00:13:28,060 --> 00:13:29,380 Inkling完全开源 328 00:13:29,380 --> 00:13:31,540 任何人都可以免费下载部署 329 00:13:31,540 --> 00:13:33,660 企业不会为此支付一分钱 330 00:13:33,660 --> 00:13:35,600 可是真实落地场景当中 331 00:13:35,600 --> 00:13:38,900 大型企业技术负责人拿到模型权重之后 332 00:13:38,900 --> 00:13:40,700 很快会遭遇现实阻碍 333 00:13:40,700 --> 00:13:43,160 想要让模型理解自身行业业务 334 00:13:43,160 --> 00:13:44,600 读懂内部财务数据 335 00:13:44,600 --> 00:13:46,460 适配客户服务场景 336 00:13:46,460 --> 00:13:49,220 必须使用企业私有数据开展微调 337 00:13:49,220 --> 00:13:50,880 模型微调门槛极高 338 00:13:50,880 --> 00:13:53,440 研发人员需要精通整套训练架构 339 00:13:53,440 --> 00:13:55,660 持续维护GPU算力集群 340 00:13:55,660 --> 00:13:59,440 还要应对训练进程当中频繁出现的程序崩溃问题 341 00:13:59,440 --> 00:14:02,620 绝大多数传统企业并不具备对应的技术团队 342 00:14:02,620 --> 00:14:06,420 这个时候Thinking Machines Lab就可以顺势推出解决方案 343 00:14:06,420 --> 00:14:08,320 向企业推荐Tinker平台 344 00:14:08,320 --> 00:14:09,600 模型免费交付 345 00:14:09,600 --> 00:14:11,540 但是微调操作门槛过高 346 00:14:11,540 --> 00:14:12,940 借助Tinker平台 347 00:14:12,940 --> 00:14:14,980 使用者不用理解底层框架 348 00:14:14,980 --> 00:14:16,560 不用管理算力节点 349 00:14:16,560 --> 00:14:18,440 直接上传私有业务数据 350 00:14:18,440 --> 00:14:22,380 简单操作就能全自动训练专属垂直领域专家模型 351 00:14:22,380 --> 00:14:25,300 在官方演示当中还有极具冲击力的功能 352 00:14:25,300 --> 00:14:27,020 平台下达指令之后 353 00:14:27,020 --> 00:14:29,520 Inklin能够自主编写训练脚本 354 00:14:29,520 --> 00:14:31,680 自动生成配套训练数据集 355 00:14:31,680 --> 00:14:33,980 独立完成全部微调流程 356 00:14:33,980 --> 00:14:36,140 加在全新全重实现自我迭代 357 00:14:36,140 --> 00:14:38,020 达成全自动自主微调 358 00:14:38,020 --> 00:14:40,000 这正是米拉真正的野心 359 00:14:40,000 --> 00:14:42,240 她根本无意打造第二个OpenAI 360 00:14:42,240 --> 00:14:45,740 依靠售卖标准化按磁源计费的通用API盈利 361 00:14:45,740 --> 00:14:48,940 现如今基础大模型赛道早已红海厮杀 362 00:14:48,940 --> 00:14:50,860 通用模型数量众多 363 00:14:50,860 --> 00:14:53,220 不少开源产品在垂直场景 364 00:14:53,220 --> 00:14:54,740 性能超越必源方案 365 00:14:54,740 --> 00:14:56,160 Thinking Machines Lab 366 00:14:56,160 --> 00:14:58,180 真实商业模式十分清晰 367 00:14:58,180 --> 00:14:59,780 依靠可控性强 368 00:14:59,780 --> 00:15:00,720 运行稳定 369 00:15:00,720 --> 00:15:02,580 完全免费的开源模型 370 00:15:02,580 --> 00:15:03,780 作为引流载体 371 00:15:03,780 --> 00:15:07,220 吸引大量想要搭建私有人工智能系统的企业 372 00:15:07,220 --> 00:15:10,680 随后依托简易自动化微调平台锁定客户 373 00:15:10,680 --> 00:15:12,820 持续收取平台服务费用 374 00:15:12,820 --> 00:15:15,420 他们不对外售卖通用智能大脑 375 00:15:15,420 --> 00:15:19,020 而是对外提供批量定制行业专家模型的生产工具 376 00:15:19,020 --> 00:15:22,540 未来倘若数万企业借助Tinker搭建专属财物 377 00:15:22,540 --> 00:15:24,360 客服务医疗智能模型 378 00:15:24,360 --> 00:15:26,180 没有人再会持续纠结 379 00:15:26,180 --> 00:15:28,380 Inkling在评测榜单排名高低 380 00:15:28,380 --> 00:15:31,180 这笔20亿美元投入真的存在浪费吗 381 00:15:31,180 --> 00:15:32,380 答案是否定的 382 00:15:32,380 --> 00:15:33,860 米拉带领顶尖团队 383 00:15:33,860 --> 00:15:36,360 跳出OpenAI必源竞争的内卷赛道 384 00:15:36,360 --> 00:15:38,520 在企业定制人工智能赛道 385 00:15:38,520 --> 00:15:40,000 筑器极高护城河 386 00:15:40,000 --> 00:15:41,540 他们没有研发失败 387 00:15:41,540 --> 00:15:45,660 而是在下一盘足以重塑全球AI商业格局的大棋 388 00:15:45,660 --> 00:15:49,180 我们接着聊硅谷行业内部很少公开讨论的细节 389 00:15:49,180 --> 00:15:51,520 Inkling这套高效混合专家 390 00:15:51,520 --> 00:15:54,800 MOE架构背后存在不少值得深究的线索 391 00:15:54,800 --> 00:15:57,760 资深AI架构师研读底层代码之后 392 00:15:57,760 --> 00:16:00,000 能够发现256个路由专业 393 00:16:00,000 --> 00:16:10,000 資深AI架構師研讀底層代碼之後,能夠發現256個路由專家單元,Sigmoid路由機制,滑動窗口注意力配比,整套工程設計思路和國內開源大模型深度求索,Deep-Seek V3高度相似。 394 00:16:10,000 --> 00:16:10,720 除此之外 395 00:16:10,720 --> 00:16:12,480 模型早期预训练阶段 396 00:16:12,480 --> 00:16:14,200 Thinking Machines Lab大量 397 00:16:14,200 --> 00:16:16,320 运用成熟开源模型生成的 398 00:16:16,320 --> 00:16:17,780 合成数据辅助训练 399 00:16:17,780 --> 00:16:20,680 大幅降低训练成本与失措周期 400 00:16:20,680 --> 00:16:22,840 一家估值120亿美元 401 00:16:22,840 --> 00:16:25,860 纯正美国背景顶尖AI实验室 402 00:16:25,860 --> 00:16:27,900 由前OpenAI核心负责人戴队 403 00:16:27,900 --> 00:16:30,140 手握英伟达高端算力资源 404 00:16:30,140 --> 00:16:31,860 最终推出旗舰模型 405 00:16:31,860 --> 00:16:34,380 大量参考国内开源模型成熟的 406 00:16:34,380 --> 00:16:35,360 工程优化方案 407 00:16:35,360 --> 00:16:37,420 很多外行看到消息第一反应 408 00:16:37,420 --> 00:16:39,020 会质疑团队创新能力 409 00:16:39,020 --> 00:16:41,080 认为只是照搬现有成果 410 00:16:41,080 --> 00:16:43,360 站在工程与商业客观视角分析 411 00:16:43,360 --> 00:16:45,500 这是非常理性务实的选择 412 00:16:45,500 --> 00:16:48,580 最近几年海外高端算力芯片供给受限 413 00:16:48,580 --> 00:16:51,720 国内AI研发团队无法像OpenAI一样 414 00:16:51,720 --> 00:16:54,380 不计成本堆叠算力暴力训练模型 415 00:16:54,380 --> 00:16:56,300 只能持续在算法效率 416 00:16:56,300 --> 00:16:57,320 内存优化 417 00:16:57,320 --> 00:16:58,920 参数调度层面深耕 418 00:16:58,920 --> 00:17:02,220 倒逼开源模型运行效率突破全新高度 419 00:17:02,220 --> 00:17:04,120 米拉团队十分清楚这一点 420 00:17:04,120 --> 00:17:06,420 工程技术不存在面子之争 421 00:17:06,420 --> 00:17:09,560 既然同行已经验证高效率优化路线可行 422 00:17:09,560 --> 00:17:11,500 开源协议允许借鉴参考 423 00:17:11,500 --> 00:17:13,660 站在前人成果之上继续迭代 424 00:17:13,660 --> 00:17:15,580 就是最高效的选择 425 00:17:15,580 --> 00:17:17,560 也正是依托成熟的优化思路 426 00:17:17,560 --> 00:17:19,700 团队大幅压缩研发周期 427 00:17:19,700 --> 00:17:23,300 快速完成9750亿参数巨型模型的 428 00:17:23,300 --> 00:17:24,660 完整训练与迭代 429 00:17:24,660 --> 00:17:25,520 讲到这里 430 00:17:25,520 --> 00:17:27,720 大家会产生一个直击本质的疑问 431 00:17:27,720 --> 00:17:29,920 既然国内开源模型效率出众 432 00:17:29,920 --> 00:17:32,360 多款国产开源模型在代码 433 00:17:32,360 --> 00:17:33,300 逻辑推理层面 434 00:17:33,300 --> 00:17:35,600 能够对标国际顶级币源产品 435 00:17:35,600 --> 00:17:37,440 使用成本极其低廉 436 00:17:37,440 --> 00:17:41,320 美国本土企业为什么不直接选用现成国产开源模型 437 00:17:41,320 --> 00:17:43,480 何必耗费资金使用Inkling 438 00:17:43,480 --> 00:17:45,900 额外采购Tinker平台服务 439 00:17:45,900 --> 00:17:50,300 这个问题刚好戳破Thinking Machines Lab整套布局最核心的商业命脉 440 00:17:50,300 --> 00:17:52,960 事实上美国资本方并不缺乏理性 441 00:17:52,960 --> 00:17:54,860 在Inkling正式发布之前 442 00:17:54,860 --> 00:17:59,000 美国企业界曾经掀起一波使用海外开源模型的热炒 443 00:17:59,000 --> 00:18:03,160 查看OPE NRO UTR平台词源路由统计数据 444 00:18:03,160 --> 00:18:06,160 能够发现高峰期有大量企业调用流量 445 00:18:06,160 --> 00:18:07,720 转向海外开源模型 446 00:18:07,720 --> 00:18:11,720 Coinbase这类头部企业都悄悄迁移智能体业务 447 00:18:11,720 --> 00:18:13,780 直接把运营成本削减一半 448 00:18:13,780 --> 00:18:14,520 好用 449 00:18:14,520 --> 00:18:15,060 低价 450 00:18:15,060 --> 00:18:16,940 资本自然会用脚投票 451 00:18:16,940 --> 00:18:19,560 全球规模顶尖对冲基金桥水 452 00:18:19,560 --> 00:18:22,840 就曾利用自身业务数据微调国内开源模型 453 00:18:22,840 --> 00:18:25,060 在金融推理评测中跑出高分 454 00:18:25,060 --> 00:18:27,900 全面超越多款高价通用必源模型 455 00:18:27,900 --> 00:18:29,540 使用成本大幅降低 456 00:18:29,540 --> 00:18:32,920 低价定制垂直模型碾压昂贵通用大模型 457 00:18:32,920 --> 00:18:35,840 这笔账所有企业决策者都能够看懂 458 00:18:35,840 --> 00:18:38,080 但是繁荣周期并没有持续太久 459 00:18:38,080 --> 00:18:40,980 地缘政策带来的监管压力快速到来 460 00:18:40,980 --> 00:18:43,920 由于海外开源模型普及速度超出预期 461 00:18:43,920 --> 00:18:46,160 美国监管机构开始高度警惕 462 00:18:46,160 --> 00:18:49,680 持续向使用海外模型的企业发送警示信函 463 00:18:49,680 --> 00:18:51,800 国会相关委员会启动调研 464 00:18:51,800 --> 00:18:54,400 甚至有部门草拟采购限制条款 465 00:18:54,400 --> 00:18:56,280 手握海量隐私业务数据 466 00:18:56,280 --> 00:18:58,840 流水规模动辄百亿的华尔街投行 467 00:18:58,840 --> 00:19:02,760 医疗集团与科技大厂根本无法承受这类合规风险 468 00:19:02,760 --> 00:19:05,300 即便模型性能出色 价格低廉 469 00:19:05,300 --> 00:19:07,920 一旦被贴上数据安全隐患标签 470 00:19:07,920 --> 00:19:11,460 对于企业而言就是随时可能引爆的经营风险 471 00:19:11,460 --> 00:19:14,660 与此同时 曾经支撑西方开源生态的Meta 472 00:19:14,660 --> 00:19:16,860 近期持续收紧授权政策 473 00:19:16,860 --> 00:19:21,240 后续Lama系列模型逐步转向更加封闭的商业授权模式 474 00:19:21,240 --> 00:19:25,640 彻底终结了西方企业免费商用开源大模型的红利时代 475 00:19:25,640 --> 00:19:27,240 就在这个行业窗口期 476 00:19:27,240 --> 00:19:29,980 西方企业市场出现巨大需求真空 477 00:19:29,980 --> 00:19:33,100 企业迫切需要开源模型降低算力开销 478 00:19:33,100 --> 00:19:35,560 希望依靠微调保护私有数据 479 00:19:35,560 --> 00:19:37,740 但是不敢选用海外免费模型 480 00:19:37,740 --> 00:19:40,820 本土老牌开源路线又不断抬高准入门槛 481 00:19:40,820 --> 00:19:43,280 此刻再回看Inkling的发布时机 482 00:19:43,280 --> 00:19:44,740 一切都豁然开朗 483 00:19:44,740 --> 00:19:48,200 Inkling本身不需要争夺全球最强模型的名号 484 00:19:48,200 --> 00:19:49,380 它的真实定位 485 00:19:49,380 --> 00:19:51,800 是当下西方企业市场为数不多 486 00:19:51,800 --> 00:19:52,900 合规安全 487 00:19:52,900 --> 00:19:54,580 纯正美国本土血统 488 00:19:54,580 --> 00:19:57,380 且综合性能达标的开源备选方案 489 00:19:57,380 --> 00:20:00,720 米拉团队不仅交付一套安全可控的底层模型 490 00:20:00,720 --> 00:20:03,520 还同步推出轻量化版本Inclean Small 491 00:20:03,520 --> 00:20:07,500 这款缩小版本总参数规模压缩至2760亿 492 00:20:07,500 --> 00:20:09,840 单次推理仅激活120亿参数 493 00:20:09,840 --> 00:20:12,880 轻量化架构适配中小企业硬件设备 494 00:20:12,880 --> 00:20:15,600 在专项垂直推理场景中表现亮眼 495 00:20:15,600 --> 00:20:18,420 专门面向有本地私有化部署需求 496 00:20:18,420 --> 00:20:21,040 服务器硬件资源有限的中小企业 497 00:20:21,040 --> 00:20:23,920 硬件配置不足可以选用轻量化版本 498 00:20:23,920 --> 00:20:26,740 缺少技术人员直接采购Tinker平台 499 00:20:26,740 --> 00:20:28,660 担忧跨境数据合规风险 500 00:20:28,660 --> 00:20:30,920 它又是本土正规团队开发 501 00:20:30,920 --> 00:20:33,520 代码完整公开托管在哈格面部 502 00:20:33,520 --> 00:20:35,240 全部流程透明可查 503 00:20:35,240 --> 00:20:36,980 把所有线索串联起来 504 00:20:36,980 --> 00:20:40,140 Thinking Machines Lab完整商业版图清晰浮现 505 00:20:40,140 --> 00:20:41,960 离开OpenAI的MamalaD 506 00:20:41,960 --> 00:20:44,880 从一开始就没有打算在早已竞争白热化的 507 00:20:44,880 --> 00:20:46,180 通用大模型赛道 508 00:20:46,180 --> 00:20:48,040 复刻第二个ChatGPT 509 00:20:48,040 --> 00:20:51,780 那条赛道只能无休止消耗投资人资金持续内卷 510 00:20:51,780 --> 00:20:53,540 他依靠20亿美元资金 511 00:20:53,540 --> 00:20:55,760 在OpenAI币源商业帝国旁边 512 00:20:55,760 --> 00:20:56,960 开辟全新赛道 513 00:20:56,960 --> 00:20:59,000 精准预判人工智能未来形态 514 00:20:59,000 --> 00:21:01,680 不会是无所不能的通用超级大脑 515 00:21:01,680 --> 00:21:03,960 而是千千万万扎根细分行业 516 00:21:03,960 --> 00:21:07,900 低成本运行数据归属企业自身的领域专用模型 517 00:21:07,900 --> 00:21:10,640 借鉴经过市场验证的高效开源架构 518 00:21:10,640 --> 00:21:13,840 搭配团队成熟的成本控制方案打造Inkling 519 00:21:13,840 --> 00:21:15,100 作为引流载体 520 00:21:15,100 --> 00:21:17,720 再借助地缘环境催生的合规焦虑 521 00:21:17,720 --> 00:21:20,040 吸引急需解决方案的企业客户 522 00:21:20,040 --> 00:21:22,400 进入Tinker微调服务生态 523 00:21:22,400 --> 00:21:24,460 花费20亿美元打造一款榜单 524 00:21:24,460 --> 00:21:26,160 无法跻身前五的模型 525 00:21:26,160 --> 00:21:28,500 表面看上去像是一场失败投资 526 00:21:28,500 --> 00:21:31,300 但是倘若这20亿美元换来未来10年 527 00:21:31,300 --> 00:21:34,120 全球企业及AI定制市场核心入口 528 00:21:34,120 --> 00:21:36,620 这就是教科书级别的商业布局 529 00:21:36,620 --> 00:21:39,960 当所有人紧盯评测排行榜相互竞争的时候 530 00:21:39,960 --> 00:21:43,620 已经有人提前锁定人工智能淘金浪潮当中 531 00:21:43,620 --> 00:21:45,480 合规产子的独家供应权 532 00:21:45,480 --> 00:21:47,800 这也是120亿美元估值背后 533 00:21:47,800 --> 00:21:50,240 最值得细细思考的行业真相 534 00:21:50,240 --> 00:21:51,840 其实我们在延伸一层 535 00:21:51,840 --> 00:21:55,080 你就能看懂硅谷资本更深层的赌具逻辑 536 00:21:55,080 --> 00:21:58,000 近两年无数AI创业公司扎堆入场 537 00:21:58,000 --> 00:22:00,220 大多都在跟风做通用大模型 538 00:22:00,220 --> 00:22:01,500 做对话机器人 539 00:22:01,500 --> 00:22:03,140 做花哨的多模态演示 540 00:22:03,140 --> 00:22:05,840 真正能落地赚钱的寥寥无几 541 00:22:05,840 --> 00:22:07,820 绝大多数AI初创企业 542 00:22:07,820 --> 00:22:09,480 盈利模式极度单一 543 00:22:09,480 --> 00:22:10,980 要么卖API调用 544 00:22:10,980 --> 00:22:12,260 要么卖会员套餐 545 00:22:12,260 --> 00:22:14,760 陷入无尽的价格战和算力内卷 546 00:22:14,760 --> 00:22:16,280 稍万融资就倒闭 547 00:22:16,280 --> 00:22:17,560 存活率极低 548 00:22:17,800 --> 00:22:20,320 而SY NTNG Mac King ES Lab 549 00:22:20,320 --> 00:22:21,840 彻底避开了这片红海 550 00:22:21,840 --> 00:22:23,940 它不跟任何人抢C端用户 551 00:22:23,940 --> 00:22:24,900 不卷跑分 552 00:22:24,900 --> 00:22:26,140 不卷花式功能 553 00:22:26,140 --> 00:22:27,860 直接切入最赚钱 554 00:22:27,860 --> 00:22:28,500 最稳定 555 00:22:28,500 --> 00:22:31,000 刚需最强的弊端企业服务市场 556 00:22:31,000 --> 00:22:32,720 对于大型企业来说 557 00:22:32,720 --> 00:22:33,600 AI成本 558 00:22:33,600 --> 00:22:34,540 数据安全 559 00:22:34,540 --> 00:22:35,620 私有化部署 560 00:22:35,620 --> 00:22:36,640 定制化适配 561 00:22:36,640 --> 00:22:38,340 是比模型聪不聪明 562 00:22:38,340 --> 00:22:39,560 更重要的刚需 563 00:22:39,560 --> 00:22:41,260 哪怕通用能力平庸 564 00:22:41,260 --> 00:22:42,240 只要足够稳 565 00:22:42,240 --> 00:22:43,180 足够便宜 566 00:22:43,180 --> 00:22:44,120 足够安全 567 00:22:44,120 --> 00:22:45,220 足够好微调 568 00:22:45,220 --> 00:22:46,540 就是企业愿意 569 00:22:46,540 --> 00:22:48,240 长期付费的核心资产 570 00:22:48,240 --> 00:22:49,940 这也是为什么A16Z 571 00:22:49,940 --> 00:22:51,820 敢于在没有成熟产品 572 00:22:51,820 --> 00:22:53,700 没有实际营收的情况下 573 00:22:53,700 --> 00:22:54,740 豪至20亿美金 574 00:22:54,740 --> 00:22:57,040 资本赌的不是一款模型的输赢 575 00:22:57,040 --> 00:22:59,360 而是赌未来企业AI私有化 576 00:22:59,360 --> 00:23:00,160 定制化 577 00:23:00,160 --> 00:23:02,020 合规化的万亿级风口 578 00:23:02,020 --> 00:23:04,320 反观行业内很多盲目堆算力 579 00:23:04,320 --> 00:23:05,100 卷参数 580 00:23:05,100 --> 00:23:06,220 卷跑分的团队 581 00:23:06,220 --> 00:23:07,480 看似热度爆棚 582 00:23:07,480 --> 00:23:08,220 数据亮眼 583 00:23:08,220 --> 00:23:10,680 实则完全偏离了真实商业需求 584 00:23:10,680 --> 00:23:11,860 最终只能沦为 585 00:23:11,860 --> 00:23:13,380 昙花一现的行业过客 586 00:23:13,380 --> 00:23:14,340 除此之外 587 00:23:14,340 --> 00:23:15,560 米拉团队的布局 588 00:23:15,560 --> 00:23:17,440 还有一个极具远见的细节 589 00:23:17,440 --> 00:23:19,040 那就是彻底放弃 590 00:23:19,040 --> 00:23:20,540 避原锁库的传统玩法 591 00:23:20,540 --> 00:23:22,540 坚持全开源 全透明 592 00:23:22,540 --> 00:23:24,600 在当下AI舆论环境中 593 00:23:24,600 --> 00:23:27,020 避原模型随时面临授权变动 594 00:23:27,020 --> 00:23:28,900 服务关停 接口涨价 595 00:23:28,900 --> 00:23:30,180 功能阉割的风险 596 00:23:30,180 --> 00:23:32,340 企业用户极其没有安全感 597 00:23:32,340 --> 00:23:34,820 而Inkling的Apache 2.0协议 598 00:23:34,820 --> 00:23:36,900 给予了企业最大的自由度 599 00:23:36,900 --> 00:23:38,840 可商用 可二次开发 600 00:23:38,840 --> 00:23:41,340 可私有化部署 无板权捆绑 601 00:23:41,340 --> 00:23:43,140 这种极致开放的姿态 602 00:23:43,140 --> 00:23:45,860 刚好击中了当下企业最核心的诉求 603 00:23:45,860 --> 00:23:47,080 免费拿模型 604 00:23:47,080 --> 00:23:48,180 免费改模型 605 00:23:48,180 --> 00:23:49,020 免费部署 606 00:23:49,020 --> 00:23:51,360 最后只为精细化微调服务 607 00:23:51,360 --> 00:23:53,060 自动化运为工具付费 608 00:23:53,060 --> 00:23:54,300 这种先让利 609 00:23:54,300 --> 00:23:55,980 后所刻的商业模式 610 00:23:55,980 --> 00:23:58,300 远比一次性售卖API要长久 611 00:23:58,300 --> 00:23:59,160 稳固得多 612 00:24:00,000 --> 00:24:12,440 很多人始終想不通一個核心問題,為什麼無數大廠耗費巨資打造的頂級通用大模型,落地企業服務時頻頻翻車,反而是Inclin這種跑分中庸的模型,精準踩中了行業痛點。 613 00:24:12,440 --> 00:24:15,400 答案其实藏在企业真实的AI落地困境里 614 00:24:15,400 --> 00:24:17,160 绝大多数传统企业 615 00:24:17,160 --> 00:24:18,100 金融机构 616 00:24:18,100 --> 00:24:19,080 医疗集团 617 00:24:19,080 --> 00:24:20,220 制造业工厂 618 00:24:20,220 --> 00:24:22,200 根本不需要模型创作诗歌 619 00:24:22,200 --> 00:24:23,020 撰写文案 620 00:24:23,020 --> 00:24:24,140 完成趣味问答 621 00:24:24,140 --> 00:24:26,540 也不需要极致的顶尖推力跑分 622 00:24:26,540 --> 00:24:28,600 他们的核心需求极其朴素 623 00:24:28,600 --> 00:24:30,400 就是稳定低成本 624 00:24:30,400 --> 00:24:30,980 可管控 625 00:24:30,980 --> 00:24:31,680 不幻觉 626 00:24:31,680 --> 00:24:33,480 能对接自身业务数据 627 00:24:33,480 --> 00:24:35,280 同时全程保障数据 628 00:24:35,280 --> 00:24:35,920 不外泄 629 00:24:35,920 --> 00:24:36,920 过去几年 630 00:24:36,920 --> 00:24:40,220 大量企业跟风接入各类顶尖臂源大模型API 631 00:24:40,220 --> 00:24:42,780 看似快速完成了AI数字化升级 632 00:24:42,780 --> 00:24:44,540 实则埋下了无数隐患 633 00:24:44,540 --> 00:24:47,000 首先是持续走高的调用成本 634 00:24:47,000 --> 00:24:51,100 企业日均数万数十万次的重复性基础调用 635 00:24:51,100 --> 00:24:52,700 累积下来的算力开支 636 00:24:52,700 --> 00:24:54,800 每年动辄数百万甚至上千万 637 00:24:54,800 --> 00:24:56,600 长期投入完全没有上线 638 00:24:56,600 --> 00:24:58,560 其次是数据泄露风险 639 00:24:58,560 --> 00:25:00,460 企业内部的客户资料 640 00:25:00,460 --> 00:25:01,460 财务报表 641 00:25:01,460 --> 00:25:02,540 生产数据 642 00:25:02,540 --> 00:25:03,560 并立信息 643 00:25:03,560 --> 00:25:06,700 一旦上传至第三方B元模型云端服务器 644 00:25:06,700 --> 00:25:08,600 就彻底脱离了自身管控 645 00:25:08,600 --> 00:25:10,760 随时可能面临数据抓取 646 00:25:10,760 --> 00:25:11,960 缓存留存 647 00:25:11,960 --> 00:25:13,440 合规溯源的风险 648 00:25:13,440 --> 00:25:14,620 更关键的是 649 00:25:14,620 --> 00:25:16,780 通用大模型的过度聪明 650 00:25:16,780 --> 00:25:18,800 反而成为企业落地的累赘 651 00:25:18,800 --> 00:25:21,040 很多企业业务只需要标准化 652 00:25:21,040 --> 00:25:22,400 流程化的固定输出 653 00:25:22,400 --> 00:25:24,220 不需要模型自由发挥 654 00:25:24,220 --> 00:25:25,340 发散创作 655 00:25:25,340 --> 00:25:28,040 通用模型自带的高自由度推理能力 656 00:25:28,040 --> 00:25:29,700 不仅会增加算力损耗 657 00:25:29,700 --> 00:25:32,300 还容易输出不符合企业规章制度 658 00:25:32,300 --> 00:25:33,480 行业规范的内容 659 00:25:33,480 --> 00:25:35,480 需要人工二次审核修正 660 00:25:35,480 --> 00:25:37,180 反而降低了工作效率 661 00:25:37,180 --> 00:25:38,920 而Inclin的全套设计 662 00:25:38,920 --> 00:25:41,480 恰好针对性解决了这些行业顽疾 663 00:25:41,480 --> 00:25:44,160 可调节的思考力度杜绝算力浪费 664 00:25:44,160 --> 00:25:46,100 极致诚实的输出特性 665 00:25:46,100 --> 00:25:47,960 规避业务幻觉风险 666 00:25:47,960 --> 00:25:50,560 轻量化部署适配各类企业硬件条件 667 00:25:50,560 --> 00:25:52,100 完全开源的属性 668 00:25:52,100 --> 00:25:53,800 满足私有化部署需求 669 00:25:53,800 --> 00:25:56,260 再搭配Tinker全自动微调平台 670 00:25:56,260 --> 00:25:59,160 彻底抹平了企业AI落地的技术鸿沟 671 00:25:59,160 --> 00:26:02,080 哪怕是没有专业算法团队的传统企业 672 00:26:02,080 --> 00:26:05,060 也能轻松搭建专属行业AI系统 673 00:26:05,060 --> 00:26:07,520 这种精准贴合商业落地的产品逻辑 674 00:26:07,520 --> 00:26:11,320 也是它能够碾压一众跑分更强模型的核心底气 675 00:26:11,320 --> 00:26:12,280 除此之外 676 00:26:12,280 --> 00:26:16,020 我们还能从行业迭代周期看懂这套布局的前瞻性 677 00:26:16,020 --> 00:26:19,520 如今通用大模型的技术迭代已经进入瓶颈期 678 00:26:19,520 --> 00:26:22,000 性能提升的边际效益持续递减 679 00:26:22,000 --> 00:26:25,740 单纯比拼参数和算力已经没有实质商业意义 680 00:26:25,740 --> 00:26:30,460 未来AI行业的竞争早已从技术参数竞赛转向商业落地竞赛 681 00:26:30,460 --> 00:26:32,540 谁能更低成本更合规 682 00:26:32,540 --> 00:26:34,420 更便捷地帮助各行各业 683 00:26:34,420 --> 00:26:35,940 完成AI私有化 684 00:26:35,940 --> 00:26:36,940 定制化改造 685 00:26:36,940 --> 00:26:38,740 谁就能掌握行业话语权 686 00:26:38,740 --> 00:26:41,320 米拉团队放弃无意义的跑分内卷 687 00:26:41,320 --> 00:26:42,880 深耕企业服务赛道 688 00:26:42,880 --> 00:26:46,300 本质是提前预判了AI行业的终局形态 689 00:26:46,300 --> 00:26:47,880 放弃短期的流量热度 690 00:26:47,880 --> 00:26:50,100 锁定了长期的行业垄断地位 691 00:26:50,100 --> 00:26:51,340 看完这期视频 692 00:26:51,340 --> 00:26:54,780 你如何看待大模型一味比拼跑分的内卷现状 693 00:26:54,780 --> 00:26:57,040 又觉得未来企业AI定制服务 694 00:26:57,040 --> 00:26:58,540 会走向怎样的格局 695 00:26:58,540 --> 00:27:00,580 欢迎在评论区理性交流 696 00:27:00,580 --> 00:27:03,680 如果你觉得本期深度行业解读干货充足 697 00:27:03,680 --> 00:27:05,680 记得点赞并且订阅频道 698 00:27:05,680 --> 00:27:08,600 后续持续跟踪全球AI巨头的战略博弈 699 00:27:08,600 --> 00:27:10,700 不要错过更多前沿产业分析 700 00:27:10,700 --> 00:27:12,040 我们下期再见