1 00:00:00,000 --> 00:00:30,000 一台8級字節記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記記 2 00:00:30,000 --> 00:00:33,960 A426B、A4B跑到了Apple Silicon Mac上 3 00:00:33,960 --> 00:00:36,340 而且運行時權重加4K 4 00:00:36,340 --> 00:00:39,140 KV cache只佔大約二級字節內存 5 00:00:39,140 --> 00:00:40,940 這個數字最狠的地方 6 00:00:40,940 --> 00:00:42,580 不是省了一點內存 7 00:00:42,580 --> 00:00:46,320 而是把本地AI的門檻直接往下砸了一層 8 00:00:46,320 --> 00:00:49,880 原本14.3GB左右的模型安裝體積 9 00:00:49,880 --> 00:00:51,960 不需要整包塞進內存 10 00:00:51,960 --> 00:00:54,920 項目只把常駐核心留在內存裡 11 00:00:54,920 --> 00:00:58,000 把大部分專家權重放在SSD上 12 00:00:58,000 --> 00:01:00,000 需要哪一塊再讀哪一塊 13 00:01:00,000 --> 00:01:03,500 視頻裡在M3 Mac上跑到23.4 tokens 14 00:01:03,500 --> 00:01:07,240 項目ReadMe裡也記錄了8GBM2 MacBook Air 15 00:01:07,240 --> 00:01:10,400 可以跑到5.1到6.3 tokens 16 00:01:10,400 --> 00:01:13,800 這個速度不是數據中心級別 17 00:01:13,800 --> 00:01:15,440 但已經不是玩具 18 00:01:15,440 --> 00:01:17,400 這件事真正值得看 19 00:01:17,400 --> 00:01:20,160 不是某個開發者做了一個炫技項目 20 00:01:20,160 --> 00:01:22,860 而是本地AI的路線開始變了 21 00:01:22,860 --> 00:01:25,800 過去本地大模型的思路很簡單 22 00:01:25,800 --> 00:01:27,960 模型越大 硬件越貴 23 00:01:27,960 --> 00:01:29,360 想跑7B 24 00:01:29,360 --> 00:01:31,940 准备一块不错的消费级显卡 25 00:01:31,940 --> 00:01:33,600 想跑30B 26 00:01:33,600 --> 00:01:35,220 开始考虑大显存 27 00:01:35,220 --> 00:01:37,160 想跑更大的MOE 28 00:01:37,160 --> 00:01:38,760 就看云服务账单 29 00:01:38,760 --> 00:01:41,400 用户被迫接受一个隐含规则 30 00:01:41,400 --> 00:01:42,920 AI越聪明 31 00:01:42,920 --> 00:01:44,880 越远离个人电脑 32 00:01:44,880 --> 00:01:46,140 越强的模型 33 00:01:46,140 --> 00:01:47,760 越集中在云厂商 34 00:01:47,760 --> 00:01:50,440 GPU集群和数据中心手里 35 00:01:50,440 --> 00:01:53,540 TurboFuelFair给出的反方向答案是 36 00:01:53,540 --> 00:01:55,180 模型可以很大 37 00:01:55,180 --> 00:01:57,280 但每一秒真正用到的部分 38 00:01:57,280 --> 00:01:58,160 可能很小 39 00:01:58,160 --> 00:01:59,840 这要从Moei讲起 40 00:01:59,840 --> 00:02:02,260 Moei叫mixture of experts 41 00:02:02,260 --> 00:02:03,960 混合专家模型 42 00:02:03,960 --> 00:02:04,960 普通模型 43 00:02:04,960 --> 00:02:07,100 像一个巨大的统一车间 44 00:02:07,100 --> 00:02:08,580 每个Token进来 45 00:02:08,580 --> 00:02:10,560 很多权重都要参与计算 46 00:02:10,560 --> 00:02:14,600 MOE更像一座有128个小工位的工厂 47 00:02:14,600 --> 00:02:15,820 每个Token进来 48 00:02:15,820 --> 00:02:19,020 Router会判断这次该找哪几个专家处理 49 00:02:19,020 --> 00:02:20,740 Gamma 4 26B 50 00:02:20,740 --> 00:02:21,640 A4B 51 00:02:21,640 --> 00:02:23,640 总参数是26B 52 00:02:23,640 --> 00:02:25,400 但每个Token实际激活的 53 00:02:25,400 --> 00:02:27,640 大約是3.88B參數 54 00:02:27,640 --> 00:02:29,260 模型名義上很大 55 00:02:29,260 --> 00:02:32,680 但每一部真正幹活的只是其中一部分 56 00:02:32,680 --> 00:02:36,080 過去這件事的好處主要體現在雲端 57 00:02:36,080 --> 00:02:38,820 模型公司可以用更大的總參數 58 00:02:38,820 --> 00:02:41,280 保持相對可控的推理成本 59 00:02:41,280 --> 00:02:45,820 可是Turbo Fieldfare把這個特性拿到了本地機器上 60 00:02:45,820 --> 00:02:48,600 既然每個Token只用幾個專家 61 00:02:48,600 --> 00:02:51,860 那為什麼要把所有專家都常駐內存 62 00:02:51,860 --> 00:02:54,320 於是他把模型拆成兩堆 63 00:02:54,320 --> 00:02:57,000 第一堆是每個Token都要用的東西 64 00:02:57,000 --> 00:02:58,500 比如Attention 65 00:02:58,500 --> 00:02:59,560 Router 66 00:02:59,560 --> 00:03:00,720 Embedding 67 00:03:00,720 --> 00:03:02,200 Shared Expert 68 00:03:02,200 --> 00:03:03,860 還有KVcash 69 00:03:03,860 --> 00:03:06,460 這一部分大約1.35級字節 70 00:03:06,460 --> 00:03:08,640 加上運行需要的緩存 71 00:03:08,640 --> 00:03:09,960 留在內存裡 72 00:03:09,960 --> 00:03:11,940 第二堆是專家權重 73 00:03:11,940 --> 00:03:13,140 30層 74 00:03:13,140 --> 00:03:15,300 每層128個Expert 75 00:03:15,300 --> 00:03:17,300 每個Expert只有幾MB 76 00:03:17,300 --> 00:03:19,120 加起來是主要體積 77 00:03:19,120 --> 00:03:20,980 這一堆不常駐內存 78 00:03:20,980 --> 00:03:22,820 只放在SSD上 79 00:03:22,820 --> 00:03:24,260 每生成一個Token 80 00:03:24,260 --> 00:03:26,020 模型先完成Attention 81 00:03:26,020 --> 00:03:29,760 再由Router決定當前程要用哪8個專家 82 00:03:29,760 --> 00:03:31,420 CPU接到名單之後 83 00:03:31,420 --> 00:03:33,900 去SSD讀取對應Expert 84 00:03:33,900 --> 00:03:36,700 放到GPU能看到的內存區域裡 85 00:03:36,700 --> 00:03:38,260 Metal接著算 86 00:03:38,260 --> 00:03:40,140 這個過程聽起來麻煩 87 00:03:40,140 --> 00:03:42,300 但關鍵在Apple Silicon 88 00:03:42,300 --> 00:03:43,580 傳統PC上 89 00:03:43,580 --> 00:03:47,340 CPU和獨立GPU通常分兩套內存 90 00:03:47,340 --> 00:03:50,620 SSD讀出來的數據先進系統內存 91 00:03:50,620 --> 00:03:53,820 再通過PCIe總線搬到顯卡VRAM 92 00:03:53,820 --> 00:03:57,820 模型如果每個Token都要頻繁從SSD拉權重 93 00:03:57,820 --> 00:04:01,200 再搬進VRAM中間的拷貝和總線延遲 94 00:04:01,200 --> 00:04:03,260 會直接把速度打穿 95 00:04:03,260 --> 00:04:04,700 數據不是不會動 96 00:04:04,700 --> 00:04:06,860 而是動得太慢、動得太貴 97 00:04:06,860 --> 00:04:11,000 Apple Silicon的統一內存架構把這個問題變小了 98 00:04:11,000 --> 00:04:14,540 CPU和GPU看的是同一塊物理內存 99 00:04:14,540 --> 00:04:17,040 CPU從SSD讀進來的數據 100 00:04:17,040 --> 00:04:20,480 可以直接成為GPU要用的Metal Buffer 101 00:04:20,480 --> 00:04:21,880 少了一次拷貝 102 00:04:21,880 --> 00:04:23,780 少了一段總線搬運 103 00:04:23,780 --> 00:04:25,880 也少了傳統讀顯架構裡 104 00:04:25,880 --> 00:04:27,820 最難受的VRAM牆 105 00:04:27,820 --> 00:04:30,980 這就是為什麼這個項目特別像一把鑰匙 106 00:04:30,980 --> 00:04:34,280 它不是證明蘋果芯片算力天下無敵 107 00:04:34,280 --> 00:04:36,580 而是證明Apple Silicon的結構 108 00:04:36,580 --> 00:04:40,480 剛好適合一種邊讀邊算的本地推理路線 109 00:04:40,480 --> 00:04:41,580 還不只這個 110 00:04:41,580 --> 00:04:44,380 Turbo Fieldfare沒有把磁盤上的權重 111 00:04:44,380 --> 00:04:46,280 用普通格式存著 112 00:04:46,280 --> 00:04:48,780 等讀取之後再解包 113 00:04:48,780 --> 00:04:51,340 再轉成GPU需要的佈局 114 00:04:51,340 --> 00:04:55,700 他在安裝階段就把模型重新打包成Double格式 115 00:04:55,700 --> 00:05:00,040 盡量讓磁盤裡的數據就是Metal kernel可以消費的樣子 116 00:05:00,040 --> 00:05:02,600 讀文件就是加載權重 117 00:05:02,600 --> 00:05:06,960 中間少一次轉換就少一次內存浪費和時間浪費 118 00:05:06,960 --> 00:05:08,240 他還用了緩存 119 00:05:08,240 --> 00:05:10,540 每一層128個expert 120 00:05:10,540 --> 00:05:13,620 不可能每次都從SSD讀 121 00:05:13,620 --> 00:05:16,180 項目給每層留了16個expert 122 00:05:16,180 --> 00:05:19,000 槽位常用的expert放在內存裡 123 00:05:19,000 --> 00:05:21,900 Router如果選中已經緩存的expert 124 00:05:21,900 --> 00:05:23,600 立刻就能用 125 00:05:23,600 --> 00:05:24,880 如果沒命中 126 00:05:24,880 --> 00:05:26,680 再從SSD讀取 127 00:05:26,680 --> 00:05:28,340 把不常用的擠出去 128 00:05:28,340 --> 00:05:29,500 他用的是 129 00:05:29,500 --> 00:05:30,780 LFU 130 00:05:30,780 --> 00:05:32,820 Least Frequently Used 131 00:05:32,820 --> 00:05:35,760 也就是踢掉使用頻率最低的專家 132 00:05:35,760 --> 00:05:38,200 而不是簡單踢掉最久沒用的 133 00:05:38,200 --> 00:05:39,820 這個選擇很重要 134 00:05:39,820 --> 00:05:42,940 因為MOE的路由並不是完全隨機 135 00:05:42,940 --> 00:05:46,520 有些專家在很多Token上都會被反覆選中 136 00:05:46,520 --> 00:05:48,320 有些專家很少出現 137 00:05:48,320 --> 00:05:50,620 按頻率留下熱門專家 138 00:05:50,620 --> 00:05:54,200 比按時間留下最近專家更適合這個任務 139 00:05:54,200 --> 00:05:55,740 項目壓住的是 140 00:05:55,740 --> 00:05:58,560 語言深層的專家選擇有規律 141 00:05:58,560 --> 00:06:00,100 只要規律足夠強 142 00:06:00,100 --> 00:06:02,900 SSD讀取就不會把速度拖死 143 00:06:02,900 --> 00:06:05,980 所以它能做到一個很奇怪的效果 144 00:06:05,980 --> 00:06:10,080 明明模型大部分權重還在SSD上 145 00:06:10,080 --> 00:06:12,840 实际体验却没有慢到不能用 146 00:06:12,840 --> 00:06:15,620 这件事对普通用户意味着什么 147 00:06:15,620 --> 00:06:16,300 第一 148 00:06:16,300 --> 00:06:19,080 本地AI不一定只能靠堆硬件 149 00:06:19,080 --> 00:06:21,160 过去用户想跑更好的模型 150 00:06:21,160 --> 00:06:23,540 唯一办法是买更大内存 151 00:06:23,540 --> 00:06:24,780 更大显存 152 00:06:24,780 --> 00:06:25,900 更贵显卡 153 00:06:25,900 --> 00:06:29,560 这个逻辑会把个人电脑推向数据中心的反面 154 00:06:29,560 --> 00:06:31,140 云越来越强 155 00:06:31,140 --> 00:06:33,140 个人设备越来越像终端 156 00:06:33,140 --> 00:06:36,140 Turbo Field Fair展示的是另一条路 157 00:06:36,140 --> 00:06:37,840 如果模型结构 158 00:06:37,840 --> 00:06:38,760 文件布局 159 00:06:38,760 --> 00:06:43,160 操作系統、芯片內存架構和運行時配合的足夠好 160 00:06:43,160 --> 00:06:46,480 小機器也能吃下一部分大模型能力 161 00:06:46,480 --> 00:06:48,100 它不是免費午餐 162 00:06:48,100 --> 00:06:50,280 SSD讀取有延遲模型 163 00:06:50,280 --> 00:06:54,060 限制在特定結構速度和雲端旗艦模型沒法比 164 00:06:54,060 --> 00:06:55,940 但它證明了一個方向 165 00:06:55,940 --> 00:06:59,200 優化路線不是只剩買更貴的GPU 166 00:06:59,200 --> 00:07:00,020 第二 167 00:07:00,020 --> 00:07:03,600 Apple Silicon的AI價值可能被低估了 168 00:07:03,600 --> 00:07:05,340 過去評價AI硬件 169 00:07:05,340 --> 00:07:07,040 大家喜歡看TOPS 170 00:07:07,040 --> 00:07:08,300 看GPU 171 00:07:08,300 --> 00:07:09,720 看顯存大小 172 00:07:09,720 --> 00:07:12,600 蘋果在這場敘事裡經常顯得尷尬 173 00:07:12,600 --> 00:07:14,220 它的Neural Engine很強 174 00:07:14,220 --> 00:07:16,140 但開發生態不如Cuda 175 00:07:16,140 --> 00:07:17,700 統一內存很漂亮 176 00:07:17,700 --> 00:07:19,660 但高配內存價格很貴 177 00:07:19,660 --> 00:07:21,920 Mac很適合創作和開發 178 00:07:21,920 --> 00:07:23,940 但說到本地大模型 179 00:07:23,940 --> 00:07:26,080 很多人還是先想到NVIDIA 180 00:07:26,080 --> 00:07:29,040 Turbo FeelFair把問題換了一個角度 181 00:07:29,040 --> 00:07:32,000 蘋果真正的優勢不一定是單點算力 182 00:07:32,000 --> 00:07:36,280 而是CPU、GPU、內存、SSD、Metal 183 00:07:36,280 --> 00:07:38,200 和系統API的整合 184 00:07:38,200 --> 00:07:40,720 只要工作負債設計的夠貼合 185 00:07:40,720 --> 00:07:43,680 它可以用更少的數據搬運換性能 186 00:07:43,680 --> 00:07:45,440 這和手機時代很像 187 00:07:45,440 --> 00:07:48,600 蘋果不一定每個硬件參數都最大 188 00:07:48,600 --> 00:07:54,140 但它能把芯片、系統、應用和框架整在一起 189 00:07:54,140 --> 00:07:56,380 讓開發者吃到一體化紅利 190 00:08:04,854 --> 00:08:10,994 寫作、翻譯、代碼補權、文件整理、私人知識庫、 191 00:08:11,254 --> 00:08:14,334 離線助手、隱私敏感的資料分析 192 00:08:14,574 --> 00:08:17,134 如果這些場景能在普通Mac上運行 193 00:08:17,654 --> 00:08:19,194 蘋果就不是旁觀者 194 00:08:19,694 --> 00:08:20,214 第三 195 00:08:20,474 --> 00:08:23,034 雲端AI的壟斷感會被削弱 196 00:08:23,294 --> 00:08:25,854 過去AI服務天然集中在雲端 197 00:08:26,094 --> 00:08:27,134 因為模型太大 198 00:08:27,374 --> 00:08:28,154 推理太貴 199 00:08:28,414 --> 00:08:29,694 用戶設備跑不動 200 00:08:29,934 --> 00:08:31,734 集中在雲端就帶來幾個問題 201 00:08:31,994 --> 00:08:33,274 數據要上傳 202 00:08:33,534 --> 00:08:34,554 隱私要交出去 203 00:08:34,554 --> 00:08:36,154 订阅费要长期付 204 00:08:36,154 --> 00:08:37,874 网络断了就没法用 205 00:08:37,874 --> 00:08:39,354 模型公司改规则 206 00:08:39,354 --> 00:08:40,834 用户只能接受 207 00:08:40,834 --> 00:08:41,654 本地AI 208 00:08:41,654 --> 00:08:43,094 如果慢慢可用 209 00:08:43,094 --> 00:08:45,074 用户会多一个选择 210 00:08:45,074 --> 00:08:46,954 不是所有任务都要上云 211 00:08:46,954 --> 00:08:48,074 私人笔记 212 00:08:48,074 --> 00:08:49,494 公司内部文档 213 00:08:49,494 --> 00:08:50,914 离线代码助手 214 00:08:50,914 --> 00:08:51,974 本地搜索 215 00:08:51,974 --> 00:08:53,174 个人自动化 216 00:08:53,174 --> 00:08:56,114 这些场景其实很适合在设备端完成 217 00:08:56,114 --> 00:08:57,934 云端负责最强模型 218 00:08:57,934 --> 00:08:59,474 本地负责高频 219 00:08:59,474 --> 00:09:00,274 隐私 220 00:09:00,274 --> 00:09:02,054 低成本和机式响应 221 00:09:02,054 --> 00:09:03,754 这会改变产品形态 222 00:09:03,754 --> 00:09:07,594 未來AI助手可能不是一個純雲端聊天窗口 223 00:09:07,594 --> 00:09:09,294 而是一套混合系統 224 00:09:09,294 --> 00:09:10,994 簡單任務本地完成 225 00:09:10,994 --> 00:09:12,634 複雜任務再上雲 226 00:09:12,634 --> 00:09:14,734 敏感文件本地分析 227 00:09:14,734 --> 00:09:16,674 公開資料雲端補充 228 00:09:16,674 --> 00:09:18,914 離線場景用本地模型 229 00:09:18,914 --> 00:09:20,774 聯網場景用大模型 230 00:09:20,774 --> 00:09:22,854 用戶感覺不到背後切換 231 00:09:22,854 --> 00:09:24,614 只知道電腦更聰明了 232 00:09:24,614 --> 00:09:26,934 這也是蘋果真正想要的方向 233 00:09:26,934 --> 00:09:29,154 蘋果不會輕易把用戶數據 234 00:09:29,154 --> 00:09:30,974 全部交給第三方雲模型 235 00:09:30,974 --> 00:09:34,814 他更喜欢把智能功能藏进设备和系统 236 00:09:34,814 --> 00:09:39,274 Apple Intelligence的战略一直强调设备端处理和隐私边界 237 00:09:39,274 --> 00:09:40,534 问题在于 238 00:09:40,534 --> 00:09:42,834 设备端模型能力如果太弱 239 00:09:42,834 --> 00:09:45,074 体验就会被云端模型拉开 240 00:09:45,074 --> 00:09:46,494 Turbo Fieldfare 241 00:09:46,494 --> 00:09:50,054 这种项目说明设备端的能力上限 242 00:09:50,054 --> 00:09:51,634 还有很多工程空间 243 00:09:51,634 --> 00:09:53,834 这里要把苹果的处境讲透 244 00:09:53,834 --> 00:09:55,574 AI这轮浪潮里 245 00:09:55,574 --> 00:09:57,354 苹果一直显得慢半排 246 00:09:57,354 --> 00:09:59,474 OpenAI抢走聊天入口 247 00:09:59,474 --> 00:10:01,514 Google抢搜索和Android 248 00:10:01,514 --> 00:10:04,914 Microsoft把Copilot塞进Windows和Office 249 00:10:04,914 --> 00:10:06,734 NVIDIA拿走算力叙事 250 00:10:06,734 --> 00:10:09,394 苹果虽然发布Apple Intelligence 251 00:10:09,394 --> 00:10:11,894 但市场反应一直不算兴奋 252 00:10:11,894 --> 00:10:13,194 原因很简单 253 00:10:13,194 --> 00:10:16,894 苹果不是靠开放API和云模型赚钱的公司 254 00:10:16,894 --> 00:10:19,914 它擅长的是把能力变成系统体验 255 00:10:19,914 --> 00:10:22,994 可大模型初期最耀眼的能力都在云端 256 00:10:22,994 --> 00:10:25,514 苹果的优势一时很难展示 257 00:10:25,514 --> 00:10:27,914 Turbo Fieldfare这类项目的意义 258 00:10:27,914 --> 00:10:28,894 就在于 259 00:10:28,894 --> 00:10:31,454 它让苹果的优势重新有用 260 00:10:31,454 --> 00:10:34,714 如果AI的未来只是谁有最大数据中心 261 00:10:34,714 --> 00:10:36,754 苹果确实不占足场 262 00:10:36,754 --> 00:10:39,314 它没有NVIDIA那种GPU生态 263 00:10:39,314 --> 00:10:43,354 也没有Microsoft Azure或Google Cloud那种云入口 264 00:10:43,354 --> 00:10:47,094 但如果未来一部分AI工作要回到设备端 265 00:10:47,094 --> 00:10:49,374 苹果手里的筹码就多了 266 00:10:49,374 --> 00:10:50,654 它有统一内存 267 00:10:50,654 --> 00:10:51,834 有自研芯片 268 00:10:51,834 --> 00:10:52,654 有Metal 269 00:10:52,654 --> 00:10:54,494 有强控制的操作系统 270 00:10:54,494 --> 00:10:56,094 有高端用户设备 271 00:10:56,094 --> 00:10:59,514 也有一群愿意花钱买稳定体验的用户 272 00:10:59,514 --> 00:11:02,514 这件事最适合用一个普通场景理解 273 00:11:02,514 --> 00:11:05,194 一个用户坐在飞机上没有网路 274 00:11:05,194 --> 00:11:07,534 想让电脑整理本地笔记 275 00:11:07,534 --> 00:11:08,814 总结PDF 276 00:11:08,814 --> 00:11:10,394 查找项目文档 277 00:11:10,394 --> 00:11:12,214 生成一段邮件草稿 278 00:11:12,214 --> 00:11:14,414 如果所有AI都依赖云端 279 00:11:14,414 --> 00:11:16,814 这些任务马上变成半残废 280 00:11:16,814 --> 00:11:20,194 可是如果Mac本地能跑一个够用的模型 281 00:11:20,194 --> 00:11:22,314 很多事情就能直接做 282 00:11:22,314 --> 00:11:23,934 速度不一定顶级 283 00:11:23,934 --> 00:11:25,674 能力不一定最强 284 00:11:25,674 --> 00:11:29,094 但它能离线、私密、低延迟 285 00:11:29,094 --> 00:11:32,094 而且不需要每次把资料传出去 286 00:11:32,094 --> 00:11:33,654 对企业也是一样 287 00:11:33,654 --> 00:11:36,014 很多公司不是不想用AI 288 00:11:36,014 --> 00:11:39,654 而是不敢把内部资料全部丢给外部API 289 00:11:39,654 --> 00:11:45,074 法律合同、原代码、客户数据、医疗资料、财务表格 290 00:11:45,074 --> 00:11:46,654 这些内容一旦上传 291 00:11:46,654 --> 00:11:51,154 就牵涉权限、审计、地区合规和数据泄漏风险 292 00:11:51,254 --> 00:11:55,514 如果本地或私有设备端模型能处理一部分任务 293 00:11:55,514 --> 00:11:57,894 企业的AI使用门槛会下降 294 00:11:57,894 --> 00:12:01,474 它不需要替代GPT5这种旗舰模型 295 00:12:01,474 --> 00:12:05,014 只要把70%的日常任务吃下来 296 00:12:05,014 --> 00:12:06,674 价值就已经很大 297 00:12:06,674 --> 00:12:08,754 这也是为什么二级字节内存 298 00:12:08,754 --> 00:12:10,674 这个数字有流量潜力 299 00:12:10,674 --> 00:12:13,554 它让观众立刻听懂一个变化 300 00:12:13,554 --> 00:12:16,714 AI不再只属于云端巨头 301 00:12:16,714 --> 00:12:19,534 哪怕这个结论现在还不能完全成立 302 00:12:19,534 --> 00:12:21,494 它已经打开想象空间 303 00:12:21,494 --> 00:12:23,694 但这件事不能吹过头 304 00:12:23,694 --> 00:12:26,074 Turbo Fieldfare不是通用魔法 305 00:12:26,074 --> 00:12:30,634 它目前主要针对Gemma 426BA4B文本推理 306 00:12:30,634 --> 00:12:33,274 不支持图像、音频、视频 307 00:12:33,274 --> 00:12:35,694 也不是一个完整的Agent系统 308 00:12:35,694 --> 00:12:37,774 项目Redmi也写得很清楚 309 00:12:37,774 --> 00:12:39,614 它是Model Specific 310 00:12:39,614 --> 00:12:43,714 不是MLX或Lama.CPP那种通用包装器 311 00:12:43,714 --> 00:12:45,594 换模型不一定能照搬 312 00:12:45,594 --> 00:12:47,654 换硬件也不一定成立 313 00:12:47,654 --> 00:12:49,294 它需要Apple Silicon 314 00:12:49,294 --> 00:12:51,234 需要MacOS 26 315 00:12:51,234 --> 00:12:52,374 Metal 4 316 00:12:52,374 --> 00:12:56,534 Swift 6.2还需要足够的SSD空间 317 00:12:56,534 --> 00:12:58,354 它还有一个限时限制 318 00:12:58,354 --> 00:13:00,354 SSD不是内存 319 00:13:00,354 --> 00:13:03,354 SSD再快也比内存慢得多 320 00:13:03,354 --> 00:13:05,794 频繁读取权重会带来延迟 321 00:13:05,794 --> 00:13:08,674 也可能增加能耗和存储磨损 322 00:13:08,674 --> 00:13:10,634 缓存命中率如果不好 323 00:13:10,634 --> 00:13:11,914 速度就会掉 324 00:13:11,914 --> 00:13:13,194 掌上下文 325 00:13:13,194 --> 00:13:14,334 复杂提示 326 00:13:14,334 --> 00:13:16,394 多轮对话并发请求 327 00:13:16,394 --> 00:13:18,074 都会挑战这套设计 328 00:13:18,074 --> 00:13:20,014 视频里的23TOKENS 329 00:13:20,014 --> 00:13:22,054 是M3MAX上的演示 330 00:13:22,054 --> 00:13:23,414 8集字節MR 331 00:13:23,414 --> 00:13:26,014 MacBook Air的5到6Token 332 00:13:26,014 --> 00:13:28,934 更接近低配用戶会看到的体验 333 00:13:28,934 --> 00:13:31,074 能用不等于湿滑 334 00:13:31,074 --> 00:13:32,774 所以正确的判断 335 00:13:32,774 --> 00:13:33,994 不是MacBook 336 00:13:33,994 --> 00:13:35,734 从此取代云GPU 337 00:13:35,734 --> 00:13:38,634 而是本地AI的下限被抬高了 338 00:13:38,634 --> 00:13:39,834 这已经很重要 339 00:13:39,834 --> 00:13:42,674 还要补一个容易被忽略的成本账 340 00:13:42,674 --> 00:13:45,114 云端AI看起来省心 341 00:13:45,114 --> 00:13:47,594 但它的成本是持续性的 342 00:13:47,594 --> 00:13:49,314 用户每个月付订阅 343 00:13:49,314 --> 00:13:51,274 开发者按Token付费 344 00:13:51,274 --> 00:13:53,854 企业按席位和调用量付钱 345 00:13:53,854 --> 00:13:56,054 用的越多账单越高 346 00:13:56,054 --> 00:13:58,894 本地AI的成本更像买设备 347 00:13:58,894 --> 00:14:00,954 前期花钱买Mac 348 00:14:00,954 --> 00:14:03,594 后面用本地算力跑任务 349 00:14:03,594 --> 00:14:04,914 对高频任务来说 350 00:14:04,914 --> 00:14:07,334 本地推理会越来越有吸引力 351 00:14:07,334 --> 00:14:09,814 当然本地也不是不要成本 352 00:14:09,814 --> 00:14:12,894 它吃电、吃存储、吃内存 353 00:14:12,894 --> 00:14:14,834 也吃开发者优化时间 354 00:14:14,834 --> 00:14:18,434 可它的账单不再完全掌握在模型公司手里 355 00:14:18,434 --> 00:14:20,234 用户买了机器之后 356 00:14:20,234 --> 00:14:23,074 至少有一部分智能能力可以自己用 357 00:14:23,074 --> 00:14:25,054 这种心理差异很重要 358 00:14:25,054 --> 00:14:26,114 订阅时代 359 00:14:26,114 --> 00:14:29,494 用户越来越讨厌每个功能都按月收费 360 00:14:29,494 --> 00:14:32,194 本地AI如果能提供购用体验 361 00:14:32,194 --> 00:14:34,654 就会成为反订阅情绪的出口 362 00:14:34,654 --> 00:14:36,694 这会影响应用开发 363 00:14:36,694 --> 00:14:38,414 现在很多AI app 364 00:14:38,414 --> 00:14:40,934 只是套一层云端API用户 365 00:14:40,934 --> 00:14:42,674 输入文字服务器 366 00:14:42,674 --> 00:14:44,014 转发给模型 367 00:14:44,014 --> 00:14:46,254 再把结果显示回来 368 00:14:46,254 --> 00:14:48,394 这种产品门槛低替代也快 369 00:14:48,394 --> 00:14:50,234 未来真正有壁垒的应用 370 00:14:50,234 --> 00:14:51,834 可能会把本地模型 371 00:14:51,834 --> 00:14:52,914 云端模型 372 00:14:52,914 --> 00:14:53,914 本地文件 373 00:14:53,914 --> 00:14:54,974 隐私权限 374 00:14:54,974 --> 00:14:57,034 系统操作结合起来 375 00:14:57,034 --> 00:14:58,654 Mac上的本地AI应用 376 00:14:58,654 --> 00:15:00,434 尤其适合这么做 377 00:15:00,434 --> 00:15:03,294 因为它能直接贴近用户的文件 378 00:15:03,294 --> 00:15:04,134 日历 379 00:15:04,134 --> 00:15:04,994 邮件 380 00:15:04,994 --> 00:15:05,834 代码 381 00:15:05,834 --> 00:15:07,874 仓库和创作软件 382 00:15:07,874 --> 00:15:08,954 这样一来 383 00:15:08,954 --> 00:15:10,734 竞争重点就从 384 00:15:10,734 --> 00:15:12,314 谁接了最强API 385 00:15:12,314 --> 00:15:14,894 变成谁把AI放进真实工作流 386 00:15:14,894 --> 00:15:16,154 Turbo Field Fair 387 00:15:16,154 --> 00:15:17,934 没有解决所有产品问题 388 00:15:17,934 --> 00:15:20,654 但它说明底层可行性在提升 389 00:15:20,654 --> 00:15:22,354 底层每提升一点 390 00:15:22,354 --> 00:15:24,314 应用层就多一批可能 391 00:15:24,314 --> 00:15:28,114 AI行业过去几年一直在讲更大模型 392 00:15:28,114 --> 00:15:29,154 更大集群 393 00:15:29,154 --> 00:15:30,194 更大融资 394 00:15:30,194 --> 00:15:31,614 模型越来越强 395 00:15:31,614 --> 00:15:33,974 但普通用户越来越像租客 396 00:15:33,974 --> 00:15:34,914 账号 397 00:15:34,914 --> 00:15:35,754 订阅 398 00:15:35,754 --> 00:15:36,854 API 399 00:15:36,854 --> 00:15:38,114 云端限制 400 00:15:38,114 --> 00:15:39,234 数据上传 401 00:15:39,234 --> 00:15:41,574 所有能力都隔着一层平台 402 00:15:41,574 --> 00:15:43,214 Turbo Field Fair 403 00:15:43,214 --> 00:15:45,634 这种项目提醒了一件事 404 00:15:45,634 --> 00:15:46,854 AI的未来 405 00:15:46,854 --> 00:15:48,854 不一定只有超級數據中心 406 00:15:48,854 --> 00:15:51,614 也可以有一部分回到個人電腦 407 00:15:51,614 --> 00:15:53,314 這對開發者也有啟發 408 00:15:53,314 --> 00:15:55,134 以後做本地AI應用 409 00:15:55,134 --> 00:15:56,854 不能只問模型能不能跑 410 00:16:11,854 --> 00:16:14,974 誰能少搬一次數據誰就多一點性能 411 00:16:14,974 --> 00:16:16,974 誰能少佔一點內存 412 00:16:16,974 --> 00:16:18,894 誰就多一批用戶 413 00:16:18,894 --> 00:16:22,094 誰能把模型結構和硬件結構對齊 414 00:16:22,094 --> 00:16:24,654 誰就能把不可能變成勉強可用 415 00:16:24,654 --> 00:16:28,014 這也是蘋果生態裡可能出現機會的地方 416 00:16:28,014 --> 00:16:29,974 如果開發者圍繞Metal 417 00:16:29,974 --> 00:16:31,294 統一內存 418 00:16:31,294 --> 00:16:32,574 Neural Engine 419 00:16:32,574 --> 00:16:34,174 本地文件鎖影 420 00:16:34,174 --> 00:16:35,374 Spotlight 421 00:16:35,374 --> 00:16:37,374 Shortcuts做AI工具 422 00:16:37,374 --> 00:16:41,134 Mac可能會變成一個很獨特的本地AI開發平台 423 00:16:41,134 --> 00:16:43,714 不是為了和雲端模型硬碰硬 424 00:16:43,714 --> 00:16:46,334 而是做那些雲端不適合做的任務 425 00:16:46,334 --> 00:16:49,214 私密、離線、低延遲 426 00:16:49,214 --> 00:16:51,134 貼近個人文件系統 427 00:16:51,134 --> 00:16:54,314 真正的競爭不是本地和雲端誰消滅誰 428 00:16:54,314 --> 00:16:56,314 而是誰掌握默認入口 429 00:16:56,314 --> 00:16:58,414 雲端模型會繼續強 430 00:16:58,414 --> 00:17:02,254 因為訓練和最強推理都離不開巨量算力 431 00:17:02,254 --> 00:17:04,414 可是本地模型一旦夠用 432 00:17:04,414 --> 00:17:06,394 就會吃掉大量日常任務 433 00:17:06,394 --> 00:17:09,054 用戶不需要每次都請最強模型 434 00:17:09,054 --> 00:17:12,814 寫一封邮件 整理會議記錄 搜索本地文檔 435 00:17:12,814 --> 00:17:15,534 解釋一段代碼 生成一個小腳本 436 00:17:15,534 --> 00:17:19,554 夠快 夠私密 夠便宜 比絕對最強更重要 437 00:17:19,554 --> 00:17:23,134 這就是Turbo Fuel Fair事件的流量價值 438 00:17:23,134 --> 00:17:26,974 表面看 它只是內存減少7倍的技術新聞 439 00:17:26,974 --> 00:17:30,634 往深一層看 它在挑戰一個行業共識 440 00:17:30,634 --> 00:17:33,054 大模型必須被雲端壟斷 441 00:17:33,054 --> 00:17:36,814 它沒有推翻雲端AI 但它撕開了一條縫 442 00:17:36,814 --> 00:17:41,494 接下來要看的不是這個項目本身能不能變成大眾產品 443 00:17:41,494 --> 00:17:44,814 而是它代表的工程路線會不會擴散 444 00:17:44,814 --> 00:17:49,094 更多MOE模型會不會被專門打包成本地流氏格式 445 00:17:49,094 --> 00:17:53,834 更多Mac應用會不會接入本地OpenAI Compatible Server 446 00:17:53,834 --> 00:17:57,374 Apple會不會把類似思路放進系統級框架 447 00:17:57,374 --> 00:18:02,774 開發者會不會開始為8級字節、16級字節設備認真優化 448 00:18:02,774 --> 00:18:06,974 而不是默認要求64級字節內存和大顯卡 449 00:18:06,974 --> 00:18:08,774 如果這些事情發生 450 00:18:08,774 --> 00:18:12,874 本地AI就會從即刻演示變成產品基礎設施 451 00:18:12,874 --> 00:18:15,874 還有一個變量是模型公司本身 452 00:18:15,874 --> 00:18:18,274 如果更多模型採用MOE 453 00:18:18,274 --> 00:18:20,674 如果更多模型公開權重 454 00:18:20,674 --> 00:18:23,874 如果更多小模型追上日常任務能力 455 00:18:23,874 --> 00:18:26,174 本地推理就會更快普及 456 00:18:26,174 --> 00:18:29,474 Turbo Field Fair依賴Gamma-4這種結構 457 00:18:29,474 --> 00:18:31,774 不代表所有模型都能這麼跑 458 00:18:31,774 --> 00:18:34,674 但AI行業已經在往稀疏激活 459 00:18:34,674 --> 00:18:38,614 專用小模型、端側模型、模型路由方向走 460 00:18:38,614 --> 00:18:41,214 雲端超級模型負責難題 461 00:18:41,214 --> 00:18:43,314 端側模型負責日常 462 00:18:43,314 --> 00:18:45,854 多個模型組合起來完成任務 463 00:18:45,854 --> 00:18:48,654 這個方向對NVIDIA是提醒 464 00:18:48,654 --> 00:18:50,254 對蘋果是機會 465 00:18:50,254 --> 00:18:51,954 對開發者是新戰場 466 00:18:51,954 --> 00:18:55,234 NVIDIA仍然會統治訓練和高端推理 467 00:18:55,234 --> 00:18:58,834 可是如果越來越多日常推理回到端側 468 00:18:58,834 --> 00:19:02,634 市場對所有AI都必須上癮GPU的想像 469 00:19:02,634 --> 00:19:03,534 會降溫 470 00:19:03,534 --> 00:19:06,134 蘋果不一定搶走數據中心的錢 471 00:19:06,134 --> 00:19:08,834 但可以搶回個人設備的智能入口 472 00:19:08,834 --> 00:19:11,474 開發者如果能把本地模型用好 473 00:19:11,474 --> 00:19:14,554 就不用完全被API成本牽著走 474 00:19:14,554 --> 00:19:17,714 觀眾繼續聽下去應該帶走的判斷是 475 00:19:17,714 --> 00:19:19,914 這不是一個小工具新聞 476 00:19:19,914 --> 00:19:22,754 而是AI權力結構的小變化 477 00:19:22,754 --> 00:19:24,354 以前能力在雲端 478 00:19:24,354 --> 00:19:25,994 用戶只是調用者 479 00:19:25,994 --> 00:19:28,634 現在一部分能力開始回到設備 480 00:19:28,634 --> 00:19:31,074 用戶重新擁有一點控制權 481 00:19:31,074 --> 00:19:32,674 這一點現在還小 482 00:19:32,674 --> 00:19:34,214 但方向很清楚 483 00:19:34,214 --> 00:19:36,374 只要端側模型繼續變強 484 00:19:36,374 --> 00:19:39,774 本地AI就會從能跑走向好用 485 00:19:39,774 --> 00:19:42,374 再從好用走向默認存在 486 00:19:42,374 --> 00:19:44,814 接下來還要看一個更現實的變量 487 00:19:44,814 --> 00:19:46,054 內存配置 488 00:19:46,054 --> 00:19:48,654 蘋果這些年一直被吐槽入門 489 00:19:48,654 --> 00:19:50,054 Mac內存太小 490 00:19:50,054 --> 00:19:51,494 升級內存太貴 491 00:19:51,494 --> 00:19:53,494 過去這個潮點主要影響 492 00:19:53,494 --> 00:19:56,434 檢視頻、跑虛擬機、開大型項目 493 00:19:56,434 --> 00:19:58,034 到了本地AI時代 494 00:19:58,034 --> 00:20:00,334 它會變成更核心的購買理由 495 00:20:00,334 --> 00:20:03,914 8級字節能跑不代表8級字節最舒服 496 00:20:03,914 --> 00:20:07,514 16級字節會成為更合理的AI入門線 497 00:20:07,514 --> 00:20:09,174 32級字節 498 00:20:09,174 --> 00:20:14,014 64級字節會變成開發者和重度用戶的新分界 499 00:20:14,014 --> 00:20:16,594 Turbo Fuel Fair把門檻打低 500 00:20:16,594 --> 00:20:18,594 不代表硬件需求消失 501 00:20:18,594 --> 00:20:21,654 而是讓更多人第一次有資格進場 502 00:20:21,654 --> 00:20:24,494 這對蘋果的產品策略很微妙 503 00:20:24,494 --> 00:20:28,334 如果本地AI真的變成Mac的重要賣點 504 00:20:28,334 --> 00:20:33,074 蘋果就會有更強理由推動用戶買更高內存版本 505 00:20:33,074 --> 00:20:37,034 用戶過去買內存是為了今天的軟件 506 00:20:37,034 --> 00:20:41,394 未來買內存可能是為了未來幾年的本地模型 507 00:20:41,394 --> 00:20:43,174 蘋果當然喜歡這個股市 508 00:20:43,174 --> 00:20:45,734 因為它能提高Mac的平均售價 509 00:20:45,734 --> 00:20:47,534 但用戶也會更敏感 510 00:20:47,534 --> 00:20:49,834 既然AI要在設備端跑入門 511 00:20:49,834 --> 00:20:51,634 配置就不能太寒酸 512 00:20:51,634 --> 00:20:54,034 苹果如果繼續把內存升級 513 00:20:54,034 --> 00:20:55,674 價格定得很高 514 00:20:55,674 --> 00:20:58,414 反而會限制本地AI的普及 515 00:20:58,414 --> 00:21:00,734 另一個變量是SSD 516 00:21:00,734 --> 00:21:03,014 Turbo Fieldfare 的路線 517 00:21:03,014 --> 00:21:04,914 把SSD從單純存儲 518 00:21:04,914 --> 00:21:06,954 變成推理鏈路的一部分 519 00:21:06,954 --> 00:21:11,154 SSD速度、壽命、文件佈局、系統緩存 520 00:21:11,154 --> 00:21:12,394 都會影響體驗 521 00:21:12,394 --> 00:21:13,794 過去用戶選電腦 522 00:21:13,794 --> 00:21:16,194 看SSD主要是容量 523 00:21:16,194 --> 00:21:18,494 未來本地AI應用多了 524 00:21:18,494 --> 00:21:21,634 SSD讀寫性能也會被重新關注 525 00:21:21,634 --> 00:21:25,454 模型權重不再只是躺在硬盤裡的大文件 526 00:21:25,454 --> 00:21:29,754 而是生成每個Token時可能被反覆訪問的工作材料 527 00:21:29,754 --> 00:21:32,494 這會把電腦硬件評價體系改掉 528 00:21:32,494 --> 00:21:35,794 過去AI電腦宣傳喜歡堆TOPS 529 00:21:35,794 --> 00:21:39,094 以後真正懂行的人會看一整套鏈路 530 00:21:39,094 --> 00:21:40,694 內存帶寬夠不夠 531 00:21:40,694 --> 00:21:43,994 CPU和GPU是否共享內存 532 00:21:43,994 --> 00:21:46,994 SSD讀取延遲如何系統 533 00:21:46,994 --> 00:21:49,674 API能不能減少拷貝模型 534 00:21:49,674 --> 00:21:51,614 格式是不是貼合硬件 535 00:21:51,614 --> 00:21:54,394 單看一個算力數字很容易被騙 536 00:21:54,394 --> 00:21:57,174 AI推理的瓶頸可能不在算力 537 00:21:57,174 --> 00:21:58,674 而在數據搬運 538 00:21:58,674 --> 00:22:02,554 這也是Turbo Fieldfare給普通觀眾上的一刻 539 00:22:02,554 --> 00:22:05,414 AI不是只有模型聰不聰明 540 00:22:05,414 --> 00:22:07,194 還有數據怎麼流動 541 00:22:07,194 --> 00:22:10,614 同一個模型放在不同硬件結構上 542 00:22:10,614 --> 00:22:12,554 體驗可能完全不一樣 543 00:22:12,554 --> 00:22:15,134 未來優秀的本地AI產品 544 00:22:15,134 --> 00:22:18,214 背後一定不是簡單下載一個模型 545 00:22:18,214 --> 00:22:21,014 而是圍繞設備做深度優化 546 00:22:21,014 --> 00:22:22,694 最後給一個明確判斷 547 00:22:22,694 --> 00:22:25,634 蘋果芯片這次贏的不是模型參數 548 00:22:25,634 --> 00:22:27,674 也不是Benchmark排名 549 00:22:27,674 --> 00:22:30,234 而是設備端AI的敘事權 550 00:22:30,234 --> 00:22:33,814 過去AI敘事被NVIDIA和雲廠商拿走 551 00:22:33,814 --> 00:22:38,434 大家談的都是GPU、集群、數據中心和API 552 00:22:38,434 --> 00:22:40,474 Turbo Fieldfare 553 00:22:40,474 --> 00:22:43,034 讓另一個問題重新回到桌面 554 00:22:43,034 --> 00:22:45,534 如果模型不必全部進內存 555 00:22:45,534 --> 00:22:48,234 如果CPU和GPU共享內存 556 00:22:48,234 --> 00:22:50,574 如果SSD能參與推理 557 00:22:50,574 --> 00:22:53,754 如果應用能直接貼著硬件寫本地 558 00:22:53,754 --> 00:22:56,114 設備還能不能變成AI的主廠 559 00:22:56,114 --> 00:22:57,174 檔案還沒訂 560 00:22:57,174 --> 00:23:00,814 但這次Mac不再只是調用雲模型的屏幕 561 00:23:00,814 --> 00:23:04,614 它開始像一台真正能存在AI的個人機器 562 00:23:04,614 --> 00:23:06,594 如果這個趨勢繼續 563 00:23:06,594 --> 00:23:09,754 未來買電腦時就會多一個新問題 564 00:23:09,754 --> 00:23:12,294 這台機器能不能把個人資料 565 00:23:12,294 --> 00:23:15,054 工作流和本地模型連起來 566 00:23:15,054 --> 00:23:17,274 過去電腦拼的是性能 567 00:23:17,274 --> 00:23:19,794 後來拼的是續航和生態 568 00:23:19,794 --> 00:23:22,234 接下來可能要拼本地智能密度 569 00:23:22,234 --> 00:23:25,434 蘋果芯片這次露出的牌就在這裡 570 00:23:25,434 --> 00:23:28,634 誰能讓AI更貼近用戶自己的設備 571 00:23:28,634 --> 00:23:30,014 自己的文件 572 00:23:30,014 --> 00:23:31,714 自己的隱私邊界 573 00:23:31,714 --> 00:23:34,814 誰就能在雲端巨頭之外拿回一塊入口